Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.

32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A N I N T RO D U C T I O N TO R E S E RVO I R S I M U L AT I O N
U S I N G M AT L A B / G N U O C TAV E

This book provides a self-contained introduction to the simulation of flow and transport
in porous media, written by a developer of numerical methods. The reader will learn
how to implement reservoir simulation models and computational algorithms in a robust
and efficient manner. The book contains a large number of numerical examples, all fully
equipped with online code and data, allowing the reader to reproduce results and use them
as a starting point for their own work. All of the examples in the book are based on the
MATLAB Reservoir Simulation Toolbox (MRST), an open-source toolbox popular in both
academic institutions and the petroleum industry. The book can also be seen as a user guide
to the MRST software. It will prove invaluable for researchers, professionals, and advanced
students using reservoir simulation methods. This title is also available as Open Access on
Cambridge Core at https://doi.org/9781108591416.

knut-andreas lie is chief scientist at SINTEF in Oslo, Norway, where he heads a research
group in computational methods for geoscience applications. Over the last 20 years he
has developed commercial and in-house software solutions for the international petroleum
industry. In particular, he is a founding father of two open-source community software
tools: MRST and OPM. Lie holds a PhD in mathematics from the Norwegian University
of Science and Technology in Trondheim, where he also has a position as adjunct pro-
fessor. He has authored 140 scientific papers and supervised 65 MSc and PhD students.
He currently serves as executive editor for the SPE Journal of the Society of Petroleum
Engineers.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A N I N T RO D U C T I O N TO R E S E RVO I R
S I M U L AT I O N U S I N G
M AT L A B / G N U O C TAV E
User Guide for the MATLAB Reservoir Simulation
Toolbox (MRST)

K N U T- A N D R E A S L I E
SINTEF

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 University Printing House, Cambridge CB2 8BS, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre, New Delhi – 110025, India
79 Anson Road, #06–04/06, Singapore 079906

Cambridge University Press is part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit of
education, learning, and research at the highest international levels of excellence.

www.cambridge.org
Information on this title: www.cambridge.org/9781108492430
DOI: 10.1017/9781108591416
© Knut-Andreas Lie 2019
This work is in copyright. It is subject to statutory exceptions and to the provisions of relevant licensing
agreements; with the exception of the Creative Commons version the link, which is provided below,
no reproduction of any part of this work may take place without the written permission of
Cambridge University Press.
An online version of this work is published at doi.org/9781108591416 under a Creative Commons Open
Access license CC-BY-NC-ND 4.0, which permits reuse, distribution, and reproduction in any medium
for noncommercial purposes providing appropriate credit to the original work is given. You may not
distribute derivative works without permission. To view a copy of this license, visit
https://creativecommons.org/licenses/by-nc-nd/4.0
All versions of this work may contain content reproduced under license from third parties.
Permission to reproduce this third-party content must be obtained from these third-parties directly.
When citing this work, please include a reference to the DOI 10.1017/9781108591416
First published 2019
Printed in the United Kingdom by TJ International Ltd., Padstow, Cornwall
A catalogue record for this publication is available from the British Library.
Library of Congress Cataloging-in-Publication Data
Names: Lie, Knut-Andreas, author.
Title: An introduction to reservoir simulation using MATLAB/GNU Octave : user
guide for the MATLAB reservoir simulation toolbox (MRST) / Knut-Andreas Lie, SINTEF.
Description: Cambridge, United Kingdom ; New York, NY, USA : Cambridge
University Press, 2019. | Includes bibliographical references and index.
Identifiers: LCCN 2018049101 | ISBN 9781108492430 (hardback : alk. paper)
Subjects: LCSH: Hydrocarbon reservoirs–Computer simulation. | Oil reservoir
engineering–Data processing. | MATLAB. | GNU Octave.
Classification: LCC TN870.53 .L54 2019 | DDC 622/.338–dc23
LC record available at https://lccn.loc.gov/2018049101
ISBN 978-1-108-49243-0 Hardback
Cambridge University Press has no responsibility for the persistence or accuracy
of URLs for external or third-party internet websites referred to in this publication
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Contents

Preface page xiii

1 Introduction 1
1.1 Petroleum Recovery 3
1.2 Reservoir Simulation 6
1.3 Outline of the Book 9
1.4 The First Encounter with MRST 14
Part I Geological Models and Grids 19
2 Modeling Reservoir Rocks 21
2.1 Formation of Sedimentary Rocks 21
2.2 Creation of Crude Oil and Natural Gas 26
2.3 Multiscale Modeling of Permeable Rocks 28
2.3.1 Geological Characterization 29
2.3.2 Representative Elementary Volumes 31
2.3.3 Microscopic Models: The Pore Scale 33
2.3.4 Mesoscopic Models 34
2.4 Modeling Rock Properties 34
2.4.1 Porosity 35
2.4.2 Permeability 36
2.4.3 Other Parameters 38
2.5 Property Modeling in MRST 39
2.5.1 Homogeneous Models 40
2.5.2 Random and Lognormal Models 40
2.5.3 The 10th SPE Comparative Solution Project: Model 2 41
2.5.4 The Johansen Formation 44
2.5.5 SAIGUP: Shallow-Marine Reservoirs 46
3 Grids in Subsurface Modeling 55
3.1 Structured Grids 57
3.2 Unstructured Grids 62

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 vi Contents

3.2.1 Delaunay Tessellation 63

3.2.2 Voronoi Diagrams 66
3.2.3 General Tessellations 68
3.2.4 Using an External Mesh Generator 69
3.3 Stratigraphic Grids 72
3.3.1 Corner-Point Grids 73
3.3.2 2.5D Unstructured Grids 85
3.4 Grid Structure in MRST 88
3.5 Examples of More Complex Grids 96
3.5.1 SAIGUP: Model of a Shallow-Marine Reservoir 97
3.5.2 Composite Grids 101
3.5.3 Control-Point and Boundary Conformal Grids 103
3.5.4 Multiblock Grids 104
Part II Single-Phase Flow 111
4 Mathematical Models for Single-Phase Flow 113
4.1 Fundamental Concept: Darcy’s Law 113
4.2 General Flow Equations for Single-Phase Flow 115
4.3 Auxiliary Conditions and Equations 120
4.3.1 Boundary and Initial Conditions 120
4.3.2 Injection and Production Wells 121
4.3.3 Field Lines and Time-of-Flight 128
4.3.4 Tracers and Volume Partitions 129
4.4 Basic Finite-Volume Discretizations 131
4.4.1 Two-Point Flux-Approximation 131
4.4.2 Discrete div and grad Operators 135
4.4.3 Time-of-Flight and Tracer 141
5 Incompressible Solvers for Single-Phase Flow 143
5.1 Basic Data Structures in a Simulation Model 144
5.1.1 Fluid Properties 144
5.1.2 Reservoir States 145
5.1.3 Fluid Sources 145
5.1.4 Boundary Conditions 146
5.1.5 Wells 147
5.2 Incompressible Two-Point Pressure Solver 149
5.3 Upwind Solver for Time-of-Flight and Tracer 153
5.4 Simulation Examples 156
5.4.1 Quarter Five-Spot 157
5.4.2 Boundary Conditions 160
5.4.3 Structured versus Unstructured Stencils 165
5.4.4 Using Peaceman Well Models 169

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Contents vii

6 Consistent Discretizations on Polyhedral Grids 174

6.1 The TPFA Method Is Not Consistent 174
6.2 The Mixed Finite-Element Method 177
6.2.1 Continuous Formulation 178
6.2.2 Discrete Formulation 180
6.2.3 Hybrid Formulation 182
6.3 Finite-Volume Methods on Mixed Hybrid Form 185
6.4 The Mimetic Method 188
6.5 Monotonicity 197
6.6 Discussion 199
7 Compressible Flow and Rapid Prototyping 202
7.1 Implicit Discretization 202
7.2 A Simulator Based on Automatic Differentiation 204
7.2.1 Model Setup and Initial State 205
7.2.2 Discrete Operators and Equations 206
7.2.3 Well Model 208
7.2.4 The Simulation Loop 209
7.3 Pressure-Dependent Viscosity 213
7.4 Non-Newtonian Fluid 215
7.5 Thermal Effects 220
Part III Multiphase Flow 229
8 Mathematical Models for Multiphase Flow 231
8.1 New Physical Properties and Phenomena 232
8.1.1 Saturation 232
8.1.2 Wettability 234
8.1.3 Capillary Pressure 235
8.1.4 Relative Permeability 239
8.2 Flow Equations for Multiphase Flow 243
8.2.1 Single-Component Phases 244
8.2.2 Multicomponent Phases 245
8.2.3 Black-Oil Models 246
8.3 Model Reformulations for Immiscible Two-Phase Flow 248
8.3.1 Pressure Formulation 248
8.3.2 Fractional-Flow Formulation in Phase Pressure 249
8.3.3 Fractional-Flow Formulation in Global Pressure 254
8.3.4 Fractional-Flow Formulation in Phase Potential 255
8.3.5 Richards’ Equation 256
8.4 The Buckley–Leverett Theory of 1D Displacements 258
8.4.1 Horizontal Displacement 258
8.4.2 Gravity Segregation 264
8.4.3 Front Tracking: Semi-Analytical Solutions 266

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 viii Contents

9 Discretizing Hyperbolic Transport Equations 272

9.1 A New Solution Concept: Entropy-Weak Solutions 272
9.2 Conservative Finite-Volume Methods 274
9.3 Centered versus Upwind Schemes 275
9.3.1 Centered Schemes 275
9.3.2 Upwind or Godunov Schemes 277
9.3.3 Comparison of Centered and Upwind Schemes 279
9.3.4 Implicit Schemes 283
9.4 Discretization on Unstructured Polyhedral Grids 286
10 Solvers for Incompressible Immiscible Flow 289
10.1 Fluid Objects for Multiphase Flow 290
10.2 Sequential Solution Procedures 292
10.2.1 Pressure Solvers 293
10.2.2 Saturation Solvers 294
10.3 Simulation Examples 297
10.3.1 Buckley–Leverett Displacement 298
10.3.2 Inverted Gravity Column 301
10.3.3 Homogeneous Quarter Five-Spot 303
10.3.4 Heterogeneous Quarter Five-Spot: Viscous Fingering 307
10.3.5 Buoyant Migration of CO2 in a Sloping Sandbox 311
10.3.6 Water Coning and Gravity Override 313
10.3.7 The Effect of Capillary Forces – Capillary Fringe 319
10.3.8 Norne: Simplified Simulation of a Real-Field Model 323
10.4 Numerical Errors 326
10.4.1 Splitting Errors 326
10.4.2 Grid Orientation Errors 330
11 Compressible Multiphase Flow 337
11.1 Industry-Standard Simulation 337
11.2 Two-Phase Flow without Mass Transfer 342
11.3 Three-Phase Relative Permeabilities 348
11.3.1 Relative Permeability Models from ECLIPSE 100 349
11.3.2 Evaluating Relative Permeabilities in MRST 353
11.3.3 The SPE 1, SPE 3, and SPE 9 Benchmark Cases 355
11.3.4 A Simple Three-Phase Simulator 358
11.4 PVT Behavior of Petroleum Fluids 359
11.4.1 Phase Diagrams 360
11.4.2 Reservoir Types and Their Phase Behavior during Recovery 364
11.4.3 PVT and Fluid Properties in Black-Oil Models 370
11.5 Phase Behavior in ECLIPSE Input Decks 379
11.6 The Black-Oil Equations 388

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Contents ix

11.6.1 The Water Component 389

11.6.2 The Oil Component 390
11.6.3 The Gas Component 392
11.6.4 Appearance and Disappearance of Phases 392
11.7 Well Models 394
11.7.1 Inflow-Performance Relationships 394
11.7.2 Multisegment Wells 395
11.8 Black-Oil Simulation with MRST 399
11.8.1 Simulating the SPE 1 Benchmark Case 399
11.8.2 Comparison against a Commercial Simulator 405
11.8.3 Limitations and Potential Pitfalls 406
12 The AD-OO Framework for Reservoir Simulation 413
12.1 Overview of the Simulator Framework 414
12.2 Model Hierarchy 420
12.2.1 PhysicalModel – Generic Physical Models 422
12.2.2 ReservoirModel – Basic Reservoir Models 426
12.2.3 Black-Oil Models 430
12.2.4 Models of Wells and Production Facilities 433
12.3 Solving the Discrete Model Equations 434
12.3.1 Assembly of Linearized Systems 434
12.3.2 Nonlinear Solvers 436
12.3.3 Selection of Time-Steps 439
12.3.4 Linear Solvers 440
12.4 Simulation Examples 449
12.4.1 Depletion of a Closed/Open Compartment 449
12.4.2 An Undersaturated Sector Model 451
12.4.3 SPE 1 Instrumented with Inflow Valves 455
12.4.4 The SPE 9 Benchmark Case 460
12.5 Improving Convergence and Reducing Runtime 470
Part IV Reservoir Engineering Workflows 475
13 Flow Diagnostics 477
13.1 Flow Patterns and Volumetric Connections 478
13.1.1 Volumetric Partitions 479
13.1.2 Time-of-Flight Per Partition Region: Improved Accuracy 482
13.1.3 Well Allocation Factors 482
13.2 Measures of Dynamic Heterogeneity 483
13.2.1 Flow and Storage Capacity 483
13.2.2 Lorenz Coefficient and Sweep Efficiency 486
13.3 Residence-Time Distributions 489
13.4 Case Studies 495

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 x Contents

13.4.1 Tarbert Formation: Volumetric Connections 495

13.4.2 Heterogeneity and Optimized Well Placement 501
13.5 Interactive Flow Diagnostics Tools 505
13.5.1 Synthetic 2D Example: Improving Areal Sweep 510
13.5.2 SAIGUP: Flow Patterns and Volumetric Connections 514
14 Grid Coarsening 518
14.1 Grid Partitions 518
14.1.1 Uniform Partitions 519
14.1.2 Connected Partitions 520
14.1.3 Composite Partitions 522
14.2 Coarse Grid Representation in MRST 524
14.2.1 Subdivision of Coarse Faces 526
14.3 Partitioning Stratigraphic Grids 529
14.3.1 The Johansen Aquifer 529
14.3.2 The SAIGUP Model 532
14.3.3 Near Well Refinement for CaseB4 536
14.4 More Advanced Coarsening Methods 540
14.5 A General Framework for Agglomerating Cells 541
14.5.1 Creating Initial Partitions 541
14.5.2 Connectivity Checks and Repair Algorithms 542
14.5.3 Indicator Functions 544
14.5.4 Merge Blocks 545
14.5.5 Refine Blocks 547
14.5.6 Examples 550
14.6 Multilevel Hierarchical Coarsening 553
14.7 General Advice and Simple Guidelines 556
15 Upscaling Petrophysical Properties 558
15.1 Upscaling for Reservoir Simulation 559
15.2 Upscaling Additive Properties 561
15.3 Upscaling Absolute Permeability 563
15.3.1 Averaging Methods 564
15.3.2 Flow-Based Upscaling 569
15.4 Upscaling Transmissibility 575
15.5 Global and Local–Global Upscaling 578
15.6 Upscaling Examples 580
15.6.1 Flow Diagnostics Quality Measure 581
15.6.2 A Model with Two Facies 581
15.6.3 SPE 10 with Six Wells 585
15.6.4 Complete Workflow Example 589
15.6.5 General Advice and Simple Guidelines 595

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Contents xi

Appendix The MATLAB Reservoir Simulation Toolbox 597

A.1 Getting Started with the Software 598
A.1.1 Core Functionality and Add-on Modules 598
A.1.2 Downloading and Installing 601
A.1.3 Exploring Functionality and Getting Help 602
A.1.4 Release Policy and Version Numbers 606
A.1.5 Software Requirements and Backward Compatibility 606
A.1.6 Terms of Usage 608
A.2 Public Data Sets and Test Cases 609
A.3 More About Modules and Advanced Functionality 610
A.3.1 Operating the Module System 611
A.3.2 What Characterizes a Module? 612
A.3.3 List of Modules 613
A.4 Rapid Prototyping Using MATLAB and MRST 620
A.5 Automatic Differentiation in MRST 623

References 631
Index 650
Usage of MRST Functions 658

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Preface

There are many books that describe mathematical models for flow in porous media and
present numerical methods used to discretize and solve the corresponding systems of partial
differential equations; a comprehensive list can be found in the References. However,
neither of these books fully describes how you should implement the models and numerical
methods to form a robust and efficient simulator. Some books may present algorithms and
data structures, but most leave it up to you to figure out all the nitty-gritty details you
need in order to get your implementation up and running. Likewise, you may read papers
presenting models or computational methods that may be exactly what you need for your
work. After the initial enthusiasm, however, you very often end up quite disappointed, or at
least I do when I realize that the authors have not presented all the details of their methods,
or that it will probably take me weeks or months to get my own implementation working.
In this book, I try to be a bit different and give a reasonably self-contained introduc-
tion to the simulation of flow and transport in porous media that also discusses how to
implement the models and algorithms in a robust and efficient manner. In the presentation,
I have tried to let the discussion of models and numerical methods go hand in hand with
numerical examples that come fully equipped with codes and data, so that you can rerun
and reproduce the results yourself and use them as a starting point for your own research
and experiments. You will get most out of the book if you continuously switch between
reading and experimenting with the many code snippets and tutorial examples on your
own computer. All examples in the book are based on the MATLAB Reservoir Simulation
Toolbox (MRST), which has been developed by my group and published online as free
open-source code under the GNU General Public License since 2009.
The book can alternatively be seen as a comprehensive user-guide to MRST. Over the
years, MRST has become surprisingly popular. At the time of writing (July 2018), the soft-
ware has more than 17,000 unique downloads, 120 students have used MRST in their mas-
ter or PhD theses, and more than 190 papers written by authors outside of SINTEF include
numerical experiments run in MRST. This book tries to give an in-depth introduction to
MRST and explain the two different programming paradigms you can find in the software.
The book is up to date with respect to the latest developments in data structures and syntax,
both for the original procedural approaches that primarily focus on incompressible flow, as

xiii

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 xiv Preface

well as for the more recent object-oriented, automatic-differentiation (AD-OO) framework

we developed to simulate compressible, multiphase flow.
The book has grown much longer than I anticipated when I started writing. Initially, my
ambition was to provide introductory material on single-phase and two-phase incompress-
ible flow, and discuss how such models could be implemented in a flexible and efficient
manner by use of MATLAB. While I was writing, the software expanded rapidly and I had
a hard time keeping pace. Inspired in part by the many people who have downloaded (and
cited) the preliminary editions I have published online, but also as a result of numerous
requests, I decided to expand the book and include a detailed discussion of the physics
underlying industry-standard black-oil simulators, and give a thorough introduction to how
such models have been implemented in MRST. Now that this is done, it is time to stop.
However, MRST has more to offer, including solvers for various enhanced oil recovery
(EOR) models, compositional flow, fractured media, geomechanics, geochemistry, as well
as a comprehensive set of tools for modeling CO2 storage in large-scale aquifer systems,
but documentation of these will have to be another book, or perhaps also another author. At
the moment, I am more than happy that I finally managed to finish this book.
First of all, I am very grateful to Equinor for the generous grant that enabled Gold Open
Access publication of this book. I would also like to thank my current and former colleagues
at SINTEF with whom I have collaborated over many year to develop MRST; primarily
Olav Møyner, Halvor Møll Nilsen, Stein Krogstad, Jostein R. Natvig, Odd Andersen, Bård
Skaflestad, and Xavier Raynaud. I have also had a great number of inspiring discussions
with Alf Birger Rustad and Vegard Kippe at Equinor over the past decade. The chapter
on flow diagnostics is partially the result of many discussions with Brad Mallison from
Chevron. I am grateful to the University of Bergen and the Norwegian University of Science
and Technology (NTNU) for funding through my Professor II positions. Victor Calo and
Yalchin Efendiev invited me to KAUST, where important parts of the chapters on grids and
petrophysics were written. Likewise, Margot Gerritsen invited me to Stanford and gave
me the opportunity to develop Jolts videos that complement the first chapters of the book.
Significant parts of the chapter on black-oil models were written during the three weeks
I participated in the Long Program on Computational Issues in Oil Field Applications at
UCLA in 2017. Let me also thank all colleagues and students who have given suggestions,
pointed out errors and misprints, asked questions, and given me inspiration to continue
working. Even though your name is not mentioned explicitly, I have not forgotten your
important contributions. Last, but not least, I thank my wife Anne: the many evenings,
late nights, and early mornings it took to write this book would never have been possible
without your support and understanding.
Finally to the reader: As you may understand from the book, I strongly believe in
reproducible computational science and sharing your work. I hope the book manages to
convince you that the combination of a high-level scripting language like MATLAB/GNU
Octave and development of reusable open-source software is an excellent approach to both
being productive and having a significant impact on others; I elaborate more on this in an
essay I wrote to my former supervisor, Helge Holden, for his 60th birthday [189]. Whereas

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Preface xv

the book is static, the MRST software will continue to develop in new directions as a
result of continued research and contributions from external users. Your help is invaluable
to making this happen; please do not hesitate to contact me if you have suggestions for
improvements or new functionality, or if you have developed your independent add-on
functionality and want advice on how to best share it with the general community.

Knut-Andreas Lie
[email protected]

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 1
Introduction

Subsurface flow phenomena cover some of the most important technological challenges
of our time. The road toward sustainable use and management of the earth’s freshwater
reserves necessarily involves modeling of hydrological systems in order to understand fluid
movement in groundwater basins, quantify limits of sustainable use, monitor transport of
pollutants in the subsurface, and appraise schemes for groundwater remediation. An equally
important problem is the reduction of the emission of greenhouse gases like CO2 into the
atmosphere. Carbon sequestration in subsurface rock formations has been suggested as a
possible means to that end. The primary concern is how fast the injected CO2 will escape
back to the atmosphere. Repositories do not necessarily need to store CO2 forever, just
long enough to allow the natural carbon cycle to reduce the atmospheric CO2 to near pre-
industrial levels. Nevertheless, making qualified estimates of leakage rates from potential
storage facilities is a nontrivial task, and demands interdisciplinary research and software
based on state-of-the-art numerical methods for modeling subsurface flow. Other obvious
concerns are whether the injected CO2 will leak into freshwater aquifers, or migrate to
habitated or different legislative areas.
A third challenge is petroleum recovery. The civilized world will likely continue to
depend on utilization of petroleum resources both as an energy carrier and as raw mate-
rial for a wide variety of petrochemical products (fertilizers, detergents, dyes, synthetic
rubber, plastics, etc.) in the foreseeable future. In the North Sea region, the recovery of
existing and discovery of new conventional petroleum resources has declined significantly
in recent years. Optimal utilization of known resources is therefore of utter importance
to meet the demands for petroleum and lessen the pressure on exploration in vulnerable
areas like in the arctic regions. Meanwhile, there has been a dramatic increase in the
utilization of unconventional petroleum resources (shale gas/oil), and there is a strong need
to understand how these can be produced in an economic way that minimizes harm to the
environment.
Reliable computer modeling of subsurface flow is much needed to overcome these three
challenges, but is also needed to exploit deep geothermal energy or to inject compressed gas

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2 Introduction

in the subsurface to store energy, ensure safe storage of nuclear waste, improve remediation
technologies to remove contaminants from the subsurface, etc. Indeed, the need for tools
that help us understand flow processes in the subsurface is probably greater than ever, and
increasing. More than fifty years of prior research in this area has led to some degree of
agreement in terms of how subsurface flow processes can be modeled adequately with
numerical simulation technology.
This book introduces and discusses mathematical models describing flow processes
in porous rocks on a macroscopic scale. The presentation focuses primarily on physical
processes that take place during hydrocarbon recovery. Simulation of such processes is
often referred to as reservoir simulation. This means that even though the mathematical
models, numerical methods, and software implementations presented can be applied to
any of the applications just mentioned, the specific examples use vocabulary, physical
scales, and balances of driving forces specific to petroleum recovery. As an example of this
vocabulary, we can consider the ability of a porous medium to transmit fluids. In petroleum
engineering this is typically given in terms of the “permeability,” which is a pure rock
property, whereas one in water resource engineering is more concerned with the “hydraulic
conductivity” that also takes the viscosity and density of the fluid into account. In CO2
sequestration you can see both quantities used.
As an example of physical scales, let us compare oil recovery by water flooding and the
long-term geological storage of CO2 . Petroleum resources can only accumulate when the
natural upward movement of hydrocarbons relative to water is prevented by confinements
in the overlying rocks. An oil reservoir is therefore usually a relatively closed system with
a spatial extent of hundreds to thousands of meters, from which oil can be recovered over
tens of years. The most promising candidates for geological CO2 storage are huge aquifer
systems that stretch out for hundreds of kilometers. The injection of CO2 is mainly driven
by pressure differences as in oil recovery. However, as CO2 moves into the aquifer and
the effects of the injection pressure ceases, the fluid movement is gradually dominated by
buoyant forces that cause the lighter CO2 phase to migrate upward in the open aquifer
system. This process can potentially continue over thousands of years. The basic flow
physics and governing equations are the same in both cases, but the balances between
physical forces are different, and this should be accounted for when formulating the overall
mathematical models and appropriate numerical methods.
Models and numerical methods developed to study subsurface flow are also applica-
ble to other natural and man-made porous media such as soils, biological tissues and
plants, electrochemical devices like batteries and fuel cells, concrete and other porous
constructional materials, food and sanitary products, textiles, industrial filtering processes,
polymer composites, water desalination, etc. Porous media models are also used to describe
human physiology, e.g., waterways in the brain or flow in the capillary part of the
vascular system. An interesting example is in-tissue drug delivery, where the challenge
is to minimize the volume swept by the injected fluid, unlike in petroleum recovery,
where one seeks to maximize the volumetric sweep to push out as much petroleum as
possible.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 1.1 Petroleum Recovery 3

1.1 Petroleum Recovery

We start by a conceptual sketch of the basic mechanisms by which hydrocarbon can be
recovered from a subsurface reservoir. In good reservoir rock, the void spaces between
mineral grains form networks of connected pores that can store and transmit large amounts
of fluids. Petroleum discoveries vary in size from small pockets of hydrocarbon that may
be buried just a few meters beneath the surface of the earth and can easily be produced,
to huge reservoirs1 stretching out several square kilometers beneath remote and stormy
seas. However, as a simple mental picture, you can think of a hydrocarbon reservoir as a
bent, rigid sponge that is confined inside an insulating material and has all its pores filled
with hydrocarbons, which may appear in the form of a liquid oleic and a gaseous phase as
illustrated in Figure 1.1. Natural gas will be dissolved in oil under high volumetric pressure
like carbon dioxide inside a soda can. If the pressure inside the pristine reservoir is above
the bubble point, the oil is undersaturated and still able to dissolve more gas. If the pressure
is below the bubble point, the oil will be fully saturated with gas and any excess gas will
form a gas cap on top of the oil since it is lighter.
To extract oil from the reservoir, we start by drilling a well into the oil zone. If the
pristine pressure inside the reservoir is sufficiently high, it will push oil up to the surface
and start what we will refer to as primary production. Alternatively, one may use a pump
to lower the pressure in the wellbore below the point where oil starts flowing. How large

Figure 1.1 Conceptual illustration of a petroleum reservoir during primary production. Over millions
of years, hydrocarbons have accumulated under a caprock that has a low ability to transmit fluids
and thus prevents their upward movement. Inside the trap, the fluids will distribute according to
density, with light hydrocarbons in a gaseous phase on top, heavier hydrocarbons in an oleic phase
in the middle, and brine at the bottom. If the pressure difference between the oil zone and the
wellbore is sufficiently high, the oleic phase will flow naturally out of the reservoir. As the oleic
phase is produced, the pressure inside the reservoir will decline, which in turn may introduce other
mechanisms that help to maintain pressure and push more oil out of the well.

1 The largest reservoir in the world is found in Ghawar in the Saudi Arabian desert and is approximately 230 km long, 30 km
wide, and 90 m thick.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 4 Introduction

pressure differential one needs for hydrocarbons to flow, depends on the permeability of
the rock; the higher the permeability, the easier the hydrocarbons will flow toward the well.
As hydrocarbons are extracted, the pressure inside the reservoir will decay and the recov-
ery will gradually decline. However, declining pressure will often induce other physical
processes that contribute to maintain recovery:

• In a water drive, the pore space below the hydrocarbons is filled with salt water, which is
slightly compressible and hence will expand a little as the reservoir pressure is lowered.
If the total water volume is large compared with the oil zone, even a small expansion will
create significant water volumes that will push oil toward the well and hence contribute to
maintain pressure. Sometimes the underlying water is part of a large aquifer system that
has a natural influx that replenishes the extracted oil by water and maintains pressure.
• Solution gas drive works like when you shake and open a soda can. Initially, the pristine
oil will be in a pure liquid state and contain no free gas. Extraction of fluids will gradually
lower the reservoir pressure below the bubble point, which causes free gas to develop and
form expanding gas bubbles that force oil into the well. Inside the well, the gas bubbles
rise with the oil and make the combined fluid lighter and hence easier to push upward to
the surface. At a certain point, however, the bubbles may reach a critical volume fraction
and start to flow as a single gas phase, which has lower viscosity than the oil and hence
moves faster. This rapidly depletes the energy stored inside the reservoir and causes the
production to falter. Gas coming out of solution can also migrate to the top of the structure
and form a gas cap above the oil; the gas cap pushes down on the liquid oil and hence
helps to maintain pressure.
• In a gas cap drive, the reservoir contains more gas than what can be dissolved in the
oil. When pressure is lowered, the gas cap expands and pushes oil into the well. Over
time, the gas cap will gradually infiltrate the oil and cause the well to produce increasing
amounts of gas.
• If a reservoir is highly permeable, gravity will force oil to move downward relative to gas
and upward relative to water. This is called gravity drive.
• A combination drive has water below the oil zone and a gas cap above that both will push
oil to the well at the same time as reservoir pressure is reduced.
These natural (or primary) drive mechanisms can only maintain the pressure for a limited
period and the production will gradually falter as the reservoir pressure declines. How fast
the pressure declines and how much hydrocarbons one can extract before the production
ceases, varies with the drive mechanism. Solution gas drives can have a relatively rapid
decline, whereas water and gas cap drives are able to maintain production for longer peri-
ods. Normally only 30% of the oil can be extracted using primary drive mechanisms.
To keep up the production and increase the recovery factor, many reservoirs employ
some kind of engineered drive mechanisms for secondary production. Figure 1.2 illustrates
two examples of voidage replacement in which water and/or gas is injected to support
pressure in the reservoir. Water can also be injected to sweep the reservoir, displace the oil,
and push it toward the wells. In some cases, one may choose to inject produced formation

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 1.1 Petroleum Recovery 5

Figure 1.2 Conceptual illustration of voidage replacement, which is an example of a secondary

production strategy in which gas and/or water is injected to maintain the reservoir pressure.

water, which is contaminated with hydrocarbons and solid particles and hence must be
disposed of in some manner. Alternatively, one can extract formation water from a nearby
aquifer. In offshore production it is also common to inject seawater. A common problem for
all waterflooding methods is to maximize the sweep efficiency so that water does not move
rapidly through high-flow zones in the reservoir and leave behind large volumes of unswept,
mobile oil. Maintaining good sweep efficiency is particularly challenging for reservoirs
with high-viscosity oil. If the injected water has lower viscosity than the resident oil, it
will tend to form viscous fingers that rapidly expand through the oil and cause early water
breakthrough in producers. (Think of water being poured into a cup of honey.) To improve
the sweep efficiency, one can add polymers to the water to increase its viscosity and
improve the mobility ratio between the injected and displaced fluids. Polymers have also
been used to create flow diversions by plugging high-flow zones so that the injected fluid
contacts and displaces more oil. For heavy oils, adverse mobility ratios can be improved by
using steam injection or some other thermal method to heat the oil to reduce its viscosity.
Polymer injection and steam injection are examples of methods for so-called enhanced
oil recovery (EOR), also called tertiary production. Another example is miscible and chem-
ical injection, in which one injects solvents or surfactants that mix with the oleic phase
in the reservoir to make it flow more readily. The solvent may be a gas such as carbon
dioxide or nitrogen. However, the most common approach is to inject natural gas produced
from the reservoir when there is no market that will accept the gas. Surfactants are similar
to detergents used for laundry. Alkaline or caustic solutions, for instance, can react with
organic acids occuring naturally in the reservoir to produce soap. The effect of all these
substances is that they reduce the interfacial tension between water and oil, which enables
small droplets of oil that were previously immobile to flow (more) freely. This is the same
type of process that takes place when you use detergent to remove waxy and greasy stains
from textiles. A limiting factor of these methods is that the chemicals are quickly adsorbed
and lost into the reservoir rock.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 6 Introduction

Often, one will want to combine methods that improve the sweep efficiency of mobile
oil with methods that mobilize immobile oil. Miscible gas injection, for instance, can be
used after a waterflood to flush out residually trapped oil and establish new pathways to
the production wells. Water-alternating-gas (WAG) is the most successful and widely used
EOR method. Injecting large volumes of gas is expensive, and by injecting alternating slugs
of water, one reduces the injected volume of gas required to maintain pressure. Similarly,
the presence of mobile water reduces the tendency of the injected gas to finger through the
less mobile oil. In polymer flooding, it is common to add surfactants to mobilize immobile
oil by reducing or removing the interface tension between oil and water, and likewise, add
alkaline solutions to reduce the adsorption of chemicals onto the rock faces.
Whereas the mechanisms of all the previously described methods for EOR are reason-
ably well studied and understood, there are other methods whose mechanisms are much
debated. This includes injection of low-salinity water, which is not well understood, even
though it has proved to be highly effective in certain cases. Another example is microbial
EOR, which relies on microbes that digest long hydrocarbon molecules to form biosur-
factants or emit carbon dioxide that will reduce interfacial tension and mobilize immobile
oil. Microbial activity can either by achieved by injecting bacterial cultures mixed with a
food source, or by injecting nutrients that will activate microbes that already reside in the
reservoir.
Use of secondary recovery mechanisms has been highly successful. On the Norwegian
continental shelf, for instance, water flooding and miscible gas injection have helped to
increase the average recovery factor to 46%, which is significantly higher than the world-
wide average of 22%. In other parts of the world, chemical methods have proved to be very
efficient for onshore reservoirs having relatively short distances between wells. For offshore
fields, however, the potential benefits of using chemical methods are much debated. First of
all, it is not obvious that such methods will be effective for reservoirs characterized by large
inter-well distances, as rapid adsorption onto the pore walls generally makes it difficult to
transport the active ingredients long distances into a reservoir. Chemicals are also costly,
need to be transported in large quantities, and consume platform space.
Even small improvements in recovery rates can lead to huge economic benefits for the
owners of a petroleum asset and for this reason much research and engineering work is
devoted to improving the understanding of mobilization and displacement mechanisms,
and to design improved methods for primary and enhanced oil recovery. Mathematical
modeling and numerical reservoir simulation play key roles in this endeavor.

1.2 Reservoir Simulation

Reservoir simulation is the means by which we use a numerical model of the geological
and petrophysical characteristics of a subsurface reservoir, the (multiphase) fluid system,
and the production equipment (wells and surface facilities) to analyze and predict how
fluids flow through the reservoir rock to the stock tank or transport pipeline over time. It is

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 1.2 Reservoir Simulation 7

Figure 1.3 The three main constituents of a reservoir simulation model.

generally very difficult to observe and understand dynamic fluid behavior inside a reservoir,
describe the physical processes, and measure all the parameters that influence the flow
behavior. Predicting how a reservoir will produce over time and respond to different drive
and displacement mechanisms therefore has a large degree of uncertainty attached. Simula-
tion of petroleum reservoirs started in the mid 1950s as a means to quantify and reduce this
uncertainty, and has become an important tool for qualitative and quantitative prediction of
the flow of fluid phases. Reservoir simulation is a complement to field observations, pilot
and laboratory tests, well testing, and analytical models, and is used by reservoir engineers
to investigate displacement processes, compare and contrast the characteristics of different
production scenarios, or as part of inverse modeling to calibrate reservoir parameters by
integrating static and dynamic (production) data. In the big picture, reservoir simulation is
mostly used to guide two different types of decisions: (i) to optimize development plans for
new fields, and (ii) to assist with operational and investment decisions.
To describe the subsurface flow processes mathematically, our simulation model will
be made up of three main constituents. First, we need a mathematical flow model that
describes how fluids flow in a porous medium. These models are typically given as a set of
partial differential equations describing the mass conservation of fluid phases, accompanied
by a suitable set of constitutive relations that describe the relationship among different
physical quantities. Second, we need a geological model that describes the given porous
rock formation (the reservoir). The geological model is realized as a grid populated with
petrophysical properties that are used as input to the flow model, and together they make
up the reservoir simulation model. Last, but not least, we need a model for the wells and
production facilities that provide pressure and fluid communication between the reservoir
and the surface.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8 Introduction

Accurate prediction of reservoir flow scenarios is a difficult task. One reason is that we
can never get a complete and accurate characterization of the rock parameters that influence
the flow pattern. Even if we did, we would not be able to run simulations that exploit all
available information, since this would require a tremendous amount of computer resources
that far exceed the capabilities of modern multiprocessor computers. On the other hand, we
neither need nor do we seek a simultaneous description of the flow scenario on all scales
down to the pore scale. For reservoir management it is usually sufficient to describe the
general trends in the reservoir flow pattern.
In the early days of the computer, reservoir simulation models were built from two-
dimensional slices with 102 –103 Cartesian grid cells representing the whole reservoir. In
contrast, contemporary reservoir characterization methods can model the porous rock for-
mations by the means of grid-blocks down to the meter scale. This gives three-dimensional
models consisting of millions of cells. Stratigraphic grid models, based on extrusion of 2D
areal grids to form volumetric descriptions, have been popular for many years and are the
current industry standard. However, more complex methods based on unstructured grids
are gaining in popularity.
Despite an astonishing increase in computer power and intensive research on computa-
tion techniques, commercial reservoir simulators are rarely used to run simulations directly
on geological grid models. Instead, coarse grid models with grid-blocks that are typically
ten to hundred times larger are built using some kind of upscaling of the geophysical
parameters. How one should perform this upscaling is not trivial. In fact, upscaling has
been, and probably still is, one of the most active research areas in the oil industry. This
effort reflects the general opinion that given the ever increasing size and complexity of
the geological reservoir models, one cannot generally expect to run simulations directly on
geological models in the foreseeable future.
Along with the development of better computers, new and more robust upscaling
techniques, and more detailed reservoir and fluid characterization, there has also been
an equally significant development in the area of numerical methods. State-of-the-art
simulators employ numerical methods that can take advantage of multiple processors,
distributed memory computer architectures, adaptive grid refinement strategies, and
iterative techniques with linear complexity. For the simulation, there exists a wide variety
of different numerical schemes that all have their pros and cons. With all these techniques
available, we see a trend where methods from the research forefront are being tuned to a
special set of applications and mathematical models, as opposed to traditional methods that
were developed for a large class of differential equations.
Altogether, these developments have enabled geologists and reservoir engineers to build
increasingly complex geological and reservoir models. Continuing to improve such models
to gain a better understanding of the reservoir and reduce uncertainty is obviously impor-
tant. Nevertheless, you should not forget that the purpose of a simulation study usually is to
help your company make better decisions, and thus the value of the study generally depends
on to what extent it influences decisions and leads to higher profit, e.g., by increasing
recovery and/or reducing capital expenditures (CAPEX) or operational expenses (OPEX).

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 1.3 Outline of the Book 9

Developing a very advanced model that has all the physical effects imaginable included
may therefore not be necessary to answer a specific question. Indeed, to paraphrase the
famous Occam’s principle: a simulation model should be as simple as possible, but not
simpler. In the coming years, we will probably see a larger degree of hybrid modeling that
combines models based on physical principles with data-driven approaches.

1.3 Outline of the Book

The book is intended to serve several purposes. First of all, you can use it as a self-contained
introduction to the physics of flow in porous media, its basic mathematical theory, and the
numerical methods used to solve the underlying differential equations. Hopefully, the book
will also give you a hands-on introduction to practical modeling of flow in porous media,
focusing in particular on models and problems that are relevant to the petroleum industry.
The discussion of mathematical models and numerical methods is accompanied by a large
number of illustrative examples, ranging from idealized and highly simplified setups to
cases involving models of real-life reservoirs. Last, but not least, the book will introduce
you to a widely used open-source software, and teach you some of the principles that have
been used in developing it.

The Software Aspect: User Guide, Examples, and Exercises

All examples in the book have been created using the MATLAB Reservoir Simulation
Toolbox (MRST). This open-source software consists on one hand of a set of reservoir
simulators and workflow tools that you can modify to suit your own purposes, and on the
other hand it can be seen as an enhancement of MATLAB/GNU Octave toward reservoir
modeling in the form of a large a collection of flexible and efficient software libraries
and data structures, which you can use to design your own simulators or computational
workflows. The use of MRST permeates more traditional textbook material, and the book
can therefore be seen as a user guide to MRST, in which you get introduced to the software
gradually, example by example. Appendix A gives a more focused introduction to MRST,
which starts by explaining how the software is organized, how you can install it, explore its
functionality, and find help much in the same way as in MATLAB/GNU Octave.
The book can alternatively be viewed as a discussion by example of how a numerical
scripting language like MATLAB/GNU Octave can for be used for rapid prototyping,
testing, and verification on realistic reservoir problems with a high degree of complexity.
Through the many examples, I also try to gradually teach you some of the techniques and
programming concepts that have been used to create MRST, so that you later can use similar
ideas to ensure flexibility and high efficiency in your own programs.
In the introductory part of the book that covers grids, petrophysical parameters and basic
discretizations and solvers for single-phase flow, I have tried to make all code examples as
self-contained as possible. To this end, all code lines necessary to produce the numerical
results and figures are presented and discuss in detail. However, occasionally I omit minor
details that either have been discussed elsewhere or should be part of your basic MATLAB

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10 Introduction

repertoire. As we move to more complex examples, in particular for multiphase flow and
reservoir engineering workflows, it is no longer expedient to discuss scripts in full detail.
In most cases, however, complete scripts containing all code lines necessary to run the
examples can be found in a dedicated book module in MRST. I strongly encourage that you
use your own computer to run as many as possible of the examples in the book, as well as
other examples and tutorials that are distributed with the software. Your understanding
will be further enhanced if you also modify the examples, e.g., by changing the input
parameters, or extend them to solve problems that are related, but (slightly) different. To
help you in this direction, I have included a number of computer exercises that modify and
extend some of the examples, combine ideas from two or more examples, or investigate in
more detail aspects that are not covered by any of the worked examples. For some of the
exercises you can find solution proposals in the book module.
Finally, I point out that MRST is an open-source software, and if reading this book
gives you ideas about new functionality, or you discover things that are not working as they
should or could, you are free to improve the toolbox and expand it in new directions. If
you do so, I strongly encourage you to pay us back by releasing your code publicly for the
benefit of the reservoir simulation community.

Part I: Geological Models and Grids

The first part of the book discusses how to represent a geological medium as a discrete
computer model that can be used to study the flow of one or more fluid phases. Chapter
2 gives you a crash course in petroleum geology and geological modeling, written by a
non-geologist. In this chapter, I try to explain key processes that lead to the formation
of a hydrocarbon reservoir, discuss modeling of permeable rocks across multiple spatial
scales, and introduce you to the basic physical properties used to describe porous media in
general. For many purposes, reservoir geology can be represented as a collection of maps
and surfaces. However, if the geological model is to be used as input to a macroscale fluid
simulation, we must assume a continuum hypothesis and represent the reservoir in terms of
a volumetric grid, in which each cell is equipped with a set of petrophysical properties. On
top of this grid, one can then impose mathematical models that describe the macroscopic
continuum physics of one or more fluid phases flowing through the microscopic network of
pores and throats between mineral grains that are present on the subscale inside the porous
rock of each grid block.
Chapter 3 describes in more detail how to represent and generate grids, with special
emphasis on the types of grids commonly used in reservoir simulation. The chapter presents
a wide variety of examples to illustrate and explain different grid formats, from simple
structured grids, via unstructured grids based on Delaunay tessellations and Voronoi dia-
grams, to stratigraphic grids represented on the industry-standard corner-point format or
as 2.5D and 3D unstructured grids. The examples also demonstrate various methods for
generating grids that represent plausible reservoir models, and they discuss some of the
realistic data sets that can be downloaded along with the MRST software.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 1.3 Outline of the Book 11

Through Chapters 2 and 3, you will be introduced to the data structures used to represent
grids and petrophysical data in MRST. Understanding these basic data structures, and
the various methods used to create and manipulate them, is fundamental if you want to
understand the inner workings of a majority of the routines implemented in MRST or use
the software as a platform to implement your own computational methods. Through the
many examples, you will also be introduced to various functionality in MRST for plotting
data associated with cells and faces (interface between two neighboring cells) as well as
various strategies for traversing the grid and picking subsets of data; these techniques will
prove very useful later in the book.

Part II: Single-Phase Flow

The second part of the book is devoted entirely to single-phase flow and will introduce
you to many of the key concepts for modeling flow in porous media, including basic
flow equations, closure relationships, auxiliary conditions and models, spatial and temporal
discretizations, and linear and nonlinear solvers.
Chapter 4 starts by introducing the two fundamental principles necessary to describe
flow in porous media: conservation of mass and Darcy’s law. Mass conservation is a fun-
damental physical property, whereas Darcy’s law is a phenomenological description of
the conservation of momentum that introduces permeability, a rock property, to relate
volumetric flow rate to pressure differentials. The chapter outlines different forms these
two equations can take in various special cases. To form a full model, the basic flow
equations must be extended with various constitutive laws and extra equations describing
external forces that drive fluid flow; these can either be boundary conditions and/or wells
that inject or produce fluids. The last section of Chapter 4 introduces the classical two-
point finite-volume method, which is the current industry standard for discretizing flow
equations. In particular, we demonstrate how to write this discretization in a very compact
way by introducing discrete analogues of the divergence and gradient operators. These
operators will be used extensively later when developing solvers for compressible single
and multiphase flow with the AD-OO framework.
Chapter 5 focuses on the special case of an incompressible fluid flowing in a completely
rigid medium, for which the flow model can be written as a Poisson-type partial differ-
ential equation (PDE) with a varying coefficient. We start by introducing the various data
structures that are necessary to make a full simulator, including fluid properties, reservoir
states describing the primary unknowns, fluid sources, boundary conditions, and models
of injection and production wells. We then discuss in detail the implementation of a two-
point pressure solver, as well as upwind solvers for linear advection equations, which, e.g.,
can be used to compute time lines (time-of-flight) in the reservoir and steady-state tracer
distributions that delineate the reservoir into various influence regions. We end the chapter
with four simulation examples: the first introduces the classical quarter five-spot pattern,
which is a standard test case in reservoir simulation, while the next three explain how
to incorporate boundary conditions and Peaceman well models and discuss the difference
between using structured and unstructured grids.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12 Introduction

Grids describing real reservoirs typically have unstructured topology and irregular cell
geometries with high aspect ratios. Two-point discretizations are unfortunately only con-
sistent if a certain relationship between the grid and the permeability tensor is satisfied,
which is quite restrictive and difficult to fulfill when modeling complex geology. Chapter 6
discusses methods that are consistent for general polyhedral cell geometries. This includes
mixed finite-element methods, multipoint flux approximation methods, and mimetic finite-
difference methods that are still being researched by academia.
Chapter 7, the last in Part II, is devoted to compressible flow, which in the general case
is modeled by a nonlinear, time-dependent, parabolic PDE. Using this relatively simple
model, we introduce many of the concepts that will later be used to develop multiphase
simulators of full industry-standard complexity. To discretize the transient flow equation,
we combine the two-point method introduced for incompressible flow with an implicit
temporal discretization. The standard approach to solving the nonlinear system of discrete
equations arising from complex multiphase models is to compute the Jacobian matrix of
first derivatives for the nonlinear system, and use Newton’s method to successively find
better approximations to the solution. Deriving and implementing analytic expressions
for Jacobian matrices is both error-prone and time-consuming, in particular if the flow
equations contain complex fluid models, well descriptions, thermodynamical behavior,
etc. In MRST, we have therefore chosen to construct Jacobian matrices using automatic
differentiation, which is a technique to numerically evaluate the derivatives of functions
specified by a computer program to working precision accuracy. Combining this technique
with discrete averaging and differential operators enables you to write very compact sim-
ulator codes in which flow models are implemented almost in the same form as they are
written in the underlying mathematical model. This greatly simplifies the task of writing
new simulators: all you have to do is to implement the new model equations in residual
form and specify which variables should be considered as primary unknowns. When the
software then evaluates all the elementary operations necessary to compute the residual, it
simultaneously uses elementary differentiation rules to compute the analytical derivative
of each elementary operation at the specific function value. These values are then gathered
using the standard chain rule and assembled into a block matrix containing the correct par-
tial derivatives with respect to all primary variables. To demonstrate the utility and power of
the resulting framework, we discuss how one can quickly change functional dependencies
in the single-phase pressure solver and extend it to include new functional dependencies
like pressure-dependent viscosity, thermal effects, or non-Newtonian fluid rheology.

Part III: Multiphase Flow

The third part of the book outlines how to extend the ideas from Part II to multiphase
flow. Chapter 8 starts by introducing new physical phenomena and properties that appear
for multiphase flows, including fluid saturations, wettability and capillary pressure, relative
permeability, etc. With this introduced, we move on to outline the general flow equations
for multiphase flow, before we discuss various model reformulations and (semi-)analytical
solutions for the special case of immiscible, two-phase flow. The main difference between

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 1.3 Outline of the Book 13

single and multiphase flow is that we now, in addition to an equation for fluid pressure or
density, get additional equations for the transport of fluid phases and/or chemical species.
These equations are generally parabolic, but will often simplify to or behave like hyperbolic
equations. Chapter 9 therefore introduces basic concept from the theory of hyperbolic
conservation laws and introduces a few classical schemes. We also introduce the basic
discretization that will be used for transport equations later in the book.
Chapter 10 follows along the same lines as Chapter 5 and explains how incompressible
solvers developed for single-phase flow can be extended to account for multiphase flow
effects using the so-called fractional-flow formulation introduced in Chapter 8 and simu-
lated using sequential methods in which pressure effects and transport of fluid saturations
and/or component concentrations are computed in separate steps. The chapter includes a
number of test cases highlighting various effect of multiphase flow and illustrating error
mechanisms inherent in sequential simulation methods.
Once the basics of incompressible, multiphase flow has been discussed, we move on
to discuss more advanced multiphase flow models, focusing primarily on the black-oil
formulation, which can be found in contemporary commercial simulators. The black-oil
equations lump all hydrocarbons into two pseudo-components: light components that exist
in a gas phase at surface conditions and heavier components that exist in a liquid oil phase.
At reservoir conditions, the two components form a gaseous and an oleic phase, but can
also dissolve fully or partially in the other phase. The dissolution depends on pressure
(and temperature), and the simplified pressure, volume, and temperature (PVT) behavior
of the resulting fluids follows analytical or tabulated relationships. Together with water,
the hydrocarbon system forms a three-phase, three-component system, whose behavior can
be surprisingly intricate. Chapter 11 gives an in-depth discussion of the underlying phys-
ical principles, whereas Chapter 12 discusses how one can use variants of the automatic-
differentiation methods introduced in Chapter 7, combined with modern principles of object
orientation, to develop robust simulators. The two chapters also include a number of illus-
trative simulation cases.

Part IV: Reservoir Engineering Workflows

The fourth and last part of the book is devoted to discussing additional computational
methods and tools that can be used to address common tasks within reservoir engineering
workflows.
Chapter 13 is devoted to flow diagnostics, which are simple numerical experiments that
can be used to probe a reservoir to understand flow paths and communication patterns.
Herein, all types of flow diagnostics will be based on the computation of time-of-flight,
which defines natural time lines in the porous medium, and steady-state distribution of
numerical tracers, which delineate the reservoir into subregions that can be uniquely asso-
ciated with distinct sources of the inflow/outflow. Both quantities will be computed using
finite-volume methods introduced in Chapters 4 and 5, and you do not have to read the
chapters in between to understand the essential ideas of Chapter 13. The concept of flow
diagnostics also includes several measures of dynamic heterogeneity, which can be used

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14 Introduction

as simple proxies for more comprehensive multiphase simulations in various reservoir

engineering workflows including placement of wells, rate optimization, etc.
As you will see in Chapter 2, porous rocks are heterogeneous at a large variety of length
scales. There is therefore a general trend toward building complex, high-resolution models
for geological characterization to represent small-scale geological structures. Likewise,
large ensembles of equiprobable models are routinely generated to systematically quantify
model uncertainty. In addition, many companies develop hierarchies of models that cover
a wide range of physical scales to systematically propagate the effects of small-scale
geological variations up to the reservoir scale. In either case, one quickly ends up with
geological models that contain more details than what can or should be used in subsequent
flow simulation studies. Hence, there is a strong need for mathematical and numerical
techniques for formulating reduced models, or communicating effective parameters and
properties between models of different spatial resolution. Such methods are discussed
in Chapters 14 and 15. Chapter 14 introduces data structures and various methods for
partitioning a fine-scale grid into a coarse-scale grid. This chapter hardly assumes any
familiarity with flow solvers and can be read directly after Chapters 2 and 3. However,
before you continue to read about upscaling in Chapter 15, which refers to the process in
which petrophysical properties in all cells that make up a coarse block are averaged into
a single effective value for each coarse block, I suggest that you read the chapters about
incompressible, single-phase flow.
After this quick tour of the book, you are probably eager to start digging into the
material. Before continuing to the next chapter, we present a simple example that will give
you a first taste of simulating flow in porous media.

1.4 The First Encounter with MRST

The purpose of this first example is to show the ten code lines needed to set up and
solve simple flow problem. We also show how to visualize the geological model and the
computed flow solution. To this end, we consider a very simple problem: compute the
pressure variation p(z) inside a [0,1] × [0,1] × [0,30] m3 rock column for a single-phase
fluid of constant density ρ assuming a datum pressure p(z0 ) = 100 bar at the top of the
column. The solution is trivial and follows by integration,
 z
p(z) = p(z0 ) + gρ dz = p(z0 ) + gρ(z − z0 ). (1.1)
z0

Here, g is the gravity constant along the z axis, which by convention in reservoir simulation
is assumed to point downwards so that z increases toward greater depth.
We will now compute the same solution using a simple incompressible flow solver. To
this end, we need two of the three constituents discussed in Section 1.2: (i) a flow model
describing the fluid behavior and (ii) a model describing the reservoir rock. As explained
already, the flow model consists of an equation describing conservation of mass and Darcy’s
law, which essentially represents conservation of momentum:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 1.4 The First Encounter with MRST 15
K 
∇ · v = 0, v = − ∇p − ρg∇z , (1.2)
μ
Darcy’s law relates the flow rate v to the gradient of the flow potential p−ρgz. The constant
of proportionality is given as the ratio between the macroscopic rock permeability K and
the fluid viscosity μ. By eliminating v, we can reduce (1.2) to an elliptic Poisson equation,
whose solution is given by (1.1) for constant μ and K.
Most examples in MRST feature some geological model, which by convention is created
first. The basic part of a geological model is a grid describing the geometry of the reservoir
rock, here chosen as a regular 1 × 1 × 30 Cartesian grid. The grid is generated in two steps
using standard routines from MRST:
G = cartGrid([1, 1, 30], [1, 1, 30]*meter^3);
G = computeGeometry(G);

The first function constructs the grid topology and cell geometries. The second function
computes derived quantities such as cell volumes and centroids, areas, and normals of all
cell faces, which you need if you want to solve flow equations on the grid. More details
about grids are given in Chapter 3. To plot the grid, we can use the command
plotGrid(G); view(3);

A geological model should also describe petrophysical and other geological properties. In
our case, the only parameter is the permeability, which is set to 100 millidarcy (md or mD).
MRST also requires porosity (i.e., the void fraction of the bulk volume) to be given for all
models. Since it is not used here, we set it to a typical value of 0.2:
rock = makeRock(G, 0.1*darcy(), 0.2);

MRST works in SI units and we must therefore be careful to specify the correct units for all
physical quantities. The result of this construction is a MATLAB structure that contains two
vectors of 30 elements each, giving the permeability and porosity in each cell. Petrophysical
properties are discussed in more detail in Chapter 2. The grid colored by permeability
values is plotted as follows:
cla, plotCellData(G, rock.perm)

To start representing the flow part of the model, we first need a structure to hold the reservoir
state, which in the basic form will consist of the pressure and the flux across cell faces. We
can construct an empty container with all values set to zero as follows
sol = initResSol(G, 0.0);

Since we are solving (1.2) on a finite domain, we must also describe conditions on all
boundaries. The default assumption is that there will be no flow over the reservoir boundary
v · n = 0). In our case, we also need to prescribe p = 100 bar at the top of the column
(
bc = pside([], G, 'TOP', 100.*barsa());

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 16 Introduction

To discretize Darcy’s law in (1.2), we use a standard two-point flux-approximation scheme,

which relates the flux between two neighboring cells i and j to their pressure difference,
vij = −Tij (pi − pj ). The constant of proportionality Tij , is called the transmissibility
and is computed as the harmonic average of two other constants Ti,j and Tj,i associated
with cell i and j , respectively. Each of these quantities on depend on geometrical grid
properties and the permeability tensor of a single cell and can be computed once and for all
independent of the particular flow model once we have defined the grid and petrophysical
parameters:
hT = computeTrans(G, rock);

This scheme will be discussed in more detail in Section 4.4. More advanced discretizations
can be found in other add-on modules and are discussed further in Chapter 6.
The next thing we need to define is the fluid properties. Unlike grids, petrophysical data,
and boundary conditions, data structures for representing fluid properties are not part of
the basic functionality of MRST. The reason is that the way fluid properties are specified
is tightly coupled with the mathematical and numerical formulation of the flow equations,
and may differ a lot between different types of simulators. Here, we have assumed incom-
pressible flow and can therefore use fluid models from the incomp add-on module,
mrstModule add incomp
gravity reset on
fluid = initSingleFluid('mu', 1*centi*poise, 'rho', 1014*kilogram/meter^3);

The second line ensures that gravity is set to its default value, whereas the third line defines
the fluid to have a density of 1014 kg/m3 , which is representative for water. The constant
dynamic viscosity does not affect the pressure distribution and is arbitrarily set to 1 cP.
More details about the fluid object will be given in Section 5.1.1. All we need to know
here is that we can query it for viscosity and density by the following call:
[mu,rho] = fluid.properties();

We now have all we need to set up and solve a discrete version of (1.2). The first
sub-equation says that the fluxes across all faces of each cell should sum to zero, i.e.,

j Tij (pi − pj ) = 0, for all i. For Cartesian grids and homogeneous permeability, this
scheme coincides with the classical seven-point finite-difference scheme for Poisson’s
problem and is the only discretization in the incomp module. As a final step, we use the
transmissibilities, the fluid object, and the boundary conditions to assemble and solve the
discrete system:
sol = incompTPFA(sol, G, hT, fluid,'bc', bc);

Having computed the solution, we plot the pressure given in units “bar,” which equals
0.1 MPa and is referred to as “barsa” in MRST since “bar” is a built-in command in
MATLAB:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 1.4 The First Encounter with MRST 17
Pressure [bar]

102.5
5

10 102

15
101.5

20

101

25

100.5
30
1
1
0.5
0.5 100
0 0

Figure 1.4 Hydrostatic pressure distribution in a gravity column computed by MRST. This example
is taken from the MRST tutorial flowSolverTutorial1.m (gravityColumn.m in older versions
of MRST).

plotFaces(G, 1:G.faces.num, convertTo(sol.facePressure, barsa()));

set(gca, 'ZDir', 'reverse'), title('Pressure [bar]')
view(3), colorbar, set(gca,'DataAspect',[1 1 10])

From the plot shown in Figure 1.4, we see that our solution correctly reproduces the linear
pressure increase given in (1.1).
Before you continue reading, I encourage you to consult Appendix A, which describes
MRST in more detail and explains how to install and get started with the software. Having
the software operational is not a prerequisite for reading this book, but will contribute
significantly to increase your understanding. And in case you should worry, you do not
need to type in all the individual MATLAB/MRST commands; complete source codes for
almost all the cases discussed later in the book can be found in a special add-on module,
which you activate by calling mrstModule add book.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Part I
Geological Models and Grids

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2
Modeling Reservoir Rocks

Aquifers and natural petroleum reservoirs consist of a subvolume of sedimentary rocks that
have sufficient porosity and permeability to store and transmit fluids. This chapter gives an
overview of how such rocks are modeled, so as to become part of a simulation model. We
start by describing briefly how sedimentary rocks and hydrocarbons are formed. In doing
so, we introduce some geological concepts that you may encounter while working with
subsurface modeling. We then move on to describe how rocks that contain hydrocarbons
or aquifer systems are modeled in terms of a volumetric grid and a set of accompanying
discrete petrophysical properties. Finally, we discuss how rock properties are represented in
MRST, and show several examples of rock models with varying complexity, ranging from
an idealized shoebox rock body with homogeneous properties, via the widely used SPE
10 model, to two realistic models, one synthetic and one representing a large-scale aquifer
from the North Sea.

2.1 Formation of Sedimentary Rocks

Sedimentary rocks are created by mineral or organic particles that are deposited and accu-
mulated on the Earth’s surface or within bodies of water to create layer upon layer of
sediments. The sedimentary rocks found in reservoirs come from sedimentary basins, inside
which large-scale sedimentation processes have taken place. Sedimentary basins are formed
as the result of stretching and breaking of the continental crust. As the crust is stretched,
hot rocks deeper in the earth come closer to the surface. When the stretching stops, the hot
rocks start to cool, which causes the crustal rock to gradually subside and move downward
to create a basin. Such processes are taking place today. The Great Rift Valley of Africa
is a good example of a so-called rift basin, where a rift splits the continental plate so that
opposite sides of the valley are moving a millimeter apart each year. This gradually creates
a basin inside which a new ocean may appear over the next hundred million years.
On the part of the Earth’s surface that lies above sea level, wind and flowing water
will remove soil and rock from the crust and transport it to another place where it is
deposited when the forces transporting the mineral particles can no longer overcome gravity
and friction. Mineral particles can also be transported and deposited by other geophysical

21

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 22 Modeling Reservoir Rocks

Figure 2.1 Illustration of various forms of depositional environments. Aeolian sediments are created
by wind, evaporite minerals are left behind as water evaporates, fluvial sediments are deposited by
rivers, and lacustrine sediments are deposited in lakes. Alluvial refers to the case where the geological
process is not well described.
The illustration is based on an image by PePeEfe on Wikimedia Commons. The modified image is published under the Creative
Commons Attribution-ShareAlike 3.0 Unported license.

processes, like mass movement and glaciers. Over millions of years, layers of sediments
will build up in deep waters, in shallow-marine waters along the coastline, or on land in
lakes, rivers, sand deltas, and lagoons, see Figure 2.1. The sediments accumulate a depth
of a few centimeters every one hundred years to form sedimentary beds (or strata) that may
extend many kilometers in the lateral directions.
Over time, the weight of the upper layers of accumulating sediment will push the lower
layers downward. A combination of heat from the Earth’s center and pressure from the
overburden will cause the sediments to undergo various chemical, physical, and biological
processes (commonly referred to as diagenesis). These processes cause the sediments to
consolidate so that the loose materials form a compact, solid substance that may come
in varying forms. This process is called lithification. Sedimentary rocks have a layered
structure with different mixtures of rock types with varying grain size, mineral types,
and clay content. The composition of a rock depends strongly upon a combination of
geological processes, the type of sediments that are transported, and the environment in
which it is deposited. Sandstone are formed by mineral particles that are broken off from
a solid material by weathering and erosion in a source area and transported by water to a
place where they settle and accumulate. Limestone is formed by the skeletons of marine
organisms such as corals, which have a high calcium content. Sandstone and limestone will
typically contain relative large void spaces between mineral particles in which fluids can
move easily. Mudrocks, on the other hand, are formed by compression of clay, mud, and
silt, and will consist of relatively fine-grained particles. These rocks, which are sometimes
also referred to as shale, will have small pores and therefore have limited ability to transmit

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2.1 Formation of Sedimentary Rocks 23

Figure 2.2 Illustration of a meandering river. Because water flows faster along the outer than along
the inner curve of the bend, the river will erode along the outer bank and deposit at the inner bank.
This causes the river to move sideways over time. Occasionally, the river will overflow its banks and
cover the lower-lying flood plain with water and mud deposits that makes the soil fertile. Flood plains
are therefore very important for agriculture.

fluids. Chemical rocks are formed by minerals that precipitate from a solution, e.g., salts
that are left behind when oceans evaporate. Like mudrocks, salts are impermeable to fluids.
Sediments will also generally contain organic particles that originate from the remains of
plants, living creatures, and small organisms living in water.
To understand a particular rock formation, one must understand the prehistoric sedimen-
tary environment from which it originates. It is common to distinguish between continental
and marine environments. The primary source of sediments in a continental environment
is rivers, inside which the moving water has high energy so that sediments will mainly be
composed of fragments of preexisting minerals and rock, called clasts. Resulting rocks are
therefore often referred to as clastic rocks. A flood plain is the area of the land adjacent to
the river bank that will be covered by water when a river breaks its bank and floods during
periods of high water discharge. Flood plains are created by bends in the river (meanders)
that erode sideways; see Figure 2.2. A flood plain is the natural place for a river to diminish
its kinetic energy. As the water moves, particles of rock are carried along by frictional drag
of water on the particle, and water must flow at a certain velocity to suspend and transport a
particle. Large particles are therefore deposited where the water flows rapidly, whereas finer
particles will settle where the current is weak. River systems also create alluvial fans, which
are fan-shaped sediment deposits built up by streams that carry debris from a single source.
Other examples of continental environments are lagoons, lakes, and swamps that contain
quiet water in which fine-grained sediments mingled with organic material are deposited.
For marine rocks formed in a sea or ocean, one usually makes the distinction between
deep and shallow environments. In deep waters (200 m or more), the water moves relatively
slowly over the bottom, and deposits will mainly consist of fine clay and skeletons of small
microorganisms. However, if the sea bottom is slightly inclined, the sediments can become
unstable and induce sudden currents of sediment-laden water to move rapidly downslope.
These so-called turbidity currents are caused by density differences between the suspended

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 24 Modeling Reservoir Rocks

sediments and the surrounding water and will generally be highly turbulent. Because of the
high density, a turbidity current can transport larger particles than what pure water would be
able to at the same velocity. These currents will therefore cause what can be considered as
instantaneous deposits of large amounts of clastic sediments into deep oceans. The resulting
rocks are called turbidite.
Shallow-marine environments are found near the coastline and contain greater kinetic
energy caused by wave activity. Clastic sediments will therefore generally be coarser than
in deep-marine environments and will consists of small-grained sand, clay, and silt that has
been washed out and transported from areas of higher water energy on the continent. Far
from the continent, the transport of clastic sediments is small, and deposits are dominated
by biological activity. In warm waters, there are multitudes of small organisms that build
carbonate skeletons, and, when deposited together with mud, these skeletons will turn into
limestone, which is an example of a carbonate rock. Carbonate rocks can also be made from
the skeletons of larger organisms living on coral reefs. Carbonates are soluble in slightly
acidic water, and this may create karst rocks that contain large regions of void space (caves,
caverns).
To model sedimentary rocks in a simulation model, one must be able to describe the
geometry and the physical properties of different rock layers. A bed denotes the smallest
unit of rock that is distinguishable from an adjacent rock layer unit above or below it, and
can be seen as bands of different colors or textures in hillsides, cliffs, river banks, road
cuts, etc. Each band represents a specific sedimentary environment, or mode of deposition,
and can range from a few centimeters to several meters thick, often varying in the lateral
direction. A sedimentary rock is characterized by its bedding, i.e., sequence of beds and
lamina (less-pronounced layers). The bedding process is typically horizontal, but beds
may also be deposited at a small angle, and parts of the beds may be weathered down
or completely eroded away during deposition, allowing newer beds to form at an angle
with older ones. Figure 2.3 shows two photos of sedimentary rock outcrops from Svalbard,
which is one of the few places in Northern Europe where one can observe large-scale
outcrops of sedimentary rocks. Figure 2.4 shows two more pictures of layered structures
on meter to kilometer scale and centimeter scale, respectively.

Figure 2.3 Outcrops of sedimentary rocks from Svalbard, Norway. The length scale is a few hundred
meters.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2.1 Formation of Sedimentary Rocks 25

ne nd)
dsto n sa
San ndblow
(wi
o n)
le ago
Sha amp/l
(sw le
sha
ne,
esto
n e ,lim flat)
l
dsto t i d a
San g o o n /
(la

Figure 2.4 Layered geological structures occur on both large and small scales in sedimentary rocks.
The left picture is from Grand Canyon, the right picture is courtesy of Silje Støren Berg, University
of Bergen.

Each sedimentary environment has its own characteristic deposits and forms what is
called a depositional facies, i.e., a body of rock with distinct characteristics. Different
sedimentary environments usually exist alongside each other in a natural succession. Small
stones, gravel, and large sand particles are heavy and are deposited at the river bottom,
whereas small sand particles are easily transported and are found at river banks and on the
surrounding plains along with mud and clay. Following a rock layer of a given age, one
will therefore see changes in the facies (rock type). Similarly, depositional environments
change with time: shorelines move with changes in the sea level and land level or because
of formation of river deltas, rivers change their course because of erosion or flooding, etc.
Dramatic events like floods create more abrupt changes. At a given position, the accumu-
lated sequence of beds will therefore contain different facies.
As time passes by, more and more sediments accumulate and the stack of beds piles
up. Simultaneously, severe geological activity takes place: Cracking of continental plates
and volcanic activity will change our eventual reservoir from a relatively smooth, layered
sedimentary basin into a complex structure. Previously, continuous layers of sediments
are compressed and pushed against each other to form arches, which are referred to as
anticlines, and depressions, which are referred to as synclines. If the deposits contain large
accumulations of salts, these will tend to flow upward through the surrounding layers to
form large domes, since salts are lighter than other mineral particles. Likewise, as the
sediments may be stretched, cut, shifted, or twisted in various directions, their movement
may introduce fractures and faults. Fractures are cracks or breakage in the rock, across
which there has been no movement. Faults are fractures or discontinuities in a volume of
rock, across which movement in the crust has caused rock volumes on opposite sides to be
displaced relative to each other. Some faults are small and localized, whereas others are part
of the vast system of boundaries between tectonic plates that crisscrosses the crust of the
Earth. Faults are described in terms of their strike, which is the compass direction in which
the fault intersects the horizontal plane, and their dip, which is the angle that the fault plane
makes with the horizontal, measured perpendicular to the strike. Faults are further classified

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 26 Modeling Reservoir Rocks

Normal dip–slip fault Reverse dip–slip fault Strike–slip fault

Figure 2.5 Fault types: In strike–slip faults, shear stress causes the rock volumes to slide past each
other to the left and right in the lateral direction with very little vertical motion. In dip–slip faults, the
rock volumes move predominantly in the vertical direction. A normal dip–slip fault occurs when the
crust is extended and is sometimes called a divergent fault. Reverse dip–slip faults occur as a result of
compressive shortening of the crust, and are sometimes called convergent faults, since rock volumes
on opposite sides of the fault plane move horizontally toward one another.

by their slip, which is the displacement vector that describes the relative movement of rock
volumes on opposite sides of the fault plane. The dip is usually separated into its vertical
component, called throw, and its horizontal component, called heave. Figure 2.5 illustrates
these different types of faults.

2.2 Creation of Crude Oil and Natural Gas

Deposits not only consist of sand grains, mud, and small rock particles, but will also contain
remains of water-based plankton, algae, and other organisms that die and fall to the bottom
where they are covered in mud. During their life, these organisms have absorbed heat from
the sunlight and when they die and mix with the sediments, they take energy with them. As
the sediments get buried deeper and deeper, the increasing heat from the Earth’s core and
pressure from the overburden will compress and “cook” the organic material that consists
of cellulose, fatty oils, proteins, starches, sugar, waxes, and so on into an intermediate
waxy product called kerogen. Whereas organic material contains carbon, hydrogen, and
oxygen, the kerogen contains less oxygen and has a higher ratio of hydrogen to carbon.
The maturation process that eventually turns kerogen into crude oil and natural gas depends
highly upon temperature, and may take from a million years at 170◦ C to a hundred million
of years at 100◦ C. Most of the oil and natural gas we extract today has been formed from
the remains of prehistoric algae and zooplankton living in the ocean. Coal, on the other
hand, is formed from the remains of dead plants. The rock in which this “cooking” process
takes place is commonly referred to as the source rock. The chances of forming a source
rock containing a significant amount of oil and gas increases if there has been an event that
caused mass death of microorganisms in an ocean basin.
Pressure from sediments lying above the source rock will force the hydrocarbons to
migrate upward through the void space in these newer sediments. The lightest hydrocarbons

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2.2 Creation of Crude Oil and Natural Gas 27

(methane and ethane) that form natural gas usually escape quickly, while the liquid oils
move more slowly toward the surface. In this process, the natural gas will separate from
the oil, and the oil will separate from the resident brine (salty water). At certain sites,
the migrating hydrocarbons will be collected and trapped in structural or stratigraphic
traps. Structural traps are created by tectonic activity that forms layers of sediments into
anticlines, domes, and folds. Stratigraphic traps form because of changes in facies (e.g.,
in clay content) within the bed itself or when the bed is sealed by an impermeable layer
such as mudstone, which consists of small and densely packed particles, and thus has
strong capillary forces that cannot easily be overcome by the buoyancy forces that drive
migrating hydrocarbons upward. If a layer of mudstone is folded into an anticline above
a more permeable sandstone or limestone, it will therefore act as a caprock that prevents
the upward migration of hydrocarbons. Stratigraphic traps are especially important in the
search for hydrocarbons; approximately 4/5 of the world’s oil and gas reserves are found
in anticlinal traps.
Figure 2.6 illustrates the various forms in which migrating oil and natural gas can be
trapped and accumulate to form petroleum reservoirs. Anticlines are typically produced
when lateral pressure causes strata to fold, but can also result from draping and subsequent
compaction of sediments accumulating over local elevations in the topography. Strati-
graphic traps accumulate hydrocarbons because of changes in the rock type. For example,
sand banks deposited by a meandering river is later covered by mud from the flood plain.
Similarly, stratigraphic traps can form when rock layers are tilted and eroded away and
subsequently covered by other low-permeable strata. Fault traps are formed when strata
are moved in opposite directions along a fault line so that permeable rocks come in contact
with impermeable rocks. The fault itself can also be an effective trap if it contains clays that

Figure 2.6 Various forms of traps in which migrating hydrocarbons can accumulate.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 28 Modeling Reservoir Rocks

are smeared as the layers of rock move past each other. Salt domes are created by buried
salt deposits that rise unevenly through the surrounding strata. Oil can either rest against
the salt itself, or the salt induces chemical reactions in the surrounding rock that makes it
impermeable.
The first evidence of hydrocarbons beneath the earth’s surface were found in so-called
seeps, which are found in many areas throughout the world. Seeps are formed when the seal
above a trap is breached so that oil and gas can migrate all the way through the geological
layers and escape to the surface. As the hydrocarbons break through the surface, the lighter
components will continue to escape to the atmosphere and leave behind heavier products
like bitumen, pitch, asphalt, and tar have been exploited by mankind since paleolithic times.
In 1859, Edwin L. Drake, also known as Colonel Drake, drilled the world’s first exploration
well into an anticline in Titusville, Pennsylvania to find oil 69.5 ft below the surface. Since
then, an enormous amount of wells have been drilled to extract oil from onshore reservoirs
in the Middle East, North America, etc. Drilling of submerged oil wells started just before
the turn of the nineteenth century: from platforms built on piles in the fresh waters of Grand
Lake St. Marys (Ohio) in 1891, and from piers extending into the salt water of the Santa
Barbara Channel (California) in 1896. Today, hydrocarbons are recovered from offshore
reservoirs that are located thousands of meters below the sea bed and in waters with a
depth up to 2,600 m in the Gulf of Mexico, the North Sea, and offshore from Brazil.

2.3 Multiscale Modeling of Permeable Rocks

All sedimentary rocks consist of a solid matrix with an interconnected void. The void pore
space allows the rocks to store and transmit fluids. The ability to store fluids is determined
by the volume fraction of pores (rock porosity), and the ability to transmit fluids (rock
permeability) is given by the interconnection of the pores.
Rock formations found in natural petroleum reservoirs are typically heterogeneous at
all length scales, from the micrometer scale of pore channels between the solid particles
making up the rock to the kilometer scale of a full reservoir formation. On the scale of
individual grains, there can be large variation in grain sizes, giving a broad distribution
of void volumes and interconnections. Moving up the scale, laminae may exhibit large
contrasts on the mm–cm scale in the ability to store and transmit fluids because of alter-
nating layers of coarse and fine-grained material. Laminae are stacked to form beds, which
are the smallest stratigraphic units. The thickness of beds varies from millimeters to tens of
meters, and different beds are separated by thin layers with significantly lower permeability.
Beds are, in turn, grouped and stacked into parasequences or sequences (parallel layers that
have undergone similar geologic history). Parasequences represent the deposition of marine
sediments, during periods of high sea level, and tend to be somewhere in the range from 1
to 100 m thick and have a horizontal extent of several kilometers.
The trends and heterogeneity of parasequences depend on the depositional environ-
ment. For instance, whereas shallow-marine deposits may lead to rather smoothly varying

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2.3 Multiscale Modeling of Permeable Rocks 29

Figure 2.7 Illustration of the hierarchy of flow models used in subsurface modeling. The length scales
are the vertical sizes of typical elements in the models.

permeability distributions with correlation lengths in the order of 10–100 m, fluvial reser-
voirs may contain intertwined patterns of sand bodies on a background with high clay
content; see Figure 2.12. The reservoir geology can also consist of other structures such as
shale layers (impermeable clays), which are the most abundant sedimentary rocks. Frac-
tures and faults, on the other hand, are created by stresses in the rock and may extend from
a few centimeters to tens or hundreds of meters. Faults may have a significantly higher or
lower ability to transmit fluids than the surrounding rocks, depending upon whether the
void space has been filled with clay material.
All these different length scales can have a profound impact on fluid flow. However, it
is generally not possible to account for all pertinent scales that impact the flow in a single
model. Instead, one has to create a hierarchy of models for studying phenomena occurring
at reduced spans of scales. This is illustrated in Figure 2.7. Microscopic models represent
the void spaces between individual grains and are used to provide porosity, permeability,
electrical, and elastic properties of rocks from core samples and drill cuttings. Mesoscopic
models are used to upscale these basic rock properties from the mm–cm scale of internal
laminations, through the lithofacies scale (∼50 cm), to the macroscopic facies association
scale (∼100 m) of geological models. In this book, we will primarily focus on another scale,
simulation models, which represent the last scale in the model hierarchy. Simulation models
are obtained by upscaling geological models and are either introduced out of necessity
because geological models contain more details than a flow simulator can cope with, or out
of convenience to provide faster calculation of flow responses.

2.3.1 Geological Characterization

To model a reservoir on a macroscopic scale, we basically need to represent its geology at a
level of detail that is sufficient for the purpose the model is built to serve: to visualize how

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 30 Modeling Reservoir Rocks

different experts perceive the reservoir, to provide estimates of hydrocarbon volumes, to

assist well planning and geosteering, or as input to geophysical analysis (seismic modeling,
rock mechanics) or flow simulations. For flow simulation, which is our primary concern in
this book, we need a volumetric description that decomposes the reservoir into a set of grid
cells (small 3D polygonal volumes) that are petrophysically and/or geometrically distinct
from each other. With a slight abuse of terminology, we will refer to this as the geological
model, which we distinguish from models describing the reservoir fluids and the forces that
cause their movement.
Geological models are generally built in a sequence of steps, using a combination of
stratigraphy (the study of rock layers and layering), sedimentology (study of sedimen-
tary rocks), structural geology (the study of how rock layers are deformed over time by
geological activity), diagenesis (the study of chemical, physical, and biological processes
that transform sediments to rock), and interpretation of measured data. The starting point
is usually a seismic interpretation, from which one obtains a representation of faults and
geological horizons that bound different geological units. The seismic interpretation is used
alongside a conceptual model in which geologists express how they believe the reservoir
looks like based on studies of geological history and geological outcrops. The result can be
expressed as a geometric model that consists of vertical or inclined surfaces representing
faults, and horizontal or slightly sloping surfaces representing horizons that subdivide
the reservoir volume into different geological units (zones). This zonation is obtained by
combining seismic interpretation that describes the gross geometry of the reservoir with
stratigraphic modeling and thickness information (isochores) obtained from well logs that
define the internal layering. Once a model of the structural and stratigraphic architecture
of the reservoir is established, the surface description can be turned into a 3D grid that is
populated with cell and face properties that reflect the geological framework.
Unfortunately, building a geological model for a reservoir is like finishing a puzzle
with most of the pieces missing. The amount of data available is limited due to the costs
of acquiring it, and the data that is obtained is measured on scales that may be quite
disparate from the geological features one needs to model. Seismic surveys give a sort
of X-ray image of the reservoir, but are both expensive and time-consuming and can only
give limited resolution; you cannot expect to see structures thinner than ten meters from
seismic data. Information on finer scales is available from various measuring tools lowered
into the wells to gather information of the rock in near-well region, e.g., by radiating the
reservoir and measuring the response. Well logs have quite limited resolution, rarely down
to centimeter scale. The most detailed information is available from rock samples (cores)
extracted from the well. The industry uses X-ray, CT scan, as well as electron microscopes
to gather high-resolution information from the cores, and the data resolution is only limited
by the apparatus at hand. However, information from cores and well logs can only tell you
how the rock looks like near the well, and extrapolating this information to the rest of the
reservoir is subject to great uncertainty. Moreover, due to high costs, one cannot expect
well logs and cores to be taken from every well. All these techniques give separately small
contributions that can help build a geological model. However, in the end we still have very

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2.3 Multiscale Modeling of Permeable Rocks 31

limited information available, considering that a petroleum reservoir can have complex
geological features that span across all types of length scales from a few millimeters to
several kilometers.
In summary, the process of making a geological model is generally strongly under-
determined. It is therefore customary to use a combination of deterministic and probabilis-
tic modeling to estimate the subsurface characteristics between the wells. Deterministic
modeling is used to specify large-scale structures such as faults, correlation, trends, and
layering, which are used as input and controls to geostatistical techniques [258, 57, 82] that
build detailed grid models satisfying statistical properties assumed for the petrophysical
heterogeneity. In recent years, process models that seek to mimic deposition and subse-
quent structural changes have also gained more popularity since they tend to produce less
artifacts in the geocellular models. Trends and heterogeneity in petrophysical properties
depend strongly on the structure of sedimentary deposits, and high-resolution petrophysical
realizations are thus rarely not built directly. Instead, one starts by building a rock model
that is based on the structural model and consists of a set of discrete rock bodies (facies)
that are specified to improve petrophysical classification and spatial shape. For carbonates,
the modeled facies are highly related to diagenesis, while the facies modeled for sandstone
reservoirs are typically derived from the depositional facies. By supplying knowledge of
the depositional environment (fluvial, shallow-marine, deep-marine, etc.) and conditioning
this to observed data, one can determine the geometry of the facies and how they are mixed.
To populate the modeled facies with petrophysical properties, it is common to replace
the facies with a volumetric grid, and use stochastic simulation to generate multiple discrete
realizations of the petrophysical properties defined on this grid. Each grid model has a plau-
sible heterogeneity and can contain millions of cells. The collection of all realizations gives
a measure of the uncertainty involved in the modeling. Hence, if the sample of realizations
(and the upscaling procedure that converts the geological models into simulation models)
is unbiased, then it is possible to supply predicted production characteristics, such as the
cumulative oil production, obtained from simulation studies with a measure of uncertainty.
You can find a lot more details about the complex process of characterizing a reservoir
and making geological models in the excellent textbook by Ringrose and Bentley [265].
The rest of this section will first introduce the concept of representative elementary vol-
umes, which underlies the continuum models used to describe subsurface flow and trans-
port, and then briefly discuss microscopic and mesoscopic modeling. This discussion is not
essential to understand what follows in the next few chapters, and if you are not interested,
you can therefore safely skip to Section 2.4, which discusses macroscopic modeling of
reservoir rocks.

2.3.2 Representative Elementary Volumes

Choosing appropriate modeling scales is often done by intuition and experience, and it
is hard to give general guidelines. An important concept in choosing model scales is the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 32 Modeling Reservoir Rocks

Vv
Porosity: φ =
V v + Vr

Figure 2.8 The concept of a representative elementary volume (REV), here illustrated for porosity,
which measures the fraction of void space to bulk volume.

notion of representative elementary volumes (REVs), which is the smallest volume over
which a measurement can be made and be representative of the whole. This concept is
based on the idea that petrophysical flow properties are constant on some intervals of scale;
see Figure 2.8. Representative elementary volumes, if they exist, mark transitions between
scales of heterogeneity, and present natural length scales for modeling.
To identify a range of length scales where REVs exist, e.g., for porosity, we move along
the length-scale axis from the micrometer-scale of pores toward the kilometer-scale of the
reservoir. At the pore scale, the porosity is a rapidly oscillating function equal to zero (in
solid rock) or to one (in the pores). Hence, obviously no REVs can exist at this scale. At the
next characteristic length scale, the core scale level, we find laminae deposits. Because the
laminae consist of alternating layers of coarse and fine grained material, we cannot expect
to find a common porosity value for the different rock structures. Moving further along the
length-scale axis, we may find long thin layers, perhaps extending throughout the entire
horizontal length of the reservoirs. Each of these individual layers may be nearly homo-
geneous because they are created by the same geological process, and probably contain
approximately the same rock types. Hence, at this scale it sounds reasonable to speak of
an REV. If we move to the high end of the length-scale axis, we start to group more and
more layers into families with different sedimentary structures, and REVs for porosity will
probably not exist.
The discussion gives grounds to claim that reservoir rock structures contain scales where
REVs may exist. From a general point of view, however, the existence of REVs in porous
media is highly disputable. A faulted reservoir, for instance, can have faults distributed
continuously both in length and aperture throughout the reservoir, and will typically have
no REVs. Moreover, no two reservoirs are identical, so it is difficult to capitalize from

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2.3 Multiscale Modeling of Permeable Rocks 33

previous experience. Indeed, porous formations in reservoirs may vary greatly, also in terms
of scales. Nevertheless, the concept of REVs can serve as a guideline when deciding what
scales to model.

2.3.3 Microscopic Models: The Pore Scale

Pore-scale models, as illustrated on the left-hand side of Figure 2.7, may be about the
size of a sugar cube and are based on measurements from core plugs obtained from well
trajectories during drilling. These rock samples are necessarily confined (in dimension) by
the radius of the well, although they lengthwise are only confined by the length of the well.
Three such rock samples are shown in Figure 2.9. The main methods for obtaining pore-
scale models from a rock sample are by studying thin slices using an electron microscope
with micrometer resolution or by CT scans. In the following, we will give a simplified
overview of flow modeling on this scale.
At the pore scale, the porous medium is either represented by a volumetric grid or by a
graph (see e.g., [242]). A graph is a pair (V ,E), where V is a set whose elements are called
vertices (or nodes), and E is a subset of V ×V whose elements are called edges. The vertices
are taken to represent pores, and the edges represent pore throats (i.e., connections between
pores). The flow process, in which one fluid invades the void space filled by another fluid,
is generally described as an invasion–percolation process dominated by capillary forces, in
which a fluid phase can invade a pore only if a neighboring pore is already invaded. Another
approach to multiphase modeling is through the use of the lattice Boltzmann method, which
represents the fluids as a set of particles that propagate and collide according to a set of
rules defined for interactions between particles of the same fluid phase, between particles
of different fluid phases, and between the fluids and the walls of the void space. A further
presentation of pore-scale modeling is beyond our scope here, but the interested reader is
encouraged to consult, e.g., [242, 44] and references therein.
From an analytical point of view, pore-scale modeling is very important as it represents
flow at the fundamental scale (or more loosely, where the flow really takes place), and

Figure 2.9 Three core plugs with a diameter of 3.81 cm and a height of 5 cm.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 34 Modeling Reservoir Rocks

hence provides the proper framework for understanding the fundamentals of porous media
flow. From a practical point of view, pore-scale modeling has a huge potential. Modeling
flow at all other scales can be seen as averaging of flow at the pore scale, and properties
describing the flow at larger scales are usually a mixture of pore-scale properties. At larger
scales, the complexity of flow modeling is often overwhelming, with large uncertainties in
determining flow parameters. Hence being able to single out and estimate the various fac-
tors determining flow parameters is invaluable, and pore-scale models can be instrumental
in this respect. However, to extrapolate properties from the pore scale to an entire reservoir
is very challenging, even if the entire pore space of the reservoir was known. In real life
you will of course not be anywhere close to knowing the entire pore space of a reservoir.

2.3.4 Mesoscopic Models

Models based on flow experiments on core plugs are by far the most common mesoscopic
models. The main equations describing flow are continuity of fluid phases and Darcy’s
law, which states that flow rate is proportional to pressure drop. The purpose of core-plug
experiments is to determine capillary pressure curves and the proportionality constant in
Darcy’s law that measures the ability to transmit fluids; see (1.2) in Section 1.4. To this
end, the sides of the core are insulated and flow is driven through the core. By measuring
flow rate versus pressure drop, one can estimate the proportionality constant for both single-
phase and multiphase flows.
In conventional reservoir modeling, the effective properties from core-scale flow exper-
iments are extrapolated to the macroscopic geological model, or directly to the simulation
model. Cores should therefore ideally be representative for the heterogeneous structures in
a typical grid block of the geological model, but flow experiments are usually performed
on relatively homogeneous cores that rarely exceed one meter in length. Flow at the core
scale is also more influenced by capillary forces than flow on a reservoir scale.
As a supplement to core-flooding experiments, it has in recent years become popular to
build 3D grid models to represent small-scale geological details like the bedding structure
and lithology (composition and texture). One example of such a model is shown in Figure
2.7. Effective flow properties for the 3D model can now be estimated in the same way as
for core plugs by replacing the flow experiment by flow simulations using rock properties
that are, e.g., based on the input from microscopic models. This way, one can incorporate
fine-scale geological details from lamina into the macroscopic reservoir models. However,
the process of extrapolating information from cores to build a geological model is largely
underdetermined and must generally be supplied with geological data from other sources.

2.4 Modeling Rock Properties

Describing the flow through a porous rock structure is largely a question of the scale of
interest. The size of the rock bodies forming a typical petroleum reservoir will be from
10 to 100 meters in the vertical direction and several hundred meters or a few kilometers

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2.4 Modeling Rock Properties 35

in the lateral direction. On this scale, it is clearly impossible to describe the storage and
transport in individual pores and pore channels, as discussed in Section 2.3.3, or through
individual laminae, as in Section 2.3.4. To obtain a description of the reservoir geology, one
builds models that attempt to reproduce the true geological heterogeneity in the reservoir
rock at the macroscopic scale by introducing macroscopic petrophysical properties based
on a continuum hypothesis and volume averaging over a sufficiently large REV. These
petrophysical properties are engineering quantities that are used as input to flow simulators,
and are not true geological or geophysical properties of the underlying media.
A geological model is a conceptual, three-dimensional representation of a reservoir,
whose main purpose is to give a spatial representation of the ability to store and trans-
mit fluids, besides giving location and geometry of the reservoir. The rock body itself is
modeled in terms of a volumetric grid, in which the layered structure of sedimentary beds
and the geometry of faults and large-scale fractures in the reservoir are represented by the
geometry and topology of the grid cells. Cell sizes in geocellular models can vary a lot,
but typical sizes are in the range of 0.1–1 m in the vertical direction and 10–50 m in the
horizontal direction. The petrophysical properties of the rock are represented as constant
values inside each grid cell (porosity and permeability) or as values attached to cell faces
(fault multipliers, fracture apertures, etc.). In the following sections, we will describe the
main rock properties in more detail. More details about the grid modeling will follow in
Chapter 3.

2.4.1 Porosity
The porosity φ of a porous medium is defined as the fraction of the bulk volume that
represents void space, which means that 0 ≤ φ < 1. Likewise, 1 − φ is the fraction
occupied by solid material (rock matrix). The void space generally consists of two parts:
interconnected pore space available to fluid flow, and disconnected pores (dead ends) that
are unavailable to flow. Only the first part is interesting for flow simulation, and porosity
is therefore usually a shorthand for the “effective porosity” that measures the fraction of
connected void space to bulk volume.
For a completely rigid medium, porosity is a static, dimensionless quantity that can be
measured in the absence of flow. Porosity is mainly determined by the pore and grain-size
distribution. Rocks with nonuniform grain size usually have smaller porosity than rocks
with a uniform grain size, because smaller grains tend to fill pores formed by larger grains.
For a bed of solid spheres of uniform diameter, the porosity depends on the packing, varying
between 0.2595 for a rhomboidal packing to 0.4764 for cubic packing. When sediments
are first deposited in water, they usually have a porosity of approximately 0.5, but as they
are buried, the water is gradually squeezed out and the void space between the mineral
particles decreases as the sediments are consolidated into rocks. For sandstone and lime-
stone, φ is in the range 0.05–0.5, although values outside this range may be infrequently
observed. Sandstone porosity is usually determined by the sedimentological process by

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 36 Modeling Reservoir Rocks

which the rock was deposited, whereas carbonate porosity is mainly a result of changes
taking place after deposition. Increase compaction (and cementation) causes porosity to
decrease with depth in sedimentary rocks. The porosity can also be reduced by minerals
that are deposited as water moves through pore spaces. For sandstone, the loss in porosity
is small, whereas shales loose their porosity very quickly. Shales are therefore unlikely
to be good reservoir rocks in the conventional sense, and will instead act like caprocks
having porosities that are orders of magnitude lower than those found in good sandstone and
carbonates.
For non rigid rocks, the porosity is usually modeled as a pressure-dependent parameter.
That is, one says that the rock is compressible, having a rock compressibility defined by
1 dφ d ln(φ)
cr = = , (2.1)
φ dp dp
where p is the overall reservoir pressure. Compressibility can be significant in some cases,
e.g., as evidenced by the subsidence observed in the Ekofisk area in the North Sea. For a
rock with constant compressibility, (2.1) can be integrated to give

φ(p) = φ0 ecr (p−p0 ), (2.2)

and for simplified models, it is common to use a linearization so that:

 
φ = φ0 1 + cr (p − p0 ) . (2.3)

Because the dimension of the pores is very small compared to any interesting scale for
reservoir simulation, one normally assumes that porosity is a piecewise continuous spatial
function. However, ongoing research aims to understand better the relation between flow
models on pore scale and on reservoir scale.

2.4.2 Permeability
Permeability is a basic flow property of a porous medium and measures its ability to trans-
mit a single fluid when the void space is completely filled with this fluid. This means that
permeability, unlike porosity, is a parameter that cannot be defined apart from fluid flow.
The precise definition of the (absolute, specific, or intrinsic) permeability K is as the pro-
portionality factor between the flow rate and an applied pressure or potential gradient ∇,
K
v = − ∇. (2.4)
μ
This relationship is called Darcy’s law after the french hydrologist Henry Darcy, who first
observed it in 1856 while studying flow of water through beds of sand [77]. In (2.4), μ is
the fluid viscosity and v is the superficial velocity, i.e., the flow rate divided by the cross-
sectional area perpendicular to the flow. This should not be confused with the interstitial
velocity u = φ −1 v, i.e., the rate at which an actual fluid particle moves through the
medium. We will come back to a more detailed discussion of Darcy’s law in Section 4.1.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2.4 Modeling Rock Properties 37

The SI-unit for permeability is m2 , which reflects the fact that permeability is determined
entirely by the geometric characteristics of the pores. However, it is more common to
use the unit “darcy” (D). The precise definition of 1 D ≈ 0.987 · 10−12 m2 involves
transmission of a 1 cP fluid through a homogeneous rock at a speed of 1 cm/s due to a
pressure gradient of 1 atm/cm. Translated to reservoir conditions, 1 D is a relatively high
permeability and it is therefore customary to specify permeabilities in millidarcys (md or
mD). Rock formations like sandstones tend to have many large or well-connected pores
and therefore transmit fluids readily. They are therefore described as permeable. Other
formations, like shales, may have smaller, fewer, or less interconnected pores and are
hence described as impermeable. Conventional reservoirs have permeabilities ranging from
0.1 mD to 20 D for liquid flow, and down to 10 mD for gases. In recent years, however, there
has been an increasing interest in unconventional resources, that is, gas and oil locked in
extraordinarily impermeable and hard rocks, with permeability values ranging from 0.1 mD
and down to 1 μD or lower. “Tight” reservoirs are defined as those having permeability
less than 0.1 mD. Compared with conventional resources, the potential volumes of tight
gas, shale gas, and shale oil are enormous, but cannot be easily produced at economic rates
unless stimulated, e.g., using a pressurized fluid to break up the rock to create new flow
paths (hydraulic fracturing). This book focuses exclusively on simulation of conventional
resources.
In (2.4), we tacitly assumed that K is a scalar quantity. However, permeability will
generally be a full tensor,

⎡ ⎤
Kxx Kxy Kxz
⎣
K = Kyx Kyy Kyz ⎦. (2.5)
Kzx Kzy Kzz

Here, the diagonal terms {Kxx ,Kyy ,Kzz } represent how the flow rate in one axial
direction depends on the pressure drop in the same direction. The off-diagonal terms
{Kxy ,Kxz,Kyx ,Kyz,Kzx ,Kzy } account for dependence between flow rate in one axial
direction and the pressure drop in perpendicular directions. A full tensor is needed to
model local flow in directions at an angle to the coordinate axes. Let us for instance
consider a layered system, for which the dominant direction of flow will generally be along
the layers. However, if the layers form an angle to the coordinate axes, a pressure drop
in one coordinate direction will produce flow at an angle to this direction. This type of
flow can only be modeled correctly with a permeability tensor with nonzero off-diagonal
terms. If the permeability can be represented by a scalar function K( x ), we say that the
permeability is isotropic as opposed to the anisotropic case where we need a full tensor
K(x ). To model a physical system, the anisotropic permeability tensor must be symmetric
because of the Onsager principle of reciprocal relations and positive definite because
the flow component parallel to the pressure drop should be in the same direction as the
pressure drop. As a result, a full-tensor permeability K may be diagonalized by a change of
basis.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 38 Modeling Reservoir Rocks

Since the porous medium is formed by deposition of sediments over thousands of years,
there is often a significant difference between permeability in the vertical and lateral direc-
tions, but no difference between the permeabilities in the two lateral directions. The per-
meability is obviously also a function of porosity. Assuming a laminar flow (low Reynolds
numbers) in a set of capillary tubes, one can derive the Carman–Kozeny relation,
1 φ3
K= , (2.6)
8τ A2v (1 − φ)2
which relates permeability to porosity φ, but also shows that the permeability depends on
local rock texture described by tortuosity τ and specific surface area Av . The tortuosity
is defined as the squared ratio of the mean arc–chord length of flow paths, i.e., the ratio
between the length of a flow path and the distance between its ends. The specific surface
area is an intrinsic and characteristic property of any porous medium that measures the
internal surface of the medium per unit volume. Clay minerals, for instance, have large
specific surface areas and hence low permeability. The quantities τ and Av can be cal-
culated for simple geometries, e.g., for engineered beds of particles and fibers, but are
rarely measured for reservoir rocks. Moreover, the relationship in (2.6) is highly idealized
and only gives satisfactory results for media consisting of grains that are approximately
spherical and have a narrow size distribution. For consolidated media and cases in which
rock particles are far from spherical and have a broad size distribution, the simple Carman–
Kozeny equation does not apply. Instead, permeability is typically obtained through macro-
scopic flow measurements.
Permeability is generally heterogeneous in space because of different sorting of par-
ticles, degree of cementation (filling of clay), and transitions between different rock for-
mations. Indeed, the permeability may vary rapidly over several orders of magnitude;
local variations in the range 1 mD to 10 D are not unusual in field and sector models.
The heterogeneous structure of a porous rock formation is a result of the deposition and
geological history and will therefore vary strongly from one formation to another, as we
will see in a few of the examples in Section 2.5.
Production of fluids may also change the permeability. When temperature and pressure
are changed, microfractures may open and significantly change the permeability. Further-
more, since the definition of permeability involves a certain fluid, different fluids will
experience different permeability in the same rock sample. Such rock–fluid interactions
are discussed in Chapter 8.

2.4.3 Other Parameters

Not all the bulk volume in a reservoir zone consist of reservoir rocks. To account for the fact
that some portion of a cell may contain impermeable shale, it is common to introduce the
so-called net-to-gross (N/G) property, which is a number in the range 0–1 that represents
the fraction of reservoir rock in a cell. To get the effective porosity of a given cell, one must
multiply the porosity and N/G value of the cell. (The N/G values also act as multipliers

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2.5 Property Modeling in MRST 39

for lateral transmissibilities, which we will come back to later in the book.) A zero value
means that the corresponding cell only contains shale (either because the porosity, the N/G
value, or both are zero), and such cells are by convention typically not included in the active
model.
Faults can either act as conduits for fluid flow in subsurface reservoirs or create flow
barriers and introduce compartmentalization that severely affects fluid distribution and/or
reduces recovery. On a reservoir scale, faults are generally volumetric objects that can be
described in terms of displacement and petrophysical alteration of the surrounding host
rock. However, lack of geological resolution in simulation models means that fault zones
are commonly modeled as surfaces that explicitly approximate the faults’ geometrical prop-
erties. To model the hydraulic properties of faults, one introduces so-called multipliers that
alter the ability to transmit fluid between two neighboring cells. Multipliers can also model
other types of subscale features that affect the communication between grid blocks, e.g.,
thin mud layers resulting from a flooding event, which may partially cover the sand bodies
and reduce vertical communication. It is also common to (ab)use multipliers to increase or
decrease the flow in certain parts of the model to calibrate simulated reservoir responses
to historic data (production curves from wells, etc.). More details about multipliers will be
given later in the book.

2.5 Property Modeling in MRST

All flow and transport solvers in MRST assume that rock parameters are represented as
fields in a structure array (struct), which by convention is called rock. You do not need
to follow this convention, but rock must contain two subfields called poro and perm. The
porosity field rock.poro is a vector with one value for each active cell in the associated
grid model. The permeability field rock.perm can either contain a single column for an
isotropic permeability, two or three columns for a diagonal permeability (in 2D and 3D,
respectively), or six columns for a symmetric, full-tensor permeability. In the latter case,
cell number i has the permeability tensor
⎡ ⎤

 K1 (i) K2 (i) K3 (i)
K1 (i) K2 (i)
Ki = , Ki = ⎣K2 (i) K4 (i) K5 (i)⎦ ,
K2 (i) K3 (i)
K3 (i) K5 (i) K6 (i)

where Kj (i) is the entry in column j and row i of rock.perm. Full-tensor, nonsymmetric
permeabilities are currently not supported. In addition to porosity and permeability, MRST
supports a field called ntg representing the net-to-gross ratio, which either consists of a
scalar or a single column with one value per active cell.
The rest of this section demonstrate how to generate and specify rock parameters in
MRST and briefly outlines a few realistic models. The discussion will also expose you to a
standard visualization capabilities in MRST. Scripts containing the code lines necessary to
reproduce all figures can be found in the rock subdirectory of the book module.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 40 Modeling Reservoir Rocks

2.5.1 Homogeneous Models

Homogeneous models are very simple to specify. To exemplify, consider a square 10 × 10
grid model with a uniform porosity of 0.2 and isotropic permeability equal 200 mD:
G = cartGrid([10 10]);
rock = makeRock(G, 200*milli*darcy, 0.2);

Because MRST works in SI units, it is important to convert from the field unit “darcy”
to the SI unit “meters2 .” Here, we did this by multiplying with milli and darcy, which
are two functions that return the corresponding conversion factors. Alternatively, we could
have used the conversion function convertFrom(200, milli*darcy). Homogeneous,
anisotropic permeability can be specified in the same way:
rock = makeRock(G, [100 100 10].*milli*darcy, 0.2);

2.5.2 Random and Lognormal Models

Given the difficulty of measuring rock properties, it is common to use geostatistical meth-
ods to make realizations of porosity and permeability. MRST contains two very simpli-
fied methods for generating geostatistical realizations. For more realistic geostatistics, you
should use GSLIB [82] or a commercial geomodeling software.
In our first example, we will generate the porosity φ as a Gaussian field. To get a crude
approximation to the permeability–porosity relationship, we assume that our medium is
made up of uniform spherical grains of diameter dp = 10 μm, for which the specific surface
area is Av = 6/dp . Using the Carman–Kozeny relation (2.6), we can then calculate the
isotropic permeability K from

1 φ 3 dp2
K= ,
72τ (1 − φ)2

where we further assume that τ = 0.81. As a simple approximation to a Gaussian porosity

field, we generate a field of independent normally distributed variables using MATLAB’s
built-in rand function, convolve it with a Gaussian kernel, and scale the result to the
interval [0.2,0.4]:
G = cartGrid([50 20]);
p = gaussianField(G.cartDims, [0.2 0.4], [11 3], 2.5);
K = p.^3.*(1e-5)^2./(0.81*72*(1-p).^2);
rock = makeRock(G, K(:), p(:));

The left plot of Figure 2.10 shows the resulting porosity field, whereas the right plot
shows the permeability obtained for a 3D realization with 50 × 20 × 10 cells generated in
the same way.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2.5 Property Modeling in MRST 41

20

15

10

0
0 5 10 15 20 25 30 35 40 45 50

0.2 0.22 0.24 0.26 0.28 0.3 0.32 0.34 0.36 0.38 0.4

plotCellData(G,rock.poro,'EdgeColor','none'); plotCellData(G,convertTo(rock.perm,milli*darcy));
colorbar('horiz'); axis equal tight; colorbar('horiz'); axis equal tight; view(3);

Figure 2.10 The left plot shows a 50 × 20 porosity field generated as a Gaussian field with a larger
filter size in x-direction than in the y-direction. The right plot shows the permeability field computed
from the Carman–Kozeny relation for a similar 50 × 20 × 10 porosity realization computed with filter
size [3,3,3].

In the second example, we use the same methodology to generate layered realizations,
defined so that the permeability is lognormally distributed in each geological layer and
independent of the other layers. Each layer can be represented by several grid cells in the
vertical direction. Rather than using a simple Cartesian grid, we generate a stratigraphic
grid with wavy geological faces and a single fault. Chapter 3 describes this type of industry-
standard grids in more detail.
G = processGRDECL(simpleGrdecl([50 30 10], 0.12));
K = logNormLayers(G.cartDims, [100 400 50 350], 'indices', [1 2 5 7 11]);

Here, we have specified four geological layers of different thickness. From top to bottom,
which is the default numbering in stratigraphic grids, the first layer is one cell thick and has
a mean permeability value of 100 mD, the second layer is three cells thick and has mean
permeability of 400 mD, the third layer is two cells thick and has mean value 50 mD, and
the fourth layer is four cells thick and has mean value 350 mD. To specify this, we use an
indirection map; that is, if Km is the n-vector of mean permeabilities and L is the (n + 1)-
vector of indices, the value Km(i) is assigned to vertical layers number L(i) to L(i+1)-1.
Figure 2.11 shows the resulting permeability field.

2.5.3 The 10th SPE Comparative Solution Project: Model 2

The Society of Petroleum Engineers (SPE) has developed a series of benchmarks for
comparing computational methods and simulators. The first nine benchmarks focus on
black-oil, compositional, dual-porosity, thermal, and miscible simulations, as well as
horizontal wells and gridding techniques. The 10th SPE Comparative Solution Project [71]

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 42 Modeling Reservoir Rocks

plotCellData(G,log10(K),'EdgeColor','k');
view(45,30);
axis tight off,
set(gca,'DataAspect',[0.5 1 1])
h=colorbar('horiz');
ticks=25*2.^[0:5];
set(h,'XTick', log10(ticks), ...
'XTickLabel', ticks);

Figure 2.11 A stratigraphic grid with a single fault and four geological layers, each with a lognormal
permeability distribution.

was posed as a benchmark for upscaling methods, but the second data set of this benchmark
has later become very popular as a general test case within the academic community. The
data set is a 3D geostatistical realization from the Jurassic Upper Brent formations from
the North Sea, which, e.g., contains the giant fields of Statfjord, Gullfaks, Oseberg, and
Snorre. The main features of the model are the permeability and porosity fields given on a
60×220×85 Cartesian grid, in which each cells is of size 20 ft×10 ft×2 ft. In this specific
model, the top 35 cell layers having a total height of 70 ft represent the shallow-marine
Tarbert formation, and the lower 50 layers having a height of 100 ft represent the fluvial
Ness formation. The original geostatistical model was developed in the PUNQ project
[111], and later the horizontal dimensions were scaled by a factor 1/3 to make it more
heterogeneous. The model is structurally simple but is highly heterogeneous, and, for this
reason, some describe it as a “simulator-killer.” On the other hand, the fact that the flow
is dictated by the strong heterogeneity means that streamline methods will be particularly
efficient for this model [4].
The SPE 10 data set is used in a large number of publications and is publicly avail-
able from the SPE website (http://www.spe.org/web/csp/). MRST supplies a module called
spe10 that downloads, reorganizes, and stores the data set in *.mat files for later use.
The module also contains routines that extract (subsets of) the petrophysical data and set
up simulation models and appropriate data structures representing grids, petrophysics, and
wells. Alternatively, the data set can be downloaded using mrstDatasetGUI. We start by
loading the petrophysical data
% load SPE 10 data set
mrstModule add spe10;
rock = getSPE10rock(); p=rock.poro; K=rock.perm;

Because of the simple Cartesian grid topology, we can use standard MATLAB functionality
to visualize the heterogeneity in the permeability and porosity (full details can be found in
the script rocks/showSPE10.m):

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2.5 Property Modeling in MRST 43

slice( reshape(p,60,220,85), [1 220], 60, [1 85]);

shading flat, axis equal off, set(gca,'zdir','reverse'), box on;
colorbar('horiz');

% show Kx
slice( reshape(log10(K(:,1)),60,220,85), [1 220], 60, [1 85]);
shading flat, axis equal off, set(gca,'zdir','reverse'), box on;
h=colorbar('horiz');
set(h,'XTickLabel',10.^[get(h,'XTick')]);
set(h,'YTick',mean(get(h,'YLim')),'YTickLabel','mD');

Figure 2.12 shows porosity and permeability; the permeability tensor is diagonal with
equal permeability in the two horizontal coordinate directions. Both formations are char-
acterized by large permeability variations, 8–12 orders of magnitude, but are qualitatively
different. The Tarbert consists of sandstone, siltstone, and shales, and comes from a tidally
influenced, transgressive, shallow-marine deposit; in other words, a deposit that has taken
place close to the coastline; see Figure 2.1. The formation has good communication in the
vertical and horizontal directions. The fluvial Ness formation has been deposited by rivers
or running water in a delta-plain, continental environment (see Figures 2.1 and 2.2), leading
to a spaghetti of well-sorted, high-permeable sandstone channels with good communication

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

mD mD
0.001 0.01 0.1 1 10 100 1000 10000 1e−07 1e−06 1e−05 0.0001 0.001 0.01 0.1 1 10 100 1000

Figure 2.12 Rock properties for the SPE 10 model. The upper plot shows the porosity, the lower
left the horizontal permeability, and the lower right the vertical permeability. (The permeabilities are
shown using a logarithmic color scale.)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 44 Modeling Reservoir Rocks
4
x 10
3
Ness
Tarbert
2.5

1.5

0.5

0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

4
x 10
3
Tarbert
Ness
2.5

1.5

0.5

0
−8 −6 −4 −2 0 2 4

Figure 2.13 Histogram of rock properties for the SPE 10 model: φ (upper plot), log Kx (lower left),
and log Kz (lower right) The Tarbert formation is shown in blue and the Ness formation in red.

(long correlation lengths) imposed on a low-permeable background of shales and coal,

which gives low communication between different sand bodies. The porosity field has a
large span of values, and approximately 2.5% of the cells have zero porosity and should be
considered as being inactive.
Figure 2.13 shows histograms of the porosity and the logarithm of the horizontal and
vertical permeabilities. The nonzero porosity values and the horizontal permeability of the
Tarbert formation appear to follow a normal and lognormal distribution, respectively. The
vertical permeability follows a bimodal distribution. For the Ness formation, the nonzero
porosities and the horizontal permeability follow bi-modal normal and lognormal distribu-
tions, respectively, as is to be expected for a fluvial formation. The vertical permeability is
trimodal. To five significant digits, the ratio of vertical to horizontal permeability has two
unique values 10−4 and 0.3, which together with the K index can be used to distinguish
the three different rock types (facies) in the model: sheet sand and mudstone in the Tarbert
formation, and channel sand and mudstone in Upper Ness.

2.5.4 The Johansen Formation

The Johansen formation is located in the deeper part of the Sognefjord delta, 40–90 km
offshore Mongstad on the west coast of Norway. Some years ago, a gas power plant with

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2.5 Property Modeling in MRST 45

carbon capture and storage was planned at Mongstad, and the water-bearing Johansen
formation was a possible candidate for storing the captured CO2 . The Johansen formation is
part of the Dunlin group, and is interpreted as a large sandstone delta 2,200–3,100 m below
sea level, which is limited above by the Dunlin shale and below by the Amundsen shale.
The average thickness of the formation is roughly 100 m, and the aquifer extends laterally
up to 100 km in the north–south direction and 60 km in the east–west direction. The aquifer
has good sand quality and lies at a depth where CO2 would undoubtedly be in supercritical
phase, and would thus be ideal for carbon storage. With average porosities of approximately
25%, this implies that the theoretical storage capacity of the Johansen formation is more
than one gigatonne of CO2 [97]. The Troll field, one of the largest gas fields in the North
Sea, is located some 500 m above the northwestern parts of the Johansen formation.
A previous research project generated several models of the Johansen formation using
so-called corner-point grids. We will discuss this format in more detail in Section 3.3.1.
For now, it is sufficient to note that corner-point grids consist of (distorted) hexahedral
cells organized according to an underlying Cartesian topology. Altogether, there are five
models: one full-field model (149 × 189 × 16 grid), three homogeneous sector models
(100×100×n for n = 11,16,21), and one heterogeneous sector model (100×100× 11). All
models are publicly available on the MRST website and can, for instance, be downloaded
to MRST using the mrstDatasetGUI function. Herein, we consider the heterogeneous
sector model. You can find all statements used in the following analysis in the script
rocks/showJohansenNPD.m.
We start by reading the grid geometry from an industry-standard input format, which
we discuss in more detail in the next section, and then construct a corresponding MRST
grid structure. The rock properties are given as plain ASCII files that each contains 110,000
values. Some of these values correspond to logical cell indices that are not part of the actual
simulation model and can be discarded if we use the indirection map G.cells.indexMap to
extract the correct subset:
G = processGRDECL(readGRDECL('NPD5.grdecl'));
p = load('NPD5_Porosity.txt')'; p = p(G.cells.indexMap);
K = load('NPD5_Permeability.txt')'; K=K(G.cells.indexMap);

In the model, the Johansen formation is represented by five layers of cells, whereas the
low-permeable Dunlin shale above and the Amundsen shale below are represented by five
and one cell layers. The Johansen formation consists of approximately 80% sandstone and
20% claystone, whereas the Amundsen formation consists of siltstones and shales; see [97]
for more details. In the left plot of Figure 2.14, the Johansen sand is clearly distinguished as
a wedge shape that is pinched out in the front part of the model and splits the shales laterally
in two at the back. The right plot shows the good reservoir rocks distinguished as cells with
a porosity value larger than 0.1. The permeability tensor is assumed to be diagonal, with
the vertical permeability equaling one-tenth of the horizontal permeability. Hence, only the
x-component Kx is given in the data file. Figure 2.15 shows three different plots of the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 46 Modeling Reservoir Rocks

args = {'EdgeAlpha'; 0.1; 'EdgeColor'; 'k'}; plotGrid(G,'FaceColor','none', args{:});

plotCellData(G, p, args{:}); view(-45,15), plotCellData(G, p, find(p>0.1), args{:})
axis tight off, zoom(1.15), view(-15,40); axis tight off, zoom(1.15),
caxis([0.1 0.3]), colorbar; caxis([0.1 0.3]), colorbar;

Figure 2.14 Porosity for the Johansen data set “NPD5.” The left plot shows porosity for the whole
model, whereas in the right plot we have masked the low-porosity cells in the Amundsen and Dunlin
formations.

Figure 2.15 Permeability for the Johansen data set “NPD5.” The left plot shows permeability for the
whole model, the middle plot shows the Johansen sand and Amundsen shale, whereas the right plot
only shows permeability of the Johansen sand.

permeability. The first plot presents the whole permeability field with a logarithmic color
scale. In the second plot, we have filtered out the Dunlin shale above Johansen but not the
Amundsen shale below. The third plot depicts the permeability in the Johansen formation
using a linear color scale, which clearly shows the depth trend that was used to model the
heterogeneity.

2.5.5 SAIGUP: Shallow-Marine Reservoirs

Most commercial simulators use a combination of an input language and a set of data files
to describe and set up a simulation model of a reservoir. In this book, we only discuss the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2.5 Property Modeling in MRST 47

ECLIPSE input format, which has emerged as an industry standard for describing static
and dynamic properties of a reservoir system, from the reservoir rock, via production and
injection wells and up to connected top-side facilities. ECLIPSE input decks use keywords
to signify and separate the different data elements that comprise a full model. These key-
words define a detailed language you can use to specify how the data elements should be
put together and modify each other to form a full spatio-temporal model of a reservoir. In
the most general form, an ECLIPSE input deck consists of eight sections of keywords that
must come in a prescribed order. However, some of the sections are optional and may not
always be present. The order of the keywords within each section is arbitrary, except in
the section that defines wells and gives operating schedule, etc. Altogether, the ECLIPSE
format consists of thousands of keywords, and describing them all is far beyond the scope
of this book. Other simulators use a somewhat different syntax, but the principles of the
overall input description is often similar.
In the following, we briefly outline some of the most common keywords from the
GRID section that describes the reservoir geometry and petrophysical properties. To pro-
vide you with a basic understanding of the required input for simulating real-life reservoir
models, we focus on the model ingredients and not on the specific syntax. For brevity,
we also do not go through all MATLAB and MRST statements used to visualize the
different data elements. All the necessary details can be found in the script rocks/show
SAIGUP.m.
As our example of a realistic petroleum reservoir, we consider a model from the
SAIGUP study [203], whose purpose was to conduct a sensitivity analysis of how
geological uncertainties impacts production forecasting in clastic reservoirs. To this end,
the study generated a large suite of geostatistical realizations and structural models to
represent a wide span of shallow-marine sedimentological reservoirs. The SAIGUP study
focused on shoreface reservoirs in which the deposition of sediments is caused by variation
in sea level, so that facies form belts in a systematic pattern (river deposits create curved
facies belts, wave deposits create parallel belts, etc.). Sediments are in general deposited
when the sea level is increasing. No sediments are deposited during decreasing sea levels;
instead, the receding sea may affect the appearing shoreline and cause the creation of a
barrier. All models are synthetic, but contain representative examples of the complexities
seen in real-life reservoirs.
One of the many SAIGUP realizations is publicly available from the MRST website
and comes in the form of a GZip-compressed TAR file (SAIGUP.tar.gz) that contains
the structural model as well as petrophysical parameters. You can download the data set
using the mrstDatasetGUI function. Here, however, we unpack the data set manually for
completeness of presentation. Assuming that we have already downloaded the archive file
SAIGUP.tar.gz with the model realization, we extract the data files and place them in a
standardized path relative to the root directory of MRST:
untar('SAIGUP.tar.gz', fullfile(ROOTDIR, 'examples', 'data', 'SAIGUP'))

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 48 Modeling Reservoir Rocks

This will create a new directory containing seventeen data files that comprise the structural
model, various petrophysical parameters, etc.:
028_A11.EDITNNC 028.MULTX 028.PERMX 028.SATNUM SAIGUP.GRDECL
028_A11.EDITNNC.001 028.MULTY 028.PERMY SAIGUP_A1.ZCORN
028_A11.TRANX 028.MULTZ 028.PERMZ SAIGUP.ACTNUM
028_A11.TRANY 028.NTG 028.PORO SAIGUP.COORD

The main file is SAIGUP.GRDECL, which lists the sequence of keywords that specifies how
the data elements found in the other files should be put together to make a complete model
of the reservoir rock. The remaining files represent different keywords: the grid geometry
is given in files SAIGUP_A1.ZCORN and SAIGUP.COORD, the porosity in 028.PORO, the
permeability tensor in the three 028.PERM* files, net-to-gross properties in 028.NTG, the
list of active cells in SAIGUP.ACTNUM, transmissibility multipliers that modify the flow
connections between different cells are given in 028.MULT*, etc. For now, we rely entirely
on MRST’s routines for reading ECLIPSE input files; more details about corner-point grids
and the ECLIPSE input format will follow later in the book, starting in Chapter 3.
The SAIGUP.GRDECL file contains seven of the eight possible sections of a full input
deck. The deckformat module in MRST contains a comprehensive set of input routines that
enable you to read the most important keywords and options supported in these sections.
For the SAIGUP model, it is mainly the sections describing static reservoir properties
that contain complete and useful information, and we will therefore use the much simpler
function readGRDECL from MRST core to read and interpret the GRID section of the input
deck:
grdecl = readGRDECL(fullfile(ROOTDIR, 'examples', ...
'data', 'SAIGUP','SAIGUP.GRDECL'));

This statement parses the input file and stores the content of all keywords it recognizes in
the structure grdecl:
grdecl =
cartDims: [40 120 20]
COORD: [29766x1 double]
ZCORN: [768000x1 double]
ACTNUM: [96000x1 int32]
PERMX: [96000x1 double]
PERMY: [96000x1 double]
PERMZ: [96000x1 double]
MULTX: [96000x1 double]
MULTY: [96000x1 double]
MULTZ: [96000x1 double]
PORO: [96000x1 double]
NTG: [96000x1 double]
SATNUM: [96000x1 double]

The first four data fields describe the grid, and we will come back to these in Section
3.3.1. In the following, we discuss the next eight data fields, which contain the petrophys-
ical parameters. We also look briefly at the last data field, which delineates the reservoir
into different (user-defined) rock types that can be used to associate different rock-fluid
properties.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2.5 Property Modeling in MRST 49

Figure 2.16 The structural SAIGUP model. The left plot shows the full model with faults marked
in red and inactive cells marked in yellow, whereas the right plot shows only the active parts of the
model seen from the opposite direction.

Recall that MRST uses strict SI conventions in all of its internal calculations. The
SAIGUP model follows the ECLIPSE “METRIC” conventions (permeabilities in mD and
so on). We thus use functions getUnitSystem and convertInputUnits to convert the input
data to MRST’s internal unit conventions.
usys = getUnitSystem('METRIC');
grdecl = convertInputUnits(grdecl, usys);

Having converted the units properly, we generate a space-filling grid and extract the petro-
physical properties:
G = processGRDECL(grdecl);
G = computeGeometry(G);
rock = grdecl2Rock(grdecl, G.cells.indexMap);

The first statement takes the description of the grid geometry and constructs an unstruc-
tured MRST grid represented with the data structure outlined in Section 3.4. The sec-
ond statement computes a few geometric primitives like cell volumes, centroids, etc., as
discussed on page 93. The third statement constructs a rock object containing porosity,
permeability, and net-to-gross.
For completeness, we first show a bit more details of the structural model in Figure 2.16.
The left plot shows the whole 40×120×20 grid model,1 where we in particular should note
the disconnected cells marked in yellow that are not part of the active model. The large fault
throw that disconnects the two parts is most likely a modeling artifact introduced to clearly
distinguish the active and inactive parts of the model. A shoreface reservoir is bounded by
faults and geological horizons, but faults also appear inside the reservoir, as the right plot
in Figure 2.16 shows. Faults and barriers will typically have a pronounced effect on the

1 To not confuse the reader, I emphasize that only the active part of the model is read with the processGRDECL routine. How to
also include the inactive part, will be explained in more detail in Chapter 3.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 50 Modeling Reservoir Rocks

p = reshape(grdecl.PORO, G.cartDims); args = {'EdgeAlpha'; 0.1; 'EdgeColor'; 'k'};

slice(p, 1, 1, 1); view(-135,30), shading flat, plotCellData(G,rock.poro, args{:});
axis equal off, axis tight off; set(gca,'DataAspect',[1 1 0.1]);
set(gca,'ydir','reverse','zdir','reverse') view(-65,55); zoom(1.4); camdolly(0,-0.2,0)
colorbar('horiz'); caxis([0.01 0.3]); colorbar('horiz'); caxis([0.1 0.3])

Figure 2.17 Porosity for the SAIGUP model. The left plot shows porosity as generated by
geostatistics in logical ij k space. The right plot shows the porosity mapped to the structural model
shown in Figure 2.16.

flow pattern, and having an accurate representation is important to produce reliable flow
predictions.
The petrophysical parameters for the model were generated by a geostatistical algorithm
on a regular 40 × 120 × 20 Cartesian grid, as illustrated in the left plot of Figure 2.17, and
then mapped onto the structural model, as shown in the plot to the right. A bit simplified,
one can view the Cartesian grid model as representing the rock body at geological “time
zero,” when the sediments have been deposited and have formed a stack of horizontal
grid layers. From geological time zero and up to now, geological activity has introduced
faults and deformed the layers, resulting in the structural model seen in the left plot of
Figure 2.17.
Having seen the structural model, we continue to study the petrophysical parameters.
The grid cells are thought to be larger than the laminae of our imaginary reservoir, hence
each grid block will generally contain both reservoir rock (with sufficient permeability) and
impermeable shale. This is modeled using the net-to-gross ratio, rock.ntg, shown in Figure
2.18 along with the horizontal and vertical permeability. The plotting routines are exactly
the same as for the porosity in Figure 2.17, but with different data and slightly different
specification of the colorbar. From the figure, we clearly see that the model has a large
content of shale and thus low permeability along the top, but we also see high-permeable
sand bodies that cut through the low-permeable top. In general, the permeabilities seem to
correlate well with the sand content given by the net-to-gross parameter.
Some parts of the sand bodies are partially covered by mud that strongly reduces the ver-
tical communication, most likely because of flooding events. These mud-draped surfaces
occur on a sub-grid scale and are modeled through a multiplier value (MULTZ) associated

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2.5 Property Modeling in MRST 51

Figure 2.18 The upper plots show the permeability for the SAIGUP model, using a logarithmic color
scale, with horizontal permeability to the left and vertical permeability to the right. The lower-left plot
shows net-to-gross, i.e., the fraction of reservoir rock per bulk volume. The lower-right plot shows
regions of the reservoir where reduced vertical communication is modeled by vertical multiplier
values less than unity.

with each cell, which takes values between zero and one and can be used to manipulate
the effective communication (the transmissibility) between a given cell (i,j,k) and the cell
immediately above (i,j,k + 1). For completeness, we remark that the horizontal multiplier
values (MULTX and MULTY) play a similar role for vertical faces, but are equal one in (almost)
all cells for this particular realization.
To further investigate the heterogeneity of the model, we next look at histograms of
the porosity and the permeabilities, as we did for the SPE 10 example (the MATLAB
statements are omitted since they are almost identical). In Figure 2.19, we clearly see that
the distributions of porosity and horizontal permeability are multi-modal in the sense that
five different modes can be distinguished, corresponding to the five different facies used in
the petrophysical modeling.
It is common modeling practice that different rock types are assigned different rock-
fluid properties (relative permeability and capillary functions), more details about such
properties will be given later in the book. Different rock types are represented using the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 52 Modeling Reservoir Rocks
3500
Horizontal
Vertical
3000

2500

2000

1500

1000

500

0
−4 −3 −2 −1 0 1 2 3 4

Figure 2.19 Histogram of the porosity (left) and the logarithm of the horizontal and vertical
permeability (right) for the shallow-marine SAIGUP model. Since the reservoir contains five different
facies, the histograms are multimodal. See also Figure 2.20.

SATNUM keyword, and by inspection of the SATNUM field in the input data, we see that
the model contains six different rock types as depicted in Figure 2.20. For completeness,
the figure also shows the permeability distribution inside each rock type. Interestingly, the
permeability distribution is multimodal for at least two of the rock types.
Finally, to demonstrate the large difference in heterogeneity resulting from different
depositional environments, we compare the realization studied so far in this example with
another realization; Figure 2.21 shows porosities and rock-type distribution. Whereas
the original realization corresponds to a depositional environment with a flat shoreline,
the other realization has a two-lobed shoreline, giving distinctively different facies belts.
The figure also clearly demonstrates how the porosity (which depends on the grain-size
distribution and packing) varies with the rock types. This can be confirmed by a quick
analysis:
for i=1:6, pavg(i) = mean(rock.poro(SN==i));
navg(i) = mean(rock.ntg(SN==i)); end

pavg = 0.0615 0.1883 0.1462 0.1145 0.0237 0.1924

navg = 0.5555 0.8421 0.7554 0.6179 0.3888 0.7793

In other words, rock types two and six are good sands with high porosity, three and four
have intermediate porosity, whereas one and five correspond to less quality sand with a
high clay content and hence low porosity.

Computer exercises

2.5.1 Look at the correlation between the porosity and the permeability for the SPE 10
data set. Do you see any artifacts, and if so, how would you explain them? (Hint:
plot φ versus log K.)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 2.5 Property Modeling in MRST 53

% Key statements
SN = grdecl.SATNUM(G.cells.indexMap);
plotCellData(G,SN, args{:});
colorbar('horiz'); caxis([0.5 6.5])
j = jet(60); colormap(j(1:10:end,:))
plotCellData(G,SN, find(SN==1), args{:});
plotCellData(G,SN, find(SN==5), args{:});

Figure 2.20 The upper-left plot shows the rock type distribution for SAIGUP. The right column
shows the six rock types grouped in pairs; from top to bottom, rock types number 1 and 5, 2 and 4,
and 3 and 6. The bottom part of the figure shows histograms of the lateral permeability in units [mD]
for each of the six rock types found in the model.

2.5.2 Download the CaseB4 models that represent a sector model with intersecting faults.
Pick at least one model realization and set homogeneous and random petrophysical
data as discussed in Sections 2.5.1 and 2.5.2.
2.5.3 The 58 × 48 × 10 permeability field given in rock1.mat in the book module
contains an unusual geological structure. Can you find what it is? (Hint: if you
want to explore the model interactively, you can try to use plotToolbar from the
mrst-gui module).
2.5.4 Download the BedModels1 and BedModel2 data sets that represent sedimentary beds
similar to the facies model shown in Figure 2.7. Use the techniques introduced in
Sections 2.5.3–2.5.5 to familiarize yourself with these models:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 54 Modeling Reservoir Rocks

Figure 2.21 Comparison of porosity (left) and the distribution of rock types (right) for two different
SAIGUP realizations.

• look at porosities and permeabilities in physical space

• compare with the same quantities in ij k space
• find models that have facies information and look at the distribution of petro-
physical properties inside each facies
2.5.5 Modify the simpleGravityColumn example from Section 1.4 so that it uses the geo-
metry and petrophysical data in the mortarTestModel or periodic Tilted models
from the BedModels1 data set instead. Can you explain what you observe?

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3
Grids in Subsurface Modeling

The basic geological description of a petroleum reservoir or an aquifer system will typically
consist of two sets of surfaces. Geological horizons are lateral surfaces describing the
bedding planes that delimit the rock strata, whereas faults are vertical or inclined surfaces
along which the strata may have been displaced by geological processes. In this chapter, we
discuss how to turn the basic geological description into a discrete model that can be used
to formulate various computational methods, e.g., for solving the equations that describe
fluid flow.
A grid is a tessellation of a planar or volumetric object by a set of contiguous simple
shapes referred to as cells. Grids can be described and distinguished by their geometry,
which gives the shape of the cells that form the grid, and their topology that tells how the
individual cells are connected. In 2D, a cell is in general a closed polygon for which the
geometry is defined by a set of vertices and a set of edges that connect pairs of vertices
and define the interface between two neighboring cells. In 3D, a cell is a closed polyhedron
for which the geometry is defined by a set of vertices, a set of edges that connect pairs
of vertices, and a set of faces (surfaces delimited by a subset of the edges) that define the
interface between two different cells; see Figure 3.1. Herein, we assume that all cells are
nonoverlapping, so that each point in the planar/volumetric object represented by the grid
is either inside a single cell, lies on an interface or edge, or is a vertex. Two cells that share
a common face are said to be connected. Likewise, one can also define connections based
on edges and vertices. The topology of a grid is defined by the total set of connections,
which is sometimes also called the connectivity of the grid.
When implementing grids in modeling software, one always has the choice between
generality and efficiency. To represent an arbitrary grid, it is necessary to explicitly store
the geometry of each cell in terms of vertices, edges, and faces, as well as storing the
connectivity among cells, faces, edges, and vertices. However, as we will see later, huge
simplifications can be made for particular classes of grids by exploiting regularity in the
geometry and structures in the topology. Consider, for instance, a planar grid consisting
of rectangular cells of equal size. Here, the topology can be represented by two indices,
and one only needs to specify a reference point and the two side lengths of the rectan-
gle to describe the geometry. This way, one ensures minimal memory usage and optimal
efficiency when accessing the grid. On the other hand, exploiting the simplified description

55

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 56 Grids in Subsurface Modeling

Figure 3.1 Illustration of a single cell, vertices and edges, and cell faces.

explicitly in your flow or transport solver inevitably means that the solver must be reimple-
mented if you later decide to use another grid format.
The most important goal for our development of MRST is to provide a toolbox that
both allows and enables the use of various grid types. To avoid having a large number of
different, and potentially incompatible, grid representations, we have therefore chosen to
store all grid types using a general unstructured format in which cells, faces, vertices, and
connections between cells and faces are explicitly represented. This means that we, for the
sake of generality, have sacrificed some of the efficiency one can obtain by exploiting spe-
cial structures in a particular grid type, and instead have focused on obtaining a flexible grid
description that is not overly inefficient. Moreover, our grid structure can be extended by
other properties that are required by various discretization schemes for flow and transport
simulations. A particular discretization may need the volume or the centroid (grid-point,
midpoint, or generating point) of each cell. Likewise, for cell faces one may need to know
the face areas, the face normals, and the face centroids. Although these properties can be
computed from the geometry (and topology) of the grid, it is often useful to precompute
and include them explicitly in the grid representation.
The first third of this chapter is devoted to standard grid formats that are available in
MRST. We introduce examples of structured grids, including regular Cartesian, rectilinear,
and curvilinear grids, and briefly discuss unstructured grids, including Delaunay trian-
gulations and Voronoi grids. The purpose of our discussion is to demonstrate the basic
grid functionality in MRST and show some key principles you can use to implement new
structured and unstructured grid formats. In the second part of the chapter, we discuss
industry-standard formats for stratigraphic grids that are based on extrusion of 2D shapes
(corner-point, prismatic, and 2.5D PEBI grids). Although these grids have an inherent
logical structure, representation of faults, erosion, pinch-outs, and so on lead to cells that
can have quite irregular shapes and an (almost) arbitrary number of faces. In the last part of
the chapter, we discuss how the grids introduced in the first two parts of the chapter can be
partitioned to form flexible coarse descriptions that preserve the geometry of the underlying
fine grids. The ability to represent a wide range of grids, structured or unstructured on the
fine and/or coarse scale, is a strength of MRST compared to the majority of research codes
arising from academic institutions.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.1 Structured Grids 57

A number of videos that complement the material presented in this chapter can be found
in the second MRST Jolt [189]. This Jolt introduces different types of grids, discusses how
such grids can be represented, and outlines functionality you can use to generate your own
grids.

3.1 Structured Grids

As we saw in this chapter’s introduction, a grid is a tessellation of a planar or volumetric
object by a set of simple shapes. In a structured grid, only one basic shape is allowed
and this basic shape is laid out in a regular repeating pattern so that the topology of the
grid is constant in space. The most typical structured grids are based on quadrilaterals in
2D and hexahedrons in 3D, but in principle it is also possible to construct grids with a
fixed topology using certain other shapes. Structured grids can be generalized to so-called
multiblock grids (or hybrid grids), in which each block consists of basic shapes that are laid
out in a regular repeating pattern.

Regular Cartesian Grids

The simplest form of a structured grid consists of unit squares in 2D and unit cubes in 3D,
so that all vertices in the grid are integer points. More generally, a regular Cartesian grid
can be defined as consisting of congruent rectangles in 2D and rectilinear parallelepipeds
in 3D, etc. Hence, the vertices have coordinates (i1 x1,i2 x2, . . . ) and the cells can be
referenced using the multi-index (i1,i2, . . . ). Herein, we only consider finite Cartesian grids
that consist of a finite number n2 × n2 × · · · × nk of cells that cover a bounded domain
[0,L1 ] × [0,L2 ] × · · · × [0,Lk ].
Regular Cartesian grids can be represented very compactly by storing ni and Li for
each dimension. In MRST, however, Cartesian grids are represented as if they were fully
unstructured using a general grid structure that will be described in more detail in Section
3.4. Cartesian grids therefore have special constructors,
G = cartGrid([nx, ny], [Lx Ly]);
G = cartGrid([nx, ny, nz], [Lx Ly Lz]);

that set up the data structures representing the basic geometry and topology of the grid. The
second argument is optional.

Rectilinear Grids
A rectilinear grid (also called a tensor grid) consists of rectilinear shapes (rectangles or
parallelepipeds) that are not necessarily congruent to each other. In other words, whereas a
regular Cartesian grid has a uniform spacing between its vertices, the grid spacing can vary
along the coordinate directions in a rectilinear grid. The cells can still be referenced using a
multi-index (i1,i2, . . . ), but the mapping from indices to vertex coordinates is nonuniform.
In MRST, one can construct a rectilinear grid by specifying vectors with the grid vertices
along the coordinate directions:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 58 Grids in Subsurface Modeling

Figure 3.2 Example of a rectilinear grid.

G = tensorGrid(x, y);
G = tensorGrid(x, y, z);

This syntax is the same as for the MATLAB functions meshgrid and ndgrid.
As an example of a rectilinear grid, we construct a 2D grid that covers the domain
[−1,1] × [0,1] and is graded toward x = 0 and y = 1 as shown in Figure 3.2.
dx = 1-0.5*cos((-1:0.1:1)*pi);
x = -1.15+0.1*cumsum(dx);
y = 0:0.05:1;
G = tensorGrid(x, sqrt(y));
plotGrid(G); axis([-1.05 1.05 -0.05 1.05]);

Curvilinear Grids
A curvilinear grid is a grid with the same topological structure as a regular Cartesian grid,
but in which the cells are quadrilaterals rather than rectangles in 2D and cuboids rather than
parallelepipeds in 3D. The grid is given by the coordinates of the vertices, but there exists
a mapping that will transform the curvilinear grid to a uniform Cartesian grid so that each
cell can still be referenced using a multi-index (i1,i2, . . . ).
For the time being, MRST has no constructor for curvilinear grids. Instead, you can
create such grids by first instantiating a regular Cartesian grid or a rectilinear grid and then
manipulating the vertices. This method is quite simple as long as there is a one-to-one
mapping between the curvilinear grid in physical space and the logically Cartesian grid in
reference space. The method will not work if the mapping is not one-to-one so that vertices
with different indices coincide in physical space. In this case, the user should create an

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.1 Structured Grids 59

Figure 3.3 The middle plot shows a rough grid created by perturbing all internal nodes of the regular
6×12 Cartesian grid in the left plot. The right plot shows a curvilinear grid created using the function
twister that uses a combination of sin functions to perturb a rectilinear grid. The color is determined
by the cell volumes.

ECLIPSE input file [270, 33] with keywords COORD[XYZ] (see Section 3.3.1) and use the
function buildCoordGrid to create the grid.
To illustrate the discussion, we show two examples of how to create curvilinear grids.
In the first example, we create a rough grid by perturbing all internal nodes of a regular
Cartesian grid (see Figure 3.3):
nx = 6; ny=12;
G = cartGrid([nx, ny]);
subplot(1,2,1); plotGrid(G);
c = G.nodes.coords;
I = any(c==0,2) | any(c(:,1)==nx,2) | any(c(:,2)==ny,2);
G.nodes.coords(I,:) = c(I,:) + 0.6*rand(sum(I),2)-0.3;
subplot(1,2,2); plotGrid(G);

In the second example, we use the routine twister to perturb the internal vertices. The
function maps the grid back to the unit square, perturbs the vertices according to the
mapping
 
(xi ,yi ) → xi + f (xi ,yi ),yi − f (xi ,yi ) , f (x,y) = 0.03 sin(π x) sin 3π(y − 12 ) ,

and then maps the grid back to its original domain. The right plot of Figure 3.3 shows the
resulting grid. To illuminate the effect of the mapping, we have colored the cells according
to their volume, which has been computed using the function computeGeometry, which we
will come back to later in the chapter.
G = cartGrid([30, 20]);
G.nodes.coords = twister(G.nodes.coords);
G = computeGeometry(G);
plotCellData(G, G.cells.volumes, 'EdgeColor', 'k'), colorbar

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 60 Grids in Subsurface Modeling

Fictitious Domains
One obvious drawback with Cartesian and rectilinear grids, as just defined, is that they can
only represent rectangular domains in 2D and cubic domains in 3D. Curvilinear grids, on
the other hand, can represent more general shapes by introducing an appropriate map-
ping, and can be used in combination with rectangular/cubic grids in multiblock grids
for efficient representation of realistic reservoir geometries. However, finding a mapping
that conforms to a given boundary is often difficult, in particular for complex geologies,
and using a mapping in the interior of the domain will inadvertently lead to cells with
rough geometries that deviate far from being rectilinear. Such cells may in turn introduce
problems if the grid is to be used in a subsequent numerical discretization, as we will see
later.
As an alternative, complex geometries can be easily modeled using structured grids by a
so-called fictitious domain method. In this method, the complex domain is embedded into
a larger “fictitious” domain of simple shape (a rectangle or cube) using, e.g., a Boolean
indicator value in each cell to tell whether the cell is part of the domain or not. The
observant reader will notice that we already have encountered the use of this technique
for the SAIGUP dataset (Figure 2.16) and the Johansen dataset in Chapter 2. In some
cases, one can also adapt the structured grid by moving the nearest vertices to the domain
boundary.
MRST has support for fictitious domain methods through the function removeCells,
which we will demonstrate in the next example, where we create a regular Cartesian grid
that fills the volume of an ellipsoid:

x = linspace(-2,2,21);
G = tensorGrid(x,x,x);
subplot(1,2,1); plotGrid(G);view(3); axis equal

subplot(1,2,2); plotGrid(G,'FaceColor','none');
G = computeGeometry(G);
c = G.cells.centroids;
r = c(:,1).^2 + 0.25*c(:,2).^2+0.25*c(:,3).^2;
G = removeCells(G, r>1);
plotGrid(G); view(-70,70); axis equal;

Worth observing here is the use of computeGeometry to compute cell centroids that are not
part of the basic geometry representation in MRST. Figure 3.4 shows plots of the grid before
and after removing the inactive parts. Because of the fully unstructured representation
used in MRST, calling computeGeometry actually removes the inactive cells from the grid
structure, but from the outside, the structure behaves as if we had used a fictitious domain
method.
You can find more examples of how you can make structured grids and populate them
with petrophysical properties in the fourth video of the second MRST Jolt [189].

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.1 Structured Grids 61

−2

−1

0 2

−2
1 1
0
2
2 0
2
2
2
0 1 −1
0
0
−2 −2 −1
−2 −2

Figure 3.4 Example of a regular Cartesian grid representing a domain in the form of an ellipsoid.
The underlying logical Cartesian grid is shown in the left plot and as a wireframe in the right plot.
The active part of the model is shown in yellow color in the right plot.

Computer exercises
3.1.1 Make the grid shown:

Hint: the grid spacing in the x-direction is given by x(1 − 12 cos(π x)) and the
colors signify cell volumes.
3.1.2 Metaballs are commonly used in computer graphics to generate organic-looking
objects. Each metaball is defined as a smooth function that has finite support. One
example is
 x |2 4
 |
m( x,r) = 1 − min 2 ,1 .
r
Metaballs can be used to define objects implicitly, e.g., as all the points x that satisfy

m( x − xi ,ri ) ≤ C, C ∈ IR+
i

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 62 Grids in Subsurface Modeling

Use this approach and try to make grids similar to the ones shown here:

3.1.3 A simple way to make test models with funny geometries is to use the method of
fictitious domains and let an image define the domain of interest. In the figures,
the image was taken from penny, which is one of the standard data sets that are
distributed with MATLAB, and then used to define the geometry of the grid and
assign permeability values

Pick your own favorite image or make one in a drawing program and use imread
to load the image into MATLAB as a 3D array, which you can use to define your
geometry and petrophysical values. If you do not have an image at hand, you can
use penny or spine. For penny, in particular, you may have to experiment a bit with
the threshold used to define your domain to ensure that all cells are connected, i.e.,
that the grid you obtain consists of only one piece.

3.2 Unstructured Grids

An unstructured grid consists of a set of simple shapes that are laid out in an irregular
pattern so that any number of cells can meet at a single vertex. The topology of the grid
will therefore change throughout space. An unstructured grid can generally consist of a
combination of polyhedral cells with varying numbers of faces, as we will see in this
section. However, the most common forms of unstructured grids are based on triangles
in 2D and tetrahedrons in 3D. These grids are very flexible and are relatively easy to adapt
to complex domains and structures, or refine to provide increased local resolution.
Unlike structured grids, unstructured grids cannot generally be efficiently referenced
using a structured multi-index. Instead, one must describe a list of connectivities that
specifies the way a given set of vertices make up individual cells and cell faces, and how
these cells are connected to each other via faces, edges, and vertices.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.2 Unstructured Grids 63

To understand the properties and construction of unstructured grids, we start by a

brief discussion of two concepts from computational geometry: Delaunay tessellation
and Voronoi diagrams. Both these concepts are supported by standard functionality in
MATLAB.

3.2.1 Delaunay Tessellation

A tessellation of a set of generating points P = {xi }ni=1 is defined as a set of simplices that
completely fills the convex hull of P. The convex hull H of P is the convex minimal set
that contains P and can be described constructively as the set of convex combinations of a
finite subset of points from P,
  
   
H (P) = 
λi xi xi ∈ P,λi ∈ R,λi ≥ 0, λi = 1,1 ≤  ≤ n .
i=1 i=1

Delaunay tessellation is by far the most common method of generating a tessellation based
on a set of generating points. In 2D, the Delaunay tessellation consists of a set of trian-
gles defined so that three points form the corners of a Delaunay triangle only when the
circumcircle that passes through them contains no other generating points; see Figure 3.5.
The definition using circumcircles can readily be generalized to higher dimensions using
simplices and hyperspheres.
The center of the circumcircle is called the circumcenter of the triangle. We will come
back to this quantity when discussing Voronoi diagrams in the next subsection. When four
(or more) points lie on the same circle, the Delaunay triangulation is not unique. As an
example, consider four points defining a rectangle. Using either of the two diagonals will
give two triangles satisfying the Delaunay condition.
The Delaunay triangulation can alternatively be defined using the so-called max-min
angle criterion, which states that the Delaunay triangulation is the one that maximizes the
minimum angle of all angles in a triangulation; see Figure 3.6. Likewise, the Delaunay tri-
angulation minimizes the largest circumcircle and minimizes the largest min-containment
circle, which is the smallest circle that contains a given triangle. Additionally, the closest
two generating points are connected by an edge of a Delaunay triangulation. This is called

Figure 3.5 Two triangles and their circumcircles.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 64 Grids in Subsurface Modeling

Figure 3.6 Example of two triangulations of the same five points; the triangulation to the right
satisfies the min-max criterion.

Figure 3.7 Two different Delaunay tessellations of a rectangular point mesh.

the closest-pair property, and the two neighboring points are referred to as natural neigh-
bors. This way, the Delaunay triangulation can be seen as the natural tessellation of a set of
generating points.
Delaunay tessellation is a popular research topic and there exists a large body of litera-
ture on theoretical aspects and computer algorithms. Likewise, there are a large number of
software implementations available on the Internet. For this reason, MRST does not have
any routines for generating tessellations based on simplexes. Instead, we have provided
simple routines for mapping a set of points and edges, as generated by MATLAB’s Delau-
nay triangulation routines, to the internal data structure used to represent grids in MRST.
How these routines work will be illustrated in terms of a few simple examples.
In the first example, we triangulate a rectangular mesh and then convert the result by use
of the MRST routine triangleGrid:
[x,y] = meshgrid(1:10,1:8);
t = delaunay(x(:),y(:));
G = triangleGrid([x(:) y(:)],t);
plot(x(:),y(:),'o','MarkerSize',8);
plotGrid(G,'FaceColor','none');

Depending on what version you have of MATLAB, the 2D Delaunay routine delaunay
will produce one of the triangulations shown in Figure 3.7. In older versions of MATLAB,
the implementation of delaunay was based on Qhull (see www.qhull.org), which pro-
duces the unstructured triangulation shown in the right plot. MATLAB 7.9 and newer has

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.2 Unstructured Grids 65

improved routines for 2D and 3D computational geometry, and here delaunay will produce
the structured triangulation shown in the left plot. However, the n-D tessellation routine
delaunayn([x(:)y(:)]) is still based on Qhull and will generally produce an unstructured
tessellation, as shown in the right plot.
If the set of generating points is structured, e.g., as one would obtain by calling either
meshgrid or ndgrid, it is straightforward to make a structured triangulation. The following
skeleton of a function makes a 2D triangulation and can easily be extended by the interested
reader to 3D:
function t = mesh2tri(n,m)
[I,J]=ndgrid(1:n-1, 1:m-1); p1=sub2ind([n,m],I(:),J(:));
[I,J]=ndgrid(2:n , 1:m-1); p2=sub2ind([n,m],I(:),J(:));
[I,J]=ndgrid(1:n-1, 2:m ); p3=sub2ind([n,m],I(:),J(:));
[I,J]=ndgrid(2:n , 1:m-1); p4=sub2ind([n,m],I(:),J(:));
[I,J]=ndgrid(2:n , 2:m ); p5=sub2ind([n,m],I(:),J(:));
[I,J]=ndgrid(1:n-1, 2:m ); p6=sub2ind([n,m],I(:),J(:));
t = [p1 p2 p3; p4 p5 p6];

In Figure 3.8, we have used the demo case seamount supplied with MATLAB as an
example of a more complex unstructured grid
load seamount;
plot(x(:),y(:),'o');
G = triangleGrid([x(:) y(:)]);
plotGrid(G,'FaceColor',[.8 .8 .8]); axis off;

The observant reader will notice that we did not explicitly generate a triangulation before
calling triangleGrid; if the second argument is omitted, the routine uses MATLAB’s built-
in delaunay triangulation as default.
For 3D grids, MRST supplies a conversion routine tetrahedralGrid(P, T) that
constructs a valid grid definition from a set of points P (m × 3 array of node coordinates)

Figure 3.8 The left plot shows the triangular grid from the seamount demo case. The right plot
shows a tetrahedral tessellation of a 3D point mesh.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 66 Grids in Subsurface Modeling

and a tetrahedron list T (n array of node indices). The tetrahedral tessellation shown to the
right in Figure 3.8 was constructed from a set of generating points defined by perturbing a
regular hexahedral mesh:
N=7; M=5; K=3;
[x,y,z] = ndgrid(0:N,0:M,0:K);
x(2:N,2:M,:) = x(2:N,2:M,:) + 0.3*randn(N-1,M-1,K+1);
y(2:N,2:M,:) = y(2:N,2:M,:) + 0.3*randn(N-1,M-1,K+1);
G = tetrahedralGrid([x(:) y(:) z(:)]);
plotGrid(G, 'FaceColor',[.8 .8 .8]); view(-40,60); axis tight off

3.2.2 Voronoi Diagrams

The Voronoi diagram of a set of points P = {xi }ni=1 is the partitioning of Euclidean space
into n (possibly unbounded) convex polytopes1 such that each polytope contains exactly
one generating point xi and every point inside the given polytope is closer to its generating
point than any other point in P. The convex polytopes are called Voronoi cells or Voronoi
regions. Mathematically, the Voronoi region V (xi ) of a generating point xi in P can be
defined as
  
V (xi ) = x  x − xi  < x − xj  ∀j = i . (3.1)

A Voronoi region is not closed in the sense that a point that is equally close to two or more
generating points does not belong to the region defined by (3.1). Instead, these points are
said to lie on the Voronoi segments and can be included in the Voronoi cells by defining the
closure of V (xi ), using “≤” rather than “<” in (3.1).
The Voronoi regions for all generating points lying at the convex hull of P are
unbounded, all other Voronoi regions are bounded. For each pair of two points xi and xj ,
one can define a hyperplane with co-dimension one consisting of all points that lie equally
close to xi and xj . This hyperplane is the perpendicular bisector to the line segment between
xi and xj and passes through the midpoint of the line segment. The Voronoi diagram of a
set of points can be derived directly as the dual of the Delaunay triangulation of the same
points. To understand this, we consider the planar case; see Figure 3.9. For every triangle,
there is a polyhedron in which the vertices occupy complementary locations:

• The circumcenter of a Delaunay triangle corresponds to a vertex of a Voronoi cell.

• Each vertex in the Delaunay triangulation corresponds to, and is the center of, a Voronoi
cell.
Moreover, for locally orthogonal Voronoi diagrams, an edge in the Delaunay triangula-
tion corresponds to a segment in the Voronoi diagram and the two intersect each other
orthogonally. However, as we can see in Figure 3.9, this is not always the case. If the

1 A polytope is a generic term that refers to a polygon in 2D, a polyhedron in 3D, and so on.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.2 Unstructured Grids 67

Figure 3.9 Duality between Voronoi diagrams and Delaunay triangulation. From top left across
to bottom right: generating points, Delaunay triangulation, Voronoi diagram, and Voronoi diagram
(thick lines) and Delaunay triangulation (thin lines).

Figure 3.10 Three examples of Voronoi diagrams generated from 2D point lattices. From left to right:
square lattice, square lattice rotated 45 degrees, and lattice forming equilateral triangles.

circumcenter of a triangle lies outside the triangle itself, the Voronoi segment does not
intersect the corresponding Delaunay edge. To avoid this situation, one can perform a
constrained Delaunay triangulation and insert additional points where the constraint is
not met (i.e., the circumcenter is outside its triangle).
Figure 3.10 shows three examples of planar Voronoi diagrams generated from 2D point
lattices using the MATLAB function voronoi. MRST does not yet have a similar function
that generates a Voronoi grid from a point set, but offers V=pebi(T) that generates a locally
orthogonal, Voronoi grid V as a dual to a triangular grid T. The grids are constructed by
connecting the perpendicular bisectors of the edges of the Delaunay triangulation, hence
the name perpendicular bisector (PEBI) grids. To demonstrate the functionality, we first
generate a honeycombed grid similar to the one shown in the right plot in Figure 3.10:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 68 Grids in Subsurface Modeling

Figure 3.11 Two examples of Voronoi grids. The left plot shows a honeycombed PEBI grid and the
right plot shows the PEBI grid derived from the seamount demo case.

[x,y] = meshgrid([0:4]*2*cos(pi/6),0:3);
x = [x(:); x(:)+cos(pi/6)];
y = [y(:); y(:)+sin(pi/6)];
G = triangleGrid([x,y]);
plotGrid(pebi(G), 'FaceColor','none'); axis equal off

Figure 3.11 shows the resulting grid. As a second example, we reiterate the seamount
examples shown in Figure 3.8:
load seamount
V = pebi(triangleGrid([x y]));
plotGrid(V,'FaceColor',[.8 .8 .8]); axis off;

Several of the examples discussed in this section can also be found in the fifth video of the
second MRST Jolt [189]. Since pebi is a 2D code, we cannot apply it directly to the 3D
tetrahedral grid shown in Figure 3.8 to generate a dual 3D Voronoi grid. The upr module
[42, 169] offers several more advanced tools for generating 2D and 3D Voronoi grids; we
will return to this in Section 3.5.3. Likewise, Section 3.3.2 discusses how 2D Voronoi grids
can be extruded to 3D giving so-called 2.5D grid; this is a standard approach in geological
modeling to preserve geological layers.

3.2.3 General Tessellations

Tessellations come in many other forms than the Delaunay and Voronoi types discussed
in the two previous sections. Tessellations are more commonly referred to as tilings and
are patterns made up of geometric forms (tiles) that are repeated over and over without
overlapping or leaving any gaps. Such tilings can be found in many patterns of nature,
like in honeycombs, giraffe skin, pineapples, snake skin, tortoise shells, to name a few.
Tessellations have also been extensively used for artistic purposes since ancient times,
from the decorative tiles of Ancient Rome and Islamic art to the amazing artwork of
M. C. Esher. For completeness (and fun), MRST offers the function tessellationGrid
that can take a tessellation consisting of symmetric n-polygonals and turn it into a correct
grid structure. While this may not be very useful in modeling petroleum reservoirs, it can

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.2 Unstructured Grids 69

easily be used to generate irregular grids that can be used to stress test various discretization
methods. To exemplify, the following commands will make a standard n × m Cartesian
mesh:
[x,y] = meshgrid(linspace(0,1,n+1),linspace(0,1,m+1));
I = reshape(1:(n+1)*(m+1),m+1,n+1);
T = [reshape(I(1:end-1,1:end-1),[],1)'; reshape(I(1:end-1,2:end ),[],1)';
reshape(I(2:end, 2:end ),[],1)'; reshape(I(2:end, 1:end-1),[],1)']';
G = tessellationGrid([x(:) y(:)], T);

Here, the vertices and cells are numbered first in the y direction and then in the x-direction,
so that the first two lines in T read [1 m+2 m+3 2; 2 m+3 m+4 3], and so on. There are
obviously many ways to make more general tilings. The script showTessellation in the
book module shows some simple but amazing examples.

3.2.4 Using an External Mesh Generator

Using a Delaunay triangulation as discussed earlier, we can generate grids that fit a given
set of vertices. However, in most applications, the vertex points are not given a priori and all
one wants is a reasonable grid that fits an exterior outline and possibly also a set of interior
boundaries and/or features, which in the case of subsurface media could be faults, fractures,
or geological horizons. To this end, you will typically have to use a mesh generator. There
are a large number of mesh generators available online, and in principle any of these
can be used in combination with MRST’s grid factory routines, as long as they produce
triangulations on the form outlined earlier. The triangle module implements a simple
MATLAB interface to the Triangle generator (www.cs.cmu.edu/~quake/triangle.html).
My personal favorite, however, is DistMesh by Persson and Strang [248]. While most
mesh generators tend to be complex and quite inaccessible codes, DistMesh is a relatively
short and simple MATLAB code written in the same spirit as MRST. The performance of
the code may not be optimal, but the user can go in and inspect all algorithms and modify
them to his or her purpose. A slightly modified version of DistMesh is, for instance, used
on the inside of the upr module for generating high-quality Voronoi grids, which will be
briefly outlined in Section 3.5.3. In the following, we use DistMesh to generate a few
examples of more complex triangular and Voronoi grids in 2D.
DistMesh is distributed under the GNU GPL license (which is the same license that
MRST uses) and can be downloaded from the software’s website. The simplest way to
integrate DistMesh with MRST is to install it as a third-party module. Assuming that you
are connected to Internet, this is done as follows:
pth = fullfile(ROOTDIR,'utils','3rdparty','distmesh');
mkdir(pth)
unzip('http://persson.berkeley.edu/distmesh/distmesh.zip', pth);
mrstPth('reregister','distmesh', pth);

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 70 Grids in Subsurface Modeling

Figure 3.12 Grids generated by the DistMesh grid generator.

and you are ready to use DistMesh. If you intend to use it many times, you should copy
the last line to the startup_user.m file in MRST’s root directory.
In DistMesh, the perimeter of the domain is represented using a signed distance func-
tion d(x,y), which by definition is set to be negative inside the region. The software
offers a number of utility functions that makes it simple to describe relatively complex
geometries, as we shall see in the following. Let us start with a simple example taken from
the DistMesh website: Consider a square domain [−1,1] × [−1,1] with a circular cutout
of radius 0.5 centered at the origin. We start by making a grid that has a uniform target size
h = 0.2
mrstModule add distmesh;
fd=@(p) ddiff(drectangle(p,-1,1,-1,1), dcircle(p,0,0,0.5));
[p,t]=distmesh2d(fd, @huniform, 0.2, [-1,-1;1,1], [-1,-1;-1,1;1,-1;1,1]);
G = triangleGrid(p, t);

Here, we have used utility functions ddif, drectangle and dcircle to compute the
signed distance from the outer and inner perimeter. Likewise, huniform is set to enforce
uniform cell size and equal distance between the initial points, here this distance is 0.2.
The fourth argument is the bounding box of our domain, and the fifth argument consists
of fixed points that the algorithm is not allowed to move. After the triangulation has been
computed by distmesh2d, we pass it to triangleGrid to make an MRST grid structure.
The resulting grid is shown to the left in Figure 3.12.
As a second test, let us make a graded grid that has a mesh size of approximately 0.05
at the inner circle and 0.2–0.35 at the outer perimeter. To enforce this, we replace the
huniform function by another function that gives the correct mesh size distribution:
fh=@(p) 0.05+0.3*dcircle(p,0,0,0.5);
[p,t]=distmesh2d(fd, fh, 0.05,[-1,-1;1,1], [-1,-1;-1,1;1,-1;1,1]);

The middle plot of Figure 3.12 shows that the resulting grid has the expected grading from
the inner boundary and outwards to the perimeter.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.2 Unstructured Grids 71

In our last example, we create a graded triangulation of a polygonal domain and then
use pebi to compute its Voronoi diagram
pv = [-1 -1; 0 -.5; 1 -1; 1 1; 0 .5; -1 1; -1 -1];
fh = @(p,x) 0.025 + 0.375*sum(p.^2,2);
[p,t] = distmesh2d(@dpoly, fh, 0.025, [-1 -1; 1 1], pv, pv);
G = pebi(triangleGrid(p, t));

Here, we use the utility function dpoly(p,pv) to compute the signed distance of any point
set p to the polygon with vertices in pv. Notice that pv must form a closed path. Likewise,
since the signed distance function and the grid density function are assumed to take the
same number of arguments, fh is created with a dummy argument x. The argument sent to
these functions is passed as the sixth argument to distmesh2d. The resulting grid is shown
to the right in Figure 3.12.
DistMesh also has routines for creating nD triangulations and triangulations of surfaces,
but these are beyond the scope of the current presentation.

Computer exercises
3.2.1 Create MRST grids from the standard data sets trimesh2d and tetmesh. How would
you assign lognormal petrophysical parameters to these grids so that the spatial
correlation is preserved?
3.2.2 All grids generated by triangleGrid are assumed to be planar so that each vertex
can be given by a 2D coordinate. However, triangular grids are commonly used to
represent nonplanar surfaces in 3D. Can you extend the function so it can construct
both 2D and 3D grids? You can use the data set trimesh3d as an example of a
triangulated 3D surface.
3.2.3 Load the triangle module and use it to mesh the polygonal outline to the right in
Figure 3.12. Try to also make other convex and concave outlines.
3.2.4 Download and install distmesh and try to make MRST grids from all the triangu-
lations shown in [248, Fig. 5.1].
3.2.5 Running the tutorial showTessellation from the book module will produce fig-
ures like these:

What would you do to fit the tessellations so that they fill a rectangular box without
leaving any gaps along the border?
3.2.6 MRST does not yet have a grid factory routine to generate structured grids with
local, nested refinement, as shown in the figure to the left:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 72 Grids in Subsurface Modeling

Try to use a combination of triangleGrid and pebi to make a good approximation

to such a grid as shown to the right in the figure. (Hint: to get rid of artifacts, one
layer of cells were removed along the outer boundary.)
3.2.7 The figure shows an unstructured hexagonal grid that has been adapted to two faults
in the interior of the domain and padded with rectangular cells near the boundary.
Try to implement a routine that generates a similar grid.

Hint: in this case, the strike direction of the faults

are ±30◦ and the start and endpoints of the faults
have been adjusted so that they coincide with the
generating points of hexagonal cells.

3.3 Stratigraphic Grids

When discussing how to model reservoir rocks, we saw that the grid is closely connected
to the description of the petrophysical characteristics of a reservoir. Not only does the
grid describe the outer geometry of the reservoir, but it is also used to represent internal
structures such as fractures and faults, as well as surfaces representing marked changes the
rock’s physical characteristics (lithography). This means that the grid is closely related to
the parameter description of the flow model and, unlike in many other disciplines, cannot be
chosen independently to provide a certain numerical accuracy. Indeed, the grid is typically
chosen by a geologist who tries to describe the rock body by as few volumetric cells as
possible and who does not necessarily care too much about potential numerical difficulties
the choice of grid may cause in subsequent flow simulations. Although grossly simplified,
this observation is important to bear in mind throughout the rest of this chapter.
The industry standard to represent reservoir geology in a flow simulator is to use a
stratigraphic grid built based on geological horizons and fault surfaces. The volumetric
grid is typically built by extruding 2D tessellations of the geological horizons in the vertical
direction or in a direction following major fault surfaces. For this reason, some stratigraphic
grids, like the PEBI grids you will meet in Section 3.3.2, are often called 2.5D rather than
3D grids. These grids may be unstructured in the lateral direction, but have a clear structure
in the vertical direction to reflect the layering of the reservoir.
Because of the role grid models play in representing geological formations, real-life
stratigraphic grids tend to be highly complex and have unstructured connections induced

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.3 Stratigraphic Grids 73

by the displacements that have occured over faults. Another characteristic feature is high
aspect ratios. Typical reservoirs extend several hundred or thousand meters in the lateral
direction, but the zones carrying hydrocarbon may be just a few tens of meters in the vertical
direction and consist of several layers with (largely) different rock properties. Getting the
stratigraphy correct is crucial, and high-resolution geological modeling will typically result
in a high number of (very) thin grid layers in the vertical direction, resulting in two or three
orders of magnitude aspect ratios.
A full exposition of stratigraphic grids is far beyond the scope of this book. In the
next two subsections, we discuss the basics of the two most commonly used forms of
stratigraphic grids. A complementary discussion is given in videos 2, 6, and 7 of the second
MRST Jolt [189].

3.3.1 Corner-Point Grids

To model the geological structures of petroleum reservoirs, the industry-standard approach
is to introduce what is called a corner-point grid [254], which we already encountered in
Section 2.5. A corner-point grid consists of a set of hexahedral cells that are topologically
aligned in a Cartesian fashion so that the cells can be numbered using a logical ij k index.
In its simplest form, a corner-point grid is specified in terms of a set of vertical or inclined
pillars defined over an areal Cartesian 2D mesh in the lateral direction. Each cell in the
volumetric grid is defined by eight logical corner points specified as two depth coordinates
on each of the four coordinate lines that define a pillar; see Figure 3.13. The grid consists of
nx ×ny ×nz grid cells and the cells are ordered with the i-index (x-axis) cycling fastest, then
the j -index (y-axis), and finally the k-index (negative z-direction). All cellwise property
data are assumed to follow the same numbering scheme.
As discussed previously, a fictitious domain approach is used to embed the reservoir in a
logically Cartesian shoebox. This means that inactive cells that are not part of the physical
model, e.g., as shown in Figure 2.16, are present in the topological ij k-numbering but are

Figure 3.13 Each cell in the corner-point grid is restricted by two points on each of the four grid lines
that make up a pillar cells. The plot shows cells and grid lines from two neighboring pillars.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 74 Grids in Subsurface Modeling

indicated by a zero porosity or net-to-gross value, as discussed in Section 2.4, or marked

by a special boolean indicator (called ACTNUM in the input files).
So far, the topology and geometry of a corner-point grid have not deviated from that
of the mapped Cartesian grids studied in the previous section. Somewhat simplified, one
may view the logical ij k numbering as a reflection of the sedimentary rock bodies as they
may have been deposited in geological time, so that cells with the same k index have been
deposited at approximately the same time. To model geological features like erosion and
pinch-outs of geological layers, the corner-point format allows point-pairs to collapse along
coordinate lines. This creates degenerate hexahedral cells that may have less than six faces,
as illustrated in Figure 3.14. The corner points can even collapse along all four lines of a
pillar, so that a cell completely disappears. This implicitly introduces a new local topology,
which is sometimes referred to as non-neighboring connections, in which cells that are
not logical ij k neighbors can be neighbors in physical space and share a common face.
Figure 3.15 shows an example of a model that contains both eroded geological layers and
fully collapsed cells. In a similar manner, (simple) vertical and inclined faults can be easily
modeled by aligning the pillars with fault surfaces and displacing the corner points defining
neighboring cells on one or both sides of the fault. This way, one creates nonmatching
geometries and non-neighboring connections in the underlying ij k topology.
To illustrate the concepts introduced so far, we consider a low-resolution version of the
model from Figure 2.11 on page 42 created by the simpleGrdecl routine, which generates
an input stream containing the basic keywords that describe a corner-point grid in the
ECLIPSE input deck [270, 33]

Figure 3.14 Examples of deformed and degenerate hexahedral cells arising in corner-point grid
models.

Figure 3.15 Side view in the xz-plane of corner-point grid with vertical pillars modeling a stack of
sedimentary beds (each layer indicated by a different color).

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.3 Stratigraphic Grids 75

grdecl = simpleGrdecl([4, 2, 3], .12, 'flat', true);

grdecl =
cartDims: [4 2 3]
COORD: [90x1 double]
ZCORN: [192x1 double]
ACTNUM: [24x1 int32]

The COORD field contains the pairs of 3D coordinates for the 15 coordinate lines that define
the 4 × 2 mesh of pillars, whereas the ZCORN field gives the z-values that determine vertical
positions uniquely along each coordinate line for the eight corner-points of the 24 cells. To
extract these data, we use two MRST routines
[X,Y,Z] = buildCornerPtPillars(grdecl,'Scale',true);
[x,y,z] = buildCornerPtNodes(grdecl);

We now plot the coordinate lines and the corner-points and mark those lines on which the
corner-points of logical ij neighbors do not coincide,
% Plot pillars
plot3(X',Y',Z','k');
set(gca,'zdir','reverse'), view(35,35), axis off, zoom(1.2);

% Plot points on coordinate lines, mark those with faults red

hold on; I=[3 8 13];
hpr = plot3(X(I,:)',Y(I,:)',Z(I,:)','r','LineWidth',2);
hpt = plot3(x(:),y(:),z(:),'o'); hold off;

From the plots in the upper row of Figure 3.16 we clearly see how the coordinate lines
change slope from the east and west side toward the fault in the middle, and how the grid
points sit like beads on a string along each coordinate line.
Cells along a pillar are now defined by connecting pairs of points from four neighboring
coordinate lines that make up a cell in the areal mesh. To see this, we plot two pillars as
well as the whole grid with the fault surface marked in blue:
% Create grid and plot two stacks of cells
G = processGRDECL(grdecl);
args = {'FaceColor'; 'r'; 'EdgeColor'; 'k'};
hcst = plotGrid(G,[1:8:24 7:8:24],'FaceAlpha', .1, args{:});

% Plot cells and fault surface

delete([hpt; hpr; hcst]);
plotGrid(G,'FaceAlpha', .15, args{:});
plotFaces(G, G.faces.tag>0,'FaceColor','b','FaceAlpha',.4);

The upper-left plot in Figure 3.17 shows the same model sampled with even fewer cells.
We have used different coloring of the cell faces on each side of the fault to highlight
nonmatching faces along the fault plane. MRST represents corner-point grids as matching

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 76 Grids in Subsurface Modeling

Figure 3.16 Specification of a corner-point grid. Starting from the coordinate lines defining pillars
(upper left), we add corner-points and identify lines containing nonmatching corner marked in red
(upper right). A stack of cells is created for each set of four lines defining a pillar (lower left), and
then the full grid is obtained (lower right). In the last plot, the fault faces have been marked in blue.

Figure 3.17 Subdivision of fault face in two three-dimensional models. In the left column, the
subfaces are all rectangular. In the right columns they are not. In both the upper plots, the faces
marked in red belong only to the cells behind the fault surface, the blue faces belong only to the cells
in front of the fault surface, and the magenta ones belong to cells on both sides. The lower plot shows
the cells behind the surface, where each cell has been given its own color.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.3 Stratigraphic Grids 77

unstructured grids obtained by subdividing all nonmatching cell faces instead of using
the more compact nonmatching hexahedral form. This means that the four cells with non-
neighboring connections across the fault plane will have seven and not six faces. For each
such cell, two of the seven faces lie along the fault plane. Here, the subdivision resulted
in new faces that all have four corners and rectangular geometry. This is generally not the
case; the right column of Figure 3.17 shows a grid where we can see cells with six, seven,
or eight faces, and faces with three, four, and five corners. Indeed, for real-life models,
subdivision of nonmatching fault faces can lead to cells having much more than six faces.
Using the inherent flexibility of the corner-point format, it is possible to construct very
complex geological models that come a long way in matching the geologist’s perception
of the underlying rock formations. Because of their many appealing features, corner-point
grids have been an industry standard for years, and the format is supported in most com-
mercial software for reservoir modeling and simulation.

A Simple Grid Generator

The internal and external geology of a real reservoir is usually describe by a large set of
horizons and fault surfaces. As explained earlier, horizons are lateral surfaces that delim-
inate the reservoir in the upward and downward direction and represent the main strati-
graphic layers. These surfaces can either be triangulated/tessellated or represented using
splines. Major fault surfaces are typically vertical or inclined, but need to be planar. Gener-
ating corner-point grids that adapt to these surfaces is a challenging task and requires spe-
cialized geomodeling software and integration of a lot of different data types. MRST does
not offer any such capabilities, and complex geomodels must therefore be created using
external software. However, we do offer a relatively simple grid generator that takes a set
of horizons as input and generates a corner-point grid by interpolating vertically between
pairs of consecutive horizons. The function converHorizonsToGrid cannot model faults,
and assumes that each horizon is represented over a rectilinear areal mesh. The horizons
need not use the same areal mesh points, but the areal meshes must have a certain minimal
overlap inside which the corner-point will be generated. The following example illustrates
the basic use of the grid generator:
[y,x,z] = peaks(15); z = z+8;
horizons = {struct('x',x,'y',y,'z',z),struct('x',x,'y',y,'z',z+8)};
grdecl = convertHorizonsToGrid(horizons,'layers', 4);
G = processGRDECL(grdecl);

Here, we use the peaks function from MATLAB to generate a surface and create two
horizons that deliminate the reservoir upward and downward. We then create a grid with
four layers of cells in the vertical direction. When no extra parameters are specified, the
areal mesh defining the corner-point grid has the same number of points as the horizons;
see Figure 3.18. When the areal domain of definition for the horizons do not coincide, the
interpolation region is set to be the minimum rectangle that contains the areal bounding
boxes of all the horizons. Inside this rectangle, the routine creates a Cartesian mesh and

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 78 Grids in Subsurface Modeling

Figure 3.18 Example of a corner-point grid generated by interpolating vertically between two
geological horizons.

(a) (b)

(c) (d)

Figure 3.19 Example of a corner-point grid generated by interpolating vertically between two
partially overlapping horizons. (a) Horizons and projected bounding boxes. (b) Interpolation region
with 31 × 31 areal mesh covers the two bounding boxes. (c) Corner-point grid generated inside the
maximum overlap of the two horizons. (d) The full 3D grid generated by the routine.

uses this to interpolate vertically. In the output grid, the routine discards all cells whose
lateral projection is not contained inside the maximum rectangle that fits inside the areal
bounding boxes of all horizons. Figure 3.19 shows illustrates how the grid resulting from
the following commands is created:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.3 Stratigraphic Grids 79

[n,m] = deal(30);
horizons = {struct('x',x,'y',y,'z',z),struct('x',x+.5,'y',2*y+1,'z',z+10)};
grdecl = convertHorizonsToGrid(horizons,'dims',[n m], 'layers', 3);
G = processGRDECL(grdecl);

Here, we have only used two horizons to simplify the illustration. Section 15.6.4 presents
a more complex case with multiple horizons.

A Synthetic Faulted Reservoir

In our second example we consider a synthetic model of two intersecting faults that make
up the letter Y in the lateral direction. The two fault surfaces are highly deviated, making
an angle far from 90 degrees with the horizontal direction. To model this scenario using
corner-point grids, we basically have two different choices. The first choice, which is quite
common, is to let the coordinate lines (and hence the extrusion direction) follow the main
fault surfaces. For highly deviated faults, like in the current case, this will lead to extruded
cells that are far from K-orthogonal and hence susceptible to grid-orientation errors in a
subsequent simulation, as will be discussed in more detail in Chapter 6. Alternatively, we
can choose a vertical extrusion direction and replace deviated fault surfaces by stair-stepped
approximations so that the faults zigzag in a direction not aligned with the grid. This will
create cells that are mostly K-orthogonal and less prone to grid-orientation errors, but will
only represent the fault approximately.
Figure 3.20 shows two different grid models, taken from the CaseB4 data set. In
the stair-stepped model, the use of cells with orthogonal faces causes the faults to be
represented as zigzag patterns. The pillar grid correctly represents the faults as inclined
planes, but has cells with degenerate geometries and cells that deviate strongly from
being orthogonal in the lateral direction. Likewise, some coordinate lines have close to 45
degrees inclination, which will likely give significant grid-orientation effects in a standard
two-point scheme.

A Simulation Model of the Norne Field

Norne is an oil and gas field lies located in the Norwegian Sea. The reservoir is found in
Jurassic sandstone at a depth of 2,500 meters below sea level, and was originally estimated
to contain 90.8 million m3 oil, mainly in the Ile and Tofte formations, and 12 billion m3 gas
in the Garn formation. The field is operated by Statoil and production started in November
1997, using a floating production, storage, and offloading (FPSO) ship connected to seven
subsea templates at a water depth of 380 meters. The oil is produced with water injection
as the main drive mechanisms, and the expected ultimate oil recovery is more than 60%,
which is very high for a subsea oil reservoir. During thirteen years of production, five 4D
seismic surveys of high quality have been recorded. In cooperation with NTNU, operator
Statoil and partners have released large amounts of subsurface data from the Norne field for

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 80 Grids in Subsurface Modeling

Deviated pillars Zoom of fault intersection

Stair-stepped Zoom of fault intersection

Figure 3.20 Modeling the intersection of two deviated faults using deviated pillars (top) and stair-
stepped approximation (bottom). CaseB4 grids courtesy of Statoil.

research and education purposes.2 More recently, the Open Porous Media (OPM) initiative
(opm-project.org) released the full simulation model as an open data set on Github
(github.com/OPM/opm-data). For several years, Norne was the only full simulation model
of a real oil reservoir that was freely available to the general public. Recently, Equinor
(former Statoil) released a much more comprehensive data set of the Volve field. The
Norne model can either be downloaded and installed using mrstDatasetGUI or directly
from the command line:
makeNorneSubsetAvailable() && makeNorneGRDECL()

Once in place, we can load the grid and the petrophysical data from the simulation model
as follows:
grdecl = readGRDECL(fullfile(getDatasetPath('norne'), 'NORNE.GRDECL'));
grdecl = convertInputUnits(grdecl, getUnitSystem('METRIC'));

The model consists of 46 × 112 × 22 corner-point cells. We start by plotting the whole
model, including inactive cells. To this end, we need to override3 the ACTNUM field before

2 The Norne Benchmark data sets are hosted and supported by the Center for Integrated Operations in the Petroleum Industry
(IO Center) at NTNU (www.ipt.ntnu.no/~norne/). The data set used herein was first released as part of “Package 2: Full field
model” (2013)
3 At this point I hasten to warn you that inactive cells often contain junk data and may generally not be inspected in this manner.
Here, however, most inactive cells are defined in a reasonable way. By not performing basic sanity checks on the resulting grid
(option 'checkgrid'=false), we manage to process the grid and produce reasonable graphical output. In general, however,
I advise that 'checkgrid' remains set in its default state of true.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.3 Stratigraphic Grids 81

plotGrid(G,'FaceColor','none','EdgeAlpha',.1); plotGrid(G, actnum(G.cells.indexMap), ...

plotFaces(G, G.faces.tag>0, ... 'FaceColor','none','EdgeAlpha',.1);
'FaceColor','red','FaceAlpha',.2, ... plotGrid(G, actnum(G.cells.indexMap), ...
'EdgeColor','r','EdgeAlpha',.1); 'FaceColor','y', 'EdgeAlpha',.1);
axis off; view(-155,80); zoom(1.7); axis off; view(-155,80); zoom(1.7);

Figure 3.21 The Norne field from the Norwegian Sea. The plots show the whole grid with fault faces
marked in red (left) and active cells marked in yellow (right).

we start processing the input, because if the ACTNUM flag is set, all inactive cells will be
ignored when the unstructured grid is built
actnum = grdecl.ACTNUM;
grdecl.ACTNUM = ones(prod(grdecl.cartDims),1);
G = processGRDECL(grdecl, 'checkgrid', false);

Having processed and converted the grid to the correct unstructured format, we first plot
the outline of the whole model and highlight all faults and the active part of the model;
see Figure 3.21. During the processing, all fault faces are tagged with a positive number.
This can be utilized to highlight the faults: we simply find all faces with a positive tag, and
color them with a specific color as shown in the left box in the figure. To continue with
the active model only, we reset the ACTNUM field to its original values so that inactive
cells are ignored when we process the ECLIPSE input stream [270, 33]. This gives a grid
with two components: the main reservoir and a detached stack of twelve cells that should
be ignored.
To examine some parts of the model in more detail, we can use the function cutGrdecl
to extract a rectangular box in index space from the ECLIPSE input stream, e.g., as follows:
cut_grdecl = cutGrdecl(grdecl, [6 15; 80 100; 1 22]);
g = processGRDECL(cut_grdecl);

Figure 3.22 shows a zoom on four different regions. The first region (red color), is
sampled near a laterally stair-stepped fault, i.e., a curved fault surface that has been
approximated by a surface that zigzags in the lateral direction. We also notice how the fault

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 82 Grids in Subsurface Modeling

a) b)

c) d)

e)

Figure 3.22 Detailed view of subsets from the Norne simulation model. a) The whole model with
active and inactive cells and four regions of interest marked in different colors. b) Zoom of the
red region with pillars and corner-points shown as red circles. c) The magenta region with coloring
according to cell volumes, which vary by a factor 700. d) The blue region in which fault faces have
been colored gray and the corresponding coordinate lines have been colored blue. e) The green cross-
section with coloring according to layer number from top to bottom.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.3 Stratigraphic Grids 83

displacement leads to cells that are nonmatching across the fault surface, and the presence
of some very thin layers (the thinnest layers may actually appear as thick lines in the plot).
The thin layers are also clearly seen in the second region (magenta color), which represents
a somewhat larger sample from an area near the tip of one of the “fingers” in the model.
Here, we clearly see how similar layers have been strongly displaced across the fault zone.
In the third (blue) region, we have colored the fault faces to clearly show the displacement
and the hole through the model in the vertical direction, which likely corresponds to a shale
layer that has been eliminated from the active model. Gaps and holes, and displacement
along fault faces, are even more evident for the vertical cross section (green region) for
which the layers have been given different colors, as in Figure 3.15. Altogether, the four
views of the model demonstrate typical patterns that you can see in realistic models. We
return to this simulation model in Section 10.3.8.

Extensions, Difficulties, and Challenges

The original corner-point format has been extended in several ways, e.g., to enable vertical
intersection of two coordinate lines in the shape of the letter Y. Coordinate lines may also
be defined by piecewise polynomial curves, resulting in what is sometimes called S-faulted
grids. Likewise, two neighboring coordinate lines can collapse so that the basic grid shape
becomes a prism rather than a hexahedron. However, there are several features you cannot
easily model, like multiple fault intersections (e.g., as in the letter F), and the industry is
thus constantly in search for improved griding methods. One example will be discussed in
the next subsection. First, however, we will discuss some difficulties and challenges, seen
from the side of a computational scientist seeking to use corner-point grids for his or her
computations.
The flexible cell geometry of the corner-point format poses several challenges for
numerical implementations. Indeed, a geocellular grid is typically chosen by a geologist
who tries to describe the rock body by as few volumetric cells as possible and who may not
be aware of potential numerical difficulties his or her choice of geometries and topologies
may cause in subsequent flow simulations.
Writing a robust grid-processing algorithm to compute geometry and topology or to
determine an equivalent matching, polyhedral grid proved to be much more challenging
than we anticipated when we started developing MRST. Displacements across faults will
lead to geometrically complex, nonconforming grids, e.g., as illustrated in Figure 3.22.
Since each face of a grid cell is specified by four (arbitrary) points, the cell interfaces in the
grid will generally be bilinear and possibly strongly curved surfaces. Geometrically, this
can lead to several complications. Cell faces on different sides of a fault may intersect each
other so that cells overlap volumetrically, or they may leave void spaces in between them.
There may be tiny overlap areas between cell faces on different sides a fault, and so on.
All these factors contribute to make fault geometries hard to interpret in a geometrically
consistent way: a subdivision into triangles is, for instance, not unique. Likewise, top and
bottom surfaces may intersect for highly curved cells with high aspect ratios, cell centroids
may be outside the cell volume, etc.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 84 Grids in Subsurface Modeling

a) b)

c) d)

Figure 3.23 Illustration of difficulties associated with real-life corner-point geometries. a) Many
faces resulting from subdivision to give matching grid at faults. b) A curved 800 × 800 × 0.25 m cell,
whose centroid lies outside the cell. c) Difficult geometries. d) Small interface between two cells.

The presence of collapsed corner-points implies that the grid cells will generally be poly-
hedral and possibly contain both triangular and bilinear faces (see Figure 3.14). Likewise,
individual cells can have many neighbors; the Norne model has cells with up to 20 neigh-
bors. Altogether, this calls for flexible discretizations that work on unstructured topologies
and are not sensitive to the geometry of each cell or the number of faces and corner points.
Simple discretization methods that only account for pressure differences between cells
sharing a common face (see Section 4.4.1 are fairly easy to implement for general unstruc-
tured grids, but more advanced discretization methods that rely on the use of dual grids or
reference elements can be quite challenging to implement on highly irregular grids. Figure
3.23 illustrates some geometrical and topological challenges seen in standard grid models.
Cell geometries will often deviate significantly from being orthogonal to adapt to com-
plex horizons and fault surfaces. This may introduce unphysical preferential flow directions
in the numerical methods (as we will see later). Stratigraphic grids will also typically
have aspect ratios that are two or three orders of magnitude, which can introduce severe
numerical difficulties because the majority of the flow in and out of a cell occurs across the
faces with the smallest area. Similarly, the possible presence of strong heterogeneities and
anisotropies in the permeability fields, e.g., as seen in the SPE 10 example in Chapter 2,
typically introduces large condition numbers in the discretized flow equations, which make
them hard to solve.
Corner-point grids generated by geological modeling typically contain too many cells.
Once created by the geologist, the grid is handed to a reservoir engineer, whose first job is
to reduce the number of cells if he or she is to have any hope of getting the model through
a simulator. The generation of good coarse grids for use in upscaling, and the upscaling
procedure itself, is generally work-intensive, error-prone, and not always sufficiently
robust, as we will come back to in Chapter 15.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.3 Stratigraphic Grids 85

3.3.2 2.5D Unstructured Grids

So-called 2.5D grids have been designed to combine the advantages of two different griding
methods: the (areal) flexibility of unstructured grids and the simple topology of Cartesian
grids in the vertical direction. These grids are constructed in much the same way as corner-
point grids, except that coordinate lines now are defined over an unstructured lateral grid so
that pillars will have polygonal base area. One starts by defining an areal tessellation on a
surface that either aligns with the lateral direction or one of the major geological horizons.
Then a coordinate line is introduced through each vertex in the areal grid. The lines can
either be vertical, inclined to gradually align with major fault planes, as shown for the
corner-point grid in Figure 3.20, or be defined so that they connect pairs of vertices in areal
tessellations of two different geological horizons.
Figure 3.24 shows the key steps in the construction of a simple 2.5D PEBI grid. Start-
ing from a Delaunay triangulation of set of generating points, we form a perpendicu-
lar bisector grid. Through each vertex in this areal tessellation, we define a coordinate
line, whose angle of inclination will change from 90 degrees for vertices on the far left
to 45 degrees for vertices on the far right. coordinate lines associated with individual
areas cells form pillars, and are next used to extrude the areal tessellation to a volumet-
ric grid. The resulting volumetric grid is unstructured in the lateral direction, but has
a layered structure in the vertical direction and can thus be indexed using an ik index
pair. This format has quite large freedom in choosing the size and shape of the grid cells
to adapt to complex features such as curved faults or to improve the areal resolution in
near-well zones.

a) b)

c) d)

Figure 3.24 Illustration of a typical process for generating 2.5D PEBI grids. a) Triangulated point
set. b) Perpendicular bisector grid. c) Coordinate lines aligned with faults. d) Volumetric extrusion.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 86 Grids in Subsurface Modeling

Figure 3.25 The left plot shows a 2.5D Voronoi grid derived from a perturbed 2D point mesh extruded
in the z-direction, whereas the right plot shows a radial grid.

Example 1: We first construct an areal Voronoi grid from a set of generating points
obtained by perturbing the vertices of a regular Cartesian grid and then use the func-
tion makeLayeredGrid to extrude this Voronoi grid to 3D along vertical pillars in the
z-direction.
N=7; M=5; [x,y] = ndgrid(0:N,0:M);
x(2:N,2:M) = x(2:N,2:M) + 0.3*randn(N-1,M-1);
y(2:N,2:M) = y(2:N,2:M) + 0.3*randn(N-1,M-1);
aG = pebi(triangleGrid([x(:) y(:)]));
G = makeLayeredGrid(aG, 3);
plotGrid(G, 'FaceColor',[.8 .8 .8]); view(-40,60); axis tight off

The resulting grid is shown in the left plot of Figure 3.25 and should be contrasted to the
3D tetrahedral tessellation shown to the right in Figure 3.8
Example 2: Radial symmetric grids graded towards the origin are commonly used to
increase resolution near wells. Figure 3.25 shows one example that can be generated as
follows:
P = [];
for r = exp(-3.5:.25:0),
[x,y,z] = cylinder(r,16); P = [P [x(1,:); y(1,:)]];
end
P = unique([P'; 0 0],'rows');
G = makeLayeredGrid(pebi(triangleGrid(P)), 5);
plotGrid(G,'FaceColor',[.8 .8 .8]); view(30,50), axis tight off

Example 3: The model shown in Figure 3.26 is a realistic representation of a reservoir in

which an unstructured areal grid has been extruded vertically to model different geological
layers. Some of the layers are very thin and appear as if they were thick lines in plot a. Near
the perimeter of the model (plot b), we notice that the lower layers (yellow to red colors)
have been eroded away in most of the grid columns, and although the vertical dimension is
strongly exaggerated, we see that the layers contain steep slopes. To a non-geologist look-
ing at subplot e, it may appear as if the reservoir was formed by sediments being deposited
along a sloping valley that ends in a flat plain. Subplots c and d show more details of the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.3 Stratigraphic Grids 87

a)

b) c)

d) e)

Figure 3.26 A real petroleum reservoir modeled by a 2.5D PEBI grid having 1,174 cells in the lateral
direction and 150 cells along each pillar. Only 90,644 out of the 176,100 cells are active. The plots
show the whole model as well as selected details. a) The whole model with three areas of interest
marked in different colors. b) Layers 41 to 99 of the red region with colors representing the k-index.
c) Horizontal permeability in the green region. d) Horizontal permeability in the magenta region.
e) Horizontal permeability along the perimeter and bottom of the model.

permeability field inside the model. The layering is particularly distinct in plot d, which
is sampled from the flatter part of the model. The cells in plot c, on the other hand, show
examples of pinch-outs. The layering provides a certain ik structure in the model, similar
to the logical ij k index for corner-point grids, where i refers to the areal numbering and k
to the different layers. Some also prefer to associate a virtual logically Cartesian grid as an
overlay to the 2.5D grid that can be used, e.g., to simplify lookup of cells in visualization.
In this setup, more than one grid cell may be associated with a cell in the virtual grid.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 88 Grids in Subsurface Modeling

Figure 3.27 Example of a virtual 37 × 20 × 150 grid used for fast lookup in a 2.5D PEBI grid with
dimensions 1174 × 150.

3.4 Grid Structure in MRST

The two previous sections introduced various structured and unstructured grid types that
can be created using the many grid-factory routines in MRST. In this section, we go
into more detail about the internal data structure used to represent all these different grid
formats. This data structure is in many ways the most fundamental part of MRST since
almost all solvers, workflow tools, and visualization routines require an instance of a grid
as input argument. By convention, instances of the grid structure are denoted G. If you are
mainly interested in using solvers and visualization routines already available in MRST,
you need no further knowledge of the grid structure beyond what has been encountered
in the examples presented so far, and you can safely skip the remains of this section. On
the other hand, if you wish to use the software to prototype new computational methods,
knowledge of the inner workings of the grid structure is essential.
To provide a uniform way of accessing different grid types, all grids are stored using
the same unstructured grid format that explicitly represents cells, faces, vertices, and con-
nections between cells and faces. The main grid structure G contains three fields – cells,
faces, and nodes – that specify individual properties for each individual cell/face/vertex
in the grid. Grids in MRST can either be volumetric or lie on a 2D or 3D surface. The
field griddim distinguishes volumetric and surface grids; all cells in a grid are polygonal
surface patches if griddim=2 and polyhedral volumetric entities otherwise. In addition, the
grid contains a field type consisting of a cell array of strings describing the history of grid
constructors and modifier functions that have been applied to create and modify the grid,
e.g., 'tensorGrid'. For grids having an underlying logical Cartesian structure, we also
include the field cartDims.
As with the rest of the basic functionality in MRST, the grid structure is thoroughly
documented in the code. To read the documentation, type
help grid_structure

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.4 Grid Structure in MRST 89

Before we continue to discuss the individual data members in the structure, we first have to
introduce you to two essential implementation tricks in MATLAB.

Indirection maps and run-length encoding. In an unstructured grid, individual cells may
have a different number of faces, and faces may have different number of edges/vertices.
A compact way to represent such data is to use a data container and an indirection map. To
illustrated, let us consider a simple example of a data object b, whose first members consist
of five, three, and four numbers. This can be stored using an indirection map m and a data
array a = [a1, . . . ,a12 ], as illustrated:

m: 1 6 9 13

a: a1 a2 a3 a4 a5 a6 a7 a8 a9 a10 a11 a12

b1 b2 b3

Using this data structure, we can quickly find all the a values that belong to a particular
bi . In some cases, we may wish to go the other way around and determine the bi element
a particular value aj corresponds to. A simple solution would be to introduce an extra row
in a that contains the values [1,1,1,1,1,2,2,2,3,3,3,3]. However, this contains a lot of
redundant information and can be stored much more compactly using run-length encoding,
which is a simple lossless data compression method that removes repeated values. In this
particular example, the run-length encoding would give a compressed array b = [1,2,3]
and a repetition count n = [5,3,4]. This data compression and decompression technique
is used abundantly at a low level in MRST to avoid storing redundant information, as you
will see shortly.

Representing grid cells. The cell structure G.cells consists of the mandatory fields:

– num: the number nc of cells in the global grid.

– facePos: an indirection map of size [num+1,1] into the faces array. Specifically, the
face information of cell i is found in the submatrix
faces(facePos(i): facePos(i+1)-1, :)

The number of faces per cell may be computed as diff(facePos), whereas the
total number of faces is nf = facePos(end)-1.
– faces: an nf × 2 array specifying the global faces connected to a given cell. Specifi-
cally, faces(i,1)==j if face with global number faces(i,2) is connected to cell
number j . The array may optionally have a third component, faces(i,3), that
for certain types of grids contains a name tag used to distinguish face directions:
West, East, South, North, Bottom, Top.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 90 Grids in Subsurface Modeling

The first column of faces is redundant: it consists of each cell index j repeated
facePos(j+1)-facePos(j) times and can therefore be reconstructed by decom-
pressing a run-length encoding with the cell indices 1:num as the encoded vector
and the number of faces per cell as repetition count. Hence, to conserve memory,
only the last two columns of faces are stored, while the first column can be
reconstructed using the statement:
rldecode(1:G.cells.num, diff(G.cells.facePos), 2) .'

This construction is used a lot throughout MRST and has therefore been imple-
mented as a utility function inside mrst-core/utils/gridtools
f2cn = gridCellNo(G);

– indexMap: an optional nc ×1 array that maps internal cell indices to external cell indices.
For models with no inactive cells, indexMap equals 1 : nc . For cases with inactive
cells, indexMap contains the indices of the active cells sorted in ascending order.
For grids with an underlying Cartesian organization, a map of cell numbers to
logical indices can be constructed using the following statements in 2D and 3D:
[ij{1:2}] = ind2sub(dims, G.cells.indexMap(:)); % In 2D
ij = [ij{:}];
[ijk{1:3}] = ind2sub(dims, G.cells.indexMap(:)); % In 3D
ijk = [ijk{:}];

In the latter case, ijk(i:) is the global (I,J,K) index of cell i.

In addition, the cell structure can contain the following optional fields that typically will be
added by a call to computeGeometry:
– volumes: an nc × 1 array of cell volumes
– centroids: an nc × d array of cell centroids in IRd

Representing cell faces. The mandatory fields of the face structure, G.faces, are quite sim-
ilar to those of the cells in the sense that they provide mappings to the lower-dimensional
objects (nodes) that delimit the faces:
– num: the number nf of global faces in the grid.
– nodePos: an indirection map of size [num+1,1] into the nodes array. Specifically, the
node information of face i is found in the submatrix
nodes(nodePos(i): nodePos(i+1)-1, :)
The number of nodes of each face may be computed using the statement diff
(nodePos). Likewise, the total number of nodes is given as nn = nodePos(end)-1.
– nodes: an nn × 2 array of vertices in the grid. If nodes(i,1)==j, the local vertex i
is part of global face number j and corresponds to global vertex nodes(i,2).
For each face the nodes are assumed to be oriented such that a right-hand rule

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.4 Grid Structure in MRST 91

determines the direction of the face normal. As for cells.faces, the first column
of nodes is redundant and can be easily reconstructed. Hence, to conserve memory,
only the last column is stored, while the first column can be constructed using the
statement:
rldecode(1:G.faces.num, diff(G.faces.nodePos), 2) .'

– neighbors: an nf × 2 array of neighboring information. Global face i is shared

by global cells neighbors(i,1) and neighbors(i,2). One of the entries in
neighbors(i,:), but not both, can be zero, to indicate that face i is an external
face that belongs to only one cell (the nonzero entry).

In addition to the mandatory fields, G.faces has optional fields containing geometry infor-
mation that typically are added by a call to computeGeometry:

– areas: an nf × 1 array of face areas.

– normals: an nf × d array of area weighted, directed face normals in IRd . The normal
on face i points from cell neighbors(i,1) to cell neighbors(i,2).
– centroids: an nf × d array of face centroids in IRd .

Moreover, G.faces can sometimes contain an nf × 1 (int8) array, G.faces.tag, with user-
defined face indicators used, e.g., to specify that the face belongs to a fault.

Representing vertices. The vertex structure, G.nodes, consists of two fields:

– num: number nn of global nodes (vertices) in the grid,

– coords: an nn ×d array of physical nodal coordinates in IRd . Global node i is at physical
coordinate coords(i,:).

The grid is constructed according to a right-handed coordinate system where the z coordi-
nate is interpreted as depth. Consequently, plotting routines such as plotGrid display the
grid with a reversed z axis. To illustrate how the grid structure works, we consider two
examples.
Example 1: We start by considering a regular 3 × 2 grid, in which we take away the
second cell in the logical numbering,
G = removeCells( cartGrid([3,2]), 2)

This produces the output

G =
cells: [1x1 struct]
faces: [1x1 struct]
nodes: [1x1 struct]
cartDims: [3 2]
type: {’tensorGrid’ ’cartGrid’ ’removeCells’}
griddim: 2

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 92 Grids in Subsurface Modeling

cells.faces = faces.nodes = faces.neighbors =

1 1 1 (West) 1 1 0 1
1 9 3 (South) 1 5 1 0
1 2 2 (East) 2 2 0 2
1 11 4 (North) 2 6 2 0
2 3 1 (West) 3 3 0 3
2 10 3 (South) 3 7 3 4
2 4 2 (East) 4 4 4 5
2 13 4 (North) 4 8 5 0
3 5 1 (West) 5 5 0 1
3 11 3 (South) 5 9 0 2
3 6 2 (East) 6 6 1 3
3 14 4 (North) 6 10 0 4
4 6 1 (West) 7 7 2 5
4 12 3 (South) 7 11 3 0
4 7 2 (East) 8 8 4 0
4 15 4 (North) 8 12 5 0
5 7 1 (West) 9 2
5 13 3 (South) 9 1
5 8 2 (East) : :
5 16 4 (North) : :

Figure 3.28 Illustration of the cell and faces fields of the grid structure: cell numbers are marked
by circles, node numbers by squares, and face numbers have no marker.

Examining the output from the call, we notice that the field G.type contains three values,
'cartGrid' indicates the creator of the grid, which again relies on 'tensorGrid', whereas
the field 'removeCells' indicates that cells have been removed from the Cartesian topol-
ogy. The resulting 2D geometry consists of five cells, twelve nodes, and sixteen faces.
All cells have four faces and hence G.cells.facePos = [1 5 9 13 17 21]. Figure
3.28 shows4 the geometry and topology of the grid, including the content of the fields
cells.faces, faces.nodes, and faces.neighbors. We notice, in particular, that all interior
faces (6, 7, 11, and 13) are represented twice in cells.faces as they belong to two different
cells. Likewise, for all exterior faces, the corresponding row in faces.neighbors has one
zero entry. Finally, being logically Cartesian, the grid structure contains a few optional
fields:
– G.cartDims equals [3 2],
– G.cells.indexMap equals [1 3 4 5 6], since the second cell in the logical numbering
has been removed from the model, and
– G.cells.faces contains a third column with tags that distinguish global directions for
the individual faces.
Example 2: We consider the Delaunay triangulation of seven points in 2D:
p = [ 0.0, 1.0, 0.9, 0.1, 0.6, 0.3, 0.75; ...
0.0, 0.0, 0.8, 0.9, 0.2, 0.6, 0.45]'; p = sortrows(p);
G = triangleGrid(p)

4 To create the plot in Figure 3.28, we first called plotGrid to plot the grid, then called computeGeometry to compute cell and
face centroids, which were used to place a marker and a text label with the cell/face number in the correct position.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.4 Grid Structure in MRST 93

cells.faces = faces.nodes = faces.neighbors =

1 10 1 1 0 2
1 8 1 2 2 3
1 7 2 1 3 5
2 1 2 3 5 0
2 2 3 1 6 2
2 5 3 4 0 6
3 3 4 1 1 3
3 7 4 7 4 1
3 2 5 2 6 4
4 8 5 3 1 8
4 12 6 2 8 5
4 9 6 6 4 7
5 3 7 3 7 8
5 4 7 4 0 7
5 11 8 3
6 9 8 5
6 6 9 3
6 5 9 6
7 13 10 4
7 14 10 5
: : : :

Figure 3.29 Illustration of the cell and faces fields of the grid structure: cell numbers are marked
by circles, node numbers by squares, and face numbers have no marker.

which produces the output

G =
faces: [1x1 struct]
cells: [1x1 struct]
nodes: [1x1 struct]
type: {’triangleGrid’}
griddim: 2

The grid contains no structured parts, and thus G consists of the three mandatory fields,
cells, faces, and nodes, that are sufficient to determine the geometry and topology of
the grid; the type tag that names its creator, and griddim that gives that it is a surface
grid. Altogether, the grid consists of eight cells, fourteen faces, and seven nodes, which are
shown in Figure 3.29, along with the contents of the fields cells.faces, faces.nodes, and
faces.neighbors. Notice, in particular, the absence of the third column in cells.faces,
since face tags associated with the axial directions generally do not make sense for a (fully)
unstructured grid. Likewise, the cells structure does not contain any indexMap, as all cells
in the model are active.

Computing geometry information. All grid-factory routines in MRST generate the basic
geometry and topology of a grid, that is, how nodes are connected to make up faces,
how faces are connected to form cells, and how cells are connected over common faces.
Whereas this information is sufficient for many purposes, more geometric information
may be required in many cases. As explained earlier, such information is provided by
the routine computeGeometry, which computes cell centroids and volumes as well as face
areas, centroids, and normals. Computing this information is straightforward for simplexes
and Cartesian grids, but is not so for general polyhedral grids that may contain curved
polygonal faces. In the following we will therefore go through how it is done in MRST.
Our discussion follows the original implementation, which in recent releases has been

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 94 Grids in Subsurface Modeling

superseded by a more optimized routine. You can find the complete original code in
computeGeometryOrig.
For each cell, the basic grid structure provides us with a list of vertices, a list of cell
faces, etc., as shown in plots a and e of Figure 3.30. The routine starts by computing face
quantities (areas, centroids, and normals). To utilize MATLAB efficiently, the computations
are programmed using vectorization so that each derived quantity is computed for all points,
all faces, and all cells in one go. To keep the current presentation as simple as possible, we
will herein only give formulas for a single face and a single cell. Let us consider a single
face given by the points p(i  m ) and let α = (α1, . . . ,αm ) denote a multi-index
 1 ), . . . , p(i
describing how these points are connected to form the outline of the faces. For the face with
global number j , the multi-index is given by the vector
G.faces.nodes(G.faces.nodePos(j):G.faces.nodePos(j+1)-1)

Let us consider two faces. Global face number two in Figure 3.30a is planar and consists of

points p(2), 
p(4), 
p(6),  with the ordering α = (2,4,8,6). Likewise, we consider global
p(8)
face number one in Figure 3.30e, which is curved and consists of points p(1),  
. . . , p(5)
with the ordering α = (4,3,2,1,5). For curved faces, we need to make a choice of how
to interpret the surface spanned by the node points. In MRST (and some commercial
simulators) this is done as follows: we start by defining a so-called hinge point ph , which
is often given as part of the input specification of the grid. If not, the hinge point can be

computed as the center point of the m points that make up the face, ph = m  k )/m.
k=1 p(α
The hinge point can now be used to tessellate the face into m triangles, as shown in Figures
3.30b and 3.30f. The triangles are defined by the points p(α  mod(k,m)+1 ), and ph for
 k ), p(α
k = 1, . . . ,m. Each triangle has a center point pck defined in the usual way as the average
of its three vertexes and a normal vector and area given by
   
nk = p(α
 mod(k,m)+1 ) − p(α
 k ) × ph − p(α
 k ) = v1k × v2k

Ak = nk · nk .

The face area, centroid, and normal are now computed as follows


m 
m 
m
Af = Ak , cf = (Af )−1 pck Ak , nf = nk . (3.2)
k=1 k=1 k=1

The result is shown in plot c of Figure 3.30, where the observant reader will see that the
centroid cf does not coincide with the hinge point ph unless the planar face is a square.
This effect is more pronounced for the curved faces of the PEBI cell in Figure 3.30g.
The computation of centroids in (3.2) is only valid if the grid does not have faces with
zero area, because otherwise the second formula would involve a division by zero and hence
incur centroids with NaN values. The reader interested in creating his/her own grid-factory
routines for grids that may contain degenerate (pinched) cells should be aware of this and
make sure that all faces with zero area are removed in a preprocessing step.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.4 Grid Structure in MRST 95

a) b)

c) d)

e) f)

g) h)

Figure 3.30 Steps in the computation of geometry information for a single corner-point cell and a
single PEBI cell using computeGeometry. a) A single corner-point cell with face numbers (squares)
and node numbers (circles). b) Tessellation of faces with vectors v1k (blue), v2k (green), and nk (red).
c) Face centroids and normal vectors computed from tessellation. d) Triangulation of cell volume
with vectors nk (blue) and crk (green). e) A single PEBI cell with face numbers (squares) and node
numbers (circles). f) Tessellation of faces with vectors v1k (blue), v2k (green), and nk (red). g) Face
centroids and normal vectors computed from tessellation. h) Triangulation of cell volume with vectors
nk (blue) and crk (green).

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 96 Grids in Subsurface Modeling

To compute the cell centroid and volume, we start by computing the center point cc of the
mf
cell, which we define as the average of the face centroids, cc = k=1 cf /mf , where mf
is the number of faces of the cell. By connecting this center point to the mt face triangles,
we define a unique triangulation of the cell volume, as shown in Figures 3.30d and 3.30h.
For each tetrahedron, we define the vector crk = pck − cc and compute the volume (which
may be negative if the center point cc lies outside the cell)
V k = 13 crk · nk .
The triangle normals nk will point outward or inward depending upon the orientation of the
points used to calculate them, and to get a correct computation we therefore must modify
the triangle normals so that they point outward. Finally, we can define the volume and the
centroid of the cell as follows

mt
3 
mt
V = V k, c = cc + V k crk . (3.3)
4V
k=1 k=1

In the original implementation, all cell quantities were computed inside a loop, which may
not be as efficient as the computation of the face quantities. This has been improved in
recent releases.

Computer exercises
3.4.1 Go back to Exercise 3.1.1. What would you do to randomly perturb all nodes in
the grid except for those that lie on an outer face whose normal vector has no
component in the y-direction?
3.4.2 Exercise 3.2.2 extended the function triangleGrid to triangulated surfaces in 3D.
Verify that computeGeometry computes cell areas, cell centroids, face centroids, and
face lengths correctly for 3D surfaces.
3.4.3 How would you write a function that purges all cells that have an invalid vertex
(with value NaN) from a grid?

3.5 Examples of More Complex Grids

To help users generate test cases, MRST supplies a number of routines for generating
example grids. We have previously encountered twister, which perturbs the x and y
coordinates in a grid. Likewise, in Section 2.5 we used simpleGrdecl to generate a simple
ECLIPSE input stream for a stratigraphic grid describing a wavy structure with a single
deviated fault. The routine has several options that enable you to specify the magnitude of
the fault displacement, flat rather than a wavy top and bottom surfaces, and vertical rather
than inclined coordinate lines; see Figure 3.31.
The routine with the somewhat cryptic name makeModel3 generates a corner-point input
stream that models parts of an anticline structure that is cut through by two faults; see Figure
3.32. Similarly, extrudedTriangleGrid generates a 2.5D prismatic grid with a laterally

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.5 Examples of More Complex Grids 97

simpleGrdecl([20, 20, 5]); simpleGrdecl([20, 20, 5], ... simpleGrdecl([20, 20, 5], ...
@(x) .05*(sin(2*pi*x)-1.5)); @(x) .25*(x-.5),'flat',true);

Figure 3.31 The simpleGrdecl routine can be used to produce faulted, two-block grids of different
shapes.

makeModel3([30,20,5]); extrudedTriangleGrid(50); extrudedTriangleGrid(50,true);

Figure 3.32 Three different example grids created by the grid example functions makeModel3 and
extrudedTriangleGrid.

curved fault in the middle. Alternatively, the routine can generate a 2.5D PEBI grid in
which the curved fault is laterally stair-stepped; see Figure 3.32.

3.5.1 SAIGUP: Model of a Shallow-Marine Reservoir

Having discussed the corner-point format in some detail, it is now time to return to the
SAIGUP model. In the following discussion, we will look at the grid representation in
more detail and show some examples of how to interact and visualize different features of
the grid (see also the last video of the second MRST Jolt on grids and petrophysical data
[189]). You can find the complete source code showGridSAIGUP. In Section 2.5, we saw
that parsing the input file creates the following structure
grdecl =
cartDims: [40 120 20]
COORD: [29766x1 double]
ZCORN: [768000x1 double]
ACTNUM: [96000x1 int32]
PERMX: [96000x1 double]
: : :

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 98 Grids in Subsurface Modeling

The first four fields describe the grid and the remaining the petrophysical propertes. Here,
we will (mostly) consider the first four fields, which represent the following information:
1. The dimension of the underlying logical Cartesian grid: ECLIPSE keyword SPECGRID,
equal 40 × 120 × 20.
2. The coordinate lines: ECLIPSE keyword COORD, top and bottom coordinate per vertex
in the logical 40 × 120 areal grid, i.e., 6 × 41 × 121 values.
3. The coordinates along the individual coordinate lines: ECLIPSE keyword ZCORN, eight
values per cell, i.e., 8 × 40 × 120 × 20 values.
4. The Boolean indicator for active cells: ECLIPSE keyword ACTNUM, one value per cell,
i.e., 40 × 120 × 20 values.
To process the ECLIPSE input stream and turn the corner-point grid into MRST’s unstruc-
tured description, we use the processGRDECL routine. The interested reader may ask the
processing routine to display diagnostic output
G = processGRDECL(grdecl, 'Verbose', true);
G = computeGeometry(G)

and consult the SAIGUP tutorial (showSAIGUP) or the technical documentation of the
processing routine for a more detailed explanation of the resulting output.
The model has been created with vertical pillars having lateral resolution of 75 meters
and vertical resolution of 4 meters, giving a typical aspect ratio of 18.75. You can see this,
e.g., by extracting the pillars and corner points and analyzing the results as follows:
[X,Y,Z] = buildCornerPtPillars(grdecl,'Scale',true);
dx = unique(diff(X)).'
[x,y,z] = buildCornerPtNodes(grdecl);
dz = unique(reshape(diff(z,1,3),1,[]))

The resulting grid has 78,720 cells that are almost equal in size (consult the histogram
hist(G.cells.volumes)), with cell volumes varying between 22,500 m3 and 24,915 m3 .
Altogether, the model has 264,305 faces: 181,649 vertical faces on the outer boundary and
between lateral neighbors, and 82,656 lateral faces on the outer boundary and between
vertical neighbors. Most of the vertical faces are not part of a fault and are therefore
parallelograms with area equal 300 m2 . However, the remaining 26,000–27,000 faces are
a result of the subdivision introduced to create a matching grid along the (stair-stepped)
faults. Figure 3.33 shows where these faces appear in the model and a histogram of their
areas: the smallest face has an area of 5.77·10−4 m2 and there are 43, 202, and 868 faces
with areas smaller than 0.01, 0.1, and 1 m2 , respectively. The processGRDECL routine has an
optional parameter 'Tolerance' that sets the minimum distance for distinguishing points
along coordinate lines (default value is zero). By setting this to parameter to 5, 10, 25, or 50
cm, the area of the smallest face is increased to 0.032, 0.027, 0.097, or 0.604 m2 , respec-
tively. In general, we advice against aggressive use of this tolerance parameter; one should

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.5 Examples of More Complex Grids 99

plotFaces(G,G.faces.tag>0 & ...

G.faces.areas>=290),'y','edgea',.1);
plotFaces(G, G.faces.areas<290,'r', 'edgea',.1);

Figure 3.33 Faces that have been subdivided for the SAIGUP mode. The left plot shows a histogram
of the faces areas. The right plot shows all fault faces (yellow) and fault faces having area less than
290 m2 (red).

instead develop robust discretization schemes and, if necessary, suitable postprocessing

methods that eliminate or ignore faces with small areas.

Visualizing subsets of the grid. Next, we show a few examples of visualizations highlight-
ing various mechanisms for interacting with the grid and accessing parts of it. As a first
example, we start by plotting the layered structure of the model. To this end, we use a
simple trick: create a matrix with ones in all cells of the logical Cartesian grid and then do
a cumulative summation in the vertical direction to get increasing values,
val = cumsum(ones(G.cartDims),3);

which we then plot using a standard call to plotCellData; see the left plot in Figure 3.34.
To reveal the layering in the interior of the model, we extract and visualize only the cells
that are adjacent to a fault:
cellList = G.faces.neighbors(G.faces.tag>0, :);
cells = unique(cellList(cellList>0));

The first statement uses logical indexing to loop through all faces and extract the neighbor-
ing cells of all faces that are marked with a tag (i.e., lie at a fault face). The list may have
repeated entries if a cell is attached to more than one fault face and contain zeros if a fault
face is part of the outer boundary. We get rid of these in the second statement, and can then
plot the result list of cells, giving the plot shown to the right in Figure 3.34
plotCellData(G,val(G.cells.indexMap),cells)

Next, let us inspect the fault structure in the lower-right corner of the plot. If we disregard
using cutGrdecl as discussed on page 81, there are basically two ways we can extract parts

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 100 Grids in Subsurface Modeling

Figure 3.34 Visualizing the layered structure of the SAIGUP model.

of the model, which both rely on the construction of a map of cell numbers of logical
indices. In the first method, we first construct a logical set for the cells in a logically
Cartesian bounding box and then use the built-in function ismember to extract the members
of cells that lie within this bounding box:
[ijk{1:3}] = ind2sub(G.cartDims, G.cells.indexMap); IJK = [ijk{:}];
[I,J,K] = meshgrid(1:9,1:30,1:20);
bndBox = find(ismember(IJK,[I(:), J(:), K(:)],'rows'));
inspect = cells(ismember(cells,bndBox));

The ismember function has an operational count of O(n log n). A faster alternative is to
use logical operations having an operational count of O(n). That is, we construct a vector
of boolean numbers that are true for the entries we want to extract and false for the
remaining entries
I = false(G.cartDims(1),1); I(1:9)=true;
J = false(G.cartDims(2),1); J(1:30)=true;
K = false(G.cartDims(3),1); K(1:20)=true;
pick = I(ijk{1}) & J(ijk{2}) & K(ijk{3});
pick2 = false(G.cells.num,1); pick2(cells) = true;
inspect = find(pick & pick2);

Both approaches produce the same index set; the resulting plot is shown in Figure 3.35. To
mark the fault faces in this subset of the model, we do the following steps
cellno = gridCellNo(G);
faces = unique(G.cells.faces(pick(cellno), 1));
inspect = faces(G.faces.tag(faces)>0);
plotFaces(G, inspect, [.7 .7 .7], 'EdgeColor','r');

The first statement constructs a list of cells attached to all faces in the model, the second
extracts a unique list of face numbers associated with the cells in the logical vector pick
(which represents the bounding box in logical index space), and the third statement extracts
the faces within this bounding box that are marked as fault faces.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.5 Examples of More Complex Grids 101

Figure 3.35 Details from the SAIGUP model showing a zoom of the fault structure in the lower-right
corner of the right plot in Figure 3.34. The left plot shows the cells attached to the fault faces, and in
the right plot the fault faces have been marked with gray color and red edges.

Logical operations are also useful in other circumstances. As an example, we will extract
a subset of cells forming a sieve that can be used to visualize the petrophysical quantities
in the interior of the model:
% Every fifth cell in the x-direction
I = false(G.cartDims(1),1); I(1:5:end)=true;
J = true(G.cartDims(2),1);
K = true(G.cartDims(3),1);
pickX = I(ijk{1}) & J(ijk{2}) & K(ijk{3});

% Every tenth cell in the y-direction

I = true(G.cartDims(1),1);
J = false(G.cartDims(2),1); J(1:10:end) = true;
pickY = I(ijk{1}) & J(ijk{2}) & K(ijk{3});

% Combine the two picks

plotCellData(G,rock.poro, pickX | pickY, 'EdgeColor','k','EdgeAlpha',.1);

3.5.2 Composite Grids

Using an unstructured grid description enables you to define composite grids whose geome-
tries and topologies vary throughout the model. That is, different types of cells or different
grid resolution may be used locally to adapt to well trajectories and flow and geological
constraints; see e.g., [123, 209, 113, 202, 51, 83, 292] and references therein. You may
already have encountered a composite grid if you did Exercise 3.2.7 on page 72, where
we sought an unstructured grid that adapted to two skew faults and was padded with
rectangular cells near the boundary. As another example, we will generate a Cartesian
grid that has a radial refinement around two wells in the interior of the domain. This
composite grid will be constructed from a set of control points using the pebi routine.
To this end, we first construct the generating points for a unit refinement, as discussed in
Figure 3.25 on page 86

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 102 Grids in Subsurface Modeling

Figure 3.36 A “sieve” plot of the porosity in the SAIGUP model. Using this technique, one can more
easily see the structure in the interior of the model.

Figure 3.37 A composite grid consisting of a regular Cartesian mesh with radial refinement around
two well positions.

Pw = [];
for r = exp(-3.5:.2:0),
[x,y,z] = cylinder(r,28); Pw = [Pw [x(1,:); y(1,:)]];
end
Pw = [Pw [0; 0]];

Then this point set is translated to the positions of the wells and glued into a standard
regular point lattice (generated using meshgrid):
Pw1 = bsxfun(@plus, Pw, [2; 2]);
Pw2 = bsxfun(@plus, Pw, [12; 6]);
[x,y] = meshgrid(0:.5:14, 0:.5:8);
P = unique([Pw1'; Pw2'; x(:) y(:)], 'rows');
G = pebi(triangleGrid(P));

Figure 3.37 shows the resulting grid. To get a good grid, it is important that the number
of points around the cylinder has a reasonable match with the density of the points in the
regular lattice. If not, the transition cells between the radial and the regular grid may exhibit

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.5 Examples of More Complex Grids 103

Figure 3.38 Examples of composite grids. The left plot shows an areal grid consisting of Cartesian,
hexagonal, and radial parts. The right plot shows the same grid extruded to 3D with two faults added.
Both faults are stair-stepped in the areal direction. The fault in the north–south direction has no
displacement and thus the adjacent cells have no extra non-neighboring connections.

quite unfeasible geometries. The observant reader will also notice the layer of small cells
at the boundary, which is an effect of the particular distribution of the generating points
(see the left plot in Figure 3.10 on page 67) and can, if necessary be avoided by a more
meticulous choice of points.
In the left plot of Figure 3.38, we have combined these two approaches to generate
an areal grid consisting of three characteristic components: Cartesian grid cells at the
outer boundary, hexagonal cells in the interior, and a radial grid with exponential radial
refinement around two wells. The right plot shows a 2.5D grid in which the areal Voronoi
grid has been extruded to 3D along vertical pillars. In addition, structural displacement has
been modeled along two areally stair-stepped faults that intersect near the west boundary.
Petrophysical parameters have been sampled from layers 40–44 of the SPE 10 data set [71].

3.5.3 Control-Point and Boundary Conformal Grids

To correctly split the reservoir volume into sub-volumes, build flow units with similar
or correlated petrophysical properties, and resolve flow patterns, it is important that a
volumetric simulation grid conforms as closely as possible to fault surfaces and horizons
representing the structural architecture of the reservoir. This can be achieved by introducing
unstructured grids whose cell interfaces adapt to curved lines in 2D and triangulated sur-
faces in 3D. Likewise, it is of interest to have grids that can adapt the centroids of certain
cells to 2D lines or 3D lines or planes. This is particularly relevant if one were to provide
accurate description of well paths and the near-well region. Modern wells are typically long
and horizontal and may have complex geometry that consists of multiple branches. Such

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 104 Grids in Subsurface Modeling

Figure 3.39 Two different 2D grids adapting to curved well paths. The hybrid grid shown to the left
has Cartesian cells away from the well path, with two levels of refinement in the near-well zone.
Polygonal cells are used on the transition between refinement levels and to track the well path. The
Voronoi grid shown to the right has gradual refinement towards the well and is computed as the dual
to a triangulation generated by DistMesh.

wells are increasingly becoming the main determining factor for reservoir flow patterns
because of their long reach, but also as a result of various techniques for modifying the
near-well region to increase injectivity or use of (intelligent) inflow devices to control fluid
production.
MRST has a third-party module called upr, which offers such adaptivity in the form of
2D Voronoi grids and/or composite grids as well as fully 3D Voronoi grids. The module
was initially developed by one of my previous students as part of his master thesis [42].
The most recent version offers two types of conformity to lower-dimensional objects: (i)
control-point alignment of cell centroids to accurately represent horizontal and multilateral
wells or fracture networks and (ii) boundary alignment of cell faces to accurately preserve
layers, fractures, faults, pinchouts, etc.
Each constraint is represented as a lower-dimensional grid, and intersections of two
objects is associated with a grid with dimension one lower than the objects. The grids are
hierarchically consistent in the sense that cell faces of a higher-dimensional grid conform to
the cells of all lower-dimensional grids. I will not go into more details about the underlying
methods; you can find more details, as well as references to other related work, in [42,
169]. Instead, I have included Figures 3.39– 3.42 to illustrate the flexibility in geometric
modeling inherent in such grids.

3.5.4 Multiblock Grids

A somewhat different approach to get grids whose geometry and topology vary throughout
the physical domain is to use multiblock grids in which different types of structured or

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.5 Examples of More Complex Grids 105

Figure 3.40 The left grid adapts to three intersecting faults, whereas the grid to the right adapts to
both cell and face constraints.

Figure 3.41 The left grid is an areal description of a complex fault network taken from Branets et al.
[46] and is generated as a 2D centroidal Voronoi diagram (CVD) found by minimizing an energy
functional. In the right grid, there are 51 randomly distributed fractures, and the resolution is graded
toward these.

unstructured subgrids are glued together. The resulting grid can be nonmatching across
block interfaces (see e.g., [313, 25, 312]) or have grid lines that are continuous (see e.g.,
[148, 180]). MRST does not have any grid-factory routine for generating advanced multi-
block grids, but offers the function glue2DGrid for gluing together rectangular blocks in

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 106 Grids in Subsurface Modeling

Figure 3.42 The left grid shows a conceptual model of a salt dome, with triangulated surface
extracted from a 3D triangulation generated by DistMesh. The plots to the right shows a 3D Voronoi
grid adapted to a curved and a planar fault. At the intersection of these, only the curved fault is
represented exactly.

2D. These grids can, if necessary, be extruded to 3D by the use of makeLayeredGrid. In the
following, we will show a few examples of such grids.
As our first example, let us generate a curvilinear grid that has a local refinement at its
center as shown in Figure 3.43 (see also Exercise 3.2.6 on page 71). To this end, we start
by generating three different block types shown in red, green, and blue colors to the left in
the figure:
G1 = cartGrid([ 5 5],[1 1]);
G2 = cartGrid([20 20],[1 1]);
G3 = cartGrid([15 5],[3 1]);

Once these are in place, we translate the blocks and glue them together and then apply the
twister function to make a curvilinear transformation of each grid line:
G = glue2DGrid(G1, translateGrid(G2,[1 0]));
G = glue2DGrid(G, translateGrid(G1,[2 0]));
G = glue2DGrid(G3, translateGrid(G, [0 1]));
G = glue2DGrid(G, translateGrid(G3,[0 2]));
G = twister(G);

Let us now replace the central block by a patch consisting of triangular cells. To this end,
we start by generating a new grid G2

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.5 Examples of More Complex Grids 107

Figure 3.43 Using a multiblock approach to construct a rectilinear grid with refinement.

[N,M]=deal(10,15);
[x,y] = ndgrid( linspace(0,1,N+1), linspace(0,1,M+1));
x(2:N,2:M) = x(2:N,2:M) + 0.3*randn(N-1,M-1)*max(diff(xv));
y(2:N,2:M) = y(2:N,2:M) + 0.3*randn(N-1,M-1)*max(diff(yv));

G2 = computeGeometry(triangleGrid([x(:) y(:)]));

The glue2DGrid routine relies on face tags as explained in Figure 3.4 on page 89 that
can be used to identify the external faces facing east, west, north, and south. Generally,
such tags do not make much sense for triangular grids and are therefore not supplied.
However, to be able to find the correct interface to glue together, we need to supply tags
on the perimeter of the triangular patch, where the normal vectors follow the axial direc-
tions and tags therefore make sense. To this end, we start by computing the true normal
vectors:
hf = G2.cells.faces(:,1);
hf2cn = gridCellNo(G2);
sgn = 2*(hf2cn == G2.faces.neighbors(hf, 1)) - 1;
N = bsxfun(@times, sgn, G2.faces.normals(hf,:));
N = bsxfun(@rdivide, N, G2.faces.areas(hf,:));
n = zeros(numel(hf),2); n(:,1)=1;

Then, all interfaces that face east are those whose dot-product with the vector (1,0) is
identical to −1:
G2.cells.faces(:,2) = zeros(size(hf));
i = sum(N.*n,2)==-1; G2.cells.faces(i,2) = 1;

Similarly, we can identify all the interfaces facing west, north, and south. The left plot in
Figure 3.44 shows the resulting multiblock grid. Likewise, the middle plot shows another
multiblock grid where the central refinement is the dual to the triangular patch and G3 has

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 108 Grids in Subsurface Modeling

Figure 3.44 Using a multiblock approach to construct rectilinear grid with triangular (left) and
polygonal refinements (middle). The right plot shows a 3D multiblock grid.

been scaled in the y-direction and refined in the x-direction so that the grid lines are no
longer matching with the grid lines of G1.
As a last example, let us use this technique to generate a 3D multiblock grid that consists
of three blocks in the vertical direction
G = glue2DGrid(G1, translateGrid(G2,[0 1]));
G = glue2DGrid(G, translateGrid(G1,[0 2]));
G = makeLayeredGrid(G, 5);
G.nodes.coords = G.nodes.coords(:,[3 1 2]);

That is, we first generate an areal grid in the xy-plane, extrude it to 3D along the z direction,
and then permute the axis so that their relative orientation is correctly preserved (notice that
simply flipping [1 2 3]→[1 3 2], for instance, will not create a functional grid).

Computer exercises
3.5.1 How would you populate the grids shown in Figures 3.31 and 3.32 with petro-
physical properties so that spatial correlation and displacement across the fault(s)
is correctly accounted for? As an illustrative example, you can try to sample petro-
physical properties from the SPE 10 data set.
3.5.2 Select at least one of the models in the data sets BedModels1 or BedModel2 and try
to find all inactive cells and then all cells that do not have six faces. Hint: it may be
instructive to visualize these models both in physical space and in index space.
3.5.3 Extend your function from Exercise 3.2.7 on page 72 to also include radial refine-
ment in near-well regions as shown in Figure 3.37.
3.5.4 Make a grid similar to the one shown to the right in Figure 3.38. Hint: although it
is not easy to see, the grid is matching across the fault, which means that you can
use the method of fictitious domain to make the fault structure.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 3.5 Examples of More Complex Grids 109

3.5.5 As pointed out in Exercise 3.2.6, MRST does not yet

have a grid factory routine to generate structured grids
with local nested refinement as shown in the figure
to the right. While it is not very difficult to generate
the necessary vertices if each refinement patch is
rectangular and matches the grid cells on the coarser
level, building the grid structure may prove to be a
challenge. Try to develop an efficient algorithm and
implement it in MRST.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Part II
Single-Phase Flow

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 4
Mathematical Models for Single-Phase Flow

If you have read the chapters of the book in chronological order, you have already encoun-
tered the equations modeling flow of a single, incompressible fluid through a porous media
twice: first in Section 1.4, which showed how vertical equilibrium inside can be computed
by use of MRST, and then in Section 2.4.2, in which we discussed the concept of rock
permeability. In this chapter, we review mathematical models for single-phase flow in more
detail, including auxiliary boundary and initial conditions, subscale models for injection
and production wells, and steady-state advection equations. We then introduce basic spatial
discretizations for the resulting equations and show how these discretizations can be formu-
lated as abstract discrete differentiation operators. A more thorough discussion of how these
discretizations can be turned into efficient solvers is deferred to Chapters 5 and 6, which
focus on the incompressible case, and Chapter 7, which discusses compressible flow and
how the discrete differentiation operators can be combined with automatic differentiation
for rapid prototyping of new solvers.

4.1 Fundamental Concept: Darcy’s Law

Macroscale modeling of single-phase flow in porous media started with the work of Henry
Darcy, a French hydraulic engineer, who in the middle of the nineteenth century was
engaged to enlarge and modernize the waterworks of the city of Dijon. To understand the
physics of flow through the sand filters that were used to clean the water supply, Darcy
designed a vertical experimental tank filled with sand, in which water was injected at the
top and allowed to flow out at the bottom; Figure 4.1 shows a conceptual illustration. Once
the sand pack is filled with water, and the inflow and outflow rates are equal, the hydraulic
head at the inlet and at the outlet can be measured using mercury-filled manometers. The
hydraulic head is given as h = E/mg = z−p/ρg, relative to a fixed datum. As water flows
through the porous medium, it will experience a loss of energy. In a series of experiments,
Darcy measured the water volumetric flow rate out of the tank and compared this rate with
the loss of hydrostatic head from top to bottom of the column. From the experiments, he
established that for the same sand pack, the discharge (flow rate) Q [m3 /s] is proportional

113

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 114 Mathematical Models for Single-Phase Flow

Figure 4.1 Conceptual illustration of Darcy’s experiment.

Figure 4.2 The macroscopic Darcy velocity represents an average of microscopic fluid fluxes.

to the cross-sectional area A [m2 ] and the difference in hydraulic head (height of the water)
ht − hb [m], and inversely proportional to the flow length of the tank L [m]. Altogether,
this can be summarized as
Q ht − hb
=κ . (4.1)
A L
The result was presented in 1856 as an appendix to [77] entitled “Determination of the
laws of flow of water through sand” and is the direct predecessor to what we today call
Darcy’s law. In (4.1), κ [m/s] denotes the hydraulic conductivity, which is a function both
of the medium and the fluid flowing through it. It follows from a dimensional analysis that
κ = ρgK/μ, where g [m/s2 ] is the gravitational acceleration, μ [kg/ms] is the dynamic
viscosity, and K [m2 ] is the intrinsic permeability of a given sand pack.
The specific discharge v = Q/A, or Darcy flux, through the sand pack represents the
volume of fluid flowing per total area of the porous medium and has dimensions [m/s].

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 4.2 General Flow Equations for Single-Phase Flow 115

Somewhat misleading, v is often referred to as a velocity. However, since only a fraction

of the cross-sectional area is available for flow (the majority of the area is blocked by sand
grains), v is not a velocity in the microscopic sense. Instead, v is the apparent macroscopic
velocity obtained by averaging the microscopic fluxes inside representative elementary
volumes (REVs), which were discussed in Section 2.3.2. The macroscopic fluid velocity,
defined as volume of fluid flowing per area occupied by fluid, is therefore given by v/φ,
where φ is the porosity associated with the REV.
Henceforth, we will, with a slight abuse of notation, refer to the specific discharge as
the Darcy velocity. In modern differential notation, first proposed by Nutting [239] and
Wyckoff et al. [320], Darcy’s law for a single-phase fluid reads,

K
v = − (∇p − gρ∇z), (4.2)
μ

where p is the fluid pressure and z is the vertical coordinate. The equation expresses con-
servation of momentum and was derived from the Navier–Stokes equations by averaging
and neglecting inertial and viscous effects by Hubbert [137] and later from Stokes flow
by Whitaker [315]. The observant reader will notice that Darcy’s law (4.2) is analogous
to Fourier’s law (1822) for heat conduction, Ohm’s law (1827) in the field of electrical
networks, or Fick’s law (1855) for fluid concentrations in diffusion theory, except that
for Darcy there are two driving forces, pressure and gravity. Notice also that Darcy’s law
assumes a reversible fluid process, which is a special case of the more general physical laws
of irreversible processes that were first described by Onsager (1931).

4.2 General Flow Equations for Single-Phase Flow

To derive a mathematical model for single-phase flow on the macroscopic scale, we first
make a continuum assumption based on the existence of REVs as discussed in the previous
section and then look at a control volume as shown in Figure 4.3. From the fundamental

n


q v

ρ
∂

Figure 4.3 Illustration of a control volume  on which one can apply the principle of conservation
to derive macroscopic continuity equations.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 116 Mathematical Models for Single-Phase Flow

law of mass conservation, we know that the accumulation of mass inside this volume must
equal the net flux over the boundaries,
  
∂
φρ d x + ρ v · n ds = ρq d x, (4.3)
∂t  ∂ 

where ρ is the density of the fluid, φ is the rock porosity, v is the macroscopic Darcy
velocity, n denotes the normal at the boundary ∂ of the computational domain , and
q denotes fluid sources and sinks, i.e., outflow and inflow of fluids per volume at certain
locations. Applying Gauss’s theorem, this conservation law can be written in the alternative
integral form

 
∂
φρ + ∇ · (ρ v) d x = ρq d x. (4.4)
 ∂t 

This equation is valid for any volume , and in particular volumes that are infinitesimally
small. Hence, it follows that the macroscopic behavior of the single-phase fluid must satisfy
the continuity equation
∂(φρ)
+ ∇ · (ρ v) = ρq. (4.5)
∂t
Equation (4.5) contains more unknowns than equations, and to derive a closed mathemat-
ical model, we need to introduce what are commonly referred to as constitutive equations
that give the relationship between different states of the system (pressure, volume, tempera-
ture, etc.) at given physical conditions. Darcy’s law, discussed in the previous section, is an
example of a constitutive relation that has been derived to provide a phenomenological rela-
tionship between the macroscale v and the fluid pressure p. In Section 2.4.1 we introduced
the rock compressibility cr = d ln(φ)/dp, which describes the relationship between the
porosity φ and the pressure p. In a similar way, we can introduce the fluid compressibility
to relate the density ρ to the fluid pressure p.
A change in density will generally cause a change in both pressure p and temperature
T . The usual way of describing these changes in thermodynamics is to consider the change
of volume V for a fixed number of particles,
   
dV 1 ∂V 1 ∂V
= dp + dT , (4.6)
V V ∂p T V ∂T p
where the subscripts T and p indicate that the change takes place under constant tem-
perature and pressure, respectively. Since ρV is constant for a fixed number of particles,
V dρ = −ρdV , and (4.6) can be written in the equivalent form
   
dρ 1 ∂ρ 1 ∂ρ
= dp + dT = cf dp + αf dT . (4.7)
ρ ρ ∂p T ρ ∂T p
Here, cf denotes the isothermal compressibility and αf the thermal expansion coefficient.
In many subsurface systems, the density changes slowly so that heat conduction keeps the
temperature constant, in which case (4.7) simplifies to

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 4.2 General Flow Equations for Single-Phase Flow 117
1 dρ d ln(ρ)
cf = = . (4.8)
ρ dp dp
The factor cf , which we henceforth will refer to as the fluid compressibility, is nonnegative
and will generally depend on both pressure and temperature, i.e., cf = cf (p,T ).
Introducing Darcy’s law and fluid and rock compressibilities in (4.5), we obtain the
parabolic equation for the fluid pressure


∂p ρK
ct φρ −∇ · (∇p − gρ∇z) = ρq, (4.9)
∂t μ
where ct = cr + cf denotes the total compressibility. Notice that this equation is generally
nonlinear since both ρ and ct may depend on p. In the following, we will look briefly
at several special cases in which the governing single-phase equation becomes a linear
equation for the primary unknown; more extensive discussions can be found in standard
textbooks like [246, chap. 1], [69, chap. 2]. For completeness, we also briefly review the
concept of an equation of state.

Incompressible flow. In the special case of an incompressible rock and fluid – that is, ρ and
φ are independent of p so that ct = 0 – (4.9) simplifies to an elliptic equation with variable
coefficients,


K
−∇ · ∇(p − gρz) = q. (4.10)
μ
If we introduce the fluid potential,  = p − gρz, (4.10) can be recognized as a variable-
coefficient Poisson’s equation −∇ · K∇ = q or as the Laplace equation ∇ · K∇ = 0 if
there are no volumetric fluid sources or sinks. In the next section, we will discuss in detail
how to discretize the second-order spatial Laplace operator L = ∇ · K∇, which is a key
technological component that will enter almost any software for simulating flow in porous
rock formations.

Constant compressibility. If the fluid compressibility is constant and independent of pres-

sure, (4.8) can be integrated from a known density ρ0 at a pressure datum p0 to give the
equation

ρ(p) = ρ0 ecf (p−p0 ), (4.11)

which applies well to most liquids that do not contain large quantities of dissolved gas. To
develop the differential equation, we first assume that the porosity and the fluid viscosity
do not depend on pressure. Going back to the definition of fluid compressibility (4.8), it
also follows from this equation that ∇p = (cf ρ)−1 ∇ρ, which we can use to eliminate ∇p
from Darcy’s law (4.2). Inserting the result into (4.5) gives us the continuity equation
∂ρ 1  
− ∇ · K∇ρ − cf gρ 2 K∇z = ρq. (4.12)
∂t μφcf

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 118 Mathematical Models for Single-Phase Flow

In the absence of gravity forces and source terms, this is a linear equation for the fluid
density that is similar to the classical heat equation with variable coefficients,
∂ρ 1
= ∇ · (K∇ρ) . (4.13)
∂t μφcf

Slightly compressible flow. If fluid compressibility is small, it is sufficient to use a linear

relationship to evaluate the exponential form as
 
ρ = ρ0 1 + cf (p − p0 ) . (4.14)

We further assume that φ has a similar functional dependence and that μ is constant. For
simplicity, we also assume that g and q are both zero. Then, we can simplify (4.9) as
follows:
  ∂p cf ρ ρ
(cr + cf )φρ = ∇p · K∇p + ∇ · (K∇p).
∂t μ μ
If cf is sufficiently small, in the sense that cf ∇p · K∇p  ∇ · (K∇p), we can neglect the
first term on the right-hand side to derive a linear equation similar to (4.13) for the fluid
pressure (introducing total compressibility c = cr + cf )
∂p 1
= ∇ · (K∇p) . (4.15)
∂t μφc

Ideal gas. If the fluid is a gas, compressibility can be derived from the gas law, which for
an ideal gas can be written in two alternative forms,

pV = nRT , ρ = p(γ − 1)e. (4.16)

In the first form, T is temperature, V is volume, R = 8.314 J K−1 mol−1 is the gas constant,
and n = m/M is the amount of substance of the gas in moles, where m is the mass and
M is the molecular weight. In the second form, γ is the adiabatic constant, i.e., the ratio of
specific heat at constant pressure and constant volume, and e is the specific internal energy
(internal energy per unit mass). In either case, it follows from (4.8) that cf = 1/p.
Gravity effects are negligible for gases and for brevity we once again we assume that φ
is a function of x only. Inserting (4.16) into (4.9) then gives
∂(ρφ) ∂p 1 (γ − 1)e
= φ(γ − 1)e = ∇ · (ρK∇p) = ∇ · (pK∇p) ,
∂t ∂t μ μ
from which it follows that
∂p φμ ∂p2  
φμ = ∇ · (pK∇p) ⇔ = ∇ · K∇p2 . (4.17)
∂t p ∂t

Equation of state. Equation (4.11), (4.14), and (4.16) are all examples of what is commonly
referred to as equations of state, which provide constitutive relationships between mass,
pressures, temperature, and volumes at thermodynamic equilibrium. Another popular form

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 4.2 General Flow Equations for Single-Phase Flow 119

of these equations are the so-called cubic equations of state, which can be written as cubic
functions of the molar volume Vm = V /n = M/ρ involving constants that depend on
pressure pc , temperature
  2 Tc , and the molar volume Vc at the critical point, i.e., the point
∂p ∂ p
at which ∂V = ∂V 2 ≡ 0. A few particular examples include the Redlich–Kwong
T T
equation of state
RT a
p= −√ ,
Vm − b T Vm (Vm + b)
5/2
0.42748R 2 Tc 0.08662RTc
a= , b= , (4.18)
pc pc
the modified version called Redlich–Kwong–Soave
RT aα
p= −√ ,
Vm − b T Vm (Vm + b)
0.427R 2 Tc2 0.08664RTc
a= , b= , (4.19)
pc pc
    2
α = 1 + 0.48508 + 1.55171ω − 0.15613ω2 (1 − T /Tc ) ,

as well as the Peng–Robinson equation of state,

RT aα
p= − 2 ,
Vm − b Vm + 2bVm − b2 )
0.4527235R 2 Tc2 0.077796RTc
a= , b= , (4.20)
pc pc
    2
α = 1 + 0.37464 + 1.54226ω − 0.26992ω2 (1 − T /Tc ) .

Here, ω denotes the acentric factor of the species, which is a measure of the centricity
(deviation from spherical form) of the molecules in the fluid. The Peng–Robinson model
is much better at predicting liquid densities than the Redlich–Kwong–Soave model, which
was developed to fit pressure data of hydrocarbon vapor phases. If we introduce an attrac-
tive factor A, a repulsive factor B, and the gas deviation factor Z (to be discussed in Section
11.4.3 on page 373)
aαp bp pV
A= , B= , Z= ,
(RT )2 RT RT
the Redlich–Kwong–Soave equation (4.19) and the Peng–Robinson equation (4.20) can be
written in alternative polynomial forms,

0 = Z 3 − Z 2 + Z(A − B − B 2 ) − AB, (4.21)

0 = Z − (1 − B)Z + (A − 2B − 3B )Z − (AB − B − B ),
3 2 2 2 3
(4.22)

which illustrates why they are called cubic equations of state.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 120 Mathematical Models for Single-Phase Flow

4.3 Auxiliary Conditions and Equations

The governing equations for single-phase flow discussed in the previous section are all
parabolic, except for the incompressible case in which the governing equation is elliptic.
For the solution to be well-posed1 inside a finite domain for any of the equations, one
needs to supply conditions that determine the behavior on the external boundary. For the
parabolic equations describing unsteady flow, we also need to impose an initial condition
that determines the initial state of the fluid system. This section discusses these conditions
in more detail. We also discuss models for representing flow in and out of the reservoir rock
through wellbores. Flow near wellbores typically takes place on a length scale much smaller
than the length scales of the global flow inside the reservoir, and it is customary to model
it using special analytical models. Finally, we also discuss a set of auxiliary equations for
describing the movement of fluid elements and/or neutral particles that follow the single-
phase flow without affecting it.

4.3.1 Boundary and Initial Conditions

In reservoir simulation one is often interested in describing closed flow systems that have
no fluid flow across its external boundaries. This is a natural assumption when study-
ing full reservoirs that have trapped and contained petroleum fluids for million of years.
Mathematically, no-flow conditions across external boundaries are modeled by specifying
homogeneous Neumann conditions,
v · n = 0 for x ∈ ∂. (4.23)
With no-flow boundary conditions, any pressure solution of (4.10) is immaterial and only
defined up to an additive constant, unless a datum value is prescribed at some internal point
or along the boundary.
It is also common that parts of the reservoir may be in communication with a larger
aquifer system that provides external pressure support, which can be modeled in terms of a
Dirichlet condition of the form
p( x,t) for x ∈ a ⊂ ∂.
x ) = pa ( (4.24)
The function pa can, for instance, be given as a hydrostatic condition. Alternatively, parts
of the boundary may have a certain prescribed influx, which can be modeled in terms of an
inhomogeneous Neumann condition,
x,t) for x ∈ a ⊂ ∂.
v · n = ua ( (4.25)
Combinations of these conditions are used when studying parts of a reservoir (e.g., sector
models). There are also cases, for instance when describing groundwater systems or CO2

1 A solution is well-posed if it exists, is unique, and depends continuously on the initial and boundary conditions.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 4.3 Auxiliary Conditions and Equations 121

sequestration in saline aquifers, where (parts of) the boundaries are open, or the system
contains a background flow. More information of how to set boundary conditions will
be given in Section 5.1.4. In the compressible case in (4.9), we also need to specify an
initial pressure distribution. Typically, this pressure distribution will be hydrostatic, as in
the gravity column we discussed briefly in Section 1.4, and hence be given by the ordinary
differential equation,

dp
= ρg, p(z0 ) = p0 . (4.26)
dz

4.3.2 Injection and Production Wells

In a typical reservoir simulation, the inflow and outflow in wells occur on a subgrid scale. In
most discretized flow models, the pressure is modeled using a single pressure value inside
each grid cell. The size of each grid cell must therefore be chosen so small that the pressure
variation inside the cell can be approximated accurately in terms of its volumetric average.
Far away from wells, spatial variations in pressure tend to be relatively slow, at least in
certain directions, and one can therefore choose cell sizes in the order of tens or hundreds of
meters, which is a reasonable size compared with the extent of the reservoir. Near the well,
however, the pressure will have large variations over short distances, and to compute a good
approximation of these pressure variations, one would need grid cells than are generally
smaller than what is computationally tractable. As a result, one ends up with a setup similar
to the illustration in Figure 4.4, where the radius of the well is typically between 1/100 and
1/1000 of the horizontal dimensions of the grid cell. The largest percentage of the pressure
drop associated with a well occurs near the well and the pressure at the well radius will thus
deviate significantly from the volumetric pressure average inside the cell. Special analytical
models are therefore developed to represent the subgrid variations in the particular flow
patterns near wells.

Figure 4.4 Illustration of a well inside a grid cell. The proportions are not fully to scale: whereas
the diameter of a well varies from 5 to 40 inches, a grid block may extend from tens to hundreds of
meters in the lateral direction and from a few decimeters to ten meters in the vertical direction.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 122 Mathematical Models for Single-Phase Flow

Inflow-Performance Relation
Normally, fluids are injected from a well at either constant surface rate or at constant
bottom-hole pressure, which is also called wellbore flowing pressure and refers to the
pressure at a certain datum point inside the wellbore. Similarly, fluids are produced at
constant bottom-hole pressure or at constant surface liquid rate. The main purpose of a
well model is then to accurately compute the pressure at the well radius when the injection
or production rate is known, or to accurately compute the flow rate in or out of the reservoir
when the pressure at well radius is known. The resulting relation between the bottom-hole
pressure and surface flow rate is often called the inflow-performance relation or IPR.
The simplest and most widely used inflow–performance relation is the linear law

qo = J (pR − pw ), (4.27)

which states that the flow rate is directly proportional to the pressure drawdown near the
well; that is, flow rate is proportional to the difference between the average reservoir pres-
sure pR in the grid cell and the pressure pw at the wellbore. The constant of proportionality
J is called the productivity index (PI) for production wells, or the well injectivity index (WI)
for injectors, and accounts for all rock and fluid properties, as well as geometric factors
that affect the flow. In MRST, we do not distinguish between productivity and injectivity
indices, and henceforth we will only use the shorthand WI.

Steady-State Flow in a Radial System

To derive the basic linear relation (4.27), we consider a vertical well that drains a rock with
uniform permeability. The well penetrates the rock completely over a height h and is open
in the radial direction. Gravity is neglected and the single-phase fluid is assumed to flow
radially inward from an infinite domain; see Figure 4.5. From mass conservation written in
cylinder coordinates, we have

1 ∂(rvB −1 )
=0 −→ v = BC/r.
r ∂r

2rw

a) Horizontal view of well b) From above

Figure 4.5 Radial flow used to derive inflow–performance relationship.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 4.3 Auxiliary Conditions and Equations 123

Here, the formation-volume factor B is the ratio between the volume of the fluid at reservoir
conditions and the volume of the fluid at surface conditions.
The integration constant C is determined by integrating around a small cylinder sur-
rounding the fluid sink,

q = B −1 v · n ds = −2π hC.

Inserting this into Darcy’s law in cylinder coordinates, we have

qB K dp
v=− =− .
2π rh μ dr
Even if several different flow patterns can be expected when fluids flow toward a verti-
cal wellbore, two-dimensional radial flow is considered to be the most representative for
vertical oil and gas wells. We now integrate this equation from the wellbore rw and to an
arbitrary radius,
 p  r
1 1
2π Kh dp = dr.
pw qμB rw r

Here, B decreases with pressure and μ increases. The composite effect is that (μB)−1
decreases almost linearly with pressure. We can therefore approximate μB by (μB)avg
evaluated at the average pressure (pw + p)/2, to obtain an explicit expression for the
pressure (dropping the subscript on μB for brevity)
qμB
p(r) = pw + ln(r/rw ). (4.28)
2π Kh

Figure 4.6 shows a plot of this solution on a scale that is somewhat larger than the
typical size of grid cells in field-scale models. Notice, in particular, that most of the pressure

Figure 4.6 Illustration of the pressure distribution inside a circular region with a radius of 100 meters
computed from (4.28), assuming the well is producing fluids from an infinite domain. Here, pa is the
volumetric pressure average inside the domain and ra is the radius at which this value is found. The
inset shows a zoom of the near-well zone.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 124 Mathematical Models for Single-Phase Flow

drawdown takes place within a relatively short distance to the wellbore, which confirms the
need for analytical submodels to avoid having to introduce very fine grids near the wellbore.

Pseudo-Steady State in a Radial System

We can also derive a linear inflow relationship of the form (4.27) under somewhat more
realistic conditions by considering pseudo-steady flow conditions. Assuming constant com-
pressibility and fluid viscosity, the single-phase flow equation written in cylinder coordi-
nates reads
 
1 ∂ ∂p μφc ∂p
r = . (4.29)
r ∂r ∂r K ∂t

At pseudo-steady state, ∂p/∂t is constant. To determine the pressure differential, we con-

sider the global mass-balance, which states that the amount of produced fluid must equal
the change of mass inside the reservoir. Writing this on integral form and differentiating
gives us

 t  p̄ Ahφc ∂p d p̄ Bq
q(t) dt = − d p̄ −→ = =−  2 .
0 p0 B ∂t dt π rd − rw2 hφc

Here, q is the volumetric rate at surface conditions, p0 is the initial reservoir pressure, and
p̄ is the average pressure inside the cylindrical drainage region at time t. In the expression
above, we have introduced the formation-volume factor B, defined as the ratio between the
volume of the fluid at reservoir conditions and the volume of the fluid at surface conditions.
Inserting the constant pressure change into (4.29) gives

 
1 d dp qμB
r =−  = −C.
r dr dr hKπ rd2 − rw2

To solve this equation, we first integrate with respect to r and use the fact that dp/dr = 0
at the outer boundary,

    dp  
dp dr |rd = 0 dp
d r =− Crdr −→ r = 12 C rd2 − r 2 .
dr dr

Then, we integrate outward from rw to determine p(r),

      
r rd2 qμB rd2 r 1 r 2 − rw2
p(r) − pw = 1
2C −r dr = ln − .
rw r 2π Kh r d − rw
2 2 rw 2 rd2 − rw2

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 4.3 Auxiliary Conditions and Equations 125

This equation can be used to develop the desired relationship between the average reservoir
pressure and the wellbore flowing pressure pw ,
  rd
1 rd 2π hφ
pa = p(r) dV = p(r)r dr
V rw π(rd2 − rw2 )hφ rw
 rd
    
2 C 2 r
= 2 pw r + r r ln − 2 r r − rw
1 2 2
dr
rd − rw2 rw 2 d rw
2 qμB
= pw +  2  [I1 + I2 ] .
rd − rw 2π hK rd2 − rw2
2

The two integrals can be computed as follows,

 rd  2
I1 = r ln (r/rw ) dr = 12 rd2 ln (rd /rw ) − 14 rd2 − rw ,
r
 wrd      
I2 = 1
2 r r 2
− r 2
w dr = 1
8 rd
4
− rw
4
− 1 2
r
4 w rd
2
− r 2
w .
rw

We can now insert this back into the equation and use the fact that rd  rw to derive our
desired expression,
⎡ ⎤
 
qμB ⎢ ⎢ rd4 rd 1 rd2 1 rd2 + rw2 1 rw2 ⎥ ⎥
pa − pw = ⎢ 2 ln − − + ⎥
2π hK ⎣ (rd − rw2 )2 rw 2 rd2 − rw2 4 rd2 − rw2 2 rd2 − rw2 ⎦
! "# $ ! "# $ ! "# $ ! "# $
≈1 ≈1 ≈1 ≈0

  
qμB rd 3
≈ ln − .
2π hK rw 4
Rearranging terms, we obtain an expression for the fluid rate of the producer
2π Kh
q= (pa − pw ) . (4.30)
μB (ln(rd /rw ) − 0.75)
For an injector, we likewise obtain
2π Kh
q= (pw − pa ) .
μB (ln(rd /rw ) − 0.75)
The relation (4.30) was developed for an ideal well under several simplifying assumptions:
homogeneous and isotropic formation of constant thickness, clean wellbore, etc. A well
will rarely experience these ideal conditions in practice. Typically, the permeability is
altered close to the wellbore under drilling and completion, the well will only be partially
completed, and so on. The actual pressure performance will therefore deviate from (4.30).
To model this, it is customary to include a skin factor S to account for extra pressure loss
due to alterations in the inflow zone. The resulting equation is
2π Kh
q= (pa − pw ) . (4.31)
μB (ln(rd /rw ) − 0.75 + S)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 126 Mathematical Models for Single-Phase Flow

Producer

Injector

Quarter five-spot

Figure 4.7 Excerpts of a repeated five-spot pattern.

Often the constant −0.75 is included in the skin factor S, which could be negative for
stimulated wells. Sometimes h is modified to ch, where c is the completion factor, i.e.,
a dimensionless number between zero and one that describes the fraction of the wellbore
open to flow.

Peaceman-Type Well Models

To use the radial model in conjunction with a reservoir model, the volumetric average
pressure in the radial model must be related to the computed cell pressure p. Exact solutions
are generally not known, since real reservoirs have complicated geometries and irregular
boundaries. Well models are therefore developed using highly idealized reservoir geome-
tries. One such example is the so-called repeated five-spot pattern, which consists of a thin,
infinitely large, horizontal reservoir with a staggered pattern of injection and production
wells as (see Figure 4.7) that repeats itself to infinity in all directions. The name comes from
the fact that each injector is surrounded by four producers, and vice versa, hence creating
tiles of five-spot patterns. If all wells operate at equal rates, the flow pattern has certain
symmetries and it is common to only consider one quarter of the five spot, as shown in
Figure 4.7, subject to no-flow boundary conditions. Muskat and Wyckoff [219] derived the
following explicit solution for the pressure drop between the injection and production wells,
qμB  
p = ln(d/rw ) − C , (4.32)
π Kh
where d is the distance between the wells, and C is given by an infinite series. The authors
originally used C = 0.6190, but a more accurate value, C = 0.61738575, was later derived
by Peaceman [245], who used (4.32) to determine an equivalent radius re at which the
cell pressure is equal to average of the exact pressure. Assuming isotropic permeabilities,
square grid blocks, single-phase flow, and a well at a center of an interior block, Peaceman
[247] showed that the equivalent radius is

re ≈ 0.20788 x y
for the two-point discretization that will be discussed in more detail in Section 4.4.1. To
give you an idea, let us present a somewhat simplified derivation in the incompressible case

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 4.3 Auxiliary Conditions and Equations 127

pN

pW p pE

y
pS

Figure 4.8 Five-point stencil on a grid with square grid blocks (x = y).

without gravity and assume that the viscosity and formation-volume factor are constant.
For a single-layered square grid (Figure 4.8), the two-point scheme then simplifies to a
standard five-point difference scheme

Kh  
− 4p − pW − pN − pE − pS = q.
μB

The solution is symmetric for the repeated five-spot pattern, so that pE = pN = pW = pS

(which explains why μB could be factored out). Hence, we can eliminate the last three
pressure values and write the five-point stencil as

qμB
p = pE − .
4Kh

Next, we assume that the radial well solution is an accurate approximation to the pressure
values in the neighboring cells,

qμB
p E = pw + ln (x/rw ) .
2π Kh

Combining the two equations, we have

qμB qμB
p = pw + ln (x/rw ) −
2π Kh 4Kh
qμB  π qμB  
= pw + ln (x/rw ) − = pw + ln e−π/2 x/rw .
2π Kh 2 2π Kh

Now, exp(−π/2) ≈ 0.20788, and hence we have derived Peaceman’s classical result.
This basic model has later been extended to cover a lot of other cases, e.g., off-center
wells, multiple wells, non-square grids, anisotropic permeability, horizontal wells; see for
instance [18, 102, 13]. For anisotropic permeabilities K =diag(Kx ,Ky ,Kz ), the permeabil-
ity is replaced by an effective permeability

Ke = Kx Ky , (4.33)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 128 Mathematical Models for Single-Phase Flow
√
which follows by a coordinate transformation x → x̂ Kx , etc. For a grid with rectangular
grid blocks (x = y), the equivalent radius reads [245]
  1/2
Ky /Kx x 2 + Kx /Ky y 2
re = 0.28  1/4  1/4 . (4.34)
Ky /Kx + Kx /Ky
The same formulas apply to horizontal wells, with the necessary permutations; that is, for
a well in the x-direction x → z and Kx → Kz .
If we include gravity forces in the well and assume hydrostatic equilibrium, the well
model reads
2π hKe  
qi = pR − pw − ρi (z − zbh )g , (4.35)
μi Bi [ln(re /rw ) + S]
where Ke is given by (4.33) and re is given by (4.34). For deviated wells, h denotes the
length of the grid block in the major direction of the wellbore and not the length of the
wellbore.
At this point we should add a word of caution. The equivalent radius of a numerical
method generally depends on how the method approximates pressure inside the grid cell
containing the well perforation. The formulas given in this subsection are, strictly speak-
ing, only valid if you use the specific two-point discretization they were developed for.
When using another discretization method, you may have to compute other values for the
equivalent radius, e.g., as discussed in [192, 197].

4.3.3 Field Lines and Time-of-Flight

Equation (4.10) together with a set of suitable and compatible boundary conditions consti-
tute all you need to determine the pressure distribution and flow velocity of an incompress-
ible fluid inside an incompressible rock. In the remainder of this section, we will discuss a
few simple concepts and auxiliary equations you can use to visualize, analyze, and improve
your understanding of the computed flow fields.
A simple way to visualize a flow field is to use field lines resulting from the vector field:
streamlines, streaklines, and pathlines. In steady flow, the three are identical. However, if
the flow is not steady, i.e., when v changes with time, they differ. Streamlines are associ-
ated with an instant snapshot of the flow field and consist of a family of curves that are
everywhere tangential to v and show the direction a fluid element will travel at this specific
point in time. That is, if x(r) is a parametric representation of a single streamline at this
instance tˆ in time, then
d x d x v(tˆ)
x, tˆ) = 0,
× v( or equivalently, = . (4.36)
dr dr v (tˆ)|
|
In other words, streamlines are calculated throughout the fluid from an instantaneous snap-
shot of the flow field. Because two streamlines from the same instance in time cannot
cross, there cannot be flow across a streamline, and if we align a coordinate along a bundle
of streamlines, the flow through them will be one-dimensional.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 4.3 Auxiliary Conditions and Equations 129

Pathlines are the trajectories individual fluid elements will follow over a certain period.
In each moment of time, the path a fluid particle takes will be determined by the streamlines
associated with the velocity field at this instance in time. If y(t) represents a single path
line starting in y0 at time t0 , then
d y
= v(
y,t), y(t0 ) = y0 . (4.37)
dt
A streakline is the line traced out by all fluid particles that have passed through a
prescribed point throughout a certain period of time. (Think of dye injected into the fluid at
a specific point.) If we let z(t,s) denote a parametrization of a streakline and z0 the specific
point through which all fluid particles have passed, then
dz
= v(z,t), z(s) = z0 . (4.38)
dt
Like streamlines, two streaklines cannot intersect each other. In summary: streamline pat-
terns change over time, but are easy to generate mathematically. Pathlines and streaklines
are recordings of the passage of time obtained through experiments.
Streamlines are far more used within reservoir simulation than pathlines and streaklines.
Instead of using the arc length r to parametrize streamlines, it is common to introduce an
alternative parametrization called time-of-flight, which takes into account the reduced
volume available for flow, i.e., the porosity φ. Time-of-flight is defined by the following
integral
 r
φ(x (s))
τ (r) = ds, (4.39)
0 |v (
x (s))|
where τ expresses the time it takes a fluid particle to travel a distance r along a streamline
(in the interstitial velocity field v/φ). Alternatively, computing the directional derivative of
τ along a streamline and applying the fundamental theorem of calculus to (4.39),
 v  d φ
·∇ τ = τ=
|
v| dr |
v|
from which it follows that τ can be expressed by a differential equation [78, 79]
v · ∇τ = φ. (4.40)
For lack of a better name, we will refer to this as the time-of-flight equation.

4.3.4 Tracers and Volume Partitions

A tracer can somewhat simplified be seen as a neutral particle that passively flow with
the fluid without altering its flow properties. The concentration of a tracer is given by a
continuity equation on the same form as (4.5),
∂(φC)
+ ∇ · (
v C) = qC . (4.41)
∂t

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 130 Mathematical Models for Single-Phase Flow

Figure 4.9 Illustration of time-of-flight, shown as gray isocontour lines, and streamlines shown as
thick black lines.

Volumetric connections within a reservoir can be determined by simulating the evolution

of artificial, nondispersive and nondiffusive tracers whose concentration does not change
upon fluid compression or expansion. A simple flow diagnostic is to set the tracer concen-
tration equal to one in a particular fluid source or at a certain part of the inflow boundary,
and compute the solution approached at steady-state conditions from the nonconservative
equation,

v · ∇C = qC , C|inflow = 1. (4.42)

The resulting tracer distribution gives the portion of the total fluid volume passing through
a point in the reservoir that can be attributed to a given fluid source or parts of the inflow
boundary. Likewise, by reversing the sign of the flow field and assigning unit tracers to a
particular fluid sink or parts of the outflow, one can compute the portion of the fluid content
at a certain point in the reservoir that eventually will arrive at the given fluid sink or outflow
boundary. By repeating this process for all parts of the inflow, we obtain a partition of the
instantaneous flow field: Let Ci denote the tracer associated with fluid source Qi . This
fluid source will influence the flow through all points x for which Ci ( x ) > 0. If C(x ) = 1,
then Qi is the only fluid source affecting the flow through x. Otherwise, the point must be
affected by at least one other fluid source Qj . The tracer distributions {Ci ( x )}m
i=1 can thus
be seen as measures of how strongly the flow through a point is influenced by various fluid
sources. We will return with detailed discussion in Chapter 13 on flow diagnostics.
You can also use streamlines and time-of-flight to define an alternative curvilinear
and flow-based coordinate system in three dimensions. To this end, we introduce the bi-
streamfunctions ψ and χ [34], for which v = ∇ψ × ∇χ . In the streamline coordinates
(τ,ψ,χ ), the gradient operator is expressed as
∂ ∂ ∂
∇(τ,ψ,χ ) = (∇τ ) + (∇ψ) + (∇χ ) . (4.43)
∂τ ∂ψ ∂χ
Moreover, a streamline  is defined by the intersection of a constant value for ψ and a
constant value for χ . Because v is orthogonal to ∇ψ and ∇χ , it follows from (4.40) that

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 4.4 Basic Finite-Volume Discretizations 131
∂ ∂
v · ∇(τ,ψ,χ ) = (
v · ∇τ )
=φ . (4.44)
∂τ ∂τ
Therefore, the coordinate transformation (x,y,z) → (τ,ψ,χ ) will reduce the three-
dimensional transport equation (4.41) to a family of one-dimensional transport equations
along each streamline [78, 164]. For incompressible flow this reads
∂C ∂C
+ = 0. (4.45)
∂t ∂τ
There is no exchange of the quantity C among streamlines, and we can thus view each
streamline as an isolated flow system. Assuming a prescribed concentration history C0 (t)
at the inflow gives a time-dependent boundary-value problem for the concentration at the
outflow. Here, the response is given as (see [78]),
C(t) = C0 (t − τ ), (4.46)
which is easily verified by inserting the expression into (4.45) and using the fact that the
solution is unique [134]. For the special case of continuous and constant injection, the
solution is particularly simple

0, t < τ,
C(t) =
C0, t > τ .

4.4 Basic Finite-Volume Discretizations

Extensive research on numerical methods for the Laplace/Poisson equation has led to a
large number of different finite-difference and finite-volume methods, as well as finite-
element methods based on standard, mixed, or discontinuous Galerkin formulations. In the
rest of this section we present the simplest example of a finite-volume discretization, the
two-point flux-approximation scheme, which is used extensively throughout industry and
also is the default discretization method in MRST. We give a detailed derivation of the
two-point method and briefly outline how to discretize the time-of-flight and the stationary
tracer equations.
The two-point method is simple to implement, gives sparse linear systems that are not
too costly to invert, and is robust in the sense that it produces monotone pressure approx-
imations. Unfortunately, the method is only conditionally consistent, and will introduce
numerical artifacts and not provide convergent solutions for increased grid resolution if
the grid does not satisfy a certain orthogonality condition with respect to the permeability
tensor. We discuss this in more detail in Chapter 6, which also introduces more advanced
and consistent discretizations that are guaranteed to be convergent on general polyhedral
grids and for anisotropic permeabilities.

4.4.1 Two-Point Flux-Approximation

To keep technical details at a minimum, we will in the following without loss of generality
consider the simplified single-phase flow equation

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 132 Mathematical Models for Single-Phase Flow

∇ · v = q, v = −K∇p, in  ⊂ Rd . (4.47)
In classical finite-difference methods, partial differential equations are approximated by
replacing the derivatives with appropriate divided differences between point-values on a
discrete set of points in the domain. Finite-volume methods, on the other hand, have a
more physical motivation and are derived from conservation of (physical) quantities over
cell volumes. Thus, in a finite-volume method the unknown quantities are represented in
terms of average values defined over a set of finite volumes. The PDE model is integrated
and required to hold in an averaged sense rather than in a pointwise sense. In this sense,
finite-difference and finite-volume methods have fundamentally different interpretation and
derivation, but the names are often used interchangeably in the scientific literature. The
main reason for this is probably that the discrete equations derived for the cell-centered val-
ues in a low-order, mass-conservative, finite-difference method are identical to the discrete
equations for the cell averages in corresponding finite-volume methods. Herein, we will
stick to this convention and not make a strict distinction between the two types of methods.
To develop a finite-volume discretization for (4.47), we start by rewriting the equation
in integral form using a single cell i in the discrete grid as control volume
 
v · n ds = q d x. (4.48)
∂i i

This is a simpler form of (4.3), where the accumulation term has disappeared because φ and
ρ are independent of time and the constant ρ has been eliminated. Equation (4.48) ensures
that mass is conserved for each grid cell. The next step is to use Darcy’s law to compute
the flux across each face of the cell,

vi,k = v · n ds, i,k = ∂i ∩ ∂k . (4.49)
ik

We refer to the faces i,k as half-faces, since they are associated with a particular grid cell
i and a certain normal vector ni,k . However, since the grid is assumed to be matching,
each interior half face will have a twin half-face k,i that has identical area Ak,i = Ai,k but
opposite normal vector nk,i = − ni,k . If we now approximate the integral over the cell face
in (4.49) by the midpoint rule, we can use Darcy’s law to write the flux as
vi,k ≈ Ai,k v(
xi,k ) · ni,k = −Ai,k (K∇p) (
xi,k ) · ni,k , (4.50)
where xi,k denotes the centroid on i,k . The idea is now to use a one-sided finite difference
to express the pressure gradient as the difference between the pressure πi,k at the face
centroid and the pressure at some point inside the cell. However, in a finite-volume method,
we only know the cell averaged value of the pressure inside the cell. We must therefore
make some additional assumption to be able to reconstruct the point values we need to
estimate the pressure gradient in Darcy’s law. If we assume that the pressure is linear (or
constant) inside each cell, the reconstructed pressure value at the cell center is identical to
the average pressure pi inside the cell, and hence it follows that (see Figure 4.10)
(pi − πi,k ) ci,k
vi,k ≈ Ai,k Ki · ni,k = Ti,k (pi − πi,k ). (4.51)
|
ci,k | 2

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 4.4 Basic Finite-Volume Discretizations 133

pk
pi k
i πi,k
ci,k ni,k

i,k

Figure 4.10 Two cells used to define the two-point finite-volume discretization of the Laplace
operator.

Here, we have introduced one-sided transmissibilities Ti,k that are associated with a single
cell and give a two-point relation between the flux across a cell face and the difference
between the pressure at the cell and face centroids. We refer to these one-sided transmissi-
bilities as half-transmissibilities, since they are associated with a half face.
To derive the final discretization, we impose continuity of fluxes across all faces, vi,k =
−vk,i = vik and continuity of face pressures πi,k = πk,i = πik . This gives us two
equations,
−1 −1
Ti,k vik = pi − πik , − Tk,i vik = pk − πik .

By eliminating the interface pressure πik , we end up with the following two-point flux-
approximation (TPFA) scheme,
 −1
−1 −1
vik = Ti,k + Tk,i (pi − pk ) = Tik (pi − pk ), (4.52)

where Tik is the transmissibility associated with the connection between the two cells. As
the name suggests, the TPFA scheme uses two “points,” the cell averages pi and pk , to
approximate the flux across the interface ik between cells i and k . In our derivation
so far, we have parametrized the cell fluxes in terms of the index of the neighboring cell.
Extending the derivation to also include fluxes on exterior faces is trivial since we either
know the flux explicitly for Neumann boundary conditions (4.23) or (4.25), or know the
interface pressure for Dirichlet conditions (4.24).
By inserting the expression for vik into (4.48), we see that the TPFA scheme for (4.47),
in compact form, seeks a set of cell averages that satisfy the following system of equations

Tik (pi − pk ) = qi , ∀i ⊂ . (4.53)
k

This system is clearly symmetric, and a solution is only defined up to an arbitrary constant,
as is also the case for the corresponding Poisson problem. We can make the system positive-
definite and preserve symmetry by specifying the pressure in a single point. In MRST, we

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 134 Mathematical Models for Single-Phase Flow

vi,k
pi πi,k pk y

xi xk

Figure 4.11 Two cells used to derive the TPFA discretization for a 2D Cartesian grid.

have chosen to set p1 = 0 by adding a positive constant to the first diagonal of the matrix
A = [aij ], where:
& 
k Tik if j = i,
aij =
−Tij if j = i.
The matrix A is sparse and will have a banded structure for structured grids: tridiagonal
for 1D grids and penta- and heptadiagonal for logically Cartesian grids in 2D and 3D,
respectively. The TPFA scheme is monotone, robust, and relatively simple to implement,
and is currently the industry standard with reservoir simulation.
Example 4.4.1 To tie the links with standard finite-difference methods on Cartesian grids,
we derive the two-point discretization for a 2D Cartesian grid with isotropic permeability.
Consider the flux in the x-direction between two cells i and k as illustrated in Figure 4.11.
As above, we impose mass conservation inside each cell. For cell i this reads:
(pi − πi,k )  1  2Ki 
vi,k = y  2 2 xi ,0 Ki (1,0)T = y pi − πi,k
1 xi
2 xi

and likewise for cell k:

(pk − πk,i )  1  2Kk 
vk,i = y  2 − 2 xk ,0 Kk (−1,0) = y
T
pk − πk,i
1 xk
2 xk

Next, we impose continuity of fluxes and face pressures,

vi,k = −vk,i = vik , πi,k = πk,i = πik ,
which gives us two equations
xi xk
vik = pi − πik , − vik = pk − πik .
2Ki y 2Kk y
Finally, we eliminate πik to obtain
 x xk −1
i
vik = 2y + (pi − pk ) ,
Ki Kk
which shows that the transmissibility is given by the harmonic average of the permeability
values in the two adjacent cells, as one would expect.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 4.4 Basic Finite-Volume Discretizations 135

In [2], two former colleagues and I showed how you can implement an efficient and self-
contained MATLAB program that in approximately 30 compact lines solves the incom-
pressible flow equation (4.47) using the two-point method outlined in this section. The
program was designed for Cartesian grids with no-flow boundary conditions only, and
relied strongly on the logical ij k numbering of grid cells. For this reason, the program has
limited applicability beyond highly idealized cases such as the SPE 10 model. However, in
its simplicity, it presents an interesting contrast to the general-purpose implementation in
MRST which is designed to handle unstructured grids, wells, and more general boundary
conditions. I encourage you to read the paper and try the accompanying program and
example scripts, which you can download from my personal website.

4.4.2 Discrete div and grad Operators

The double-index notation vi,k and vik used in the previous section is simple and easy
to comprehend when working with a single interface between two neighboring cells. It
becomes more involved when we want to introduce the same type of discretizations for
more complex models than the Poisson equation for incompressible flow. To prepare for
discussions later in the book, we will now introduce a more abstract way of writing the two-
point finite-volume discretization. The idea is to introduce discrete operators for the diver-
gence and gradient operators that mimic their continuous counterparts. This will enable us
to write the discretized version of Poisson’s equation (4.47) in the same form as its continu-
ous counterpart. These operators will necessarily have to incorporate information about the
grid topology in the same way as the transmissibility incorporates information about grid
geometry. However, once defined, we can apply them to any quantities defined over the
cells and faces of a grid without knowing anything about the underlying grid topology.
This gives us a powerful tool to write solvers in a way that effectively is grid agnostic.
We start by a quick recap of the definition of unstructured grids. As discussed in detail
in Section 3.4, the grid structure in MRST consists of three objects: cells, faces, and
nodes. Each cell is delimited by a set of faces, and each face by a set of edges, which
again are determined by a pair of nodes. Each object has given geometrical properties
(volume, areas, centroids). As before, let nc and nf denote the number of cells and faces
respectively. To define the topology of the grid, we can use two mappings. The first is given
by F : {1, . . . ,nc } → {0,1}nf and maps a cell to the set of faces that delimit this cell.
In the grid structure G, this is represented as the G.cells.faces array, where the first
column specifying the cell numbers is not stored, since it is redundant and instead must be
computed by a call f2cn = gridCellNo(G). The second mapping consists of two compo-
nents that for a given face give the corresponding neighboring cells, C1,C2 : {1, . . . ,nf } →
{1, . . . ,nc }. In the grid structure G, C1 is given by G.faces.neighbors(:,1) and C2 by
G.faces.neighbors(:,2). This is illustrated in Figure 4.12.
Let us now construct the discrete analogues of the divergence and gradient operators,
which we denote div and grad. The divergence operator div is a linear mapping from

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 136 Mathematical Models for Single-Phase Flow

Figure 4.12 Illustration of the mappings from cells to faces and from faces to cells used to define the
discrete divergence and gradient operators.

faces to cells, which typically will be applied to a discrete flux v ∈ Rnf . We write v[f ] to
denote the flux evaluated on a face f and assume that the orientation of this flux is from
C1 (f ) to C2 (f ). Hence, the total amount of matter leaving a cell c is given by the sum of
outfluxes minus the sum of influxes
 
div(v)[c] = v[f ] 1{c=C1 (f )} − v[f ] 1{c=C2 (f )} . (4.54)
f ∈F (c) f ∈F (c)

Here, element c of the vector 1{c=C1 (f )} equals one if cell c is neighbor to face f and zero
otherwise. The grad operator maps Rnc to Rnf and is defined as
grad(p)[f ] = p[C2 (f )] − p[C1 (f )], (4.55)
for any p ∈ Rnc . In the continuous case, the gradient operator is the adjoint of the diver-
gence operator (up to a sign), as we have
 
p ∇ · v d x + v · ∇p d x = 0, (4.56)
 
for vanishing boundary conditions. Let us prove that this property also holds in the discrete
case. To simplify notation, we set Sc = {1, . . . ,nc } and Sf = {1, . . . ,nf }. For any v ∈ Rnf
and p ∈ Rnc , we have
     
div(v)[c] p[c] = p[c] v[f ] 1{c=C1 (f )} − v[f ] 1{c=C2 (f )}
c∈Sc c∈Sc f ∈F (c) f ∈F (c)
 
= v[f ] p[c] 1{c=C1 (f )} 1{f ∈F (c)}
c∈Sc f ∈Sf
 
− v[f ] p[c] 1{c=C2 (f )} 1{f ∈F (c)} . (4.57)
c∈Sc f ∈Sf

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 4.4 Basic Finite-Volume Discretizations 137

We can switch the order in the sums above and obtain

 
v[f ] p[c]1{c=C1 (f )} 1{f ∈F (c)} =
c∈Sc f ∈Sf
 
v[f ] p[c] 1{c=C1 (f )} 1{f ∈F (c)} .
f ∈Sf c∈Sc

For a given face f , we have that 1{c=C1 (f )} 1{f ∈F (c)} is nonzero if and only if c = C1 (f ),
and therefore
  
1{c=C1 (f )} 1{f ∈F (c)} v[f ]p[c] = v[f ]p[C1 (f )].
f ∈Sf c∈Sc f ∈Sf

In the same way, we have

  
v[f ] p[c] 1{c=C2 (f )} 1{f ∈F (c)} = v[f ] p[C2 (f )],
c∈Sc f ∈Sf f ∈Sf

so that (4.57) yields

 
div(v)[c] p[c] + grad(p)[f ] v[f ] = 0. (4.58)
c∈Sc f ∈Sf

Until now, we have ignored boundary conditions. We can include them by introducing one
more cell number c = 0 to denote the exterior. Then, we can consider external faces and
extend the mappings C1 and C2 to Sc ∪ {0} so that a given face f is external if it satisfies
C1 (f ) = 0 or C2 (f ) = 0. Note that the grad operator only defines values on internal
faces. Now taking external faces into account, we obtain
 
div(v)[c] p[c] + grad(p)[f ] v[f ]
c∈Sc f ∈Sf
  
= p[C1 (f )] 1{C2 (f )=0} − p[C2 (f )] 1{C1 (f )=0} v[f ], (4.59)
f ∈S f \Sf

where S f denotes the extended set of faces, consisting of both internal and external faces.
Identity (4.59) is the discrete counterpart to
  
p∇ · v d x + v · ∇p d x = v · n ds.
p (4.60)
  ∂

Let us now use these operators to discretize the Poisson equation. Going back to (4.47), let
the vector v ∈ Rnf be a discrete approximation of the flux on faces; that is, for f ∈ Sf , we
have

v[f ] ≈ v(x) · nf ds,
f

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 138 Mathematical Models for Single-Phase Flow

where nf is the normal to face f with orientation given by the grid. The relation between
the discrete pressure p ∈ Rnc and the discrete flux is given by the two-point flux approxi-
mation discussed in the previous section,

v[f ] = −T [f ] grad(p)[f ] ≈ − K(x)∇p · nf ds. (4.61)
f

Here, T [f ] denotes the transmissibility associated with face f , as defined in (4.52). Hence,
the discretization of (4.47) can be written as

div(v) = q (4.62a)
v = −T grad(p), (4.62b)

where the multiplication in (4.62b) holds element-wise.

We end our discussion by two examples that demonstrate how these discrete operators
can be used to write very compact solvers that easily can be extended to solve problems of
increasing complexity with minimal modifications to the code.

Example 4.4.2 To illustrate the use of the discrete operators, let us set up and solve the
classical Poisson equation on a simple box geometry,

− div(T grad(p)) = q,  = [0,1] × [0,1], (4.63)

subject to no-flow boundary conditions with q consisting of a point source at (0,0) and a
point sink at (1,1). First, we construct a small Cartesian grid
G = computeGeometry(cartGrid([5 5],[1 1]));

for which T equals a scalar multiple of the identity matrix and is therefore dropped for
simplicity. The natural way to represent the div and grad operators is to use sparse
matrices. To this end, we use (4.58), which implies that the sparse matrix representing
div is the negative transpose of the matrix that defines grad using (4.55). Moreover, since
we assume no-flow boundary conditions, we only need to let C1 and C2 account for internal
connections:

C = G.faces.neighbors;
C = C(all(C = 0, 2), :); ∂
nf = size(C,1); ∂x
nc = G.cells.num; D=
D = sparse([(1:nf)'; (1:nf)'], C, ...
ones(nf,1)*[-1 1], nf, nc); ∂
grad = @(x) D*x; ∂y
div = @(x) -D'*x;

Once we have the discrete operators, we can write (4.63) in residual form, f (p) =
Ap + q = 0 and use automatic differentiation as discussed in Example A.5.3 to obtain A
by computing ∂f /∂p

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 4.4 Basic Finite-Volume Discretizations 139

p = initVariablesADI(zeros(nc,1));
q = zeros(nc, 1); % source term
q(1) = 1; q(nc) = -1; % -> quarter five-spot

eq = div(grad(p))+q; % equation
eq(1) = eq(1) + p(1); % make solution unique
p = -eq.jac{1}\eq.val; % solve equation

Next, we try to solve the same type of flow problem on a non-rectangular domain. That
is, we still consider the unit square, but remove two half-circles of radius 0.4 centered at
(0.5,0) and (0.5,1) respectively. To construct the corresponding grid, we use the fictitious
grid approach from Section 3.1 (see Exercise 3.1.1):

G = cartGrid([20 20],[1 1]);

G = computeGeometry(G);
r1 = sum(bsxfun(@minus,G.cells.centroids,[0.5 1]).^2,2);
r2 = sum(bsxfun(@minus,G.cells.centroids,[0.5 0]).^2,2);
G = extractSubgrid(G, (r1>0.16) & (r2>0.16));

The construction of the discrete operators is agnostic to the exact layout of the grid, and
because the transmissibility matrix T is still a multiple of the identity matrix, since the grid
cells are equidistant squares, we can simply reuse the exact same set-up as in the previous
code box:

% Grid information
C = G.faces.neighbors; ∂
: ∂x

% Operators
D = sparse([(1:nf)'; (1:nf)'], C, ...
ones(nf,1)*[-1 1], nf, nc); ∂
: ∂y

% Assemble and solve equations

p = initVariablesADI(zeros(nc,1));
q = zeros(nc, 1);
q(1) = 1; q(nc) = -1;
eq = div(grad(p))+q;
eq(1) = eq(1) + p(1);
p = -eq.jac{1}\eq.val;
plotCellData(G,p);

Notice that the D matrix has almost the same sparsity structure as in our first problem,
except that the nonzero bands now are curved because the number of cells in each colum-
n/row of the grid changes throughout the domain.

Example 4.4.3 To illustrate the power of the combination of an unstructured grid format
and discrete differential operators, we also go through how you can use this technique to

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 140 Mathematical Models for Single-Phase Flow

solve the Poisson equation on an unstructured grid. As an example, we use the Voronoi
grid generated from the seamount data set shown in Figure 3.11 on page 68. Now comes
the important point: Because the discrete differential operators are defined in terms of the
two general matrices C1 and C2 that describe the internal connections in the grid, their
construction remains exactly the same as for the Cartesian grid:

load seamount
G = pebi(triangleGrid([x(:) y(:)], delaunay(x,y)));
G = computeGeometry(G);

C =
G.faces.neighbors;
C C(all(C = 0, 2), :);
=
nf =
size(C,1);
nc =
G.cells.num;
D =
sparse([(1:nf)'; (1:nf)'], C, ...
ones(nf,1)*[-1 1], nf, nc);
grad = @(x) D*x;

Here, the directional derivatives do not follow the axial directions and hence D will have
a general sparse structure and not the banded structure we saw for the Cartesian grids in
the previous example. Moreover, because the cell centers are no longer equidistant points
on a uniform mesh, the diagonal entries in the transmissibility matrix will not be the same
constant for all cells and hence cannot be scaled out of the discrete system. For histor-
ical reasons, MRST only supplies a routine for computing half-transmissibilities defined
in (4.51) on page 132. These are defined for all faces in the grid. Since we have assumed
no-flow boundary conditions, we only need to find the half-transmissibilities associated
with the interior faces, and compute their harmonic average to get the transmissibilities
defined in (4.52):

hT = computeTrans(G, struct('perm', ones(nc,1)));

cf = G.cells.faces(:,1);
T = 1 ./ accumarray(cf, 1 ./ hT, [G.faces.num, 1]);
T = T(all(C=0,2),:);

p = initVariablesADI(zeros(nc,1));
q = zeros(nc, 1); q([135 282 17]) = [-1 .5 .5];
eq = div(T.*grad(p))+q;
eq(1) = eq(1) + p(1);
p = -eq.jac{1}\eq.val;

You may also notice that we have changed our source terms slightly so that there is now a
fluid sink at the center and fluid sources to the northwest and southeast. We will return to a
more detailed discussion of the computation of transmissibilities and assembly of discrete
linear systems in Section 5.2.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 4.4 Basic Finite-Volume Discretizations 141

4.4.3 Time-of-Flight and Tracer

The transport equation (4.40) and (4.42) are examples of a more general class of steady-
state transport problems,
∇ · (u
v ) = h(
x,u), (4.64)
where u = τ and h = φ + τ ∇ · v for time-of-flight and u = C and h = qC + C∇ · v for
the artificial tracer. To discretize the steady transport equation (4.64), we integrate it over a
single grid cell i and use Gauss’s divergence theorem to obtain
 
u v · n ds = h ( x )) d x.
x,u(
∂i i

In Section 4.4.1 we discussed how to discretize the flux over an interface ik between two
cells i and k for the case that u ≡ 1. To be consistent with the notation used above, we
will call this flux vik . If we can define an appropriate value uik at the interface ik , we can
write the flux across the interface as

u v · n ds = uik vik . (4.65)
ik

The obvious idea of setting uik = 12 (ui + uk ) gives a centered scheme that is unfortunately
notoriously unstable. To get a better approximation, we can inspect the direction informa-
tion propagates in the transport equation. If the flux vik is positive, the content of u in cell
i is continuously streaming into cell k , and it is thus natural to choose ui as the interface
value, or vice versa. This gives us the so-called upwind value

ui , if vik ≥ 0,
uik = (4.66)
uk , otherwise.
Using the upwind value can be thought of as adding extra numerical dispersion to stabilize
the resulting scheme so that it does not introduce spurious oscillations; more details about
such upwind schemes are given in Chapter 9.
For completeness, let us also write this discretization using the abstract notation defined
in the previous section. If we discretize u by a vector u ∈ Rnc and h by a vector function
h(u) ∈ Rnc , the transport equation (4.64) can be written in the discrete form
div(uv) = h(u). (4.67)
We also substitute the expression for v from (4.62b) and use (4.66) to define u at each face
f . Then, we define for each face f ∈ Sf ,
(uv)[f ] = uuw [f ] T [f ] grad(p)[f ], (4.68)
where

u[C1 (f )]), if grad(p)[f ] > 0,
u uw
[f ] = (4.69)
u[C2 (f )]), otherwise.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 142 Mathematical Models for Single-Phase Flow

Time-of-flight and tracer distributions can of course also be computed if we trace indi-
vidual streamlines by solving the ordinary differential equation (4.36). The most commonly
used method for tracing streamlines on hexahedral grids is a semi-analytical tracing algo-
rithm introduced by Pollock [253], which uses analytical expressions of the streamline
paths inside each cell, based on the assumption that the velocity field is piecewise linear
locally. Pollock’s method is only correct for regular grids but is still used for highly skewed
and irregular grids. Other approaches for tracing on unstructured grids and the associated
accuracy are discussed in [74, 256, 150, 205, 122, 204, 166]. On unstructured polygonal
grids, tracing of streamlines becomes significantly more involved. Because the general
philosophy of MRST is that solvers should work independent of grid type – so that you
can seamlessly switch from structured to fully unstructured, polygonal grids – we prefer
to use finite-volume methods rather than streamline tracing to compute time-of-flight and
tracer distributions. The disadvantages of this choice are that finite-volume methods have
lower pointwise accuracy and that averaged time-of-flight value have a less intuitive inter-
pretation. We will return to a more in-depth discussion of how these quantities can be used
to investigate timelines and volumetric connections in the reservoir in Chapter 13.

Computer exercises
4.4.1 Compare the discrete differentiation operators for selected grids from Chapter 3,
e.g., Exercises 3.1.1–3.1.3, 3.2.6, and 3.2.7. Can you explain the differences?
4.4.2 Go back to Example 4.4.2 and extend the solver to also compute time-of-flight by
using the discrete operators defined in the previous exercise in this chapter.
4.4.3 Extend the solver from Example 4.4.3 to compute one tracer associated with each
source term. Together, the two tracers should define a partition of unity for the
reservoir volume. How would you visualize it?

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 5
Incompressible Solvers for Single-Phase Flow

A simulation model can be considered to consist of three main parts; the first part describes
the reservoir rock, the second part describes the mathematical laws that govern fluid behav-
ior, and the last represents wells and other drive mechanisms. We have already discussed
how to model the reservoir rock and its petrophysical properties in Chapters 2 and 3, and
shown how the resulting models are represented in MRST using a grid object, usually called
G, that describes the geometry of the reservoir, and a rock object, usually called rock, that
describes petrophysical parameters. Likewise, Chapter 4 discussed the fundamental flow
equations for single-phase flow and presented basic numerical discretizations for elliptic
Poisson-type equations and the gradient/divergence operators that appear in flow models.
This chapter discusses the additional parts you need to make a full model and implement
a simulator. We show how to represent fluid behavior in terms of fluid objects that contain
basic properties such as density, viscosity, and compressibility. We will later extend these
fluid objects to model more complex behavior by including properties like relative perme-
ability and capillary pressure that describe interaction among multiple fluid phase and the
porous rock. Likewise, we discuss necessary data structures to represent forcing terms such
as boundary conditions, (volumetric) source terms, and models of injection and production
wells. It is also convenient to introduce a state object holding the primary unknowns and
derived quantities like pressure, fluxes, and face pressures.
There are two different ways the data objects just outlined can be combined to form a
full simulator. In Section 4.4.2, we saw how to use discrete differential operators to write
the flow equations in residual form and then employ automatic differentiation to linearize
and form a linear system. Whereas this technique is elegant and will prove highly versatile
for compressible flow models later in the book, it is an overkill for incompressible single-
phase flow, since the flow equations already are linear. In this chapter, we therefore outline
how to use a classic procedural approach to implement the discretized flow equations. We
start by outlining the data structures and constructors needed to set up fluid properties
and forcing terms, and once this is done, we move on to discuss in detail how to build two-
point discretizations and assemble and solve corresponding linear systems. For pedagogical
purposes, we present a somewhat simplified version of the basic flow solvers for incom-
pressible flow that are implemented in the add-on modules incomp and diagnostics of

143

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 144 Incompressible Solvers for Single-Phase Flow

MRST. At the end of the chapter we go through several simulation cases and give all code
lines necessary for full simulation setups with various drive mechanisms.

5.1 Basic Data Structures in a Simulation Model

The simple flow solvers we discussed in the previous chapter did not contain any fluid prop-
erties and assumed no-flow boundary conditions and point sources as the only forcing term.
In this section we outline basic data structures you can use to set up more comprehensive
single-phase simulation cases.

5.1.1 Fluid Properties

The only fluid properties we need in the basic single-phase flow equations are the viscosity
and the fluid density for incompressible models and the fluid compressibility for com-
pressible models. More complex single-phase and multiphase models require additional
fluid and rock-fluid properties. To simplify the communication of fluid properties between
flow and transport solvers, it is good practice to introduce a common API. To this end,
MRST uses so-called fluid objects that contain a predefined set of basic fluid properties
as well as function handles used to evaluate rock-fluid properties that are only relevant for
multiphase flow. This basic structure can be expanded by optional parameters and functions
to represent more advanced fluid models. The following shows how to initialize the most
basic fluid object that only requires viscosity and density as input
fluid = initSingleFluid('mu' , 1*centi*poise, ...
'rho', 1014*kilogram/meter^3);

After initialization, the fluid object contains pointers to functions that can be used to eval-
uate petrophysical properties of the fluid:
fluid =
properties: @(varargin)properties(opt,varargin{:})
saturation: @(x,varargin)x.s
relperm: @(s,varargin)relperm(s,opt,varargin{:})

Only the first function is relevant for single-phase flow, and returns the viscosity when
called with a single output argument, and the viscosity and the density when called with
two output arguments. The other two functions can be considered as dummy functions
that ensure that the single-phase fluid object is compatible with solvers written for more
advanced fluid models. The saturation function accepts a reservoir state as argument (see
Section 5.1.2) and returns the corresponding saturation (volume fraction of the fluid phase),
which will either be empty or set to unity, depending upon how the reservoir state has been
initialized. The relperm function accepts a fluid saturation as argument and returns the
relative permeability, i.e., the reduction in permeability due to the presence of other fluid
phases. This function should always be identical to one for single-phase models.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 5.1 Basic Data Structures in a Simulation Model 145

5.1.2 Reservoir States

To hold the dynamic state of the reservoir, MRST uses a special data structure. We refer
to realizations of this structure as the state objects. In its basic form, the structure contains
three elements: a vector pressure with one pressure per cell in the grid, a vector flux with
one value per face in the grid, and a vector s that should either be empty or be a vector with
a unit entry for each cell, since we only have a single fluid. The state object is initialized by
a call to the function
state = initResSol(G, p0, s0);

where p0 is the initial pressure and s0 is an optional parameter, giving the initial saturation
(which should be identical to one for single-phase models). Contrary to what the name may
imply, this function does not initialize the fluid pressure to be in hydrostatic equilibrium. If
such a condition is needed, it must be enforced explicitly by the user. In the case of wells
in the reservoir, you should use the alternative function:
state = initState(G, W, p0, s0);

This gives a state object with an additional field wellSol, which is a vector with
one entry per well. Each element in the vector is a structure that contains two fields,
wellSol.pressure and wellSol.flux. These two fields are vectors of length equal the
number of completions in the well and contain the bottom-hole pressure and flux for each
completion.

5.1.3 Fluid Sources

The simplest way to describe flow in or out from interior points in the reservoir is to use
volumetric source terms. You can create source terms as follows
src = addSource([], cells, rates);
src = addSource(src, cells, rates, 'sat', sat);

Here, the input/output values are:

• src: array of MATLAB structures describing separate sources. If the first input argument
is empty, the routine will output a single structure. Otherwise, it will append the new
structure to the array of existing sources sent as input. Each source structure contains the
following fields:
• cell: cells containing explicit sources,
• rate: rates for these explicit sources,
• value: pressure or flux value for the given condition,
• sat: fluid composition of injected fluids in cells with rate>0.

• cells: indices to the cells in the grid model in which this source term should be applied.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 146 Incompressible Solvers for Single-Phase Flow

• rates: vector of volumetric flow rates, one scalar value for each cell in cells. Note
that these values are interpreted as flux rates (typically in units of [m3 /day]) rather than
as flux density rates (which must be integrated over the cell volumes to obtain flux
rates).
• sat: optional parameter that specifies the composition of the fluid injected from this
source. In this n × m array of fluid compositions, n is the number of elements in cells
and m is the number of fluid phases. For m = 3, the columns are interpreted as: 1=“aqua,”
2=“liquid,” and 3=“vapor.” This field is for the benefit of multiphase transport solvers,
and is ignored for all sinks (at which fluids flow out of the reservoir). The default value is
sat = [], which corresponds to single-phase flow. If size(sat,1)==1, this saturation
value will be repeated for all cells specified by cells.
For convenience, rates and sat may contain a single value; this value is then used for all
faces specified in the call.
There can only be a single net source term per cell in the grid. Moreover, for incom-
pressible flow with no-flow boundary conditions, the source terms must sum to zero if the
model is to be well posed, or alternatively sum to the flux across the boundary. If not, we
would either inject more fluids than we extract, or vice versa, and hence implicitly violate
the assumption of incompressibility.

5.1.4 Boundary Conditions

As discussed in Section 4.3.1, all outer faces of a grid are assumed as no-flow boundaries
unless other conditions are specified explicitly. The basic mechanism for specifying Dirich-
let and Neumann boundary conditions is to use the function:
bc = addBC(bc, faces, type, values);
bc = addBC(bc, faces, type, values, 'sat', sat);

Here, the input values are:

• bc: array of MATLAB structures describing separate boundary conditions. If the first
input argument is empty (bc==[]), the routine will output a single structure. Otherwise,
it will append the new structure to the array of existing boundary conditions sent as input.
Each structure contains the following fields:
• face: external faces for which explicit conditions are set,
• type: cell array of strings denoting type of condition,
• value: pressure or flux value for the given condition,
• sat: composition of fluids passing through inflow faces, not used for single-phase
models.

• faces: array of external faces at which this boundary condition is applied.

• type: type of boundary condition. Supported values are 'pressure' and 'flux', or a
cell array of such strings.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 5.1 Basic Data Structures in a Simulation Model 147

• values: vector of boundary conditions, one scalar value for each face in
faces. Interpreted as a pressure value in units [Pa] when type equals
'pressure' and as a flux value in units [m3 /s] when type is 'flux'. In the latter
case, positive values in values are interpreted as injection fluxes into the reservoir, while
negative values signify extraction fluxes, i.e., fluxes out of the reservoir.
• sat: optional parameter that specifies the composition of the fluid injected across inflow
faces. Similar setup as explained for source terms in Section 5.1.3.
There can only be a single boundary condition per face in the grid. Solvers assume that
boundary conditions are given on the boundary; conditions in the interior of the domain
yield unpredictable results. Moreover, for incompressible flow and only Neumann condi-
tions, the boundary fluxes must sum to zero if the model is to be well posed. If not, we
would either inject more fluids than we extract, or vice versa, and hence implicitly violate
the assumption of incompressibility.

For convenience, MRST also offers two additional routines for setting Dirichlet and
Neumann conditions at all outer faces in a certain direction for grids having a logical I J K
numbering:
bc = pside(bc, G, side, p);
bc = fluxside(bc, G, side, flux)

The side argument is a string that must match one out of the following six alias groups:
1: 'West', 'XMin', 'Left'
2: 'East', 'XMax', 'Right'
3: 'South', 'YMin', 'Back'
4: 'North', 'YMax', 'Front'
5: 'Upper', 'ZMin', 'Top'
6: 'Lower', 'ZMax', 'Bottom'

These groups correspond to the cardinal directions mentioned as the first alternative in each
group. You should also be aware of an important difference in how fluxes are specified in
addBC and fluxside. Specifying a scalar value in addBC means that this value will be copied
to all faces the boundary condition is applied to, whereas a scalar value in fluxside sets
the cumulative flux for all faces that make up the global side to be equal the specified value.

5.1.5 Wells
Wells are similar to source terms in the sense that they describe injection or extraction of
fluids from the reservoir, but differ in the sense that they not only provide a volumetric
flow rate, but also contain a model that couples this flow rate to the difference between
the average reservoir in the grid cell and the pressure inside the wellbore. As discussed in
Section 4.3.2, this relation can be written for each perforation as

vp = J (pi − pf ), (5.1)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 148 Incompressible Solvers for Single-Phase Flow

where J is the well index, pi is the pressure in the perforated grid cell, and pf is the
flowing pressure in the wellbore. The wellbore is assumed to be in hydrostatic equilibrium
so that pf in each completion can be found from the pressure at the top of the well and the
density along the wellbore. For single-phase, incompressible flow, this hydrostatic balance
reads pf = pwh + ρzf , where pwh is the pressure at the well head and zf is the
vertical distance from this point and to the perforation. By convention, the structure used
to represent wells in MRST is called W, and consists of the following fields:

• cells: an array index to cells perforated by this well.

• type: string describing which variable is controlled (i.e., assumed to be fixed), either
'bhp' or 'rate'.
• val: the target value of the well control; pressure value for type='bhp' or rate for
type='rate'.
• r: the wellbore radius (double).
• dir: a char describing the direction of the perforation ('x', 'y' or 'z').
• WI: the well index: either the productivity index or the well injectivity index depending
on whether the well is producing or injecting.
• dZ: the height differences from the well head, which is defined as the topmost contact
(i.e., the contact with the minimum z-value counted amongst all cells perforated by this
well).
• name: string giving the name of the well.
• compi: fluid composition, only used for injectors.
• refDepth: reference depth of control mode.
• sign: defines whether the well is intended to be producer or injector.

Well structures are created by a call to the function

W = addWell(W, G, rock, cellInx);
W = addWell(W, G, rock, cellInx, 'pn', pv, ..);

Here, cellInx is a vector of indices to the cells perforated by the well, and 'pn'/pv denote
one or more keyword/value pairs that can be used to specify optional parameters in the well
model:

• type: string specifying well control, 'bhp' (default) means that the well is controlled
by bottom-hole pressure, whereas 'rate' means that the well is rate controlled.
• val: target for well control. Interpretation of this values depends upon type. For
'bhp' the value is assumed to be in unit Pascal, and for 'rate' the value is given in
unit [m3 /sec]. Default value is 0.
• radius: wellbore radius in meters. Either a single, scalar value that applies to all per-
forations, or a vector of radii, with one value for each perforation. The default radius is
0.1 m.
• dir: well direction. A single CHAR applies to all perforations, while a CHAR array
defines the direction of the corresponding perforation.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 5.2 Incompressible Two-Point Pressure Solver 149

• innerProduct: used for consistent discretizations discussed in Chapter 6.

• WI: well index. Vector of length equal the number of perforations in the well. The default
value is -1 in all perforations, whence the well index will be computed from available
data (cell geometry, petrophysical data, etc.) in grid cells containing well completions.
• Kh: permeability times thickness. Vector of length equal the number of perforations in the
well. The default value is -1 in all perforations, whence the thickness will be computed
from the geometry of each perforated cell.
• skin: skin factor for computing effective well bore radius. Scalar value or vector with
one value per perforation. Default value: 0.0 (no skin effect).
• Comp_i: fluid composition for injection wells. Vector of saturations. Default value:
Comp_i=[1,0,0] (water injection).
• Sign: well type: production (sign=-1) or injection (sign=1). Default value: [] (no type
specified).
• name: string giving the name of the well. Default value is 'Wn', where n is the number of
this well, i.e., n=numel(W)+1.
For convenience, MRST also provides the function
W = verticalWell(W, G, rock, I, J, K)
W = verticalWell(W, G, rock, I, K)

for specifying vertical wells in models described by Cartesian grids or grids that have some
kind of extruded structure. Here,

• I,J: gives the horizontal location of the well heel. In the first mode, both I and J are
given and then signify logically Cartesian indices so that I is the index along the first
logical direction, whereas J is the index along the second logical direction. This mode is
only supported for grids having an underlying Cartesian (logical) structure such as purely
Cartesian grids or corner-point grids.
In the second mode, only I is described and gives the cell index of the topmost cell in the
column through which the vertical well is completed. This mode is supported for logically
Cartesian grids containing a three-component field G.cartDims or for otherwise layered
grids that contain the fields G.numLayers and G.layerSize.
• K: a vector of layers in which this well should be completed. If isempty(K) is true, then
the well is assumed to be completed in all layers in this grid column and the vector is
replaced by 1:num_layers.

5.2 Incompressible Two-Point Pressure Solver

The two-point flux-approximation (TPFA) scheme introduced in Section 4.4.1 is imple-
mented as two different routines. The first routine,
hT = computeTrans(G,rock)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 150 Incompressible Solvers for Single-Phase Flow

computes the half-face transmissibilities and does not depend on the fluid model, the reser-
voir state, or the driving mechanisms and is hence part of is part of MRST’s core function-
ality. The second routine
state = incompTPFA(state, G, hT, fluid, 'mech1', obj1, ..)

takes the complete model description as input and assembles and solves the two-point
system. The routine is specific to incompressible flow and is thus placed in the incomp
module. Here, 'mech' specifies the drive mechanisms ('src', 'bc', and/or 'wells')
using correctly defined objects obj, as discussed in Sections 5.1.3–5.1.5.
Notice that computeTrans may fail to compute sensible transmissibilities if rock.
perm is not given in SI units. Likewise, incompTPFA may produce strange results if the
inflow and outflow specified by the boundary conditions, source terms, and wells do not
sum to zero and hence violate the assumption of incompressibility. However, if fixed
pressure is specified in wells or on parts of the outer boundary, there will be an outflow or
inflow that will balance the net rate specified elsewhere.
The remainder of this section presents details of the inner workings of the incom-
pressible solver. By going through the essential code lines needed to compute half-
transmissibilities and solve and assemble the global system, we demonstrate how simple
it is to implement the TPFA method on general polyhedral grid. If you are not interested
in these details, you can jump directly to Section 5.4, which contains several examples
demonstrating the use of the incompressible solver for single-phase flow.
To focus on the discretization and keep the discussion simple, we will not look at the full
implementation of the two-point solver in incomp. Instead, we discuss excerpts from two
simplified functions, simpleComputeTrans and simpleIncompTPFA, located in the 1phase
directory of the mrst-book module. Together, these form a simplified single-phase solver,
which has been created for pedagogical purposes.
Assume we have a standard grid G containing cell and face centroids, e.g., as computed
by the computeGeometry function discussed in Section 3.4. Then, the essential code lines
of simpleComputeTrans are as follows: First, we define the vectors ci,k from cell cen-
troids to face centroids; see Figure 5.1. To this end, we first need to determine the map

ci,k · ni,k
Ti,k = Ai,k K i
|
ci,k |2
pk
pi k
−1 −1 −1 i
Tik = [Ti,k + Tk,i ] πi,k
ci,k ni,k

Tik (pi − pk ) = qi i,k
k

Figure 5.1 Two-point discretization on general polyhedral cells

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 5.2 Incompressible Two-Point Pressure Solver 151

from faces to cell number so that the correct cell centroid is subtracted from each face
centroid.
hf = G.cells.faces(:,1);
hf2cn = gridCellNo(G);
C = G.faces.centroids(hf,:) - G.cells.centroids(hf2cn,:);

Face normals in MRST are assumed to have length equal the corresponding face areas, and
hence correspond to Ai,k ni,k in (4.51). To get the correct sign, we look at the neighboring
information specifying which cells share the face: if the current cell number is in the first
column, the face normal has positive sign. If not, it gets a negative sign:
sgn = 2*(hf2cn == G.faces.neighbors(hf, 1)) - 1;
N = bsxfun(@times, sgn, G.faces.normals(hf,:));

The permeability tensor may be stored in different formats, as discussed in Section 2.5, and
we therefore use an utility function to extract it:
[K, i, j] = permTensor(rock, G.griddim);

Finally, we compute the half transmissibilities, C T KN/C T C. To limit memory use, this is
done in a for-loop (which is rarely used in MRST):
hT = zeros(size(hf2cn));
for k=1:size(i,2),
hT = hT + C(:,i(k)) .* K(hf2cn, k) .* N(:,j(k));
end
hT = hT./ sum(C.*C,2);

The actual code has a few additional lines that perform various safeguards and consistency
checks.
Once the half transmissibilities have been computed, they can be passed to the
simpleIncompTPFA solver. The first thing this solver needs to do is adjust the half
transmissibilities to account for fluid viscosity, since they were derived for a fluid with
unit viscosity:
mob = 1./fluid.properties(state);
hT = hT .* mob(hf2cn);

Then, we loop through all faces and compute the face transmissibility as the harmonic
average of the half-transmissibilities
T = 1 ./ accumarray(hf, 1 ./ hT, [G.faces.num, 1]);

The MATLAB function accumarray constructs an array by accumulation (see

Appendix A.4). A call to a = accumarray(subs,val) will use the subscripts in subs
to create an array a based on the values val. Each element in val has a corresponding

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 152 Incompressible Solvers for Single-Phase Flow

row in subs. The function collects all elements that correspond to identical subscripts in
subs and stores their sum in the element of a corresponding to the subscript. In our case,
G.cells.faces(:,1) gives the global face number for each half face, and hence the call
to accumarray will sum the transmissibilities of the half-faces that correspond to a given
global face and store the result in the correct place in a vector of G.faces.num elements.
The function accumarray is very powerful and is used a lot in MRST in place of nested
for-loops. In fact, we also employ this function to loop over all cells in the grid and collect
and sum the transmissibilities of the faces of each cell to define the diagonal of the TPFA
matrix:
nc = G.cells.num;
i = all(G.faces.neighbors = 0, 2);
n1 = G.faces.neighbors(i,1);
n2 = G.faces.neighbors(i,2);
d = accumarray([n1; n2], repmat(T(i),[2,1]),[nc, 1]);

Now that we have computed both the diagonal and the off-diagonal elements of A, the
discretization matrix itself can be constructed by a straightforward call to MATLAB’s
sparse function:

I = [n1; n2; (1:nc)'];

J = [n2; n1; (1:nc)'];
V = [-T(i); -T(i); d]; clear d;
A = sparse(double(I), double(J), V, nc, nc);

Finally, we check if Dirichlet boundary conditions are imposed on the system, and if not,
we modify the first element of the system matrix to fix the pressure in the first cell to zero,
before solving the system:
A(1) = 2*A(1);
p = mldivide(A, rhs);

To solve the system, we rely on MATLAB’s default solver mldivide, which uses a com-
plex flow chart to check the structure of the matrix to see whether it is square, diagonal,
(permuted) triangular, tridiagonal, banded, or Hermitian, and then chooses a specialized
and efficient solver accordingly. By running the command spparms(’spumoni’,2) before
you call mldivide, you can tell MATLAB to output information about which linear solver
it chooses and the tests leading up the specific choice. For the type of sparse matrices
we consider here, the end result is a call to a direct solver from UMFPACK implementing
unsymmetric, sparse, multifrontal LU factorization [89, 88]. Such a direct solver is efficient
for small-to-medium-sized systems, but for larger systems it is more efficient to use sparse
iterative solvers such as a (preconditioned) multilevel method. The linear solver can be
passed as a function-pointer argument to both incompTPFA and simpleIncompTPFA,

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 5.3 Upwind Solver for Time-of-Flight and Tracer 153

mrstModule add agmg

state = incompTPFA(state, G, hT, fluid, 'wells', W, 'LinSolve', @(A,b) agmg(A,b,1));

Here, we have used the aggregation-based algebraic multigrid solver AGMG [238, 15],
which integrates well with MRST and is vailable at no cost for academic research and
teaching. Section 12.3.4 discusses various specialized linear solver for compressible multi-
phase flow simulations.
Once the cell pressures have been computed, we can compute pressure values at the face
centroids using the half-face transmissibilities
fp = accumarray(G.cells.faces(:,1), p(hf2cn).*hT, [G.faces.num,1])./ ...
accumarray(G.cells.faces(:,1), hT, [G.faces.num,1]);

and then construct fluxes across the interior faces

ni = G.faces.neighbors(i,:);
flux = -accumarray(find(i), T(i).*(p(ni(:,2))-p(ni(:,1))), [nf, 1]);

In the code excerpts given above, we did not account for gravity forces and general Dirichlet
or Neumann boundary conditions, which both will complicate the code beyond the scope
of the current presentation. The interested reader should consult the actual code to work out
these details. The standard computeTrans function can also be used for different representa-
tions of petrophysical parameters, and includes functionality to modify the discretization by
overriding the definition of cell and face centers and/or including multipliers that modify
the values of the half-transmissibilities; see e.g., Sections 2.4.3 and 2.5.5. Likewise, the
incompTPFA solver from the incomp module is implemented for a general, incompress-
ible flow model with multiple fluid phases with flow driven by a general combination of
boundary conditions, fluid sources, and well models.
We will shortly present several examples of how this solver can be used for flow
problems on structured and unstructured grids. However, before doing so, we outline
another flow solver from the diagnostics module, which will prove useful to visualize
flow patterns.

5.3 Upwind Solver for Time-of-Flight and Tracer

The diagnostics module, discussed in more detail in Chapter 13, provides various func-
tionality to probe a reservoir model to establish communication patterns between inflow
and outflow regions, timelines for fluid movement, and various measures of reservoir het-
erogeneity. At the hart of this module, lies the function
tof = computeTimeOfFlight(state, G, rock, 'mech1', obj1, ..)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 154 Incompressible Solvers for Single-Phase Flow

which implements the upwind, finite-volume discretization introduced in Section 4.4.3 for
solving the time-of-flight equation v · ∇τ = φ. As you probably recall, time-of-flight is
the time it takes a neutral particle to travel from the nearest fluid source or inflow bound-
ary to each point in the reservoir. Here, the 'mech' arguments represent drive mechanisms
('src', 'bc', and/or 'wells') specified in terms of specific objects obj, as discussed in
Sections 5.1.3 to 5.1.5. You can also compute the backward time-of-flight – the time it takes
to travel from any point in the reservoir to the nearest fluid sink or outflow boundary – with
the same equation if we change sign of the flow field and modify the boundary conditions
and/or source terms accordingly. In the following, we will go through the main parts of
how this discretization is implemented.
We start by identifying all volumetric sources of inflow and outflow, which may be
described as source/sink terms in src and/or as wells in W, and collect the results in a
vector q of source terms having one value per cell
[qi,qs] = deal([]);
if isempty(W),
qi = [qi; vertcat(W.cells)];
qs = [qs; vertcat(state.wellSol.flux)];
end
if isempty(src),
qi = [qi; src.cell];
qs = [qs; src.rate];
end
q = sparse(qi, 1, qs, G.cells.num, 1);

We also need to compute the accumulated inflow and outflow from boundary fluxes for each
cell. This will be done in three steps. First, we create an empty vector ff with one entry
per global face, find all faces that have Neumann conditions, and insert the corresponding
value in the correct row
ff = zeros(G.faces.num, 1);
isNeu = strcmp('flux', bc.type);
ff(bc.face(isNeu)) = bc.value(isNeu);

The flux is not specified on faces with Dirichlet boundary conditions and must be extracted
from the solution computed by the pressure solver, i.e., from the state object that holds
the reservoir state. We also need to set the correct sign so that fluxes into a cell are positive
and fluxes out of a cell are negative. The sign of the flux across an outer face is correct if
neighbors(i,1)==0, but if neighbors(i,2)==0 we need to reverse the sign (we should also
check that i is not empty)
i = bc.face(strcmp('pressure', bc.type));
ff(i) = state.flux(i) .* (2*(G.faces.neighbors(i,1)==0) - 1);

The last step is to sum all the fluxes across outer faces and collect the result in a vector qb
that has one value per cell

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 5.3 Upwind Solver for Time-of-Flight and Tracer 155

outer = all(double(G.faces.neighbors) > 0, 2);

qb = sparse(sum(G.faces.neighbors(outer,:), 2), 1, ff(is_outer), G.cells.num, 1);

Here, sum(G.faces.neighbors(outer,:), 2) gives an array containing the indices of each

cell attached to an outer face.
Once the contributions to inflow and outflow are collected, we can start building the
upwind flux discretization matrix A. The off-diagonal entries are defined such that Aj i =
max(vij ,0) and Aij = − min(vij ,0), where vij is the flux computed by the TPFA scheme
discussed in the previous section.
i = any(G.faces.neighbors==0, 2);
out = min(state.flux(i), 0);
in = max(state.flux(i), 0);

The diagonal entry equals the outflux minus the divergence of the velocity, which can be
obtained by summing the off-diagonal rows. This will give the correct equation in all cell
except for those with a positive fluid source. Here, the net outflux equals the divergence of
the velocity and we hence end up with an undetermined equation. In these cells, we can as
a reasonable approximation1 set the time-of-flight to be equal the time it takes to fill the
cell, which means that the diagonal entry should be equal the fluid rate inside the cell.
n = double(G.faces.neighbors(i,:));
inflow = accumarray([n(:, 2); n(:, 1)], [in; -out]);
d = inflow + max(q+qb, 0);

Having obtained diagonal and all the nonzero off-diagonal elements, we can assemble the
full matrix
nc = G.cells.num;
A = sparse(n(:,2), n(:,1), in, nc, nc) + sparse(n(:,1), n(:,2), -out, nc, nc);
A = -A + spdiags(d, 0, nc, nc);

We have now established the complete discretization matrix, and time-of-flight can be
computed by a simple matrix inversion
tof = A \ poreVolume(G,rock);

If there are no gravity forces and the flux has been computed by the two-point method
(or some other monotone scheme), one can show that the discretization matrix A can be
permuted to a lower-triangular form [221, 220]. In the general case, the permuted matrix
will be block triangular with irreducible diagonal blocks. Such systems can be inverted very

1 Notice, however, that to get the correct values for 1D cases, it is more natural to set time-of-flight equal half the time it takes
to fill the cell.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 156 Incompressible Solvers for Single-Phase Flow

efficiently using a permuted back-substitution algorithm as long as the irreducible diagonal

blocks are small. MATLAB is quite good at detecting such structures, and using the simple
backslash (\) operator is therefore efficient, even for quite large models. However, for
models of real petroleum assets described on stratigraphic grids (see Section 3.3), it is often
necessary to preprocess flux fields to get rid of numerical clutter that would otherwise intro-
duce large irreducible blocks inside stagnant regions. By specifying optional parameters to
computeTimeOfFlight, the function will get rid of such small cycles in the flux field and set
the time-of-flight to a prescribed upper value in all cells that have sufficiently small influx.
This tends to reduce the computational cost significantly for large models with complex
geology and/or significant compressibility effects.
The same routine can also compute stationary tracers, as discussed in Section 4.3.4.
This is done by passing an optional parameter,
tof = computeTimeOfFlight(state, G, rock, .., 'tracer',tr)

where tr is a cell-array of vectors that each gives the indices of cells that emit a unique
tracer. For incompressible flow, the discretization matrix of the tracer equation is the same
as that for time-of-flight, and all we need to do to extend the solver is to assemble the
right-hand side
numTrRHS = numel(tr);
TrRHS = zeros(nc,numTrRHS);
for i=1:numTrRHS,
TrRHS(tr{i},i) = 2*qp(tr{i});
end

Since we have doubled the rate in any cells with a positive source when constructing the
matrix A, the rate on the right-hand side must also be doubled.
With the extra right-hand sides assembled, we can solve the combined time-of-
flight/tracer problem as a linear system with multiple right-hand sides,
T = A \ [poreVolume(G,rock) TrRHS];

which means that we essentially get the tracer for free as long as the number of tracers does
not exceed the number of right-hand columns MATLAB can handle in one solve. We will
return to a more thorough discussion of the tracer partitions Chapter 13 and show how these
can be used to delineate connectivities within the reservoir. In the rest of this chapter, we
will consider time-of-flight and streamlines as a means to study flow patterns in reservoir
models.

5.4 Simulation Examples

We have now introduced you to all the functionality from the incomp module neces-
sary to solve a single-phase flow problem, as well as the time-of-flight solver from the
diagnostics module, which can be used to compute time lines in the reservoir. In the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 5.4 Simulation Examples 157

following, we discuss several examples and demonstrate step by step how to set up a flow
model, solve it, and visualize and analyze the resulting flow field. Complete codes can be
found in the 1phase directory of the book module.

5.4.1 Quarter Five-Spot

As our first example, we show how to solve −∇ · (K∇p) = q with no-flow boundary con-
ditions and two source terms at diagonally opposite corners of a 2D Cartesian grid covering
a 500 × 500 m2 area. This setup mimics a standard quarter five-spot well pattern, which
we encountered in Figure 4.7 on page 126 when discussing well models. The full code is
available in the script quarterFiveSpot.m. We use a rectangular grid with homogeneous
petrophysical data (K = 100 mD and φ = 0.2):
[nx,ny] = deal(32);
G = cartGrid([nx,ny],[500,500]);
G = computeGeometry(G);
rock = makeRock(G, 100*milli*darcy, .2);

As we saw above, all we need to develop the spatial discretization is the reservoir geometry
and the petrophysical properties. This means that we can compute the half transmissibilities
without knowing any details about the fluid properties and the boundary conditions and/or
sources/sinks that will drive the global flow:
hT = simpleComputeTrans(G, rock);

The result of this computation is a vector with one value per local face of each cell in the
grid, i.e., a vector with G.cells.faces entries.
The reservoir is horizontal and gravity forces are therefore not active. We create a fluid
with properties that are typical for water:
gravity reset off
fluid = initSingleFluid('mu' , 1*centi*poise, ...
'rho', 1014*kilogram/meter^3);

To drive flow, we use a fluid source in the southwest corner and a fluid sink in the northeast
corner. The time scale of the problem is defined by the strength of the source terms.
Here, we set these terms so that a unit time corresponds to the injection of one pore
volume of fluids. By convention, all flow solvers in MRST automatically assume no-
flow conditions on all outer (and inner) boundaries if no other conditions are specified
explicitly.
pv = sum(poreVolume(G,rock));
src = addSource([], 1, pv);
src = addSource(src, G.cells.num, -pv);

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 158 Incompressible Solvers for Single-Phase Flow

The data structure used to represent the fluid sources contains three elements:
cell: [2x1 double]
rate: [2x1 double]
sat: []

We recall that src.cell gives cell numbers where source terms are nonzero, and the
vector src.rate specifies the fluid rates, which by convention are positive for inflow into
the reservoir and negative for outflow from the reservoir. The last data element src.sat
specifies fluid saturations, which only has meaning for multiphase flow models and hence
is set to be empty here.
Strictly speaking, state structure need not be initialized for an incompressible model
in which none of the fluid properties depend on the reservoir state. However, to avoid
treatment of special cases, MRST requires that the structure is initialized and passed as
argument to the pressure solver. We therefore initialize it with a dummy pressure value of
zero and a unit fluid saturation (fraction of void volume filled by fluid), since we only have
a single fluid
state = initResSol(G, 0.0, 1.0);
display(state)

state =
pressure: [1024x1 double]
flux: [2112x1 double]
s: [1024x1 double]

This completes the setup of the model. To solve for pressure, we simply pass reservoir
state, grid, half transmissibilities, fluid model, and driving forces to the flow solver, which
assembles and solves the incompressible equation.
state = simpleIncompTPFA(state, G, hT, fluid, 'src', src);
display(state)

state =
pressure: [1024x1 double]
flux: [2112x1 double]
s: [1024x1 double]
facePressure: [2112x1 double]

As explained here, simpleIncompTPFA solves for pressure as the primary variable and then
uses transmissibilities to reconstruct the face pressure and intercell fluxes. After a call to
the pressure solver, the state object is therefore expanded by a new field facePressure
that contains pressures reconstructed at the face centroids. Figure 5.2 shows the resulting
pressure distribution. To improve the visualization of the flow field, we show streamlines.
The streamlines add-on module to MRST implements Pollock’s method [253] for semi-
analytical tracing of streamlines. Here, we use this functionality to trace streamlines for-
ward and backward, starting from the midpoint of all cells along the northwest–southeast
diagonal in the grid

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 5.4 Simulation Examples 159

plotCellData(G, state.pressure); hf=streamline(Sf);

plotGrid(G, src.cell, 'FaceColor', 'w'); hb=streamline(Sb);
axis equal tight; colormap(jet(128)); set([hf; hb],'Color','k');

Figure 5.2 Solution of the quarter five-spot problem on a 32 × 32 uniform grid. The left plot shows
the pressure distribution and in the right plot we have imposed streamlines passing through centers
of the cells on the northwest–southeast diagonal.

mrstModule add streamlines;

seed = (nx:nx-1:nx*ny).';
Sf = pollock(G, state, seed, 'substeps', 1);
Sb = pollock(G, state, seed, 'substeps', 1, 'reverse', true);

The pollock routine produces a cell array of individual streamlines, which we pass onto
MATLAB’s streamline routine for plotting, as shown to the right in Figure 5.2.
To get a better picture of how fast the fluid flows through our domain, we solve the
time-of-flight equation (4.40) subject to the condition that τ = 0 at the inflow, i.e., at all
points where q > 0. For this purpose, we use the computeTimeOfFlight solver discussed
in Section 5.3, which can compute time-of-flight both forward from inflow points and into
the reservoir,
toff = computeTimeOfFlight(state, G, rock, 'src', src);

and from outflow points and backwards into the reservoir

tofb = computeTimeOfFlight(state, G, rock, 'src', src, 'reverse', true);

Isocontours of time-of-flight define natural time lines in the reservoir. To emphasize this,
the left plot in Figure 5.3 shows time-of-flight plotted using only a few colors to make a
rough contouring effect. The sum of forward and backward time-of-flight gives the total
time it takes a fluid particle to pass from an inflow point to an outflow point. We can thus
use this total residence time to visualize high-flow and stagnant regions, as demonstrated
in the right plot of Figure 5.3.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 160 Incompressible Solvers for Single-Phase Flow
500 500

450 450

400 400

350 350

300 300

250 250

200 200

150 150

100 100

50 50

0 0
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500

plotCellData(G, toff); plotCellData(G, toff+tofb);

plotGrid(G,src.cell,'FaceColor','w'); plotGrid(G,src.cell,'FaceColor','w');
axis equal tight; axis equal tight;
colormap(jet(16)); caxis([0,1]); colormap(jet(128));

Figure 5.3 Quarter five-spot problem on a 32 × 32 grid. The left plot shows time-of-flight plotted
with a few color levels to create a crude contouring effect. The right plots total travel time clearly
distinguishing high-flow and stagnant regions.

Computer exercises
5.4.1 Run the quarter five-spot example with the following modifications:
a. Replace the Cartesian grid by a curvilinear grid, e.g., using twister or a
random perturbation of internal nodes as shown in Figure 3.3.
b. Replace the grid by the locally refined grid from Exercise 3.2.6.
c. Replace the homogeneous permeability by a heterogeneous permeability
derived from the Carman–Kozeny relation (2.6).
d. Set the domain to be a single layer of the SPE 10 model. Hint: use
getSPE10rock() to sample the petrophysical parameters and remember
to convert to SI units.
Notice that the pollock function may not work for non-Cartesian grids.
5.4.2 Construct a grid similar to the one in Exercise 3.1.1, except that the domain is given
a 90-degree flip so that axis of the cylindrical cutouts align with the z-direction.
Modify the code presented in this section so that you can compute a five-spot setup
with one injector near each corner and a producer in the narrow middle section
between the cylindrical cutouts.

5.4.2 Boundary Conditions

To demonstrate how to specify boundary conditions, we go through essential code lines of
three different examples. In all three examples, the reservoir is 50 m thick, is located at a

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 5.4 Simulation Examples 161

depth of approximately 500 m, and is restricted to a 1 × 1 km2 area. The permeability is

uniform and anisotropic, with a diagonal (1,000, 300, 10) mD tensor, and the porosity is
uniform and equal 0.2. In the first two examples, the reservoir is represented as a 20×20×5
rectangular grid, and in the third example the reservoir is given as a corner-point grid of the
same Cartesian dimension, but with an uneven uplift and four intersecting faults (as shown
in the left plot of Figure 3.32):
[nx,ny,nz] = deal(20, 20, 5);
[Lx,Ly,Lz] = deal(1000, 1000, 50);
switch setup
case 1,
G = cartGrid([nx ny nz], [Lx Ly Lz]);
case 2,
G = cartGrid([nx ny nz], [Lx Ly Lz]);
case 3,
G = processGRDECL(makeModel3([nx ny nz], [Lx Ly Lz/5]));
G.nodes.coords(:,3) = 5*(G.nodes.coords(:,3)-min(G.nodes.coords(:,3)));
end
G.nodes.coords(:,3) = G.nodes.coords(:,3) + 500;

Setting rock and fluid parameters, computing transmissibilities, and initializing the reser-
voir state can be done as explained in the previous section, and details are not included for
brevity; you find the complete scripts in boundaryConditions.

Linear Pressure Drop

In the first example (setup=1), we specify Neumann conditions with total inflow of
5,000 m3 /day on the east boundary and Dirichlet conditions with fixed pressure of 50 bar
on the west boundary:
bc = fluxside(bc, G, 'EAST', 5e3*meter^3/day);
bc = pside (bc, G, 'WEST', 50*barsa);

This completes the definition of the model, and we can pass the resulting objects to the
simpleIncompTFPA solver to compute the pressure distribution shown to the right in Figure
5.4. In the absence of gravity, these boundary conditions will result in a linear pressure drop
from east to west inside the reservoir.

Hydrostatic Boundary Conditions

In the next example, we use the same model, except that we now include the effects
of gravity and assume hydrostatic equilibrium at the outer vertical boundaries of the
model. First, we initialize the reservoir state according to hydrostatic equilibrium, which
is straightforward to compute if we for simplicity assume that the overburden pressure is
caused by a column of fluids with the exact same density as in the reservoir:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 162 Incompressible Solvers for Single-Phase Flow
61

60
500
59
500
510
58
510
520 57

Depth
520
56
530
530
55
540
540 54

550
550 1000 53
1000
800 1000 52
800 1000 600 800
600 800 600
400
400 600
400 51
400 200
200 200
200 0
y 0 [bar]
0 0 x

Figure 5.4 First example of a flow driven by boundary conditions. In the left plot, faces with
Neumann conditions are marked in blue and faces with Dirichlet conditions are marked in red. The
right plot shows the resulting pressure distribution.

state = initResSol(G, G.cells.centroids(:,3)*rho*norm(gravity), 1.0);

There are at least two different ways to specify hydrostatic boundary conditions. The
simplest approach is to use the function psideh, i.e.,
bc = psideh([], G, 'EAST', fluid); bc = psideh(bc, G, 'WEST', fluid);
bc = psideh(bc, G, 'SOUTH', fluid); bc = psideh(bc, G, 'NORTH', fluid);

Alternatively, we can do it manually ourselves. To this end, we need to extract the reservoir
perimeter defined as all exterior faces are vertical, i.e., whose normal vector has no z-
component,
f = boundaryFaces(G);
f = f(abs(G.faces.normals(f,3))<eps);

To get the hydrostatic pressure at each face, we can either compute it directly by using the
face centroids,
fp = G.faces.centroids(f,3)*rho*norm(gravity);

or we use the initial equilibrium that has already been established in the reservoir by
sampling from the cells adjacent to the boundary
cif = sum(G.faces.neighbors(f,:),2);
fp = state.pressure(cif);

The latter may be useful if the initial pressure distribution has been computed by a more
elaborate procedure than what is currently implemented in psideh. In either case, the
boundary conditions can now be set by the call

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 5.4 Simulation Examples 163
54
54

52
53.5 500

500
510 50
53
510

520
520 52.5 48

Depth
530 530
52 46
540 540

51.5
550 44
1000 550
1000
800 1000
51 800 1000
800 42
600 600 800
600 600
400 400
400 50.5 400
200 200
200 200 40
y 0 0 [bar]
0 0 [bar] x

Figure 5.5 A reservoir with hydrostatic boundary condition and fluid extracted from a sink
penetrating two cells in the upper two layers of the model. The left plot shows the boundary and
the fluid sink, while the right plot shows the resulting pressure distribution.

bc = addBC(bc, f, 'pressure', fp);

To make the problem a bit more interesting, we also include a fluid sink at the midpoint of
the upper two layers in the model,
ci = round(.5*(nx*ny-nx));
ci = [ci; ci+nx*ny];
src = addSource(src, ci, repmat(-1e3*meter^3/day,numel(ci),1));

Figure 5.5 shows the boundary conditions and source terms to the left and the resulting
pressure distribution to the right. The fluid sink causes a pressure drawdown, which has
ellipsoidal shape because of the anisotropic permeability field.

Conditions on Non-Rectangular Domain

For the reservoir shown in Figure 5.6, we cannot simply use tags for to select faces on the
east and west perimeter. The problem is that fluxside and pside define cardinal sides to
consist of all exterior faces whose normal vector point in the correct cardinal direction. This
is illustrated in the upper-left plot of Figure 5.6, where we have tried to specify boundary
conditions using the same procedure as in Figure 5.4. We can easily get rid of faces that
really lie on the north and south boundary if we use the subrange feature of fluxside and
pside to restrict the boundary conditions to a subset of the global side, as shown in the
upper-right plot of Figure 5.6

bc = fluxside([], G, 'EAST', 5e3*meter^3/day, 4:15, 1:5);

bc = pside (bc, G, 'WEST', 50*barsa, 7:17, []);

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 164 Incompressible Solvers for Single-Phase Flow

500 500

520 520

540 540

560 560

580 580

600 600

620 620

640 1000 640 1000

0 800 0 800
200 600 200 600
400 400
600 400 600 400
800 200 800 200
1000 0 1000 0

75
500 500

520 520
70

540 540

560
560 65

Depth
580
580

600 60
600

620
620
55
0 800
640 1000 200
0 600
800 400
200 600 400
600
400 800 200 50
600 400
1000 [bar]
800 200 y
1000 0 x

Figure 5.6 Specifying boundary conditions along the outer perimeter of a corner-point model. The
upper-left plot shows the use of fluxside/pside (blue/red color) and to set boundary conditions on
the east and west global boundaries. In the upper-right point, the same functions have been used with
a specification of subranges in the global sides. In the lower-left plot, we have utilized user-supplied
information to correctly set the conditions only along the perimeter. The lower-right plot shows the
resulting pressure solution.

Unfortunately, fluxside/pside still picks exterior faces at the faults, classified as pointing
east/west because of their normal. Similar problems may arise in other models because of
pinched, eroded, or inactive cells.
To find the east- and west-most faces only, we need to use advanced tactics. In our case,
the natural perimeter is defined as those faces that lie on the bounding box of the model,
on which we distribute the total flux to individual faces according to the face area. For the
Neumann condition we therefore get
x = G.faces.centroids(f,1);
[xm,xM] = deal(min(x), max(x));
ff = f(x>xM-1e-5);
bc = addBC(bc, ff, 'flux', (5e3*meter^3/day) ...
* G.faces.areas(ff)/ sum(G.faces.areas(ff)));

We can specify the Dirichlet condition in a similar manner. The lower-right plot shows the
correct linear pressure drop.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 5.4 Simulation Examples 165

Computer exercises
5.4.3 Consider a 2D box with a sink at the midpoint and inflow across the perimeter spec-
ified either in terms of a constant pressure or a constant flux. Are there differences
in the two solutions, and if so, can you explain why? Hint: use time-of-flight, total
travel time, and/or streamlines to investigate.
5.4.4 Apply the production setup from Figure 5.5, with hydrostatic boundary conditions
and fluids extracted from two cells at the midpoint of the model, to the model
depicted in Figure 5.6.
5.4.5 Compute flow for all models in data sets BedModels1 and BedModel2 subject to
linear pressure drop in the x and then in the y-direction. These models of small-
scale heterogeneity are developed to compute representative properties in simu-
lation models on a larger scale. A linear pressure drop is the most widespread
computational setup used for flow-based upscaling. What happens if you try to
specify flux conditions?
5.4.6 Consider models from the CaseB4 data set. Use appropriate boundary conditions
to drive flow across the faults and compare flow patterns computed on the pillar and
the stair-stepped grid for the two different model resolutions. Can you explain any
differences you observe?

5.4.3 Structured versus Unstructured Stencils

We have so far only discussed hexahedral grids having structured cell numbering. The two-
point schemes can also be applied to fully unstructured and polyhedral grids. To demon-
strate this, we define a non-rectangular reservoir by scaling the grid from Figure 3.8 to
cover a 1 × 1 km2 area.
load seamount;
T = triangleGrid([x(:) y(:)], delaunay(x,y));
[Tmin,Tmax] = deal(min(T.nodes.coords), max(T.nodes.coords));
T.nodes.coords = bsxfun(@times, ...
bsxfun(@minus, T.nodes.coords, Tmin), 1000./(Tmax - Tmin));
T = computeGeometry(T);

We assume a homogeneous and isotropic permeability of 100 mD and use the same fluid
properties as in the previous examples. Constant pressure of 50 bar is set at the outer
perimeter and fluid is drained at a constant rate of one pore volume over 50 years from
a well located at (450,500). (Script: stencilComparison.m.)
For comparison, we generate two Cartesian grids covering the same domain, one with
approximately the same number of cells as the triangular grid and a 10 × 10 refinement of
this grid to provide a reference solution,
G = computeGeometry(cartGrid([25 25], [1000 1000]));
inside = isPointInsideGrid(T, G.cells.centroids);
G = removeCells(G, inside);

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 166 Incompressible Solvers for Single-Phase Flow

The function isPointInsideGrid implements a simple algorithm for finding whether

points lie inside the circumference of a grid. First, all boundary faces are extracted and
then the corresponding nodes are sorted so that they form a closed polygon. Then, we
use MATLAB’s built-in function inpolygon to check if the points are inside this polygon.
To also construct a radial grid, we reuse the code from page 86 to set up points inside
[−1,1] × [−1,1] graded radially towards the origin
P = [];
for r = exp([-3.5:.2:0, 0, .1]),
[x,y] = cylinder(r,25); P = [P [x(1,:); y(1,:)]];
end
P = unique([P'; 0 0],'rows');

We scale the points and translate so that their origin coincides with the fluid sink
[Pmin,Pmax] = deal(min(P), max(P));
P = bsxfun(@minus, bsxfun(@times, ...
bsxfun(@minus, P, Pmin), 1200./(Pmax-Pmin)), [150 100]);

We remove all points outside of the triangular grid before we use the point set to first
generate a triangular and then a Voronoi grid:
inside = isPointInsideGrid(T, P);
V = computeGeometry( pebi( triangleGrid(P(inside,:)) ));

Once we have constructed the grids, the setup of the remaining part of the model
is the same in all cases. To avoid unnecessary replication of code, we collect the
grids in a cell array and use a simple for-loop to set up and simulate each model
realization:
g = {G, T, V, Gr};
for i=1:4
rock = makeRock(g{i}, 100*milli*darcy, 0.2);
hT = simpleComputeTrans(g{i}, rock);
pv = sum(poreVolume(g{i}, rock));

tmp = (g{i}.cells.centroids - repmat([450, 500],g{i}.cells.num,[])).^2;

[,ind] = min(sum(tmp,2));
src{i} = addSource(src{i}, ind, -.02*pv/year);

bc{i} = addBC([], boundaryFaces(g{i}), 'pressure', 50*barsa);

state{i} = incompTPFA(initResSol(g{i},0,1), ...

g{i}, hT, fluid, 'src', src{i}, 'bc', bc{i}, 'MatrixOutput', true);

[tof{i},A{i}] = computeTimeOfFlight(state{i}, g{i}, rock,...

'src', src{i},'bc',bc{i}, 'reverse', true);
end

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 5.4 Simulation Examples 167

50

49

48

47

46

45

44

43

42
Reference
Cartesian
41 Triangular
Radial
40
0 50 100 150 200 250 300

Figure 5.7 Comparison of the pressure solution for three different grid types: uniform Cartesian,
triangular, and a graded radial grid. The scattered points used to generate the triangulated domain
and limit the reservoir are sampled from the seamount data set and scaled to cover a 1 × 1 km2
area. Fluids are drained from the center of the domain, assuming a constant pressure of 50 bar at the
perimeter.

Figure 5.7 shows the pressure solutions computed on the four different grids, whereas
Figure 5.8 reports the sparsity patterns of the corresponding linear systems for the three
coarse grids. As expected, the Cartesian grid gives a banded matrix consisting of five
diagonals corresponding to each cell and its four neighbors in the cardinal directions. Even
though this discretization is not able to predict the complete drawdown at the center (the
reference solution predicts a pressure slightly below 40 bar), it captures the shape of the
drawdown region quite accurately; the region appears ellipsoidal because of the non-unit
aspect ratio in the plot. In particular, we see that the points in the radial plot follow those
of the fine-scale reference closely. The spread in the points as r → 300 is not a grid
orientation effect, but the result of variations in the radial distance to the fixed pressure at
the outer boundary on all four grids.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 168 Incompressible Solvers for Single-Phase Flow
Cartesian Triangular Radial
0 0 0

100 100
100

200
200
200
300
300
300
400
400
500 400

0 100 200 300 400 0 200 400 0 100 200 300 400
nz = 2330 nz = 2244 nz = 3045

Figure 5.8 Sparsity patterns for the TPFA stencils on the three different grid types shown in
Figure 5.7.

The unstructured triangular grid is more refined near the well and is hence able to pre-
dict the pressure drawdown in the near-well region more accurately. However, the overall
structure of this grid is quite irregular, as you can see from the sparsity pattern of the linear
system shown in Figure 5.8, and the irregularity gives significant grid orientation effects.
You can see this from the irregular shape of the color contours in the upper part of Figure
5.7, as well as from the spread in the scatter plot. In summary, this grid is not well suited
for resolving the radial symmetry of the pressure drawdown in the near-well region. But to
be fair, the grid was not generated for this purpose either.
Except for close to the well and close to the exterior boundary, the topology of the
radial grid is structured in the sense that each cell has four neighbors, two in the radial
direction and two in the angular direction, and the cells are regular trapezoids. This should,
in principle, give a banded sparsity pattern if the cells are ordered starting at the natural
center point and moving outward, one ring at the time. To verify this claim, you can execute
the following code:
[,q] = sort(state{3}.pressure);
spy(state{3}.A(q,q));

However, as a result of how the grid was generated by first triangulating and then forming
the dual, the cells are numbered from west to east, which explains why the sparsity pattern
is so far from being a simple banded structure. This may potentially affect the efficiency of
a linear solver, but has no impact on the accuracy of the numerical approximation, which is
good because of the grading towards the well and the symmetry inherent in the grid. Slight
differences in the radial profile compared with the Cartesian grid(s) are mainly the result of
the fact that the source term and the fixed pressure conditions are not located at the exact
same positions in the simulations.
In Figure 5.9, we also show the sparsity pattern of the linear system used to compute
the reverse time-of-flight from the well and back into the reservoir. Using the default cell
ordering, the sparsity pattern of each upwind matrix will appear as a less dense version

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 5.4 Simulation Examples 169
Cartesian Triangular Radial
0 0 0

100 100
100
200
200
200
300
300
400 300

400
500 400

0 100 200 300 400 0 200 400 0 100 200 300 400
nz = 1408 nz = 1405 nz = 1745
0 0 0

100 100
100
200
200
200
300
300
400 300

400
500 400

0 100 200 300 400 0 200 400 0 100 200 300 400
nz = 1408 nz = 1405 nz = 1745

Figure 5.9 Sparsity patterns for the upwind stencils used to compute time-of-flight on the three
different grid types shown in Figure 5.7. In the lower row, the matrices have been permuted to lower-
triangular form by sorting the cell pressures in ascending order.

of the pattern for the corresponding TPFA matrix. However, whereas the TPFA matrices
represent an elliptic equation in which information propagates in both directions across
cell interfaces, the upwind matrices are based on one-way connections arising from fluxes
between pairs of cells that are connected in the TPFA discretization. To reveal the true
nature of the system, we can permute the system by either sorting the cell pressures in
ascending order (potential ordering) or using the function dmperm to compute a Dulmage–
Mendelsohn decomposition. As pointed out in Section 5.3, the result is a lower triangular
matrix, from which it is simple to see the unidirectional propagation of information you
should expect for a hyperbolic equation having only positive characteristics.

Computer exercises
5.4.7 Compare the sparsity patterns resulting from potential ordering and use of dmperm
for both the upwind and the TPFA matrices.
5.4.8 Investigate the flow patterns in more details using forward time-of-flight, travel
time, and streamlines.
5.4.9 Replace the boundary conditions by a constant influx, or set pressure values sam-
pled from a radially symmetric pressure solution in an infinite domain.

5.4.4 Using Peaceman Well Models

Flow in and out of a wellbore takes place on a scale much smaller than a single grid cell in
typical sector and field models and is therefore commonly modeled using a semi-analytical

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 170 Incompressible Solvers for Single-Phase Flow

model of the form (4.35). This section presents two examples to demonstrate how such
models can be included in the simulation setup using data objects and utility functions
introduced in Section 5.1.5.

Box Reservoir
We consider a reservoir consisting of a homogeneous 500 × 500 × 25 m3 sand box with
an isotropic permeability of 100 mD, represented on a regular 20 × 20 × 5 Cartesian grid.
The fluid is the same as in the examples as in the earlier examples in this section. You can
find all code lines necessary to set up the model, solve the flow equations, and visualize the
results in the script firstWellExample.m. Setting up the model is quickly done once you
have gotten familiar with MRST:
[nx,ny,nz] = deal(20,20,5);
G = computeGeometry( cartGrid([nx,ny,nz], [500 500 25]) );
rock = makeRock(G, 100*milli*darcy, .2);
fluid = initSingleFluid('mu',1*centi*poise,'rho',1014*kilogram/meter^3);
hT = computeTrans(G, rock);

The reservoir is produced by a well pattern consisting of a vertical injector and a horizontal
producer. The injector is located in the southwest corner of the model and operates at a
constant rate of 3,000 m3 per day. The producer is completed in all cells along the upper
east rim and operates at a constant bottom-hole pressure of 1 bar (i.e., 105 Pascal in SI
units):
W = verticalWell([], G, rock, 1, 1, 1:nz, 'Type', 'rate', 'Comp_i', 1,...
'Val', 3e3/day(), 'Radius', .12*meter, 'name', 'I');
W = addWell(W, G, rock, nx : ny : nx*ny, 'Type', 'bhp', 'Comp_i', 1, ...
'Val', 1.0e5, 'Radius', .12*meter, 'Dir', 'y', 'name', 'P');

In addition to specifying the type of control on the well ('bhp' or 'rate'), we must
specify wellbore radius and fluid composition, which is '1' for a single phase. After
initialization, the array W contains two data objects, one for each well:

Well #1: | Well #2:

cells: [5x1 double] | cells: [20x1 double]
type: ’rate’ | type: ’bhp’
val: 0.0347 | val: 100000
r: 0.1000 | r: 0.1000
dir: [5x1 char] | dir: [20x1 char]
WI: [5x1 double] | WI: [20x1 double]
dZ: [5x1 double] | dZ: [20x1 double]
name: ’I’ | name: ’P’
compi: 1 | compi: 1
refDepth: 0 | refDepth: 0
sign: 1 | sign: []

This concludes the model specification, and we now have all the information we need to
initialize the reservoir state, and assemble and solve the system

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 5.4 Simulation Examples 171

Producer inflow profile [m3/d]

180
P
170
I 160

150

140

130

120
0 5 10 15 20

Pressure [bar] 3
Flux intensity [m /day]

Figure 5.10 Solution of a single-phase, incompressible flow problem inside a box reservoir with a
vertical injector and a horizontal producer.

gravity reset on;

resSol = initState(G, W, 0);
state = incompTPFA(state, G, hT, fluid, 'wells', W);

Figure 5.10 shows how the inflow rate decays with the distance to the injector as expected.
We can compute the flux intensity depicted in the lower-right plot using the following
command, which first maps the vector of face fluxes to a vector with one flux per half face
and then sums the absolute value of these fluxes to get a flux intensity per cell:
cf = accumarray(getCellNoFaces(G), abs(faceFlux2cellFlux(G, state.flux)));

Shallow-Marine Reservoir
In the final example, we return to the SAIGUP model from Section 3.5.1. This model
does not represent a real reservoir, but is one out of a large number of models built to
be plausible realizations that contain the types of structural and stratigraphic features one
could encounter in models of real clastic reservoirs. Continuing from Section 3.5.1, we
simply assume that the grid and the petrophysical model has been loaded and processed.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 172 Incompressible Solvers for Single-Phase Flow

The script saigupWithWells.m reports all details of the setup and also explains how to
speed up the grid processing by using two C-accelerated routines for constructing a grid
from ECLIPSE input and computing areas, centroids, normals, volumes, etc.
The permeability input is an anisotropic tensor with zero vertical permeability in a
number of cells. As a result, some parts of the reservoir may be completely sealed off
from the wells. This will cause problems for the time-of-flight solver, which requires that
all cells in the model must be flooded after some finite time that can be arbitrarily large.
To avoid this potential problem, we assign a small constant times the minimum positive
vertical permeability to the cells that have zero cross-layer permeability.
is_pos = rock.perm(:, 3) > 0;
rock.perm(is_pos, 3) = 1e-6*min(rock.perm(is_pos, 3));

Similar safeguards are implemented in most commercial simulators.

We recover fluid from the reservoir using six producers spread throughout the middle of
the reservoir; each producer operates at a fixed bottom-hole pressure of 200 bar. Eight injec-
tors located around the perimeter provide pressure support, each operating at a prescribed
and fixed rate. The wells are described by a Peaceman model as in the previous example.
To make the code as compact as possible, all wells are vertical with location specified in
the logical ij subindex available in the corner-point format. The following code specifies
the injectors:
nz = G.cartDims(3);
I = [ 3, 20, 3, 25, 3, 30, 5, 29];
J = [ 4, 3, 35, 35, 70, 70,113,113];
R = [ 1, 3, 3, 3, 2, 4, 2, 3]*500*meter^3/day;
W = [];
for i = 1 : numel(I),
W = verticalWell(W, G, rock, I(i), J(i), 1:nz, 'Type', 'rate', ...
'Val', R(i), 'Radius', .1*meter, 'Comp_i', 1, ...
'name', ['I$_{', int2str(i), '}$']);
end

The producers are specified in the same way. Figure 5.11 shows the well positions and the
pressure distribution. We see a clear pressure buildup along the east, south, and west rims
of the model. Similarly, there is a pressure drawdown in the middle of the model around
producers P2, P3, and P4. The total injection rate is set so that one pore volume will be
injected in a little less than 40 years.
This is a single-phase simulation, but let us for a while think of our setup in terms of
injection and production of different fluids (since the fluids have identical properties, we
can think of a blue fluid being injected into a black fluid). In an ideal situation, one would
wish that the blue fluid would sweep the whole reservoir before it breaks through to the
production wells, as this would maximize the displacement of the black fluid. Even in the
simple quarter five-spot examples in Section 5.4.1 (see Figure 5.3), we saw that this was
not the case, and you cannot expect that this will happen here, either. The lower plot in
Figure 5.11 shows all cells in which the total travel time (sum of forward and backward

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 5.4 Simulation Examples 173

Figure 5.11 Incompressible, single-phase simulation of the SAIGUP model. The upper plot shows
pressure distribution, and the lower plot shows cells with total travel time of less than 50 years.

time-of-flight) is less than 50 years. By looking at such a plot, you can get a quite a
good idea of regions where there is very limited communication between the injectors and
producers (i.e., areas without colors). If this was a multiphase flow problem, these areas
would typically contain bypassed or unswept oil and be candidates for infill drilling or
other mechanisms that would improve the volumetric sweep. We will come back to a more
detailed discussion of flow patterns and volumetric connections in Section 13.5.2.

Computer exercises
5.4.10 Change the parameter 'Dir' from 'y' to 'z' in the box example and rerun the case.
Can you explain why you get a different result?
5.4.11 Switch the injector in the box example to be controlled by a bottom-hole pressure
of 200 bar. Where would you place the injector to maximize production rate if you
can only perforate (complete) it in five cells?
5.4.12 Consider the SAIGUP model: can you improve the well placement and/or the
distribution of fluid rates. Hint: is it possible to utilize time-of-flight information?
5.4.13 Use the function getSPE10setup to set up an incompressible, single-phase version
of the full SPE 10 benchmark. Compute pressure, time-of-flight and tracer concen-
trations associated with each well. Hint: You may need to replace MATLAB’s stan-
dard backslash-solver by a highly-efficient iterative solver like AGMG [238, 15] to
get reasonable computational performance. Also, beware that you may run out of
memory if your computer is not sufficiently powerful.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 6
Consistent Discretizations on Polyhedral Grids

The two-point flux-approximation (TPFA) scheme is robust in the sense that it generally
gives a linear system that has a solution regardless of the variations in K and the geometrical
and topological complexity of the grid. The resulting solutions will also be monotone, but
the scheme is only consistent for certain combinations of grids and permeability tensors K.
This implies that a TPFA solution will not necessarily approach the true solution when we
increase the grid resolution. It also means that the scheme may produce different solutions
depending upon how the grid is oriented relative to the main flow directions; we will discuss
this in more detail in Chapter 10. In this chapter, we first explain the lack of consistency
for TPFA, before we introduce a few consistent schemes implemented in MRST. If you
have only limited interest in discretizations, my advice is that you read Sections 6.1 and 6.3
before continuing to Chapter 7.

6.1 The TPFA Method Is Not Consistent

To provide some background for the following discussion, we start by giving a quick
and informal recap of some important concepts from basic numerical analysis. When we
approximate the Laplace operator L = ∇ · K∇ by a numerical scheme, the approximation
can be written as follows:
Lp = Lh p + T (h),
where h is the characteristic size of the cells in our grid, Lh is the approximation to
the Laplace operator, and T is the truncation error, i.e., the difference between the true
operator and our approximation. A numerical scheme is said to be consistent if the trun-
cation error vanishes as h tends to zero, or in other words, if the difference between the
true operator and our approximation diminishes as the grid is refined. A fundamental
result in numerical analysis (Lax’s equivalence theorem) states that a consistent numerical
method is convergent if and only if the method is stable. By convergent, we mean that
our approximations will approach the true solution as the size of the grid cells approaches
zero. By stable, we mean that errors we make in approximating parameters determining
the solution (for Poisson’s equation: the source term q and the permeability K) do not give
unbounded amplification of the error in the computed solution. Here, a word of caution is

174

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 6.1 The TPFA Method Is Not Consistent 175

pi pk
i πi,k k
ci,k ni,k
K
i,k

Figure 6.1 Two cells in a Cartesian grid and a full permeability tensor whose principal axes are not
aligned with the coordinate system.

in order: a scheme does not need to be consistent to be convergent; showing consistency

and stability is only a convenient approach to prove convergence.
Example 6.1.1 To show that the TPFA scheme is generally not consistent, we start by a
simple example. Recall from Section 4.4.1 that if pi and pk denote the average pressures
in two neighboring cells i and k , then the flux across the interface ik between them is
given as
vik = Tik (pi − pk ), (6.1)
where the transmissibility Tik depends on the geometry of the two cells and the associated
permeability tensors Ki and Kk . To see that the method is not consistent, we assume that K
is a homogeneous symmetric tensor


Kxx Kxy
K= .
Kxy Kyy
Consider now the flux across an interface ik between two cells i and k in a 2D
Cartesian grid, whose normal vector ni,k points in the x-direction so that ni,k = ci,k =
(1,0); see Figure 6.1. If follows from Darcy’s law that v · n = −(Kxx ∂x p + Kxy ∂y p),
from which it is easy to see that the flux across the interface will have two components, one
orthogonal and one transverse,
 

vik = − K∇p · n ds = − Kxx ∂x p + Kxy ∂y p ds.
ik ik
The two flux components correspond to the x-derivative and the y-derivative of the pres-
sure, respectively. In the two-point approximation (6.1), the only points we use to approx-
imate the flux are the pressures pi and pk . Because these two pressures are associated
with the same y-value, their difference can only be used to estimate ∂x p and not ∂y p. This
means that the TPFA method cannot account for the transverse flux contribution Kxy ∂y p
and will hence generally not be consistent. The only exception is if Kxy ≡ 0 in all cells in
the grid, in which case there are no fluxes in the transverse direction.
To link the derivation in Example 6.1.1 to the geometry of the cells, we can use the
one-sided definition of fluxes (4.51) of the two-point scheme to write
Ai,k
vi,k ≈ Ti,k (pi − πi,k ) = ci,k ) · ni,k (pi − πi,k ).
(K
|
ci,k |2

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 176 Consistent Discretizations on Polyhedral Grids

More generally, we have that the TPFA scheme is only convergent for K-orthogonal grids.
An example of a K-orthogonal grid is a grid in which all cells are parallelepipeds in 3D or
parallelograms in 2D and satisfy the condition
ni,j · K
ni,k = 0, ∀j = k, (6.2)
where ni,j and ni,k denote normal vectors to faces of cell number i.
What about grids that are not parallelepipeds or parallelograms? For simplicity, we only
consider the 2D case. Let c = (c1,c2 ) denote the vector from the center of cell i to the
centroid of face ik ; see Figure 4.10. If we let c⊥ = (c2, − c1 ) denote a vector that is
orthogonal to c, then any constant pressure gradient inside i can be written as ∇p =
p1 c + p2 c⊥ , where p1 can be determined from pi and πi,k , and p2 is some unknown
constant. Inserted into the definition of the face flux, this gives

 
vi,k = − c) · n + p2 (K
p1 (K c⊥ ) · n ds.
ik
For the two-point scheme to be correct, the second term in the integrand must be zero.
Setting n = (n1,n2 ), we find that
0 = (K
c⊥ ) · n = (K1 c2, − K2 c1 ) · (n1,n2 ) = (K1 n1 c2 − K2 n2 c1 ) = (K
n) × c.
In other words, a sufficient condition for a polygonal grid to be K-orthogonal is that
(Kni,k )  ci,k for all cells in the grid, in which case the transverse flux contributions are all
zero and hence need not be estimated. Consequently, the TPFA method will be consistent.
The lack of consistency in the TPFA scheme for grids that are not K-orthogonal may lead
to significant grid orientation effects [8, 319], i.e., artifacts or errors in the solution that will
appear in varying degree depending upon the angles the cell faces make with the principal
directions of the permeability tensor. We have already seen an example of grid effects
in Figure 5.7 in Section 5.4.3, where an irregular triangular grid caused large deviations
from what should have been an almost symmetric pressure drawdown. To demonstrate grid
orientation effects and lack of convergence more clearly, we look at a simple example that
has been widely used to teach reservoir engineers the importance of aligning the grid with
the principal permeability directions.
Example 6.1.2 Consider a homogeneous reservoir in the form of a 2D rectangle [0,4] ×
[0,1]. Flow is driven from the north to the south side of the reservoir by a fluid source
with unit rate located at (2,0.975) along the north side, and two fluid sinks located at
(0.5,0.025) and (3.5,0.025) along the south side. The two source terms each have rate
equal one half and are symmetric to the line x = 2 and will therefore result in a symmetric
pressure distribution.
To challenge TPFA, we discretize the domain by a skewed grid that is uniform along the
north side and graded towards east along the south side:
G = cartGrid([41,20],[2,1]);
makeSkew = @(c) c(:,1) + .4*(1-(c(:,1)-1).^2).*(1-c(:,2));
G.nodes.coords(:,1) = 2*makeSkew(G.nodes.coords);

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 6.2 The Mixed Finite-Element Method 177

Figure 6.2 Solution of a symmetric flow problem in a homogeneous domain using the TPFA method
on a skew grid that is not K-orthogonal. The upper-left plot shows the pressure distribution, whereas
the lower-left plot shows streamlines and values of time-of-flight less than 0.25 PVI. The plot to the
right shows time-of-flight in the left (L) and right (R) sink for a series of refined grids of dimension
(20n + 1) × 10n for n = 1, . . . ,30. The solutions do not converge toward the analytical solution
shown as the dashed red line.

If we assume an isotropic and homogeneous permeability field, the grid is not K-

orthogonal, and we therefore cannot expect that the TPFA method will converge to the
correct physical solution. You can find a complete setup of the problem in the script
1phase/showInconsistentTPFA.m.
Figure 6.2 reports the result of a convergence study. All the computed solutions exhibit
the same lack of symmetry you see in the solution on the 41 × 20 grid. Moreover, the
large difference in travel times from the injector to the two producers does not decay
with increasing grid resolution, which confirms the expected lack of convergence for an
inconsistent scheme.

Grid orientation errors are discussed in more detail in Section 10.4.2. It is a good advice to
try to make grids so that cell faces align with the principle directions of the permeability
tensor. A main motivation for using PEBI grids is to maintain flexibility in areal represen-
tation while keeping grid orientation errors at a minimum; see [128, 121, 210].

6.2 The Mixed Finite-Element Method

There are several ways to formulate consistent discretization methods. The mixed finite-
element method [53] is based on a quite different formulation than the two-point method.
The first new idea is that instead of discretizing the second-order Poisson equation, we form
a system that consists of two discrete equations representing mass conservation and Darcy’s
law. Hence, instead of solving for pressure and computing discrete fluxes by postprocessing

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 178 Consistent Discretizations on Polyhedral Grids

the pressure solution, we solve for pressure and fluxes simultaneously. The second new
idea is to look for solutions that satisfy the flow equations in a weak sense; that is, look
for solutions that fulfill the equation when it is multiplied with a suitable test function
and integrated in space. The third new idea is to express the unknown solution as a linear
combination of a set of basis functions, i.e., piecewise polynomials with localized support.
MRST does not implement the mixed finite-element method directly in the classical
sense, but the key ideas of this method are instrumental when we develop a general class of
finite-volume discretizations in Section 6.3 and discussed in more details in the remaining
parts of the chapter. As a precursor to this discussion, we review the mixed finite-element
method in some detail.

6.2.1 Continuous Formulation

We start by restating the continuous flow equation written as a first-order system:
∇ · v = q, v = −K∇p, x ∈  ⊂ Rd (6.3)
with boundary conditions v · n = for x ∈ ∂. Gravity has
' been omitted for brevity, but can
easily be included. For compatibility, we require that  q d x = 0. Because this is a pure
Neumann boundary-value problem, pressure ' p is only defined up to an arbitrary constant,
and as an extra constraint, we require that  p d x = 0.
In the mixed method, we look for solutions in an abstract function space. To this end, we
need two spaces: L2 () is the space of square integrable functions, and H0div is a so-called
Sobolev space defined as
H0div () = {
v ∈ L2 ()d : ∇ · v ∈ L2 () and v · n on ∂}, (6.4)
i.e., as the set of square-integrable, vector-valued functions with compact support in ,
whose divergence is also square integrable. The mixed formulation of (6.3) now reads: find
a pair (p, v) that lies in L2 () × H0div () and satisfies
 
−1
u · K v d x − p ∇ · u d x = 0, ∀
u ∈ H0div (),

 
 (6.5)
w ∇ · v d x = qw d x, ∀w ∈ L2 ().
 

The first equation follows by multiplying Darcy’s law by K−1 and applying the divergence
theorem to the term involving' the pressure gradient. (Recall that u is zero on ∂ by
definition so that the term ∂ p u · n ds = 0.)
We can write equation (6.5) on a more compact form if we introduce the following three
inner products

b(·,·) : H0div × H0div → R, b(u, v) = u · K−1 v d x


c(·,·) : H0 × L → R,
div 2
c(u,p) = p ∇ · u d x


Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 6.2 The Mixed Finite-Element Method 179

(·,·) : L2 × L2 → R, (q,w) = qw d x, (6.6)


and write

u, v) − c(
b( u,p) = 0, v,w) = (q,w).
c( (6.7)

Alternatively, we can derive (6.7) by minimizing the energy functional


1
v) =
I ( v · K−1 v d x,
2 

over all mass-conservative functions v ∈ H0div (), i.e., subject to the constraint

∇ · v = q.

The common strategy for solving such minimization problems is to introduce a Lagrangian
functional, which in our case reads
 
  1
L(v,p) = 1
2v · K−1 v − p ∇ · v − q d x = b(
v, v) − c(
v,p) + (p,q),
 2
where p is a so-called Lagrangian multiplier. At a minimum of L, we must have ∂L/∂ v =
0. Looking at an increment u of v, we have the requirement that
∂L 1 
0= u = lim L(v + ε u,p) − L(
v,p) = b(
u, v) − c(
u,p)
∂ v ε→0 ε

for all u ∈ H0div (). Similarly, we can show that

∂L 1 
0= w = lim L(v,p + εw) − L(
v,p) = −c(
v,w) + (q,w)
∂p ε→0 ε

for all w ∈ L2 (). To show that the solution v is a minimal point, we consider a pertur-
bation v + u that satisfies the constraints so that u ∈ H0div and ∇ · u = 0. For the energy
functional we have that

1
v + u) =
I ( v + u) · K−1 (
( v + u) d x = I (
v ) + I (
u) + b(
u, v)
2 
= I (v ) + I (
u) + c(u,p) = I ( v ) + I (
u) > I (v ),

which proves that v is indeed a minimum of I . We can also prove that the solution (p, v)
v,w) ≤ L(
is a saddle-point of the Lagrange functional, i.e., that L( v,p) ≤ L(u,p) for all
u = v and w = p. The right inequality can be shown as follows

L( v + u, v + u) − c(

v + u,p) = 12 b( v + u,p) + (q,p)
= L(
v,p) + I (
u) + b(
u, v) − c(
u,p) = L(
v,p) + I (
u) > L(
v,p).

The left inequality follows in a similar manner.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 180 Consistent Discretizations on Polyhedral Grids

6.2.2 Discrete Formulation

To discretize the mixed formulation, (6.5), we introduce a grid h = ∪i , replace L2 ()
and H0div () by finite-dimensional subspaces U and V defined over h , and rewrite the
inner products (6.6) as sums of integrals localized to cells

b(·,·)h : V × V → R, u, v)h =
b( u · K−1 v d x,
i i

c(·,·)h : V × U → R, u,p)h =
c( p ∇ · u d x, (6.8)
i i

(·,·)h : U × U → R, (q,w)h = qw d x.
i i

To obtain a practical numerical method, the spaces U and V are typically defined as piece-
wise polynomial functions that are nonzero on a small collection of grid cells. For instance,
in the Raviart–Thomas method [262, 53] of lowest order for triangular, tetrahedral, or
regular parallelepiped grids, L2 () is replaced by

U = {p ∈ L2 (∪i ) : p|i is constant ∀i ⊂ } = span{χi },

where χi is the characteristic function of grid cell i . Likewise, H0div () is replaced by a

space V consisting of functions v ∈ H0div (∪i ) that have linear components on each grid
cell i ∈ , have normal components v · nik that are constant on each cell interface ik , and
are continuous across these interfaces. These requirements are satisfied by functions v ∈
P1 (i )d , i.e., first-order polynomials that on each grid cell take the form v(
x ) = a + B x,
where a is a constant vector and B is a matrix. These functions can be parametrized in terms
of the faces between two neighboring cells. This means that we can write V = span{ψ ik },
where each function ψ ik is defined as

1, if j l = ik ,
ψ ik ∈ P1 (i ) ∪ P1 (k )
d d
and (ψ ik · nj l )|j l =
0, otherwise.

Figure 6.3 illustrates the four nonzero basis functions for the special case of a Cartesian
grid in 2D.
To derive a fully discrete method, we use χi and ψ ik as our trial functions and express

the unknown p( x ) and v(
x ) as sums over these trial functions, p = i pi χi and v =

v ψ
ik ik ik . We now have two different degrees of freedom, pi associated with each cell
and vik associated with each interface ik , which we collect in two vectors p = {pi } and
v = {vik }. Using the same functions as test functions, we derive a linear system of the form




B −C T v 0
= . (6.9)
C 0 p q

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 6.2 The Mixed Finite-Element Method 181

Figure 6.3 Illustration of the velocity basis for the lowest-order Raviart–Thomas method for a 2D
Cartesian grid. There are four basis functions that have nonzero support in the interior of the center
cell. The basis functions correspond to degrees of freedom associated with the west, east, south, and
north cell faces.

Here, B = [bik,j l ], C = [ci,kl ], and q = [qi ], where:


bik,j l = b(ψ ik , ψ j l )h = ψ ik · K−1 ψ j l d x,
 

ci,kl = c(ψ kl ,χi )h = ∇ · ψ kl d x, (6.10)
i

qi = (q,χi )h = q d x.
i

Note that for the first-order Raviart–Thomas finite elements, we have

⎧
⎪
⎨ 1, if i = k,
⎪
ci,kl = −1, if i = l,
⎪
⎪
⎩ 0, otherwise.

The matrix entries bik,j l depend on the geometry of the grid cells and whether K is isotropic
or anisotropic. Unlike the two-point method, mixed methods are consistent and will there-
fore be convergent also on grids that are not K-orthogonal.
In [2], we presented a simple MATLAB code that in approximately seventy-five code
lines implements the lowest-order Raviart–Thomas mixed finite-element method on regular
hexahedral grids for flow problems with diagonal tensor permeability in two or three spatial
dimensions. The code is divided into three parts: assembly of the B block, assembly of
the C block, and a main routine that loads data (permeability and grid), assembles the
whole matrix, and solves the system. Figure 6.4 illustrates the sparsity pattern of the mixed
system for a nx × ny × nz Cartesian grid. The system matrix clearly has the block structure
given in (6.9). The matrix blocks B and C each have three nonzero blocks that correspond
to velocity basis functions oriented along the three axial directions. That is, degrees of
freedom at interfaces (fluxes) have been numbered in the same way as the grid cells; i.e.,
first faces orthogonal to the x-direction, then faces orthogonal to the y-direction, and finally
faces orthogonal to the z-direction, resulting in a heptadiagonal structure.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 182 Consistent Discretizations on Polyhedral Grids

A: B:

C:

Figure 6.4 Sparsity patterns for the full mixed finite-element matrix A, and the matrix blocks B and
C for a 4 × 3 × 3 Cartesian grid.

The code presented in [2] exploits the simple geometry of the Cartesian grid to integrate
the Raviart–Thomas basis functions exactly and perform a direct assembly of the matrix
blocks, giving a MATLAB code that is both compact and efficient. For more general grids,
you would typically perform an element-wise assembly by looping over all cells in the
grid, map each cell them back to a reference element, and then integrate the individual
inner production by use of a suitable numerical quadrature rule. In our experience, this
procedure is cumbersome to implement for general stratigraphic and polyhedral grids and
would typically require the use of many different reference elements or subdivision of cells.
MRST is designed to work on general polyhedral grids in 3D and does not supply any
direct implementation of mixed finite-element methods. Instead, the closest we get to a
mixed method is a finite-volume method in which the half-transmissibilities are equivalent
to the discrete inner products for the lowest-order Raviart–Thomas (RT0) method on rect-
angular cuboids; Section 6.4 on page 188 explains more details. In the rest of the current
section, we introduce an alternative formulation of the mixed method that gives a smaller
linear system that is better conditioned.

6.2.3 Hybrid Formulation

In the same way as we proved that the solution of the mixed problem is a saddle point, we
can show that the linear system (6.9) is indefinite. Indefinite systems are harder to solve and
generally require special linear solvers. We therefore introduce an alternative formulation
that, when discretized, gives a positive-definite discrete system and thereby simplifies the
computation of a discrete solution. Later in the chapter, this so-called hybrid formulation

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 6.2 The Mixed Finite-Element Method 183

will form the basis for a general family of finite-volume discretizations on polygonal and
polyhedral grids.
A hybrid formulation avoids solving a saddle-point problem by lifting the constraint that
the normal velocity must be continuous across cell faces and instead integrate (6.3) to get a
weak form that contains jump terms at the cell interfaces. Continuity of the normal velocity
component is then reintroduced by adding an extra set of equations, in which the pressure
π at the cell interfaces plays the role of Lagrange multipliers. (Recall how Lagrange multi-
pliers were used to impose mass conservation as a constraint in the minimization procedure
used to derive the mixed formulation in Section 6.2.1.) Introducing Lagrange multipliers
does not change v or p, but enables the recovery of pressure values at cell faces, in addition
to introducing a change in the structure of the weak equations. Mathematically, the mixed
1
hybrid formulation reads: find ( v,p,π) ∈ H0div (h ) × L2 (h ) × H 2 (h ) such that
   
−1
u · K v − p ∇ · u d x + π u · n ds = 0,
i i ik ik
 
w ∇ · v d x = qw d x, (6.11)
i i h

μ v · n ds = 0
ik ik

1
for all test functions u ∈ H0div (h ), w ∈ L2 (h ), and μ ∈ H 2 (h ). Here, h denotes all
1
the interior faces of the grid and H (h ) is the space spanned by the traces1 of functions
2

in H 1 (h ), i.e., the space of square integrable functions whose derivatives are also square
integrable. As for the mixed formulation, we can introduce inner products to write the weak
equations (6.11) in a more compact form,
u, v) − c(
b( u,p) + d(
u,π) = 0
v,w) = (q,w)
c( (6.12)
v,μ) = 0,
d(
where the inner products b(·,·), c(·,·), and (·,·) are defined as in (6.6), and d(·,·) is a new
inner product defined over the interior faces,
1 
d(·,·)h : H0 (h ) × H (h ) → R,
div 2 d(v,π) = π v · n ds. (6.13)
ik ik

To derive a fully discrete problem, we proceed in the exact same way as for the mixed
1
problem by first replacing the function spaces L2 , H0div , and H 2 by finite-dimensional
subspaces V and U that are spanned by piecewise polynomial basis functions with local

1 If you are not familiar with the notion of a trace operator, think of it as the values of a function along the boundary of the
domain this function is defined on.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 184 Consistent Discretizations on Polyhedral Grids

support, as discussed earlier. In the lowest-order approximation, the finite-dimensional

space  consists of functions that are constant on each face,

1, if x ∈ ik ,
 = span{μik }, x) =
μik ( (6.14)
0, otherwise.

Using these basis functions as test and trial functions, one can derive a discrete linear
system of the form,
⎡ ⎤⎡ ⎤ ⎡ ⎤
B C D v 0
⎣C T 0 0 ⎦ ⎣−p ⎦ = ⎣q ⎦ , (6.15)
DT 0 0 π 0

where the vectors v, p, and π collect the degrees of freedom associated with fluxes across
the cell interfaces, face pressures, and cell pressures, respectively, the matrix blocks B and
C are defined as in (6.10), and D has two nonzero entries per column (one entry for each
side of the cell interfaces).
The linear system (6.15) is an example of a sparse, symmetric, indefinite system, i.e.,
a system A whose quadratic form x T Ax takes both positive and negative values. Several
methods for solving such systems can be found in the literature, but these are generally not
as efficient as solving a symmetric, positive-definite system. We will therefore use a so-
called Schur-complement method to reduce the mixed hybrid system to a positive-definite
system. The Schur-complement method basically consists of using a block-wise Gaussian
elimination of (6.15) to form a positive-definite system (the Schur complement) for the face
pressures,

D T B −1 D − F T L−1 F π = F T L−1 q. (6.16)

Here, F = C T B −1 D and L = C T B −1 C. Given the face pressures, the cell pressures and
fluxes can be reconstructed by back-substitution, i.e., by solving

Lp = q + F π, Bu = Cp − Dπ. (6.17)

Unlike the mixed method, the hybrid method has a so-called explicit flux representation,
which means that the intercell fluxes can be expressed as a linear combination of neighbor-
ing values for the pressure. This property is particularly useful in fully implicit discretiza-
tions of time-dependent problems, as discussed in Chapter 7.
This is all we will discuss about mixed and mixed-hybrid methods herein. If you want
to learn more about these methods, and how to solve the corresponding linear systems,
you should consult some of the excellent books on the subject, for instance [50, 45, 53].
In the rest of the chapter, we discuss how ideas from mixed-hybrid methods can be used
to formulate finite-volume methods that are consistent and hence convergent on non-K-
orthogonal grids.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 6.3 Finite-Volume Methods on Mixed Hybrid Form 185

6.3 Finite-Volume Methods on Mixed Hybrid Form

We will now present a large class of consistent, finite-volume methods. Our formulation
borrows ideas from the mixed methods, but is much simpler to formulate and implement
for general polygonal and polyhedral grids. For simplicity, we assume that all methods can
be written on the following local form

v i = T i (ei pi − π i ). (6.18)

This is essentially the same formulation we used for the one-sided fluxes to develop the
TPFA method in (4.51), except that now v i is a vector of all one-sided fluxes associated
with cell i , π i is a vector of face pressures, T i is a matrix of one-sided transmissibilities,
and ei = (1, . . . ,1)T has one unit value per face of the cell. If we define M i = T −1 i , the
local discretization (6.18) can alternatively be written as

M i v i = e i pi − π i . (6.19)

Consistent with the discussion in Section 6.2, we refer to the matrix M i as the local inner
product. From the local discretizations (6.18) or (6.18) on each cell, we derive a linear
system of discrete global equations written on mixed or hybridized mixed form. Both
forms are supported in MRST, but herein we only discuss the hybrid form for brevity.
In the two-point method, we assembled the linear system by combining mass conservation
and Darcy’s law into one second-order discrete equation for the pressure. In the mixed
formulation, mass conservation and Darcy’s law were kept as separate first-order equations
that together formed a coupled system for pressure and face fluxes. In the hybrid formu-
lation, the continuity of pressures across cell faces was introduced as a third equation that
together with mass conservation and Darcy’s law constitute a coupled system for pressure,
face pressure, and fluxes. Not all consistent methods need to, or can be, formulated in this
form. Nevertheless, using the mixed hybrid formulation will enable us to give a uniform
presentation of a large class of schemes.
Going back to the TPFA method formulated in Section 4.4.1, it follows immediately
from (4.52) that we can write the method as in (6.18) and that the resulting matrix T i is
diagonal with entries

(T i )kk = nik · K cik |2,

cik /| (6.20)

where the length of the normal vector nik is assumed to be equal the area of the correspond-
ing face. The equivalent form (6.19) follows trivially by setting M i = T −1i . Examples of
consistent methods that can be written in the form (6.18) or (6.19) include the lowest-order
RT0 method seen in the previous section, multipoint flux approximation (MPFA) schemes
[9, 94, 7], and recently developed mimetic finite-difference methods [52], as well as the
coarse-scale formulation of several multiscale methods [95]. For all these methods, the
corresponding T i and M i will be full matrices. We will come back to more details about
specific schemes later in the chapter.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 186 Consistent Discretizations on Polyhedral Grids

To derive the global linear system on mixed hybrid form, we augment (6.19) with flux
and pressure continuity across cell faces. Assembling contributions from all cells in the
grid, gives the following linear system
⎡ ⎤⎡ ⎤ ⎡ ⎤
B C D v 0
⎣C T 0 0 ⎦ ⎣−p ⎦ = ⎣q ⎦ , (6.21)
DT 0 0 π 0

where the first row in the block-matrix equation corresponds to (6.19) for all grid cells.
Here, vector v contains the outward fluxes associated with half-faces ordered cell-wise so
that fluxes over interior interfaces in the grid appear twice, once for each half-face, with
opposite signs. Likewise, the vector p contains the cell pressures, and π the face pressures.
The matrices B and C are block diagonal with each block corresponding to a cell,
⎡ ⎤ ⎡ ⎤
M1 0 ... 0 e1 0 ... 0
⎢ 0 M2 . . . 0 ⎥ ⎢0 e2 . . . 0⎥
⎢ ⎥ ⎢ ⎥
B=⎢ . .. .. ⎥, C=⎢. .. .. ⎥ , (6.22)
⎣ .. . ... . ⎦ ⎣ .. . ... .⎦
0 0 ... Mn 0 0 ... en

where M i = T −1 i . Similarly, each column of D corresponds to a unique interface in the

grid and has two unit entries for interfaces between cells in the interior of the grid. The two
nonzero entries appear at the indexes of the corresponding internal half-faces in the cell-
wise ordering. Similarly, D has a single nonzero entry in each column that corresponds to
an interface between a cell and the exterior.
The hybrid system (6.21) is obviously much larger than the linear system for the standard
TPFA method, but can be reduced to a positive-definite system for the face pressures, as
discussed in Section 6.2.3, and then solved using either MATLAB’s standard linear solvers
or a highly efficient, third-party solver like AGMG [238, 15]. From (6.16) and (6.17),
we see that to compute the Schur complement to form the reduced linear system for face
pressures, and to reconstruct cell pressures and face fluxes, we only need B −1 . Moreover,
the matrix L is by construction diagonal, and computing fluxes is therefore an inexpensive
operation. Many schemes – including the mimetic method, the MPFA-O method, and the
standard TPFA scheme – yield algebraic approximations for the B −1 matrix. Thus, (6.21)
encompasses a family of discretization schemes whose properties are determined by the
choice of B, which we will discuss in more detail in Section 6.4.
Before digging into details about specific, consistent methods, we revisit the example
discussed on page 177 to demonstrate how use of a consistent method significantly reduces
the grid orientation effects seen in Figure 6.2. In MRST, the mimetic methods are imple-
mented in a separate module. Replacing TPFA by a mimetic solver is easy; we first load
the mimetic module by calling
mrstModule add mimetic

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 6.3 Finite-Volume Methods on Mixed Hybrid Form 187

Figure 6.5 Solution of a symmetric flow problem in a homogeneous domain using the TPFA method
(left) and the mimetic method (right) on a skew grid that is not K-orthogonal.

Then, we can replace computation of half-face transmissibilities by a call that constructs

the local mimetic half-transmissibility T i or its equivalent inner product M i = T −1
i for
each cell
% hT = computeTrans(G, rock);
S = computeMimeticIP(G, rock);

Likewise, we replace each call to the two-point solver by a call to a function that assembles
and solves the global mixed hybrid system
% state = incompTPFA(state, G, hT, fluid);
state = incompMimetic(state, G, S, fluid);

We can also use the same routine to assemble a mixed system, and in certain cases also
a TPFA-type system. In Figure 6.5 we have applied the mimetic solver to the exact same
setup as in Figure 6.2. The approximate solution computed by the mimetic method is
almost symmetric and represents a significant improvement compared with the TPFA
method. In particular, the difference in travel times between the injector and each producer
is reduced from 17% for TPFA to less than 2% for the mimetic method. Moreover, repeating
a similar grid refinement study as reported in Figure 6.2 verifies that the consistent mimetic
method converges toward the correct solution.
The two calls just outlined constitute all you need to obtain a consistent discretization
on general polygonal and polyhedral grids, and if you are not interested in getting to know
the inner details of various consistent methods, you can safely jump to the next chapter.

Computer exercises
6.3.1 Run mimeticExample1 to familiarize yourself with the mimetic solver, the mixed-
hybrid formulation, the Schur-component reduction, etc.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 188 Consistent Discretizations on Polyhedral Grids

6.3.2 Go back to Exercise 5.4.1a and verify that using the mimetic method cures the grid
orientation effects you can observe on the non-K-orthogonal grid.
6.3.3 Can you cure the grid orientation problems observed in Figure 5.7 by using a
consistent discretization?
6.3.4 Compare the solutions computed by TPFA and the mimetic method for the
SAIGUP model as set up in Section 5.4.4. In particular, you should compare
discrepancies in pressure and time-of-flight values at the well perforations. Which
solution do you trust most?
Hints: (i) Remember to use the correct inner product when setting up well models.
(ii) The mimetic linear system is much larger than the TPFA system and you
may have to use an iterative solver like AGMG to reduce runtime and memory
consumption.

6.4 The Mimetic Method

Mimetic finite-difference methods are examples of so-called compatible spatial discretiza-
tions that are constructed so that they not only provide accurate approximation of the
mathematical models but also inherit or mimic fundamental properties of the differential
operators and mathematical solutions they approximate. Examples of properties include
conservation, symmetries, vector calculus identities, etc. Such methods have become very
popular in recent years and are currently used in wide range of applications [40].
Mimetic methods can be seen as a finite-volume generalization of finite-differences or
(low-order) mixed-finite element methods to general polyhedral grids. The methods are
defined in a way that introduces a certain freedom of construction that naturally leads to a
family of methods. By carefully picking the parameters needed to fully specify a method,
one can construct mimetic methods that coincide with other known methods, or reduce to
these methods (e.g., the two-point method, the RT0 mixed finite-element method, or the
MPFA-O multipoint method) on certain types of grids.
The mimetic methods discussed herein can all be written on the equivalent forms (6.18)
or (6.19) and are constructed so that they are exact for linear pressure fields and give
a symmetric positive-definite matrix M. In addition, the methods use discrete pressures
and fluxes associated with cell and face centroids, respectively, and consequently resemble
finite-difference methods.
If we write a linear pressure field in the form p = x · a + b for a constant vector a
and scalar b, the corresponding Darcy velocity is v = −K a . Let ni,k denote the area-
weighted normal vector to ik and ci,k be the vector pointing from the centroid of cell
i to the centroid of face ik , as seen in Figure 6.6. Using Darcy’s law and the notion of
a one-sided transmissibility T i,k (which will generally not be the same as the two-point
transmissibility), the flux and pressure drop can be related as follows,
vi,k = − a = Ti,k (pi − πi,k ) = −Ti,k ci,k · a .
ni,k K (6.23)
To get these relations in one of the local forms (6.18) and (6.19), we collect all vectors ci,k
and ni,k defined for cell i as rows in two matrices C i and N i . Because relation (6.23) is

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 6.4 The Mimetic Method 189

pk
pi k
i πi,k
ci,k ni,k

i,k

Figure 6.6 Two neighboring cells with quantities used to define mimetic methods.

required to hold for all linear pressure drops, i.e., for an arbitrary vector a , we see that the
matrices M i and T i must satisfy the following consistency conditions

MNK = C, N K = T C, (6.24)

where we have dropped the subscript i that identifies cell number. These general equations
give us (quite) large freedom in how to specify a specific discretization method: any method
in which the local inner product M i , or equivalently the local transmissibility matrix T i , is
positive definite and satisfies (6.24) will give a consistent, first-order accurate discretization.
In the rest of the section, we discuss various choices of valid inner products M i or reverse
inner products T i .

General Family of Inner Products

The original article [52] that first introduced the mimetic methods considered herein, dis-
cussed inner products for discrete velocities. However, as we have seen in previous chap-
ters, it is more common in the simulation of flow in porous media to consider intercell
fluxes as the primary unknowns. We therefore consider inner products of fluxes rather than
velocities in the following. The relation between the two is trivial: an inner product of
velocities becomes an inner product for fluxes by pre- and post-multiplying by the inverse
of the area of the corresponding cell faces. In other words, if A is a diagonal matrix with
element Ajj equal the area of the j -th face, the flux inner product M flux is related to the
velocity inner product M vel through

M flux = A−1 M vel A−1 . (6.25)

To simplify the derivation of valid solutions, we start by stating a geometrical property that
relates C and N as follows (for a proof, see [52]):

C T N = V = diag(|i |). (6.26)

Next, we multiply the left matrix equation in (6.24) by K −1 C T N

C (K −1 C T N ) = (MNK) (K −1 C T N ) = MNC T N = MNV ,

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 190 Consistent Discretizations on Polyhedral Grids

from which it follows that there exists a family of valid solution of the form
1
M= CK −1 C T + M 2, (6.27)
|i |
where M 2 is a matrix defined such that M 2 N = 0, i.e., any matrix whose rows lie in
the left nullspace of N T . Moreover, to make a sensible method, we must require that M
is symmetric positive definite. In other words, any symmetric and positive-definite inner
product that fulfills these requirements can be represented in two alternative compact forms,
1 ⊥T
M= CK −1 C T + Q⊥N S M QN
|i |
1
= CK −1 C T + P ⊥ ⊥
N SM P N . (6.28)
|i |

Here, S M denotes any symmetric positive-definite matrix, Q⊥ N is an orthonormal basis for

the left nullspace of N T , and P ⊥
N is the nullspace projection I −Q T
N QN , in which QN is a
basis for spaces spanned by the columns of N . Similarly, we can derive a closed expression
for the inverse inner product T by multiplying the right matrix equation in (6.24) by V −1
from the left and by V from the right and using the identity (6.26),
T
T C = V −1 (N K) V = V −1 (N K) (C T N ) = V −1 (N KN T ) C.

By the same argument as for M, mutatis mutandis, we obtain the following family of
inverse inner products,
1 ⊥T
T = N KN T + Q⊥C S T QC
|i |
1
= N KN T + P ⊥ ⊥
C ST P C, (6.29)
|i |

where Q⊥ ⊥
C is an orthonormal basis for the left nullspace of C and P C = I − QC QC is
T T

the corresponding nullspace projection.

The matrices M and T in (6.28) and (6.29) are evidently symmetric, so we only need to
prove that they are positive definite. We start by writing the inner product as M = M 1 +M 2 ,
and observe that each of these matrices are positive semi-definite. Hence, our result follows
if we can prove that this implies that M is positive definite. Let z be an arbitrary nonzero
vector, which we split uniquely as z = N x + y, where x lies in the column space of N and
y lies in the left nullspace of N T . If y is zero, we have

zT Mz = x T N T M 1 N x
= |i |−1 x T (N T C) K −1 (C T N ) x = |i | x T K −1 x > 0

because K −1 is a positive definite matrix. If y is nonzero, we have

zT Mz = zT M 1 z + y T M 2 y > 0

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 6.4 The Mimetic Method 191

because zT M 1 z ≥ 0 and y T M 2 y > 0. An analogous argument holds for the matrix T ,

and hence we have proved that (6.28) and (6.29) give a well-defined family of consistent
discretizations.
So far, we have not put any restrictions on the matrices S M and S T that will determine
the specific methods, except for requiring that they should be positive definite. In addition,
these matrices should mimic the scaling properties of the continuous equation, which is
invariant under affine transformations of space and permeability,

x → σ x and K → σ T Kσ, (6.30)

To motivate how we should choose the matrices S M and S T to mimic these scaling prop-
erties, we look at a simple 1D example:

Example 6.4.1 Consider the grid cell x ∈ [−1,1], for which N = C = [1, − 1]T and
Q⊥ ⊥
N = QC =
√1 [1,1]T . Hence
2



1 1 1 1 1
M= [1, − 1] + SM [1,1],
2 −1 K 2 1



1 1 1 1
T = K [1, − 1] + S [1,1]. (6.31)
2 −1 2 1
The structure of the inner product should be invariant under scaling of K and thus we can
write the inner product as a one-parameter family of the form



1 1 −1 2 1 1
M= + ,
2K −1 1 t 1 1



K 1 −1 t 1 1
T = + . (6.32)
2 −1 1 2 1 1
Having established a plausible way of scaling the inner products, we are now in a position to
go through various specific choices that are implemented in the mimetic module of MRST
and look at correspondence between these methods and the standard two-point method, the
lowest-order RT0 mixed method, and the MPFA-O method. Our discussion follows [192].

General Parametric Family

Motivated by Example 6.4.1, we propose to choose the matrix S T as the diagonal of the
first matrix term in (6.29) so that these two terms scale similarly under transformations of
the type (6.30). Using this definition of S T and allowing that M should be equal to T −1
suggests the following general family of inner products that only differ in the constant in
front of the second (regularization) matrix term:
1 |i | ⊥
M= CK −1 C T + P N diag(N KN T )−1 P ⊥
N,
|i | t
1  
T = N KN T + t P ⊥
C diag(N KN T
)P ⊥
C . (6.33)
|i |

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 192 Consistent Discretizations on Polyhedral Grids

In MRST, this family of inner products is called 'ip_qfamily', and the parameter t is
supplied in a separate option:
S = computeMimeticIP(G, rock, 'InnerProduct', 'ip_qfamily', 'qparam', t);

As we will see shortly, mimetic inner products that reduce to the standard TPFA method or
the RT0 mixed finite-element method on simple grids are members of this family.

Two-Point Type Methods

A requirement of the two-point method is that the transmissibility matrix T (and hence
also the matrix M of the inner product) should be diagonal. Looking at (6.24), we see
that this is only possible if the vectors KNik and cik are parallel, which is the exact same
condition for K-orthogonality that we argued was sufficient to guarantee consistency of the
two-point method on page 176. An explicit expression for the diagonal entries of the two-
point method on K-orthogonal grids has already been given in (6.20). Strictly speaking,
this relation does not always yield a positive value for the two-point transmissibility on
any grid. For instance, for corner-point grids it is normal to define the face centroids as the
arithmetic mean of the associated corner-point nodes and the cell centroids as the arithmetic
mean of the centroids of the top and bottom cell faces, which for most grids will guarantee
a positive transmissibility.
The extension of the two-point to non-orthogonal grids is not unique. Here, we present
a mimetic method that coincides with the two-point methods in this method’s region of
validity and at the same time gives a valid mimetic inner product and hence a consis-
tent discretization for all types of grids and permeability tensors. One advantage of this
method, compared with multipoint flux-approximation methods, is that the implementation
is simpler for general unstructured grids. The most instructive way to introduce the general
method is to look at a simple example that will motivate the general construction.

Example 6.4.2 We consider the grid cell [−1,1] × [−1,1] in 2D and calculate the compo-
nents that make up the inverse inner product T for a diagonal and a full permeability tensor
(K1 and K2 , respectively) and compare with the corresponding two-point discretization. We
start by setting up the permeability and the geometric properties of the cell
K1 = eye(2); K2 = [1 .5; .5 1];
C = [-1 0; 1 0; 0 -1; 0 1]; N = 2*C; vol = 4;

Using the definition (6.20), we see that the transmissibility matrix resulting from the stan-
dard two-point discretization can be computed as:
T = diag(diag(N*K*C')./sum(C.*C,2))

The result is T=diag([2 2 2 2]) for both permeability tensors, which clearly demonstrates
that the scheme is not consistent for K2 . To construct the inverse inner product, we start by
computing the nullspace projection

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 6.4 The Mimetic Method 193

Q = orth(C);
P = eye(size(C,1)) - Q*Q';

We saw earlier that as a simple means of providing a positive definite matrix S T that scales
similarly to the first term in the definition of T in (6.29), we could choose S T as a multiple
of the diagonal of this matrix:
W = (N * K * N') ./ vol;
St = diag(diag(W));

Collecting our terms, we see that the inverse inner product will be made up of the following
two terms for the case with the diagonal permeability tensor:
W = 1 -1 0 0 P*St*P = 0.5 0.5 0 0
-1 1 0 0 0.5 0.5 0 0
0 0 1 -1 0 0 0.5 0.5
0 0 -1 1 0 0 0.5 0.5

If we now define the inverse inner product as

T = W + 2*P*St*P

the resulting method will coincide with the diagonal transmissibility matrix for diagonal
permeability and give a full matrix for the tensor permeability,
T1 = 2.0 0 0 0 T2 = 2.0 0 0.5 -0.5
0 2.0 0 0 0 2.0 -0.5 0.5
0 0 2.0 0 0.5 -0.5 2.0 0
0 0 0 2.0 -0.5 0.5 0 2.0

Altogether, we have derived a discrete inner product that generalizes the two-point method
to full tensor permeabilities on 2D Cartesian grids and gives a consistent discretization
also when the grid is not K-orthogonal.
Motivated by this example, we define a quasi-two-point inner product that simplifies to
the standard TPFA method on Cartesian grids with diagonal permeability tensor:
1 |i | ⊥
M= CK −1 C T + P N diag(N KN T )−1 P ⊥
N,
|i | 2
1  
T = N KN T + 2 P ⊥ T ⊥
C diag(N KN )P C . (6.34)
|i |
The observant reader will notice that this is a special case for t = 2 of the general family
(6.33) of inner products. In MRST, this inner product is constructed by
S = computeMimeticIP(G, rock, 'InnerProduct', 'ip_quasitpf');

For completeness, the mimetic module also supplies a standard, diagonal two-point inner
product that is not generally consistent. This is invoked by:
S = computeMimeticIP(G, rock, 'InnerProduct', 'ip_tpf');

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 194 Consistent Discretizations on Polyhedral Grids

Raviart–Thomas-Type Inner Product

To compute the mixed finite-element inner product on cells that are not simplexes or
hexahedrons aligned with the coordinate axes, the standard approach is to map the cell
back to a unit reference cell on which the mixed basis functions are defined and perform
the integration there. For a general hexahedral cell in 3D, the resulting integral would
conceptually look something like
 
x ) d x =
f ( x (ξ,η,ζ )) |J (ξ,η,ζ )| dξ dηdζ,
f (
i [0,1]3

where J = ∂(x,y,z)/∂(ξ,η,ζ ) denotes the Jacobian matrix of the coordinate transforma-

tion. The determinant of J is generally nonlinear and it is therefore not possible to develop a
mimetic inner product that equals the lowest-order, RT0 inner product on general polygonal
or polyhedral cells, and at the same time is simpler to compute.
Instead, we develop an inner product that is equivalent to RT0 on grids that are orthogo-
nal and aligned with the principal axes of the permeability tensor. To motivate the definition
of this inner product, we first look at a simple example.

Example 6.4.3 The RT0 basis functions defined on the reference element in two spatial
dimensions read
       
 1−x  x  0  0
ψ1 = , ψ2 = , ψ3 = , ψ4 = .
0 0 1−y y

We compute the elements of the corresponding inner product matrix for a diagonal perme-
ability with unit entries, K = I . This entails careful treatment of a sticky point: whereas
the mixed inner product involves velocities, the mimetic inner product involves fluxes that
are considered positive when going out of a cell and negative when going into the cell. To
get comparative results, we thus reverse the sign of ψ 1 and ψ 3 . Using symmetry and the
fact that ψ i · ψ k = 0 if i = 1,2 and k = 3,4, we only have to compute two integrals:
 1 1  1
ψ 1 · ψ 1 dxdy = (x − 1)2 dx = 1
3
0 0 0
 1 1  1
ψ 1 · ψ 2 dxdy = (x − 1)x dx = − 16 .
0 0 0

This means that the RT0 inner product in 2D reads

⎡ ⎤
1
3 − 16 0 0
⎢ − 16 1
0 ⎥
⎢ 0 ⎥
M=⎢ 3
⎥.
⎣ 0 0 1
3 − 16⎦
0 0 − 16 1
3

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 6.4 The Mimetic Method 195

Similarly as in Example 6.4.2, we can compute the two matrices used to form the mimetic
inner product
C = .5*[-1 0; 1 0; 0 -1; 0 1]; N = 2*C; vol = 1; K = eye(2);
Q = orth(N);
P = eye(size(C,1)) - Q*Q';
Sm = diag(1 ./ diag(N*K*N'))*vol;
M1 = C*(K\C')./vol
M2 = P*Sm*P

The result of this computation is

M1 = 0.25 -0.25 0 0 M2 = 0.50 0.50 0 0
-0.25 0.25 0 0 0.50 0.50 0 0
0 0 0.25 -0.25 0 0 0.50 0.50
0 0 -0.25 0.25 0 0 0.50 0.50
1
from which it follows that M 2 should be scaled by 6 if the mimetic inner product is to
coincide with the RT0 mixed inner product.
For a general grid, we define a quasi-RT0 inner product that simplifies to the standard
RT0 inner product on orthogonal grids with diagonal permeability tensors, as well as on all
other cases that can be transformed to such a grid by an affine transformation of the form
(6.30). The quasi-RT0 inner product reads
1 |i | ⊥
M= CK −1 C T + P N diag(N KN T )−1 P ⊥
N,
|i | 6
1  
T = N KN T + 6 P ⊥ T ⊥
C diag(N KN )P C . (6.35)
|i |
and is a special case of the general family (6.33) of inner products for t = 6. In MRST, this
inner product is constructed as follows
S = computeMimeticIP(G, rock, 'InnerProduct', 'ip_quasitpf');

For completeness, the mimetic module also supplies a standard RT0 inner product,
'ip_rt', that is only valid on Cartesian grids.

Default Inner Product in MRST

For historical reasons, the default discretization in the mimetic module of MRST corre-
sponds to a mimetic method with half-transmissibilities defined by the following expres-
sion,
1  6 
T = N KN T + tr(K)A (I − QQT ) A . (6.36)
|i | d
Here, A is the diagonal matrix containing face areas and Q is an orthogonal basis for the
range of AC. The inner product, referred to as 'ip_simple', was inspired by [52] and
introduced in [6] to resemble the RT0 inner product; they are equal for scalar permeability

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 196 Consistent Discretizations on Polyhedral Grids

on orthogonal grids, which can be verified by inspection. Because the inner product is
based on velocities, it involves pre- and post-multiplication of (inverse) face areas and
is in this sense different from the general class of inner products discussed in (6.33). It is
nevertheless possible to show that the eigenspace corresponding to the nonzero eigenvalues
of the second matrix term is equal to the nullspace for C.

Local-Flux Mimetic Method

Another large class of consistent discretization methods that have received a lot of attention
in recent years is the multipoint flux-approximation (MPFA) schemes [9, 94, 7]. Discussing
different variants of this method is beyond the scope of the current presentation, mainly
because efficient implementation of a general class of MPFA schemes requires some addi-
tional mappings that are not yet part of the standard grid structure outlined in Section
3.4. These mappings are available and have been used in some of the add-on modules for
geomechanics, and it should therefore not be too difficult for the competent and interested
reader to also implement other MPFA type schemes.
Having said this, there is a module mpfa in MRST that gives a simple implementation of
the MPFA-O method. In this implementation, we have utilized the fact that some variants of
the method can be formulated as a mimetic method. This was first done by [281, 199] and
is called the local-flux mimetic formulation of the MPFA method. In this approach, each
face in the grid is subdivided into a set of subfaces, one subface for each of the nodes that
make up the face. The inner product of the local-flux mimetic method gives exact results
for linear flow and is block diagonal with respect to the faces corresponding to each node
of the cell, but it is not symmetric. The block-diagonal property makes it possible to reduce
the system into a cell-centered discretization for the cell pressures. This naturally leads to a
method for calculating the MPFA transmissibilities. The crucial point is to have the corner
geometry in the grid structure and handle the problems with corners that do not have three
unique half-faces associated.
The local-flux mimetic formulation of the MPFA-O method can be used with a calling
sequence that is similar to that of the TPFA and the mimetic methods:

mrstModule add mpfa;

hT = computeMultiPointTrans(G, rock);
state = incompMPFA(state, G, hT, fluid)

By default, computeMultiPointTrans uses MATLAB to invert a sequence of small

matrices that arise in the discretization. This operation is relatively slow and the rou-
tine therefore takes an optional parameter 'invertBlocks' that can be set equal to
'MEX' to invoke C-accelerated routines for the inversion of the small matrix blocks.
This option may not work if MRST has not been properly set up to automatically compile
C-code on your computer.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 6.5 Monotonicity 197

6.5 Monotonicity
The main motivation for developing consistent discretizations is to get schemes that are
convergent for rough grids and full-tensor permeabilities and less influenced by grid ori-
entation errors. All consistent methods discussed in this chapter fulfill this requirement.
On the other hand, these methods are all linear and not guaranteed to satisfy discrete
counterparts of the maximum principle fulfilled by the continuous solution. For the Poisson
equation ∇ · K∇p = q, this principle implies that if there is a single source within the
domain, the pressure will decrease monotonically from the source toward the boundary.
The discrete solution should also follow a similar principle, in which case we say that the
underlying scheme is monotone.
If a scheme is not monotone, the discrete solutions may contain oscillations that are not
present in the underlying physical system, particularly for cases with full-tensor permeabil-
ities and large anisotropy ratios and/or high aspect ratio grids. While such oscillations are
not very serious for single-phase, incompressible flow, they may have a significant impact
on multiphase flows, which can be highly sensitive to small pressure variations; think of an
oil–gas system operating slightly above the bubble point, for which artificial gas is liberated
if the pressure falls below the bubble point; see Chapter 11. A classical result says that a
discretization scheme is monotone if it generates an M-matrix, i.e., a matrix whose off-
diagonal elements are less than or equal zero and whose eigenvalues have positive real
parts. Two-point schemes are strictly monotone if all transmissibilities are nonnegative.
Multipoint schemes, on the other hand, do not necessarily give an M-matrix, and extensive
research has gone into investigate their monotonicity properties [236, 234, 235, 12, 157]
and proving convergence [308, 26, 103, 104, 96, 314, 10, 167, 11]. In 2D, this has lead
to the somewhat negative result [157]: no linear nine-point control volume method can be
constructed for quadrilateral grids in 2D that is exact for linear solutions while remaining
monotone for general problems..
For brevity, we will only present an example that illustrates lack of monotonicity for lin-
ear, consistent schemes. We refer the interested reader to [235] and references therein for a
thorough discussion of the underlying mathematical properties that cause these unphysical
oscillations.

Example 6.5.1 (Example 2 from [192]) To illustrate the different behavior of TPFA and
linear, consistent discretizations, we consider a 2D domain of size n × n, where n is
also the number of grid cells in each spatial direction. To generate a non-orthogonal
grid, we use the twister routine, which you have encountered multiple times already.
This routine normalizes all coordinates in a rectilinear grid to the interval [0,1], then
perturbs all interior points by adding 0.03 sin(π x) sin(3π(y − 1/2)) to the x-coordinates
and subtracting the same value from the y-coordinate before transforming back to the
original domain. This creates a non-orthogonal, but smoothly varying, logically Cartesian
grid. The domain has Dirichlet boundary conditions on the left and right-hand sides, with
values one and zero, respectively, and zero Neumann conditions at the top and bottom.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 198 Consistent Discretizations on Polyhedral Grids

TPFA Mimetic MPFA-O

1.8

1.6

1.4

1.2

Figure 6.7 Grid orientation effects for a flow described by a linear pressure drop from the left to
the right boundary. Approximate solutions computed by three different incompressible solvers from
the incomp, mimetic, and mpfa modules, respectively. The permeability field is homogeneous with
anisotropy ratio 1:1,000 aligned with the grid.

An important design principle for consistent discretizations is that they, unlike the TPFA
method, should reproduce flows that result from a linear pressure drop exactly on grids that
are not K-orthogonal. We therefore start by considering a diagonal permeability tensor with
anisotropy ratio 1:1,000 aligned with the x-axis. Figure 6.7 compares the discrete solutions
computed TPFA, the local-flux mimetic (MPFA-O) method, and the mimetic method with
inner product 'ip_simple' on a 101 × 101 grid. Whereas the mimetic and MPFA-O
schemes produce the expected result, the pressure and velocity solutions computed by TPFA
show significant grid-orientation effects and are obviously wrong.
To illustrate non-monotonic behavior, we repeat the experiment with the tensor rotated
by π/6 on a grid with 21 × 21 cells; see Figure 6.8. The solution computed by TPFA is
monotone, but fails to capture the correct flow pattern in which streamlines starting near
the upper-left corner should exit near the lower-right corner. Judging from the color plots,
the pressure values computed by the mimetic and MPFA-O methods appear to be correct
and have no discernible oscillations. However, both the streamline distribution and the
velocity vectors plotted for cells along the midsection of the grid clearly show that the
resulting velocity fields are highly oscillatory, in particular for the mimetic method. If we
change the inner product from its default value ('ip_simple') to, e.g., the quasi-two-point
inner product ('ip_quasitpf'), the solution is qualitatively the same as for the MFPA-O
method.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 6.6 Discussion 199

TPFA Mimetic MPFA-O

1.8

1.6

1.4

1.2

Figure 6.8 Monotonicity illustrated with the same setup as in Figure 6.8, but with the anisotropy ratio
1:1,000, making an angle π/6 with the x-direction.

6.6 Discussion
In this chapter, we have shown that TPFA is only consistent on K-orthogonal grids and
hence may give significant grid orientation errors for tensor permeabilities and general
cell geometries. On the other hand, the method is monotone and coercive (sum of fluxes
is always positive), has a rather sparse stencil, and is hence computationally efficient. It
is also possible to reduce grid orientation errors in the TPFA method by modifying the
stencil. For tetrahedral grids satisfying the Delaunay criterion, one can design consistent
two-point schemes by computing transmissibilities using the circumcenter rather than the
cell centroid. You can also reduce numerical artifacts for general cell geometries by either
optimizing the point used to compute intercell transmissibilities and/or use fluxes from a
consistent method to adjust the transmissibilities. As a general precaution, I would recom-
mend that you compute solutions using both a two-point and a consistent method to get an
idea of the amount of grid orientation when working with stratigraphic or other unstructured
representations of complex geology. This applies particularly to multiphase flow, for which
you can get a good idea of the grid orientation errors by solving a representative single-
phase problem.
Which consistent method should you choose amongst those implemented in MRST?
The consistent MPFA method is cell-centered like the TPFA method and thus has the
same number of unknowns, but the resulting stencil is denser and the method is therefore
significantly less computationally efficient. Compared with mimetic methods, MPFA is
quite involved to implement on general polyhedral meshes, and the method is not very

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 200 Consistent Discretizations on Polyhedral Grids

robust on very skewed and deformed grids and for high aspect ratios or strong anisotropy
[267]. In my experience, the mimetic method seems more robust and would therefore be
my favorite choice, despite a higher number of unknowns.
There are also many other consistent methods than those discussed herein. Virtual ele-
ment methods [38, 39] can be seen as a variational analog of mimetic methods and have
proved to be versatile and robust for discretizing diffusion, dispersion, and mechanical
effects. Unfortunately, the methods are not locally conservative when applied to Poisson-
type flow equations, and thus require postprocessing to compute useful fluxes. MRST
offers an implementation in the recent vem module [168], but this is currently primarily
of academic interest when it comes to flow equations.
To develop schemes that are both consistent and monotone, several authors have started
looking into nonlinear two-point methods [179, 198, 226, 272]. In these methods, the flux
approximation is written
vi,k = Ti (p)p
 i − Tk (p)p
 k,
where the transmissibilities depend on one or more pressure values so that the method
becomes nonlinear. The disadvantage is that you have to solve a nonlinear system, but
this additional cost may not present a significant drawback when the method is used for
compressible and multiphase flow, where you need to use a nonlinear solver to account
for other nonlinear couplings in the model equations, as we will see later in the book. At
the time of writing, MRST only contains a preliminar implementation of nonlinear TPFA
schemes. You can find more detailed comparisons of all the methods discussed previously
in [169].
Another example of a recent and promising method is the vertex approximation gradient
(VAG) method [105, 106, 267]. Unlike other nodal methods, this method has unknowns
associated with both vertices and cell centers and relies on a local triangulation of the
grid cells to enable simple extension of finite-elements to general cell geometries. The
scheme can also be written on finite-volume form, introducing explicit cell-vertex fluxes
that connect each cell to its vertex. This method is not yet implemented in MRST.

Computer exercises
6.6.1 The two tutorials mimeticExample2 and mimeticExample3 compare the TPFA
method implemented in the incomp module with its mimetic counterpart. Do the
methods produce the same results if you perturb the grid cells using twister or
specify an anisotropic permeability field? What happens if you use a different inner
product?
6.6.2 Redo the comparison of structured and unstructured grids in Section 5.4.3 using a
mimetic solver and possibly the MPFA-O scheme. Try to also compare computa-
tional efficiency with the TPFA scheme in terms of number of unknowns, number
of nonzero elements in the matrix, and condition numbers.
6.6.3 A standard method to construct analytical solutions to the Laplace equation in 2D is
to write the unknown as the real or imaginary part of an complex analytic function

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 6.6 Discussion 201

f (z). We choose the function f (z) = (z + 12 z2 ) and set our analytical solution to be
pa (x,y) = Im(f (x + iy)) = y + xy. By prescribing pa (x,y) along the boundary,
we obtain a Dirichlet problem with known solution. Use this technique to study the
convergence of the TPFA method and at least two different consistent methods on
a sequence of rough non-Cartesian grids of increasing resolution. Try to compare
grids with cells that are always non-Cartesian at all resolutions with grids in which
the cells are non-Cartesian at coarse resolutions but gradually approach a Cartesian
shape as the grid is refined.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 7
Compressible Flow and Rapid Prototyping

In previous chapters we have outlined and explained in detail how to discretize and solve
incompressible flow problems. This chapter will teach you how to discretize the basic equa-
tions for single-phase, compressible flow by use of the discrete differential and averaging
operators that were introduced in Section 4.4.2. As briefly shown in Examples 4.4.2 and
4.4.3 in the same section, these discrete operators enable you to implement discretized
flow equations in a compact form similar to the continuous mathematical description.
Use of automatic differentiation (see Appendix A.5 for more details) then ensures that
no analytical derivatives have to be programmed explicitly as long as the discrete flow
equations and constitutive relationships are implemented as a sequence of algebraic opera-
tions. MRST makes it possible to combine discrete operators and automatic differentiation
with a flexible grid structure, a highly vectorized and interactive scripting language, and a
powerful graphical environment. This is in my opinion the main reason why the software
has proved to be an efficient tool for developing new computational methods and workflow
tools. In this chapter, I try to substantiate this claim by showing several examples of rapid
prototyping. We first develop a compact and transparent solver for compressible flow and
then extend the basic single-phase model to include pressure-dependent viscosity, non-
Newtonian fluid behavior, and temperature effects. As usual, you can find complete scripts
for all examples in a subdirectory (ad-1ph) of the book module.

7.1 Implicit Discretization

As our basic model, we consider the single-phase continuity equation,

∂ K
(φρ) + ∇ · (ρ v) = q, v = − (∇p − gρ∇z) . (7.1)
∂t μ

The primary unknown is usually the fluid pressure p. Additional equations are supplied
to provide relations between p and the other quantities in the equation, e.g., by using
an equation of state to relate fluid density to pressure ρ = ρ(p), specifying porosity as

202

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 7.1 Implicit Discretization 203

function of pressure φ(p) through a compressibility factor, and so on; see the discussion in
Section 4.2. Notice also that q is defined slightly differently in (7.1) than in (4.5).
Using the discrete operators introduced in Section 4.4.2, the basic implicit discretization
of (7.1) reads

(φρ)n+1 − (φρ)n
+ div(ρv)n+1 = q n+1, (7.2a)
t n
K  
v n+1 = − n+1 grad(pn+1 ) − gρ n+1 grad(z) . (7.2b)
μ
Here, φ ∈ Rnc denotes the vector with one porosity value per cell, v is the vector of fluxes
per face, and so on. The superscript refers to discrete times at which one wishes to compute
the unknown reservoir states and t denotes the distance between two such consecutive
points in time.
In many cases of practical interest it is possible to simplify (7.2). For instance, if the fluid
is only slightly compressible, several terms can be neglected so that the nonlinear equation
reduces to a linear equation in the unknown pressure p n+1 , which we can write on residual
form as
p n+1 − pn 1 
n
− div K grad(p n+1 ) − q n = 0. (7.3)
t ct μφ
The assumption of slight compressibility is not always applicable and for generality
we assume that φ and ρ depend nonlinearly on p so that (7.2) gives rise to a nonlinear
system of equations that needs to be solved in each time step. As we will see later in this
chapter, viscosity may also depend on pressure, flow velocity, and/or temperature, which
adds further nonlinearity to the system. If we now collect all the discrete equations, we can
write the resulting system of nonlinear equations in short vector form as

F (x n+1 ;x n ) = 0. (7.4)

Here, x n+1 is the vector of unknown state variables at the next time step and the vector of
current states x n can be seen as a parameter.
We will use the Newton–Raphson method to solve the nonlinear system (7.4): assume
that we have a guess x 0 and want to move this towards the correct solution, F (x) = 0. To
this end, we write x = x 0 + x, use a Taylor expansion for linearization, and solve for the
approximate increment δx

0 = F (x 0 + x) ≈ F (x 0 ) + ∇F (x 0 )δx.

This gives rise to an iterative scheme in which the approximate solution x i+1 in the (i + 1)-
th iteration is obtained from
dF i
(x )δx i+1 = −F (x i ), x i+1 ← x i + δx i+1 . (7.5)
dx

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 204 Compressible Flow and Rapid Prototyping

Here, J = dF /dx is the Jacobian matrix, while δx i+1 is referred to as the Newton
update at iteration number i + 1. Theoretically, the Newton process exhibits quadratic
convergence under certain smoothness and differentiability requirements on F . Obtain-
ing such convergence in practice, however, will crucially depend on having a sufficiently
accurate Jacobian matrix. For complex flow models, the computation of residual equations
typically requires evaluation of many constitutive laws that altogether make up complex
nonlinear dependencies. Analytical derivation and subsequent coding of the Jacobian can
therefore be very time-consuming and prone to human errors. Fortunately, the computation
of the Jacobian matrix can in almost all cases be broken down to nested differentiation of
elementary operations and functions and is therefore a good candidate for automation using
automatic differentiation. This will add an extra computational overhead to your code, but
in most cases the increased CPU time is completely offset by the shorter time it takes you
to develop a proof-of-concept code. Likewise, unless your model problem is very small,
the dominant computational cost of solving a nonlinear PDE comes from the linear solver
called within each Newton iteration.
The idea of using automatic differentiation to develop reservoir simulators is not new.
This technique was introduced in an early version of the commercial Intersect simulator
[80], but has mainly been pioneered through a reimplementation of the GPRS research
simulator [58]. The new simulator, called AD-GPRS, is primarily based on fully implicit
formulations [303, 325, 302], in which independent variables and residual equations are
AD structures implemented using ADETL, a library for forward-mode AD realized by
expression templates in C++ [323, 322]. This way, the Jacobi matrices needed in the
nonlinear Newton-type iterations can be constructed from implicitly computed deriva-
tives when evaluating the residual equations. In [185], the authors discuss how to use the
alternative backward-mode differentiation to improve computational efficiency. Automatic
differentiation is also used in the open-source Flow simulator from the Open Porous Media
(OPM) initiative. OPM Flow can be considered as a C++ sibling of MRST, which originally
used a similar vector-oriented AD library. This has later been replaced by a localized, cell-
based AD library for improved efficiency.

7.2 A Simulator Based on Automatic Differentiation

We will now present step-by-step how you can use the AD library in MRST to implement
an implicit solver for the compressible, single-phase continuity equation (7.1). In particular,
we revisit the discrete spatial differentiation operators from Section 4.4.2 and introduce
additional discrete averaging operators that together enable us to write the discretized
equations in an abstract residual form that resembles the semi-continuous form of
the implicit discretization in (7.2). Starting from this residual form, it is relatively
simple to obtain a linearization using automatic differentiation and set up a Newton
iteration.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 7.2 A Simulator Based on Automatic Differentiation 205

7.2.1 Model Setup and Initial State

For simplicity, we consider a homogeneous box model:
[nx,ny,nz] = deal( 10, 10, 10);
[Lx,Ly,Lz] = deal(200, 200, 50); 0

10

G = cartGrid([nx, ny, nz], [Lx, Ly, Lz]); 20

30

G = computeGeometry(G); 40

50
200

150 200
150

rock = makeRock(G, 30*milli*darcy, 0.3); 100

50

0 0
50
100

Beyond this point, our implementation is agnostic to details about the grid, except when
we specify well positions on page 206, which would typically involve more code lines for
a complex corner-point model like the Norne and SAIGUP models discussed in Sections
3.3.1 and 3.5.1.
We assume constant rock compressibility cr . Accordingly, the pore volume pv obeys the
differential equation1 cr pv = dpv/dp or
pv(p) = pvr ecr (p−pr ), (7.6)
where pvr is the pore volume at reference pressure pr . To define the relation between pore
volume and pressure, we use an anonymous function:
600.1

cr = 1e-6/barsa;
600.05
p_r = 200*barsa;
pv_r = poreVolume(G, rock); 600

599.95

pv = @(p) pv_r .* exp( cr * (p - p_r) ); 599.9

100 150 200 250 300

The fluid is assumed to have constant viscosity μ = 5 cP, and as for the rock, we assume
constant fluid compressibility c, resulting in the differential equation cρ = dρ/dp for fluid
density. Accordingly,
ρ(p) = ρr ec(p−pr ), (7.7)
where ρr is the density at reference pressure pr . With this set, we can define the equation
of state for the fluid:
940

mu = 5*centi*poise; 920

900

c = 1e-3/barsa; 880

860

rho_r = 850*kilogram/meter^3; 840

820
rhoS = 750*kilogram/meter^3; 800

780
rho = @(p) rho_r .* exp( c * (p - p_r) ); 760
100 150 200 250 300

The assumption of constant compressibility will only hold for a limited range of tem-
peratures. Moreover, surface conditions are not inside the validity range of the constant

1 To highlight the close correspondence between the computer code and the mathematical equation, we here deliberately violate
the advice to never use a compound symbol to denote a single mathematical quantity.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 206 Compressible Flow and Rapid Prototyping

show = true(G.cells.num,1);
cellInx = sub2ind(G.cartDims, ...
[I-1; I-1; I; I; I(1:2)-1], ...
[J ; J; J; J; nperf+[2;2]], ...
[K-1; K; K; K-1; K(1:2)-[0; 1]]);
show(cellInx) = false;
plotCellData(G,p_init/barsa, show,...
'EdgeColor','k');
plotWell(G,W, 'height',10);
view(-125,20), camproj perspective

Figure 7.1 Model with initial pressure and single horizontal well.

compressibility assumption. We therefore set the fluid density ρS at surface conditions

separately because we will need it later to evaluate surface volume rate in our model of the
well, which consists of a horizontal wellbore perforated in eight cells:
nperf = 8;
I = repmat(2, [nperf, 1]);
J = (1:nperf).'+1;
K = repmat(5, [nperf, 1]);
cellInx = sub2ind(G.cartDims, I, J, K);
W = addWell([ ], G, rock, cellInx, 'Name', 'producer', 'Dir', 'x');

Assuming the reservoir is initially at equilibrium implies that we must have dp/dz
= gρ(p). In our simple setup, this differential equation can be solved analytically, but
for demonstration purposes, we use one of MATLAB’s built-in ODE-solvers to compute
the hydrostatic distribution numerically, relative to a fixed datum point p(z0 ) = pr .
Without lack of generality, we set z0 = 0 since the reservoir geometry is defined relative to
this height:
gravity reset on, g = norm(gravity);
[z_0, z_max] = deal(0, max(G.cells.centroids(:,3)));
equil = ode23(@(z,p) g .* rho(p), [z_0, z_max], p_r);
p_init = reshape(deval(equil, G.cells.centroids(:,3)), [], 1);

This finishes the model setup, and at this stage we plot the reservoir with well and initial
pressure as shown in Figure 7.1.

7.2.2 Discrete Operators and Equations

We are now ready to discretize the model. As seen in Section 4.4.2, the discrete version
of the gradient operator maps from the set of cells to the set of faces. For a pressure field,
it computes the pressure difference between neighboring cells of each face. Likewise, the
discrete divergence operator is a linear mapping from the set of faces to the set of cells. For

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 7.2 A Simulator Based on Automatic Differentiation 207

a flux field, it sums the outward fluxes for each cell. The complete code needed to form the
grad and div operators has already been presented in Examples 4.4.2 and 4.4.3, but here
we repeat it in order to make the example more self-contained.
To define the discrete operators, we must first compute the map between interior faces
and cells
C = double(G.faces.neighbors);
C = C(all(C = 0, 2), :);

Exterior faces need not be included since they have zero flow, given our assumption of no-
flow boundary conditions. It now follows that grad(x) = x(C(: ,2)) − x(C(: ,1)) = Dx,
where D is a sparse matrix with values ±1 in columns C(i,2) and C(i,1) for row i. As a
linear mapping, the discrete div-function is simply the negative transpose of grad; this
follows from the discrete version of the Gauss–Green theorem, (4.58). In addition, we
define an averaging operator that for each face computes the arithmetic average of the
neighboring cells, which we will need to evaluate density values at grid faces:
n = size(C,1); ∂
∂x
D = sparse([(1:n)'; (1:n)'], C, ...
ones(n,1)*[-1 1], n, G.cells.num); ∂
∂y
grad = @(x) D*x;
div = @(x) -D'*x; ∂
avg = @(x) 0.5 * (x(C(:,1)) + x(C(:,2))); ∂z

This is all we need to define the spatial discretization for a homogeneous medium on a grid
with cubic cells. To make a generic spatial discretization that also can account for more
general cell geometries and heterogeneities, we must include transmissibilities. To this end,
we first compute one-sided transmissibilities Ti,j using the function computeTrans, which
was discussed in detail in Section 5.2, and then use harmonic averaging to obtain face-
−1 −1 −1
transmissibilities. That is, for neighboring cells i and j , we compute Tij = (Ti,j + Tj,i )
as in (4.52) on page 133.
hT = computeTrans(G, rock); % Half-transmissibilities
cf = G.cells.faces(:,1);
nf = G.faces.num;
T = 1 ./ accumarray(cf, 1 ./ hT, [nf, 1]); % Harmonic average
T = T(intInx); % Restricted to interior

Having defined the necessary discrete operators, we are in a position to use the basic
implicit discretization from (7.2). We start with Darcy’s law (7.2b),
T [f ] 
v[f ] = − grad(p) − g ρ a [f ] grad(z) , (7.8)
μ
where the density at the interface is evaluated using the arithmetic average

ρ a [f ] = 12 ρ[C1 (f )] + ρ[C2 (f )] . (7.9)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 208 Compressible Flow and Rapid Prototyping

Similarly, we can write the continuity equation for each cell c as

1  n+1  n 
φ(p)[c] ρ(p)[c] − φ(p)[c] ρ(p)[c] + div(ρ a v)[c] = 0. (7.10)
t

The two residual equations (7.8) and (7.10) are implemented as anonymous functions of
pressure:
gradz = grad(G.cells.centroids(:,3));
v = @(p) -(T/mu).*( grad(p) - g*avg(rho(p)).*gradz );

presEq = @(p,p0,dt) (1/dt)*(pv(p).*rho(p) - pv(p0).*rho(p0)) ...

+ div( avg(rho(p)).*v(p) );

In the code above, p0 is the pressure field at the previous time step (i.e., pn ), whereas p is the
pressure at the current time step (pn+1 ). Having defined the discrete expression for Darcy
fluxes, we can check that this is in agreement with our initial pressure field by computing
the magnitude of the flux, norm(v(p_init))*day. The result is 1.5 × 10−6 m3 /day, which
should convince us that the initial state of the reservoir is sufficiently close to equilibrium.

7.2.3 Well Model

The production well will appear as a source term in the pressure equation. We therefore
need to define an expression for flow rate in all cells the well is connected to the reservoir
(which we refer to as well connections). Inside the well, we assume instantaneous flow so
that the pressure drop is always hydrostatic. For a horizontal well, the hydrostatic term is
zero and could obviously be disregarded, but we include it for completeness and as a robust
precaution in case we later want to reuse the code with a different well path. Approximating
the fluid density in the well as constant, computed at bottom-hole pressure, the pressure
p c [w] in connection w of well Nw (w) is given by

p c [w] = pbh [Nw (w)] + g z[w] ρ(pbh [Nw (w)]), (7.11)

where z[w] is the vertical distance from the bottom-hole to the connection. We use the
standard Peaceman model introduced in Section 4.3.2 to relate the pressure at the well
connection to the average pressure inside the grid cell. Using the well-indices from W, the
mass flow-rate at connection c reads

ρ(p[Nc (w)]) 
q c [w] = WI[w] p c [w] − p[Nc (w)] , (7.12)
μ

where p[Nc (w)] is the pressure in cell Nc (w) containing connection w. In our code, this
model is implemented as follows:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 7.2 A Simulator Based on Automatic Differentiation 209

wc = W(1).cells; % connection grid cells

WI = W(1).WI; % well-indices pbh
dz = W(1).dZ; % depth relative to bottom-hole
qc

p_conn = @(bhp) bhp + g*dz.*rho(bhp); %connection pressures

q_conn = @(p, bhp) WI .* (rho(p(wc)) / mu) .* (p_conn(bhp) - p(wc));

We also include the total volumetric well-rate at surface conditions as a free variable. This
is simply given by summing all mass well-rates and dividing by the surface density:
rateEq = @(p, bhp, qS) qS-sum(q_conn(p, bhp))/rhoS;

With free variables p, bhp, and qS, we lack exactly one equation to close the system. This
equation should account for boundary conditions in the form of a well control. Here, we
choose to control the well by specifying a fixed bottom-hole pressure
ctrlEq = @(bhp) bhp-100*barsa;

7.2.4 The Simulation Loop

What now remains is to set up a simulation loop that will evolve the transient pressure. We
start by initializing the AD variables. For clarity, we append _ad to all variable names to
distinguish them from doubles. The initial bottom-hole pressure is set to the corresponding
grid-cell pressure.
[p_ad, bhp_ad, qS_ad] = initVariablesADI(p_init, p_init(wc(1)), 0);

This gives the following AD pairs that make up the unknowns in our system:
p_ad = ADI Properties: bhp_ad = ADI Properties: qS_ad = ADI Properties:
val: [1000x1 double] val: 2.0188e+07 val: 0
jac: {[1000x1000 double] jac: {[1x1000 double] jac: {[1x1000 double]
[1000x1 double] [1] [0]
[1000x1 double]} [0]} [1]}

∂p ∂p ∂p ∂pbh ∂pbh ∂pbh ∂qs ∂qs ∂qs

≡I ≡0 ≡0 ≡0 ≡0 qs
∂p ∂qs ∂pbh ∂p ∂qs ∂pbh ∂p ∂ ∂pbh

To solve the global flow problem, we must stack all the equations into one big system,
compute the corresponding Jacobian, and perform a Newton update. We therefore set
indices for easy access to individual variables
[p_ad, bhp_ad, qS_ad] = initVariablesADI(p_init, p_init(wc(1)), 0);
nc = G.cells.num;
[pIx, bhpIx, qSIx] = deal(1:nc, nc+1, nc+2);

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 210 Compressible Flow and Rapid Prototyping

Next, we set parameters to control the time steps in the simulation and the iterations in the
Newton solver:

[numSteps, totTime] = deal(52, 365*day); % time-steps/ total simulation time

[tol, maxits] = deal(1e-5, 10) % Newton tolerance / maximum Newton its
dt = totTime / numSteps;

Simulation results from all time steps are stored in a structure sol. For efficiency, this
structure is preallocated and initialized so that the first entry is the initial state of the
reservoir:

sol = repmat(struct('time',[], 'pressure',[], 'bhp',[], 'qS',[]), [numSteps+1, 1]);

sol(1) = struct('time', 0, 'pressure', double(p_ad), ...
'bhp', double(bhp_ad), 'qS', double(qS_ad));

We now have all we need to set up the time-stepping algorithm, which consists of an outer
and an inner loop. The outer loop updates the time step, advances the solution one step
forward in time, and stores the result in the sol structure. This procedure is repeated until
we reach the desired final time:

t = 0; step = 0;
while t < totTime,
t = t + dt; step = step + 1;
fprintf('\nTime step %d: Time %.2f -> %.2f days\n', ...
step, convertTo(t - dt, day), convertTo(t, day));
% Newton loop
[resNorm, nit] = deal(1e99, 0);
p0 = double(p_ad); % Previous step pressure
while (resNorm > tol) && (nit <= maxits)
: % Newton update
:
resNorm = norm(res);
nit = nit + 1;
fprintf(' Iteration %3d: Res = %.4e\n', nit, resNorm);
end
if nit > maxits, error('Newton solves did not converge')
else % store solution
sol(step+1) = struct('time', t, 'pressure', double(p_ad), ...
'bhp', double(bhp_ad), 'qS', double(qS_ad));
end
end

The inner loop performs the Newton iteration by computing and assembling the Jacobian
of the global system and solving the linearized residual equation to compute an iterative

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 7.2 A Simulator Based on Automatic Differentiation 211

update. The first step to this end is to evaluate the residual for the flow pressure equation
and add source terms from wells:
eq1 = presEq(p_ad, p0, dt);
eq1(wc) = eq1(wc) - q_conn(p_ad, bhp_ad);

Most of the lines we have implemented so far are fairly standard, except perhaps for the
definition of the residual equations as anonymous functions. Equivalent statements can be
found in almost any computer program solving this type of time-dependent equation by
an implicit method. Now, however, comes what is normally the tricky part: linearization
of the equations that make up the whole model and assembly of the resulting Jacobian
matrices to generate the Jacobian for the full system. And here you have the magic of
automatic differentiation: you do not have to do this at all! The computer code necessary
to evaluate all Jacobians has been defined implicitly by the functions in the AD library
in MRST, which overloads the elementary operators used to define the residual equations.
An example of a complete calling sequence for a simple calculation is shown in Figure
A.7 on page 625. The sequence of operations we use to compute the residual equations is
obviously more complex than this example, but the operators used are in fact only the three
elementary operators plus, minus, and multiply applied to scalars, vectors, and matrices, as
well as element-wise division by a scalar and evaluation of exponential functions. When
the residuals are evaluated by use of the anonymous functions defined in Sections 7.2.2 and
7.2.3, the AD library also evaluates the derivatives of each equation with respect to each
independent variable and collects the corresponding sub-Jacobians in a list. To form the
full system, we simply evaluate the residuals of the remaining equations (the rate equation
and the equation for well control) and concatenate the three equations into a cell array:
eqs = {eq1, rateEq(p_ad, bhp_ad, qS_ad), ctrlEq(bhp_ad)};
eq = cat(eqs{:});

In doing this, the AD library will correctly combine the various sub-Jacobians and set up
the Jacobian for the full system. Then, we can extract this Jacobian, solve for the Newton
increment, and update the three primary unknowns:
J = eq.jac{1}; % Jacobian
res = eq.val; % residual
upd = -(J \ res); % Newton update

% Update variables
p_ad.val = p_ad.val + upd(pIx);
bhp_ad.val = bhp_ad.val + upd(bhpIx);
qS_ad.val = qS_ad.val + upd(qSIx);

The sparsity pattern of the Jacobian is shown in the plot to the left of the code for the
Newton update. The use of a two-point scheme on a 3D Cartesian grid gives a Jacobi
matrix that has a heptadiagonal structure, except for the off-diagonal entries in the two
red rectangles. These arise from the well equation and correspond to derivatives of this
equation with respect to cell pressures.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 212 Compressible Flow and Rapid Prototyping

1000
200

avg pressure [bar]

rate [m3/day]

500
150

0 100
0 100 200 300 400
time [days]

Figure 7.2 Time evolution of the pressure solution for the compressible single-phase problem. The
plot to the left shows the well rate (blue line) and average reservoir pressure (green circles) as function
of time, and the plots to the right show the pressure after 2, 5, 10, and 20 pressure steps.

Figure 7.2 plots how the dynamic pressure evolves with time. Initially, the pressure is in
hydrostatic equilibrium as shown in Figure 7.1. When the well starts to drain the reservoir,
the pressure drawdown near the well will start to gradually propagate outward from the
well. As a result, the average pressure inside the reservoir is reduced, which again causes a
decay in the production rate.

Computer exercises
7.2.1 Apply the compressible pressure solver to the quarter five-spot problem from
Section 5.4.1.
7.2.2 Rerun compressible simulations for the three different grid models that were
derived from the seamount data set Section 5.4.3. Replace the fixed boundary
conditions by a no-flow condition.
7.2.3 Use the implementation from Section 7.2 as a template to develop a solver for
slightly compressible flow (7.3). More details about this model can be found on
page 118 in Section 4.2. How large can cf be before the assumptions of slight
compressibility become inaccurate? Use different heterogeneities, well placements,
and model geometries to investigate this.
7.2.4 Extend the compressible solver developed in this section to incorporate other
boundary conditions than no flow.
7.2.5 Try to compute time-of-flight by extending the equation set to also include the time-
of-flight equation (4.40). Hint: the time-of-flight and the pressure equations need
not be solved as a coupled system.
7.2.6 Same as the previous exercise, except that you should try to reuse the solver from
in Section 5.3. Hint: you must first reconstruct fluxes from the computed pressure
and then construct a state object to communicate with the TOF solver.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 7.3 Pressure-Dependent Viscosity 213

7.3 Pressure-Dependent Viscosity

One particular advantage of using automatic differentiation in combination with the dis-
crete differential and averaging operators is that it simplifies the testing of new models
and alternative computational approaches. In this section, we discuss two examples that
hopefully demonstrate this aspect.
In the model discussed in the previous section, the viscosity was assumed to be constant.
However, viscosity will generally increase with increasing pressures and this effect may
be significant for the high pressures seen inside a reservoir, as we will see later in the
book when discussing black-oil models in Chapter 11. To illustrate, we introduce a linear
dependence, rather than the exponential pressure-dependence used for pore volume (7.6)
and the fluid density (7.7). That is, we assume the viscosity is given by
 
μ(p) = μ0 1 + cμ (p − pr ) . (7.13)

Having a pressure dependence means that we have to change two parts of our discretization:
the approximation of the Darcy flux across a cell face (7.8) and the flow rate through a well
connection (7.12). Starting with the latter, we evaluate the viscosity using the same pressure
as was used to evaluate the density, i.e.,

ρ(p[Nc (w)]) 
q c [w] = WI[w] p c [w] − p[Nc (w)] . (7.14)
μ(p[Nc (w)])

For the Darcy flux (7.8), we have two choices: either use a simple arithmetic average as in
(7.9) to approximate the viscosity at each cell face,

T [f ] 
v[f ] = − grad(p) − g ρ a [f ] grad(z) , (7.15)
μa [f ]

or replace the quotient of the transmissibility and the face viscosity by the harmonic average
of the mobility λ = K/μ in the adjacent cells. Both choices introduce changes in the struc-
ture of the discrete nonlinear system, but because we are using automatic differentiation,
all we have to do is code the corresponding formulas. Let us look at the details of the
implementation, starting with the arithmetic approach.

Arithmetic Average
First, we introduce a new anonymous function to evaluate the relation between viscosity
and pressure:
[mu0,c_mu] = deal(5*centi*poise, 2e-3/barsa);
mu = @(p) mu0*(1+c_mu*(p-p_r));

Then, we can replace the definition of the Darcy flux (changes marked in red):
v = @(p) -(T./mu(avg(p))).*( grad(p) - g*avg(rho(p)).*gradz );

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 214 Compressible Flow and Rapid Prototyping
5.5 220
cµ = 0
c = 0.002
µ
5 200 cµ = 0.005

avg pressure [bar]

4.5 180
viscosity [cP]

4 160

3.5 140

cµ = 0
3 c = 0.002 120
µ
cµ = 0.005
2.5 100
100 120 140 160 180 200 0 50 100 150 200 250 300 350 400
pressure [bar] time [days]
1200 10000
cµ = 0
cµ = 0.002 9000
1000
cµ = 0.005
8000

cummulative production [m3]

800 7000
rate [m /day]

6000
3

600 5000

4000
400
3000
c =0
2000 µ
200 c = 0.002
µ
1000
cµ = 0.005
0 0
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
time [days] time [days]

Figure 7.3 The effect of increasing the degree of pressure-dependence for the viscosity.

and similarly for flow rate through each well connection:

q_conn = @(p,bhp) WI.*(rho(p(wc))./ mu(p(wc))) .* (p_conn(bhp) - p(wc));

Figure 7.3 illustrates the effect of increasing the pressure dependence of the viscosity. Since
the reference value is given at p = 200 bar, which is close to the initial pressure inside the
reservoir, the more we increase cμ , the lower μ will be in the pressure-drawdown zone
near the well. Hence, we see a significantly higher initial production rate for cμ = 0.005
than for cμ = 0. On the other hand, the higher value of cμ , the faster the drawdown effect
of the well will propagate into the reservoir, inducing a reduction in reservoir pressure
that eventually will cause production to cease. In terms of overall production, a stronger
pressure dependence may be more advantageous as it leads to a higher total recovery and
higher cumulative production early in the production period.

Face Mobility: Harmonic Average

A more correct approximation is to write Darcy’s law based on mobility instead of using
the quotient of the transmissibility and an averaged viscosity:


v[f ] = −[f ] grad(p) − g ρ a [f ] grad(z) . (7.16)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 7.4 Non-Newtonian Fluid 215

The face mobility [f ] can be defined in the same way as the transmissibility is defined
in terms of the half transmissibilities using harmonic averages. That is, if T [f ,c] denotes
the half transmissibility associated with face f and cell c, the face mobility [f ] for face
f can be written as
 μ[C (f )] μ[C2 (f )] −1
1
[f ] = + . (7.17)
T [f ,C1 (f )] T [f ,C2 (f )]
In MRST, the corresponding code reads:
hf2cn = getCellNoFaces(G);
nhf = numel(hf2cn);
hf2f = sparse(double(G.cells.faces(:,1)),(1:nhf)',1);
hf2if = hf2f(intInx,:);
fmob = @(mu,p) 1./(hf2if*(mu(p(hf2cn))./hT));

v = @(p) -fmob(mu,p).*( grad(p) - g*avg(rho(p)).*gradz );

Here, hf2cn represents the maps C1 and C2 , which enable us to sample the viscosity value
in the correct cell for each half-face transmissibility, whereas hf2if represents a map from
half-faces (i.e., faces seen from a single cell) to global faces (which are shared by two
cells). The map has a unit value in row i and column j if half-face j belongs to global face
i. Hence, premultiplying a vector of half-face quantities by hf2if amounts to summing the
contributions from cells C1 (f ) and C2 (f ) for each global face f .
Using the harmonic average for a homogeneous model should produce simulation results
that are identical (to machine precision) to those produced by the arithmetic average. With
heterogeneous permeability, there will be small differences in well rates and averaged
pressures for the specific parameters considered herein. For sub-samples of the SPE 10
data set, we typically observe maximum relative differences in well rates of the order 10−3 .

Computer exercises
7.3.1 Investigate the claim that the difference between using an arithmetic average of the
viscosity and a harmonic average of the fluid mobility is typically small. To this
end, you can for instance use the following sub-sample from the SPE 10 data set:
rock = getSPE10rock(41:50,101:110,1:10).

7.4 Non-Newtonian Fluid

Viscosity is the material property that measures a fluid’s resistance to flow, i.e., the resis-
tance to a change in shape, or to the movement of neighboring portions of the fluid relative
to each other. The more viscous a fluid is, the less easily it will flow. In Newtonian fluids, the
shear stress or the force applied per area tangential to the force at any point is proportional
to the strain rate (the symmetric part of the velocity gradient) at that point, and the viscosity
is the constant of proportionality. For non-Newtonian fluids, the relationship is no longer

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 216 Compressible Flow and Rapid Prototyping

linear. The most common nonlinear behavior is shear thinning, in which the viscosity of
the system decreases as the shear rate increases. An example is paint, which should flow
easily when leaving the brush, but stay on the surface and not drip once it has been applied.
The second type of nonlinearity is shear thickening, in which the viscosity increases with
increasing shear rate. A common example is the mixture of cornstarch and water. If you
search YouTube for “cornstarch pool” you will find several spectacular videos of pools
filled with this mixture. When stress is applied to the mixture, it exhibits properties like a
solid and you may be able to run across its surface. However, if you go too slow, the fluid
behaves more like a liquid and you fall in.
Solutions of large polymeric molecules are another example of shear-thinning liquids.
In enhanced oil recovery, polymer solutions may be injected into reservoirs to improve
unfavorable mobility ratios between oil and water and improve the sweep efficiency of the
injected fluid. At low flow rates, the polymer molecule chains tumble around randomly and
present large resistance to flow. When the flow velocity increases, the viscosity decreases as
the molecules gradually align themselves in the direction of increasing shear rate. A model
of the rheology is given by
   2  n−1
Kc n−1 2 2
μ = μ∞ + (μ0 − μ∞ ) 1 + γ̇ , (7.18)
μ0
where μ0 represents the Newtonian viscosity at zero shear rate, μ∞ represents the Newto-
nian viscosity at infinite shear rate, Kc represents the consistency index, and n represents
the power-law exponent (n < 1). The shear rate γ̇ in a porous medium can be approxi-
mated by
  n
3n + 1 n−1 | v|
γ̇app = 6  . (7.19)
4n Kφ
Combining (7.18) and (7.19), we can write our model for the viscosity as
  n−1   2   2n
|
v |2 2 Kc n−1 3n + 1 n−1
μ = μ0 1 + K̄c , K̄c = 36 , (7.20)
Kφ μ0 4n
where we for simplicity have assumed that μ∞ = 0.

Rapid Development of Proof-of-Concept Codes

We now demonstrate how easy it is to extend the simple simulator developed so far in
this chapter to model non-Newtonian fluid behavior (see nonNewtonianCell.m). To sim-
ulate injection, we increase the bottom-hole pressure to 300 bar. Our rheology model has
parameters:
mu0 = 100*centi*poise;
nmu = .3;
Kc = .1);
Kbc = (Kc/mu0)^(2/(nmu-1))*36*((3*nmu+1)/(4*nmu))^(2*nmu/(nmu-1));

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 7.4 Non-Newtonian Fluid 217

In principle, we could continue to solve the system using the same primary unknowns
as before. However, it has proved convenient to write (7.20) in the form μ = η μ0 , and
introduce η as an additional unknown. In each Newton step, we start by solving the equation
for the shear factor η exactly for the given pressure distribution. This is done by initializing
an AD variable for η, but not for p in etaEq so that this residual now only has one unknown,
η. This will take out the implicit nature of Darcy’s law and hence reduce the nonlinearity
and simplify the solution of the global system.
while (resNorm > tol) && (nit < maxits)
% Newton loop for eta (shear multiplier)
[resNorm2,nit2] = deal(1e99, 0);
eta_ad2 = initVariablesADI(eta_ad.val);
while (resNorm2 > tol) && (nit2 <= maxits)
eeq = etaEq(p_ad.val, eta_ad2);
res = eeq.val;
eta_ad2.val = eta_ad2.val - (eeq.jac{1} \ res);
resNorm2 = norm(res);
nit2 = nit2+1;
end
eta_ad.val = eta_ad2.val;

Once the shear factor has been computed for the values in the previous iterate, we can use
the same approach as earlier to compute a Newton update for the full system. (Here, etaEq
is treated as a system with two unknowns, p and η.)
eq1 = presEq(p_ad, p0, eta_ad, dt);
eq1(wc) = eq1(wc) - q_conn(p_ad, eta_ad, bhp_ad);
eqs = {eq1, etaEq(p_ad, eta_ad), ...
rateEq(p_ad, eta_ad, bhp_ad, qS_ad), ctrlEq(bhp_ad)};
eq = cat(eqs{:});
upd = -(eq.jac{1} \ eq.val); % Newton update

To finish the solver, we need to define the flow equations and the extra equation for the shear
multiplier. The main question now is how we should compute | v |? One solution could be to
define |
v | on each face as the flux divided by the face area. In other words, use a code like
phiK = avg(rock.perm.*rock.poro)./G.faces.areas(intInx).^2;
v = @(p, eta) -(T./(mu0*eta)).*( grad(p) - g*avg(rho(p)).*gradz );
etaEq = @(p, eta) eta - (1 + Kbc*v(p,eta).^2./phiK).^((nmu-1)/2);

Although simple, this approach has three potential issues: First, it does not tell us how
to compute the shear factor for the well perforations. Second, it disregards contributions
from any tangential components of the velocity field. Third, the number of unknowns in
the linear system increases by almost a factor six since we now have one extra unknown
per internal face. The first issue is easy to fix: To get a representative value in the well cells,
we simply average the η values from the cells’ faces. If we now recall how the discrete

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 218 Compressible Flow and Rapid Prototyping

divergence operator was defined, we realize that this operation is almost implemented for
us already: if div(x)=-D’*x computes the discrete divergence in each cell of the field x
defined at the faces, then wavg(x)=1/6*abs(D)’*x computes the average of x for each
cell. In other words, our well equation becomes:
wavg = @(eta) 1/6*abs(D(:,W.cells))'*eta;
q_conn = @(p, eta, bhp) ...
WI .* (rho(p(wc)) ./ (mu0*wavg(eta))) .* (p_conn(bhp) - p(wc));

The second issue would have to be investigated in more detail, and this is not within the
scope of this book. The third issue is simply a disadvantage.
To get a method that consumes less memory, we can compute one η value per cell. Using
the following formula, we can reconstruct an approximate velocity vi at the center of cell i

 vij 
vi = cij − ci , (7.21)
Vi
j ∈N (i)

where N(i) is the map from cell i to its neighboring cells, vij is the flux between cell i and
cell j , cij is the centroid of the corresponding face, and ci is the centroid of cell i. For a
Cartesian grid, this formula simplifies so that an approximate velocity can be obtained as
the sum of the absolute value of the flux divided by the face area over all faces that make up
a cell. Using a similar trick as we used in the wavg operator to compute η in the well cells,
our implementation follows trivially. We first define the averaging operator to compute cell
velocity
aC = bsxfun(@rdivide, 0.5*abs(D), G.faces.areas(intInx))';
cavg = @(x) aC*x;

In doing so, we also rename our old averaging operator avg as favg to avoid confusion
and make it more clear that this operator maps from cell values to face values. Then we can
define the needed equations:
phiK = rock.perm.*rock.poro;
gradz = grad(G.cells.centroids(:,3));
v = @(p, eta) -(T./(mu0*favg(eta))).*( grad(p) - g*favg(rho(p)).*gradz );
etaEq = @(p, eta)
eta - ( 1 + Kbc* cavg(v(p,eta)).^2 ./phiK ).^((nmu-1)/2);
presEq= @(p, p0, eta, dt) ...
(1/dt)*(pv(p).*rho(p) - pv(p0).*rho(p0)) + div(favg(rho(p)).*v(p, eta));

With this approach, the well equation becomes particularly simple, since all we need to do
is sample the η value from the correct cell:
q_conn = @(p, eta, bhp) ...
WI .* (rho(p(wc)) ./ (mu0*eta(wc))) .* (p_conn(bhp) - p(wc));

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 7.4 Non-Newtonian Fluid 219
120 235
Newtonian
110 Cell-based
Face-based 230
100 Face-based (not well)

90 225

avg pressure [bar]

rate [m /day]

80
220
3

70
215
60

50 210
40 Newtonian
Cell-based 205
30 Face-based
Face-based (not well)
20 200
0 100 200 300 400 0 100 200 300 400
time [days] time [days]

Figure 7.4 Single-phase injection of a highly viscous, shear-thinning fluid computed by four different
simulation methods: (i) fluid assumed to be Newtonian, (ii) shear multiplier η computed in cells, (iii)
shear multiplier computed at faces, and (iv) shear multiplier computed at faces, but η ≡ 1 used in
well model.

A potential drawback of this second approach is that it may introduce numerical smearing,
but this will, on the other hand, most likely increase the robustness of the resulting scheme.
Figure 7.4 compares the predicted flow rates and average reservoir pressure for two
different fluid models: one that assumes a standard Newtonian fluid (i.e., η ≡ 1) and
one that models shear thinning. With shear thinning, the higher pressure in the injection
well causes a decrease in the viscosity, which leads to significantly higher injection rates
than for the Newtonian fluid and hence a higher average reservoir pressure. Perhaps more
interesting is the large discrepancy in rates and pressures predicted by the face-based and
cell-based simulation algorithms. If we disregard the shear multiplier q_conn in the face-
based method, the predicted rate and pressure buildup is smaller than what is predicted by
the cell-based method, and closer to the Newtonian fluid case. We take this as evidence
that the differences between the cell and the face-based methods to a large extent can be
explained by differences in the discretized well models and their ability to capture the
formation and propagation of the strong initial transient. To further back this up, we have
included results from a simulation with ten times as many time steps in Figure 7.5, which
also includes plots of the evolution of min(η) as function of time. Whereas the face-based
method predicts a large, immediate drop in viscosity in the near-well region, the viscosity
drop predicted by the cell-based method is much smaller during the first 20–30 days. This
results in a delay in the peak of the injection rate and a much smaller injected volume.
We leave the discussion here. The parameters used in the example were chosen quite
haphazardly to demonstrate a pronounced shear-thinning effect. Which method is the most
correct for real computations is a question that goes beyond the current scope, and could

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 220 Compressible Flow and Rapid Prototyping
120 1
Newtonian
110 Cell-based 0.9
Face-based
100 Face-based (not well)
0.8

shear multiplicator [1]

90
0.7
rate [m /day]

80
0.6
3

70
0.5
60
0.4
50
0.3
40
0.2
30
100 200 300 100 200 300
time [days] time [days]

Figure 7.5 Single-phase injection of a highly viscous, shear-thinning fluid; simulation with t =
1/520 year. The right plot shows the evolution of η as a function of time: solid lines show min(η)
over all cells, dashed lines min(η) over the perforated cells, and dash-dotted lines average η value.

probably best be answered by verifying against observed data for a real case. Our point
here was mainly to demonstrate the capability for rapid implementation of proof-of-concept
codes that comes with the use of MRST. However, as the example shows, this lunch is not
completely free: you still have to understand features and limitations of the models and
discretizations you choose to implement.

Computer exercises
7.4.1 Investigate whether the large differences observed in Figures 7.4 and 7.5 between
the cell-based and face-based approaches to the non-Newtonian flow problem is a
result of insufficient grid resolution.
7.4.2 The non-Newtonian fluid has a strong transient during the first 30–100 days. Try
to implement adaptive time steps that utilize this fact. Can you come up with a
strategy that automatically choose good time steps?

7.5 Thermal Effects

As another example of rapid prototyping, we extend the single-phase flow model (7.1) to
account for thermal effects. That is, we assume that ρ(p,T ) is now a function of pressure
and temperature T and extend our model to also include conservation on energy,
∂    K 
φρ + ∇ · ρ v = q, v = − ∇p − gρ∇z , (7.22a)
∂t μ
∂     
φρEf (p,t) + (1 − φ)Er + ∇ · ρHf v − ∇ · κ∇T = qe . (7.22b)
∂t

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 7.5 Thermal Effects 221

Here, the rock and the fluid are assumed to be in local thermal equilibrium. In the energy
equation (7.22b), Ef is energy density per mass of the fluid, Hf = Ef + p/ρ is enthalpy
density per mass, Er is energy per volume of the rock, and κ is the heat conduction
coefficient of the rock. Fluid pressure p and temperature T are used as primary variables.
As in the original isothermal simulator, we must first define constitutive relationships
that express the new physical quantities in terms of the primary variables. The energy
equation includes heating of the solid rock, and we therefore start by defining a quantity
that keeps track of the solid volume, which also depends on pressure:
sv = @(p) G.cells.volumes - pv(p);

For the fluid model, we use

 
ρ(p,T ) = ρr 1 + βT (p − pr ) e−α(T −Tr ),
 
μ(p,T ) = μ0 1 + cμ (p − pr ) e−cT (T −Tr ) . (7.23)
Here, ρr = 850 is the density and μ0 = 5 cP the viscosity of the fluid at reference
kg/m3
conditions with pressure pr = 200 bar and temperature Tr = 300 K. The constants are
βT = 10−3 bar−1 , α = 5 × 10−3 K−1 , cμ = 2 × 10−3 bar−1 , and cT = 10−3 K−1 . This
translates to the following code:
[mu0,cmup] = deal( 5*centi*poise, 2e-3/barsa);
[cmut,T_r] = deal( 1e-3, 300);
mu = @(p,T) mu0*(1+cmup*(p-p_r)).*exp(-cmut*(T-T_r));

[alpha, beta] = deal(5e-3, 1e-3/barsa);

rho_r = 850*kilogram/meter^3;
rho = @(p,T) rho_r .* (1+beta*(p-p_r)) .* exp(-alpha*(T-T_r));

We use a simple linear relation for the enthalpy, which is based on the thermodynamical
relations that give
 
1 − αTr 1 ∂ρ 
dHf = cp dT + dp, α=−  , (7.24)
ρ ρ ∂T p
where cp = 4 × 103 J/kg. The code for enthalpy/energy densities reads:
Cp = 4e3;
Hf = @(p,T) Cp*T+(1-T_r*alpha).*(p-p_r)./rho(p,T);
Ef = @(p,T) Hf(p,T) - p./rho(p,T);
Er = @(T) Cp*T;

We defer discussing details of these new relationships and only note that it is important that
the thermal potentials Ef and Hf are consistent with the equation of state ρ(p,T ) to get a
physically meaningful model.
Having defined all constitutive relationships in terms of anonymous functions, we can
set up the equation for mass conservation and Darcy’s law (with transmissibility renamed
to Tp to avoid name clash with temperature):

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 222 Compressible Flow and Rapid Prototyping

v = @(p,T) -(Tp./mu(avg(p),avg(T))).*(grad(p) - avg(rho(p,T)).*gdz);

pEq = @(p,T,p0,T0,dt) ...
(1/dt)*(pv(p).*rho(p,T) - pv(p0).*rho(p0,T0)) ...
+ div( avg(rho(p,T)).*v(p,T) );

In the energy equation (7.22b), the accumulation and the heat-conduction terms are on the
same form as the operators appearing in (7.22a) and can hence be discretized in the same
way. We use an artificial “rock object” to compute transmissibilities for κ instead of K:
tmp = struct('perm',4*ones(G.cells.num,1));
hT = computeTrans(G, tmp);
Th = 1 ./ accumarray(cf, 1 ./ hT, [nf, 1]);
Th = Th(intInx);

The remaining term in (7.22b), ∇ · [ρHf v], represents advection of enthalpy and has a
differential operator on the same form as the transport equations discussed in Section 4.4.3
and must hence be discretized by an upwind scheme. To this end, we introduce a new
discrete operator that will compute the correct upwind value for the enthalpy density,

H [C1 (f )], if v[f ] > 0,
upw(H )[f ] = (7.25)
H [C2 (f )], otherwise.
With this, we can set up the energy equation in residual form in the same way as we
previously have done for Darcy’s law and mass conservation:
upw = @(x,flag) x(C(:,1)).*double(flag)+x(C(:,2)).*double(flag);
hEq = @(p, T, p0, T0, dt) ...
(1/dt)*(pv(p ).*rho(p, T ).*Ef(p ,T ) + sv(p ).*Er(T ) ...
- pv(p0).*rho(p0,T0).*Ef(p0,T0) - sv(p0).*Er(T0)) ...
+ div( upw(Hf(p,T),v(p,T)>0).*avg(rho(p,T)).*v(p,T) ) ...
+ div( -Th.*grad(T));

With this, we are almost done. As a last technical detail, we must also make sure that the
energy transfer in injection and production wells is modeled correctly using appropriate
upwind values:
qw = q_conn(p_ad, T_ad, bhp_ad);
eq1 = pEq(p_ad, T_ad, p0, T0, dt);
eq1(wc) = eq1(wc) - qw;
hq = Hf(bhp_ad,bhT).*qw;
Hcells = Hf(p_ad,T_ad);
hq(qw<0) = Hcells(wc(qw<0)).*qw(qw<0);
eq2 = hEq(p_ad,T_ad, p0, T0,dt);
eq2(wc) = eq2(wc) - hq;

Here, we evaluate the enthalpy using cell values for pressure and temperature for produc-
tion wells (for which qw<0) and pressure and temperatures at the bottom hole for injection
wells.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 7.5 Thermal Effects 223

What remains are trivial changes to the iteration loop to declare the correct vari-
ables as AD structures, evaluate the discrete equations, collect their residuals, and
update the state variables. These details can be found in the complete code given in
singlePhaseThermal.m and have been left out for brevity.

Understanding Thermal Expansion

Except for the modifications discussed in the previous subsection, the setup is the exact
same as in Section 7.2. That is, the reservoir is a 200 × 200 × 50 m3 rectangular box with
homogeneous permeability of 30 mD, constant porosity 0.3, and a rock compressibility
of 10−6 bar−1 , realized on a 10 × 10 × 10 Cartesian grid. The reservoir is realized on a
10 × 10 × 10 Cartesian grid. Fluid is drained from a horizontal well perforated in cells with
indices i = 2, j = 2, . . . ,9, and k = 5, and operating at a constant bottom-hole pressure
of 100 bar. Initially, the reservoir has constant temperature of 300 K and is in hydrostatic
equilibrium with a datum pressure of 200 bar specified in the uppermost cell centroids.
In the same way as in the isothermal case, the open well will create a pressure drawdown
that propagates into the reservoir. As more fluid is produced from the reservoir, the pressure
will gradually decay towards a steady state with pressure values between 101.2 and 104.7
bar. Figure 7.6 shows that the simulation predicts a faster pressure drawdown, and hence a
faster decay in production rates, if thermal effects are taken into account.
The change in temperature of an expanding fluid will not only depend on the initial and
final pressure, but also on the type of process in which the temperature is changed:

• In a free expansion, the internal energy is preserved and the fluid does no work. That is,
the process can be described by the following differential:
dEf dEf
p + T = 0. (7.26)
dp dT

210 −2
10
min(p) Thermal
200 avg(p) Isothermal
max(p)
190

180
−3
170 10

160

150

140
−4
10
130

120

110

100 −5
10
0 100 200 300 400 500 100 200 300 400 500

Figure 7.6 To the left, time evolution for pressure for an isothermal simulation (solid lines) and a
thermal simulation with α = 5 × 10−3 (dashed lines). To the right, decay in production rate at the
surface.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 224 Compressible Flow and Rapid Prototyping

When the fluid is an ideal gas, the temperature is constant, but otherwise the temperature
will either increase or decrease during the process depending on the initial temperature
and pressure.
• In a reversible process, the fluid is in thermodynamical equilibrium and does positive
work while the temperature decreases. The linearized function associated with this adia-
batic expansion reads
 
p 1
dE + dV = dE + p d = 0. (7.27)
ρV ρ

• In a Joule–Thomson process, the enthalpy remains constant while the fluid flows from
higher to lower pressure under steady-state conditions and without change in kinetic
energy. That is,
dHf dHf
p + T = 0. (7.28)
dp dT
Our case is a combination of these three processes and their interplay will vary with
the initial temperature and pressure as well as with the constants in the fluid model for
ρ(p,T ). To better understand a specific case, we can use (7.26–7.28) to compute the
temperature change that would take place for an observed pressure drawdown if only one
of the processes took place. Computing such linearized responses for thermodynamical
functions is particularly simple using automatic differentiation. Assuming we know the
reference state (pr ,Tr ) at which the process starts and the pressure pe after the process has
taken place, we initialize the AD variables and compute the pressure difference:
[p,T] = initVariablesADI(p_r,T_r);
dp = p_e - p_r;

Then, we can solve (7.26) or (7.28) for T and use the result to compute the temperature
change resulting from a free expansion or a Joule–Thomson expansion:
E = Ef(p,T); dEdp = E.jac{1}; dEdT = E.jac{2};
Tfr = T_r - dEdp*dp/dEdT;

hf = Hf(p,T); dHdp = hf.jac{1}; dHdT = hf.jac{2};

Tjt = T_r - dHdp*dp/dHdT;

The temperature change after a reversible (adiabatic) expansion is not described by a total
differential. In this case we have to specify that p should be kept constant. This is done by
replacing the AD variable p by an ordinary variable double(p) in the code at the specific
places where p appears in front of a differential; see (7.27).
E = Ef(p,T) + double(p)./rho(p,T);
dEdp = hf.jac{1};
dEdT = hf.jac{2};
Tab = T_r - dEdp*dp/dEdT;

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 7.5 Thermal Effects 225

The same kind of manipulation can be used to study alternative linearizations of systems of
nonlinear equations and the influence of neglecting some of the derivatives when forming
Jacobians.
To illustrate how the interplay between the three processes can change significantly and
lead to quite different temperature behavior, we compare the predicted evolution of the
temperature field for α = 5 × 10−n , n = 3,4, as shown in Figures 7.7 and 7.8. The change
in behavior between the two figures is associated with the change in sign of ∂E/∂p,
   
αT βT p − αT 1 ∂ρ 
dE = cp − dT + dp, βT =  . (7.29)
ρ ρ ρ ∂p T
In the isothermal case and for α = 5 × 10−4 , we have that αT < βT p so that ∂E/∂p > 0.
The expansion and flow of fluid will cause an instant heating near the well-bore, which is
what we see in the initial temperature increase for the maximum value in Figure 7.7. The
Joule–Thomson coefficient (αT − 1)/(cp ρ) is also negative, which means that the fluid
gets heated if it flows from high pressure to low pressure in a steady-state flow. This is seen
by observing the temperature in the well perforations. The fast pressure drop in these cells
causes an almost instant cooling effect, but soon after we see a transition in which most of
the cells containing a well perforation start having the highest temperature in the reservoir
because of heating from the moving fluids. For α = 5 × 10−3 , we have that αT > βT p
so that ∂E/∂p < 0 and likewise the Joule–Thomson coefficient is positive. The moving
fluids will induce a cooling effect and hence the minimum temperature is observed at the
well for a longer time. The weak kink in the minimum temperature curve is the result of
the point of minimum temperature moving from being at the bottom front side to the far
back of the reservoir. The cell with lowest temperature is where the fluid has done most
work, neglecting heat conduction. In the beginning, this is the cell near the well because
the pressure drop is largest there. Later, it will be the cell furthest from the well, since
this is where the fluid can expand most. The discussion in this subsection is only meant to
illustrate physical effect and does not necessarily represent realistic wells.

Computational Performance
If you are observant, you may have realized that the code presented in this chapter contains
a number of redundant function evaluations that may potentially add significantly to the
overall computational cost: in each nonlinear iteration we keep reevaluating quantities that
depend on p0 and T0 even though these stay constant for each time step. We can easily avoid
this by moving the definition of the anonymous functions evaluating the residual equations
inside the outer time loop. The main contribution to potential computational overhead,
however, comes from repeated evaluations of fluid viscosity and density. Because each
residual equation is defined as an anonymous function, v(p,T) appears three times for
each residual evaluation, once in pEq and twice in hEq. This, in turn, translates to three
calls to mu(avg(p),avg(T)) and seven calls to rho(p,T), and so on. In practice, the
number of actual function evaluations is smaller, since the MATLAB interpreter most likely
has some kind of built-in intelligence to spot and reduce redundant function evaluations.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 226 Compressible Flow and Rapid Prototyping

300.3
min(T)
avg(T)
300.2 Free expansion max(T)
wells
300.1

300

299.9

299.8

299.7

299.6 Adiabatic expansion

299.5
0 100 200 300 400 500

Figure 7.7 Time evolution of temperature for a compressible, single-phase problem with α = 5 ·
10−4 . The upper plots show four snapshots of the temperature field. The lower plot shows minimum,
average, maximum, and well-perforation values.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 7.5 Thermal Effects 227

300
min(T)
299.5 avg(T)
max(T)
299 wells
Joule−Tompson
298.5

298

297.5

297

296.5
Free expansion
296

295.5 Adiabatic expansion

295
0 100 200 300 400 500

Figure 7.8 Time evolution of temperature for a compressible, single-phase problem with α = 5 ×
10−3 . The upper plots show four snapshots of the temperature field. The lower plot shows minimum,
average, maximum, and well-perforation values.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 228 Compressible Flow and Rapid Prototyping

Nonetheless, to cure this problem, we can move the computations of residuals inside a
function so that the constitutive relationships can be computed one by one and stored in
temporary variables. The disadvantage is that we increase the complexity of the code and
move one step away from the mathematical formulas describing the method. This type
of optimization should therefore only be introduced after the code has been profiled and
redundant function evaluations have proved to have a significant computational cost.

Computer exercises
7.5.1 Perform a more systematic investigation of how changes in α affect the temperature
and pressure behavior. To this end, you should change α systematically, e.g., from
0 to 10−2 . What is the effect of changing β, the parameters cμ and cT for the
viscosity, or cp in the definition of enthalpy?
7.5.2 Use the MATLAB profiling tool to investigate to what extent the use of nested
anonymous functions causes redundant function evaluations or introduces other
types of computational overhead. Hint: to profile the CPU usage, you can use the
following call sequence
profile on, singlePhaseThermal; profile off; profile report

Try to modify the code as suggested above to reduce the CPU time. How low can
you get the ratio between the cost of constructing the linearized system and the cost
of solving it?

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Part III
Multiphase Flow

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8
Mathematical Models for Multiphase Flow

Up to now, we have only considered flow of a single fluid phase. For most applications of
reservoir simulation, however, one is interested in modeling how one fluid phase displaces
one or more other fluid phases. In hydrocarbon recovery, typical examples are water or
gas flooding, in which injected water or gas displaces the resident hydrocarbon phase(s).
Likewise, in geological carbon sequestration, the injected CO2 forms a supercritical fluid
phase that displaces the resident brine. In both cases, more than one phase will flow simul-
taneously throughout the porous medium when viewed from the scale of a representative
elementary volume, even if the fluids are immiscible and do not mix on the microscale.
To model such flows, we introduce three new physical properties of multiphase models
(saturation, relative permeability, and capillary pressure) and discuss how one can use these
to extend Darcy’s law to multiphase flow and combine it with conservation of mass for each
fluid phase to develop models describing multiphase displacements. The resulting system
of partial differential equations is parabolic in the general case, but has a mixed elliptic-
hyperbolic mathematical character. Fluid pressures tend to behave as following a near-
elliptic equation, whereas the transport of fluid phases has a strong hyperbolic character.
It is therefore common to use a so-called fractional-flow formulation to write the flow
equations as a coupled system consisting of a pressure equation describing the evolution
of one of the fluid pressures and (a system of) saturation equation(s) that describes the
transport of fluid phases. The mixed elliptic-hyperbolic nature is particularly evident in the
case of immiscible, incompressible flow. In this case, the pressure equation simplifies to a
Poisson equation of the same form we have studied in previous chapters.
This chapter describes the new physical effects that appear in multiphase flow and
discusses general multiphase models in some detail. We also derive the fractional-flow
formulation in the special case of immiscible flow and analyze the mathematical character
of the system in certain limiting cases. Chapter 9 introduces various methods for solving
hyperbolic transport equation and reviews some of the supporting theory. Then, in Chapter
10, we focus entirely on the incompressible case and show how we can easily reuse the
elliptic solvers developed in the previous chapters and combine them with a set of simple
first-order transport solvers that are implemented in the incomp module of MRST. To
simplify the discussion, this and the next two chapters will mainly focus on two-phase,

231

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 232 Mathematical Models for Multiphase Flow

immiscible systems, but the most crucial equations will be stated and developed for the
general multiphase case. In Chapters 11 and 12, we return to the general case and discuss
the compressible models and numerical methods that are used for three-phase flow in most
contemporary commercial simulators.

8.1 New Physical Properties and Phenomena

As we have seen previously, a Darcy-type continuum description of a reservoir fluid sys-
tem means that any physical quantity defined at a point x represents an average over a
representative elementary volume (REV). Let us consider a system with two or more fluid
phases that are immiscible so that no mass transfer takes place among the phases. This
means that the fluid phases will not mix and form a solution on the microscale but rather
stay as separate volumes or layers separated by a curved meniscus as illustrated in Figure
8.1. Nevertheless, when considering the flow averaged over a REV, the fluid phases will
generally not be separated by a sharp interface, and two or more phases may occupy the
same point in the continuum description. In this section, we introduce the new fundamental
concepts necessary to understand multiphase flow and formulate continuum models that
describe the simultaneous flow of two or more fluid phases taking place at the same point
in a reservoir. Unless stated otherwise, the two fluids are assumed to be oil and water when
we discuss two-phase systems.

8.1.1 Saturation
Saturation Sα is defined as the fraction of the pore volume occupied by phase α. In
the single-phase models discussed in previous chapters, we assumed that the void space
between the solid particles that make up the porous medium was completely filled by fluid.
Similarly, for multiphase models we assume that the void space is completely filled with
one or more fluid phases, so that

Figure 8.1 Averaging of a multiphase flow over a representative elementary volume (REV) to obtain
a Darcy-scale continuum model.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.1 New Physical Properties and Phenomena 233

Sα = 1. (8.1)
α

In reservoir simulation, it is most common to consider three phases: an aqueous phase

(w), an oleic phase (o), and a gaseous (g) phase. Each saturation can vary from 0, which
means that the phase is not present at all at this point in space, to 1, which means that the
phase completely fills the local pore volume. In most practical cases, however, the range of
variability is smaller.
Consider a rock originally deposited in an aqueous environment. During deposition,
the pores between rock particles are completely filled with water. Later, as hydrocarbons
start to migrate into what is to become our reservoir, the resident water will be displaced
and the saturation gradually reduced to some small value, typically 5–40%, at which the
water can no longer flow and exists as small drops trapped between mineral particles or
encapsulated by the invading hydrocarbon phases. The saturation at which water goes from
being mobile (funicular state) to being immobile (pendular state) is called the irreducible
water saturation and is usually denoted Swir or Swr . The irreducible water saturation is
determined by the topology of the pore space and the water’s affinity to wet the mineral
particles relative to that of the invading hydrocarbons; this affinity is determined by the
chemical composition of the fluids and the mineral particles.
Petroleum literature also talks of the connate water saturation, usually denoted Swc ,
which is the water saturation that exists upon discovery of the reservoir. The quantities
Swir and Swc may or may not coincide, but should not be confused. Sometimes, one also
sees the notation Swi , which may refer to any of the two. If water is later injected to displace
the oil, it is generally not possible to flush out all the oil and parts of the pore space will be
occupied by isolated oil droplets, as illustrated in Figure 8.1. The corresponding residual
oil saturation is denoted as Sor . In most systems, the water has a stronger affinity for the
rock, which means that Sor is usually higher than Swr and Swc : typically they are in the
range 10–50%.
In many models, each phase may also contain one or more components. These may be
unique hydrocarbon species like methane, ethane, propane, etc., or other chemical species
like polymers, salts, surfactants, tracers, etc. Since the number of hydrocarbon components
can be quite large, it is common to group components into pseudo-components. Because
of the varying and extreme conditions in a reservoir, the composition of the different
phases can change throughout a simulation and may sometimes be difficult to determine
uniquely. We therefore need to describe this composition. There are several ways to do
this. Herein, we use the mass fraction of component  in phase α, denoted by cα and
defined as
ρα
cα = (8.2)
ρα
where ρα denotes the bulk density of phase α and ρα the effective density of component 
in phase α. In each phase, the mass fractions should add up to unity, so that for M different
components in a system consisting of an aqueous, a gaseous, and an oleic phase, we have:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 234 Mathematical Models for Multiphase Flow


M 
M 
M

cw = cg = co = 1. (8.3)
=1 =1 =1

We will return to models having three phases and more than one component per phase
in Chapter 11. For now, however, we assume that our system consists of two immiscible
phases.

8.1.2 Wettability
At the microscale, which is significantly larger than the molecular scale, immiscible fluid
phases are separated by well-defined, infinitely thin interfaces. Because cohesion forces
between molecules are different on opposite sides, each interface has an associated surface
tension (or surface energy), which measures the forces the interface must overcome to
change its shape. In the absence of external forces, minimization of surface energy will
cause the interface of a droplet of one phase contained within another phase to assume a
spherical shape. The interface tension will keep the fluids apart, regardless of the size of
the droplet.
The microscale flow of our oil–water system is strongly affected by how the phases
attach to the interface of the solid rock. The ability of a liquid phase to maintain contact with
a solid surface is called wettability and is determined by intermolecular interactions when
the liquid and solid are brought together. Adhesive forces between different molecules in
the liquid phase and the solid rock will cause liquid droplets to spread across the mineral
surface. Likewise, cohesive forces between similar molecules within the liquid phases will
cause the droplets to avoid contact with the surface and ball up. When two fluid phases are
present in the same pore space, one phase will be more attracted to the mineral particles
than the other phase. We refer to the preferential phase as the wetting phase, while the other
is called the non-wetting phase. The balance of the adhesive and cohesive forces determines
the contact angle θ shown in Figure 8.2, which is a measure of the wettability of a system
that can be related to the interface energies by Young’s equation:

Figure 8.2 Contact angle θ and surface tension σ for two different oil–water systems. In water-wet
systems, 0 ≤ θ < 90◦ , whereas 90◦ < 180◦ in oil-wet systems.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.1 New Physical Properties and Phenomena 235

σow cos θ = σos − σws . (8.4)

Here, σow is the interface energy of the oil–water interface and σos and σws are the energies
of the oil–solid and water–solid interfaces, respectively. Hydrophilic or water-wet porous
media, in which the water shows a greater affinity than oil to stick to the rock surface,
are more widespread in nature than hydrophobic or oil-wet media. This explains why Sor
usually is larger than Swr . In a perfectly water-wet system, θ = 0 so that water spreads
evenly over the whole surface of the mineral grains. Likewise, in a perfectly oil-wet system,
θ = 180◦ , so that water forms spherical droplets at the solid surface.

8.1.3 Capillary Pressure

Because of the surface tension, the equilibrium pressure in two phases separated by a curved
interface will generally be different. The difference in phase pressures is called the capillary
pressure,
pc = pn − pw, (8.5)
and is always positive because the pressure in the non-wetting fluid is higher than the
pressure in the wetting fluid. For a water-wet reservoir, the capillary pressure is therefore
defined as pcow = po − pw , whereas one usually defines pcog = pg − po in an oil–gas
system where oil is the wetting phase.
The action of capillary pressures can cause liquids to move in narrow spaces, devoid of
or in opposition to other external forces such as gravity. To illustrate this, we consider a
thin tube immersed in a wetting and a non-wetting fluid, as shown in Figure 8.3. For the
wetting fluid, adhesive forces between the solid tube and the liquid will form a concave
meniscus and pull the liquid upward against the gravity force. It is exactly the same effect
that causes water to be drawn up into a piece of cloth or paper dipped into water. In the
non-wetting case, the intermolecular cohesion forces within the liquid exceed the adhesion
forces between the liquid and the solid so that a convex meniscus is formed and drawn
downwards relative to the liquid level outside of the tube. At equilibrium inside the capillary

Figure 8.3 Capillary tubes for a wetting liquid and air (left) and a wetting and a non-wetting liquid
(right). (The width of the capillary tube is exaggerated.)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 236 Mathematical Models for Multiphase Flow

tube, the upward and downward forces must balance each other. The force acting upward
equals

2π r(σas − σls ) = 2π r σ cos θ,

where subscripts a and l refer to air and liquid, respectively. The capillary pressure is
defined as force per unit area, or in other words,

2π r σ cos θ 2σ cos θ
pc = 2
.= (8.6)
πr r

The force acting downward can be deducted from Archimedes’ principle as π r 2 gh(ρl −ρa ).
By equating this with the action of the capillary pressure, we obtain

π r 2 gh(ρl − ρa )
pc = = ρgh. (8.7)
πr2
Void space inside a reservoir contains a large number of narrow pore throats that can be
thought of as a bundle of nonconnecting capillary tubes of different radius. As we can see
from the formulas developed thus far, the capillary pressure increases with decreasing tube
radius for a fixed interface-energy difference between two immiscible fluids. Because the
pore size is usually so small, capillary pressure will play a major role in establishing the
fluid distribution inside the reservoir. To see this, consider a hydrocarbon phase migrating
upward by buoyancy forces into a porous rock filled with water. If the hydrocarbon phase is
to enter void space in the rock, its buoyancy force must exceed a certain minimum capillary
pressure. The capillary pressure that is required to force the first droplet of oil into the rock
is called the entry pressure. If we consider the rock as a complex assortment of capillary
tubes, the first droplet will enter the widest tube, which according to (8.6) has the lowest
capillary pressure. As the pressure difference between the buoyant oil and the resident water
increases, oil will be able to enter increasingly smaller pore throats and hence reduce the
water saturation. This means that there will be a relation between saturation and capillary
pressure,

pcnw = pn − pw = Pc (Sw ), (8.8)

as illustrated in Figure 8.4. The slope of the curve is determined by the variability of
the pore sizes. If all pores are of similar size, they will all be invaded quickly once we
exceed the entry pressure and the curve will be relatively flat so that saturation decays
rapidly with increasing capillary pressure. If the pores vary a lot in size, the decrease
in saturation with increasing capillary pressure will be more gradual. As for the vertical
distribution of fluids, we see that once the fluids have reached a hydrostatic equilibrium,
the difference in densities between water and oil dictates the difference in phase pressures
and hence oil saturation increases in the upward direction, which is also illustrated in the
figure.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.1 New Physical Properties and Phenomena 237

Figure 8.4 The left plot shows a capillary pressure curve giving the relation between capillary
pressure pcow and water saturation Sw . In addition, we have included how the capillary pressure
and water saturation relate to the height above the free water level (FWL) for a system in hydrostatic
equilibrium. Notice that the entry pressure pe occurs at the oil–water contact (OWC), which is found
above the FWL for systems with nonzero entry pressures. The right plot shows hysteretic behavior
for repeated drainage and imbibition displacements.

Hysteresis, Drainage, and Imbibition

The argument earlier in this section was for the case of an invading non-wetting fluid
displacing a wetting fluid. This type of displacement is called drainage to signify that the
saturation of the wetting phase is decreasing in this type of displacement. The opposite
case, called imbibition, occurs when a wetting fluid displaces a non-wetting fluid. As an
example, let us assume that we inject water to flush out the oil in a pristine reservoir, in
which the amount of connate water equals the irreducible water saturation. During this
displacement, the water saturation will gradually increase as more water is injected. Hence,
the oil saturation decreases until we reach the residual oil saturation, at which there is only
immobile oil left. The displacement will generally not follow the same capillary curve as
the primary drainage curve, as shown in the right plot in Figure 8.4. Likewise, if another
drainage displacement takes place, starting from Sor or a larger oil saturation, this process
will generally not follow the imbibition curve. The result is an example of what is called
hysteresis, in which the behavior of a system depends both on the current state and its
previous history. The hysteretic behavior can be explained by pore-scale trapping of oil
droplets, by variations in the wetting angle between advancing and receding fluid at the
solid interface, and by the fact that whereas the drainage process is controlled by the size
of the widest non-invaded pore throat, the imbibition process is controlled by the size of
the narrowest non-invaded pore.
As we can see from Figure 8.4, a relatively large fraction of the oil will be left behind
in an immobile state after waterflooding. Several methods for enhanced oil recovery have
been developed to mobilize this immobile oil, e.g., by injecting another fluid (e.g., CO2 or
gas) that mixes with immobile oil droplets so that a larger fraction of the oil can be washed

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 238 Mathematical Models for Multiphase Flow

out along with the invading fluid. In chemical and microbial methods, one adds chemical
substances or small microorganisms to the injected fluids that alter the wetting properties
inside the pores. Simulating these processes, however, will require more advanced models
than those discussed herein; you can find more details in a textbook like [176].

Leverett J-Function
To use the relation between capillary pressure and saturation in practical modeling, it is con-
venient to express the capillary–saturation relationship Pc (Sw ) as an explicit or tabulated
function. In the petroleum industry, one usually uses flooding experiments on core samples
from the reservoir to develop empirical models based on observations of the relationship
between average pc and Sw values inside the core models. Each core sample will natu-
rally generate a different capillary curve because of differences in pore-size distribution,
porosity, and permeability. To normalize the measured data, it is common to use a so-called
Leverett J-function [184], which takes the form
,
Pc K
J (Sw ) = . (8.9)
σ cos θ φ
Here, the surface tension σ and the contact angle θ are measured in the laboratory and
√
are specific to a particular rock and fluid system. The scaling factor K/φ is proportional
to the characteristic, effective pore-throat radius. The function J can now be obtained as
a (tabulated) function of Sw by fitting rescaled observed data (pc and sw ) to a strictly
monotone J -shaped function. Then, the resulting function is used to extrapolate capillary
pressure data measured for a given rock to rocks that are similar, but have different per-
meability, porosity, and wetting properties. One cannot expect to find a J function that is
generally applicable because the parameters that affect capillary pressure vary largely with
rock type. Nevertheless, experience has shown that J -curves correlate well for a given rock
type and in reservoir models it is therefore common to derive a J -curve for each specific
rock type (facies)1 that is represented in the underlying geological model.
To set up a multiphase flow simulation, we need the initial saturation distribution inside
the reservoir. If we know the location of the OWC, the saturation higher up in the formation
can be determined by combining (8.7) and (8.9)
,
 ρgh K 
Sw = J −1 .
σ cos θ φ

Other Relationships
In other application areas than petroleum recovery, it is common to use models that express
the capillary pressure directly as an explicit function of the normalized (or effective) water
saturation,

1 By (litho)facies, we mean a part of the rock that is distinguishable by its texture, mineralogy, grain size, and the depositional
environment that produced it.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.1 New Physical Properties and Phenomena 239

Sw − Swmin
Ŝw = . (8.10)
Swmax − Swmin
Here, Swmax and Swmin are the maximum and minimum values the saturation can attain during
the displacement. For the primary drainage shown to the left in Figure 8.4, it is natural to
set Swmax = 1 and Swmin = Swr , whereas Swmax = 1 − Sor and Swmin = Swr for all subsequent
displacements. The following model was proposed by Brooks and Corey [54] to model the
relationship between capillary pressure and water saturation in partially saturated media
(i.e., in the vadoze zone where the two-phase flow consists of water and air; see Section
8.3.5):

(pc /pe )−nb , if pc > pe,
Ŝw = (8.11)
1, pc ≤ pe .
Here, pe is the entry pressure of air and nb ∈ [0.2,5] is a parameter related to the pore-size
distribution. Another classical model is the one proposed by van Genuchten [298]:
  n −mg
Ŝw = 1 + βg pc g , (8.12)

where βg is a scaling parameter related to the average size of pores and the exponents ng
and mg are related to the pore-size distribution.

8.1.4 Relative Permeability

When discussing incompressible, single-phase flow, we saw that the only petrophysical
parameter affecting how fast a fluid flows through a porous medium is the absolution
permeability K that measures the capacity of the rock to transmit fluids, or alternatively the
resistance the rock offers to flow. As described in Chapters 2 and 4, absolute permeability
is an intrinsic property of the rock and does not depend on the type of fluid that flows
through the rock. In reality, this is not true, mainly because of microscale interactions
between rock and fluid may cause particles to move, pore spaces to be plugged, clays
to swell when brought in contact with water, etc. Likewise, liquids and gases may not
necessarily experience the same permeability, because gas does not adhere to the mineral
surfaces in the same way as liquids do. This means that whereas the flow of liquids
is subject to no-slip boundary conditions, gases may experience slippage that gives a
pressure-dependent apparent permeability, which at low flow rates is higher than the
permeability experienced by liquids. This is called the Klinkenberg effect and plays a
substantial role for gas flows in low-permeable, unconventional reservoirs such as coal
seams, tight sands, and shale formations. Herein, we will not consider reservoirs where
these effects are pronounced and henceforth, we assume – as for the incompressible,
single-phase flow models in Chapter 4 – that the absolute permeability K is an intrinsic
quantity.
When more than one phase is present in the pore space, each phase α will experience
an effective permeability Keα that is less than the absolute permeability K. Looking at the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 240 Mathematical Models for Multiphase Flow

conceptual drawing in Figure 8.1, it is easy to see why this is so. The presence of another
phase will effectively present additional “obstacles,” whose interfacial tension offers resis-
tance to flow. Because interfacial tension exists between all immiscible phases, the sum of
all the effective phase permeabilities will generally be less than one, i.e.,

Keα < K.
α

To model this reduced permeability, we introduce a property called relative permeability

[219], which for an isotropic medium is defined as

krα = Kαe /K. (8.13)

Because the effective permeability is always less than or equal to the absolute permeability,
krα will take values in the interval between 0 and 1. For anisotropic media, the relationship
between the effective and absolution permeability may in principle be different for each
component of the tensors. However, it is still common to define the relative permeability as
a scalar quantity postulated to be in the form

Keα = krα K. (8.14)

Relative permeabilities will generally be functions of saturation, which means that for a
two-phase system we can write

krn = krn (Sn ) and krw = krw (Sw ).

It is important to note that the relative permeabilities generally are nonlinear functions of
the saturations, so that the sum of these functions at a specific location (with a specific
composition) is not necessarily equal to one. As for the relationship between saturation
and capillary pressure, the relative permeabilities are strongly dependent on the lithofacies.
It is therefore common practice to associate a unique pair of curves with each rock type
represented in the geological model. Relative permeabilities may also depend on pore-size
distribution, fluid viscosity, and temperature, but these factors are usually ignored in models
of conventional reservoirs.
In simplified models, it is common to assume that krα are monotone functions that
assume unique values in [0,1] for all values Sα ∈ [0,1], so that krα = 1 corresponds to the
case with fluid α occupying the entire pore space, and krα = 0 when Sα = 0; see the left
plot in Figure 8.5. In practice, krα = 0 occurs when the fluid phase becomes immobilized
for Sα ≤ Sαr , giving relative permeability curves as shown in the right plot in Figure 8.5.
The preferential wettability can be deducted from the point where two curves cross. Here,
the curves cross for Sw > 0, which indicates that the system is water-wet.
Going back to our previous discussion of hysteresis, we generally expect relative per-
meability to be different during drainage and imbibition. The drainage curve in Figure
8.6 corresponds to the primary drainage process discussed in connection with Figure 8.4,
in which a non-wetting hydrocarbon phase migrates into a water-wetting porous medium
completely saturated by water. After water has been drained to its irreducible saturation,

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.1 New Physical Properties and Phenomena 241
1 1
krw kro krw kro

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Figure 8.5 Illustration of relative permeabilities for a two-phase system. The left plot shows an
idealized system with no residual saturations, while the right plot shows more realistic curves from a
water-wet system.

1
krw kro

0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1

Figure 8.6 Illustration of relative permeability hysteresis, with drainage curves shown as thin, dashed
lines and imbibition curves shown as thick lines.

water is reinjected to flush out the oil. In this particular illustration, the relative permeability
of water krw exhibits no hysteretic behavior, whereas the imbibition and drainage curves
deviate significantly for kro .

Two-Phase Relative Permeability

It is quite common to use simple analytic relationships to represent relative permeabilities.
These are usually stated using the normalized or effective saturation Ŝw from (8.10). The
simplest model possible is a pure power-law relationship, which is sometimes called a
Corey model,
 n 0  n
krw = Ŝw w kw , kro = 1 − Ŝw o ko0, (8.15)

where the exponents nw,no ≥ 1 and the constants kα0 used for end-point scaling, should be
fit to measured data. Another popular choice is the use of the Brooks–Corey functions

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 242 Mathematical Models for Multiphase Flow
 n +n n  n   n n
krw = Ŝw 1 2 3 , kro = 1 − Ŝw 1 1 − Ŝw 2 3 , (8.16)

where n1 = 2, n2 = 1 + 2/nb , and n4 = 1 gives the Brooks–Corey–Burdine model and

n1 = η, n2 = 1 + 1/nb , and n3 = 2 gives the Brooks–Corey–Mualem model. It is also
possible to derive models that correspond with the van Genuchten capillary functions in
(8.12). In the case that mg = 1 − 1/ng , this reads

 1/m 2
krw = Ŝwκ 1 − (1 − Ŝw g )mg ,
 1/m 2m
(8.17)
kro = (1 − Ŝw )κ 1 − Ŝw g g ,

where the connectivity factor κ is a fitting parameter. This is called the van Genuchten–
Mualem model, whereas the closed-form expressions obtained for mg = 1 − 2/ng ,

 1/m 
krw = Ŝw2 1 − (1 − Ŝw g )mg ,
 1/m m
(8.18)
kro = (1 − Ŝw )2 1 − Ŝw g g ,

is called the van Genuchten–Burdine model.

Three-Phase Relative Permeability

Measuring relative permeability has traditionally been costly and complex. Recently, lab-
oratory techniques have made great progress by using computer tomography and nuclear
magnetic resonance (NMR) to scan the test cores where the actual phases are being dis-
placed. Although standard experimental procedures exist for measuring relative permeabil-
ities in two-phase systems, there is nevertheless still significant uncertainty concerning the
relevance of the experimental values found, and it is difficult to come up with reliable data
to be used in a simulator. This is mainly due to boundary effects. Particularly for three-
phase systems, there is a lot of uncertainty in reported data (see the review by [29]) and it
is not unfair to say that no reliable experimental technique exists. Thus, three-phase rela-
tive permeabilities are usually modeled using two-phase measurements, for which several
theoretical models have been proposed. Most of them are based on an idea first proposed
by Stone [282], and seek to combine sets of two-phase relative permeabilities to give three-
phase relationships.
In Stone’s original model, which was developed for water-wet porous media and is
commonly referred to as the Stone I model, the idea was that the relative permeability
of water and gas would be identical to the models measured in water/oil displacements and
oil/gas displacements, respectively. In the following, I use the convention that k wo denotes
a relationship derived from water/oil displacements and k og are relationships derived from
oil/gas displacements. Hence, we have that

go
krw (Sw ) = krw
wo
(Sw ) krg (Sg ) = krg (Sg ). (8.19)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.2 Flow Equations for Multiphase Flow 243

The three-phase relative permeability of oil, on the other hand, depends nonlinearly on both
water and gas saturations,
wo (S ) k (S ) og
Ŝo kro w ro g
kro (Sw,Sg ) = wo (S )
. (8.20)
kro wc 1 − Ŝw 1 − Ŝg

Here, we have introduced the following scaled quantities

So − Som
Ŝo = , for So ≥ Som
1 − Swc − Som
Sw − Swc
Ŝw = , for Sw ≥ Swc
1 − Swc − Som
Sg
Ŝg = .
1 − Swc − Som
Here, Swc is the connate water saturation and Som is the minimum oil saturation. To close
the model, we can for instance use the expression for Som given by Fayers and Matthews
[109]

Som = αSorw + (1 − α)Sorg , α = 1 − Sg /(1 − Swc − Sorg ), (8.21)

where Sorw and Sorg denote the residual oil saturations for waterflooding and gasflooding,
respectively.
Stone later proposed an alternative expression for the relative permeability of oil called
the Stone II model [283]. A number of modifications and alternative formulations have been
developed over the years; see, e.g., [29] for a review. MRST implements a different model
that uses a linear relationship to interpolate the two-phase curves. This is the same model
as the default choice in the commercial ECLIPSE simulator [271], and will be discussed in
more detail in Section 11.3.

8.2 Flow Equations for Multiphase Flow

Having introduced the new physical parameters and key phenomena that characterize mul-
tiphase flow of immiscible fluids, we are in a position to develop mathematical models
describing multiphase flow. To this end, we follow more or less the same steps as we did for
single-phase flow in Section 4.2. Stating the generic equations describing multiphase flow
is straightforward and will result in a system of partial differential equations that contains
more unknowns than equations. The system must hence be extended with constitutive
equations to relate the various physical quantities, as well as boundary conditions and
source terms (see Sections 4.3.1 and 4.3.2) that describe external forces driving the flow.
However, as indicated in the introduction of the chapter, the generic flow model has a
complex mathematical character and contains delicate balances of various physical forces,
whose characteristics and individual strengths vary a lot across different flow regimes. To
reveal the mathematical character, but also to make models that are more computationally

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 244 Mathematical Models for Multiphase Flow

tractable in practical simulations, you need to develop more specific models that apply
under certain assumptions on the flow regimes.

8.2.1 Single-Component Phases

To develop a generic system of flow equations for multiphase flow, we use the fundamental
principle of mass conservation. For N immiscible fluid phases that each consists of a single
component, we write one conservation equation per phase,
∂ 
φρα Sα + ∇ · ρα vα = ρα qα . (8.22)
∂t
Here, each phase can contain multiple chemical species, but these can be considered as a
single component since there is no transfer between the phases so that their composition
remains constant in time.
As for the flow of a single fluid, the primary constitutive relationship used to form a
closed model is Darcy’s law (4.2), which can be extended to multiphase flow by using the
concept of relative permeabilities discussed above
Kkrα
vα = − (∇pα − gρα ∇z). (8.23)
μα
This extension of Darcy’s law to multiphase flow is often attributed to Muskat and Wyckoff
[219] and has only been rigorously derived from first principles in the case of two fluid
phases. Equation (8.23) must therefore generally be considered as being phenomenological.
Darcy’s law is sometimes stated with the opposite sign for the gravity term, but herein we
use the convention that g is a positive constant. Likewise, it is common to introduce phase
mobilities λα = Kkrα /μα or relative phase mobilities λα = λα K to simplify the notation
slightly. From the discussion in the previous section, we also have an additional closure
relationship stating that the saturations sum to zero (8.1), as well as relations of the form
(8.8) that relate the pressures of the different phases by specifying the capillary pressures
as functions of the fluid saturations.
Most commercial reservoir simulators compute approximate solutions by inserting the
multiphase Darcy equations (8.23) into (8.22), introducing functional relationships for
capillary pressure and how ρα and φ depend on phase pressure, and then discretizing the
resulting conservation equations more or less directly. If we use the discrete derivative
operators from Section 4.4.2, combined with a backward discretization of the temporal
derivatives, the resulting system of fully implicit, discrete equations for phase α reads
(φS α ρ α )n+1 − (φS α ρ α )n
+ div(ρv)n+1
α = (ρq)n+1
α , (8.24a)
t n
Kk rα  
v n+1
α = − n+1 grad(p n+1 α ) − gρ α grad(z) ,
n+1
(8.24b)
μα
Here, φ,S α ,p α ∈ Rnc denote vectors with one porosity value, one saturation, and one
pressure value per cell, respectively, whereas v α is the vector of fluxes for phase α per

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.2 Flow Equations for Multiphase Flow 245

face, and so on. For properties depending on pressure and saturation, we have for sim-
plicity not introduced any notation to distinguish whether these are evaluated in cells or
at cell interfaces. The superscript refers to discrete times and t denotes the associated
time step.
The main advantage of using this direct discretization is that it will generally give a
reasonable approximation to the true solution as long as we are able to solve the resulting
nonlinear system at each time step. To do this, we must pick as many unknowns as we
have phases and perform some kind of linearization; we will come back in more detail
later in the book, and discuss implementation of (8.24) in Chapter 11. Here, we simply
observe that there are many ways to perform the linearization: we can choose the phase
saturations as primary unknowns, the phase pressures, the capillary pressures, or some
combinations thereof. How difficult it will be to solve for these unknowns will obviously
depend on the coupling between the two equations and nonlinearity within each equation.
The main disadvantages of discretizing (8.22) directly are that the general system conceals
its mathematical nature and that the resulting equations are not well-posed in the case of
an incompressible system. In Section 8.3, we will therefore go back to the continuous
equations and analyze the mathematical nature of the resulting system for various choices
of primary variables and simplifying assumptions.

8.2.2 Multicomponent Phases

In many cases, each phase may consist of more than one chemical species that are mixed
at the molecular level and generally share the same velocity (and temperature). This type
of flow differs from the immiscible case, since dispersion and Brownian motion will cause
the components to redistribute if there are macroscale gradients in the mass fractions. The
simplest way to model this is through a linear Fickian diffusion,

Jα = −ρα Sα Dα ∇cα , (8.25)

where cα is the mass fraction of component  in phase α, ρα is the density of phase α,
Sα is the saturation of phase α, and Dα is the diffusion tensor for component  in phase
α. Likewise, the chemical species may interact and undergo chemical reactions, but a
description of this is beyond the scope of this book.
For multicomponent, multiphase flow, we are, in principle, free to choose whether we
state the conservation of mass for components or for fluid phases. However, if we choose
the latter, we will have to include source terms in our balance equations that account for
the transfer of components between the phases. This can be a complex undertaking and
the standard approach is therefore to develop balance equations for each component. For
a system of N fluid phases and M chemical species, the mass conservation for component
 = 1, . . . ,M reads

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 246 Mathematical Models for Multiphase Flow
∂      
φ cα ρα Sα + ∇ · cα ρα vα + Jα = cα ρα qα , (8.26)
∂t α α α

where vα is the superficial phase velocity and qα is the source term. The system is closed
in the same way as for single-component phases, except that we now also have to use that
the mass fractions sum to zero, (8.3).
The generic system (8.23)–(8.26) can also describe miscible displacements in which
the composition of the fluid phases changes when the porous medium undergoes pressure
and saturation changes and one has to account for all, or a large majority, of the chemical
species that are present in the flow system. For immiscible or partially miscible systems, it is
common to introduce simplifications by lumping multiple species into pseudo-components,
or even disregard that a fluid phase may be composed of different chemical species, as we
will see next.

8.2.3 Black-Oil Models

The most common approach to simulate oil and gas recovery is to use the so-called black-
oil equations, which essentially constitute a special multicomponent model with no diffu-
sion among the components. The name refers to the assumption that the chemical species
can be lumped together to form two pseudo-components at surface conditions, a heavy
hydrocarbon component called “oil” and a light hydrocarbon component called “gas.” At
reservoir conditions, the two components can be partially or completely dissolved in each
other depending on pressure and temperature, forming either one or two phases: a liquid
oleic phase and a gaseous phase. In addition to the two hydrocarbon phases, the framework
includes an aqueous phase that in the simplest models of this class is assumed to consist of
only water. In more comprehensive models, the hydrocarbon components are also allowed
to dissolve in the aqueous phase and the water component may be dissolved or vaporized in
the two hydrocarbon phases. The composition of each hydrocarbon component, however,
remains constant for all times.
We henceforth assume three phases (oleic, gaseous, and aqueous) and three components
(oil, gas, and water). By looking at our model (8.23)–(8.26), we see that we so far have
introduced 27 unknown physical quantities: 9 mass fractions cα and 3 of each of the
following six quantities: ρα , Sα , vα , pα , μα , and krα . In addition, the porosity will typically
depend on pressure as discussed in Section 2.4.1. To determine these 27 unknowns, we have
3 continuity equations (8.26), an algebraic relation for the saturations (8.1), 3 algebraic
relations for the mass fractions (8.3), and Darcy’s law (8.23) for each of the 3 phases.
Altogether, this constitutes only 10 equations. Thus, we need to add 17 extra closure
relations to make a complete model.
The first five of these can be obtained immediately from our discussion in Sections 8.1.3
and 8.1.4, where we saw that the relative permeabilities krα are functions of the phase
saturations. Likewise, capillary pressures are functions of saturation and can be used to
relate phase pressures as follows:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.2 Flow Equations for Multiphase Flow 247

po − pw = Pcow (Sw,So ), pg − po = Pcgo (So,Sg ).

The required functional forms are normally obtained from a combination of physical exper-
iments, small-scale numerical simulations, and analytical modeling based on bundle-of-
tubes arguments, etc. Viscosities are either assumed to be constant or can be established as
pressure-dependent functions through laboratory experiments, which gives us another three
equations. The remaining nine closure relations are obtained from mixture rules and PVT
models that are generalizations of the equations of state discussed in Section 4.2.
By convention, the black-oil equations are formulated as conservation for the oil, water,
and gas components. To this end, we employ a simple PVT model that uses pressure-
dependent functions to relate fluid volumes at reservoir and surface conditions. Specifically,
we use the formation-volume factors B = V /Vs that relate the volumes V and Vs
occupied by a bulk of component  at reservoir and surface conditions, respectively. The
formation-volume factors, and their reciprocal shrinkage factors b = Vs /V , which
I personally prefer to use for notational convenience, are assumed to depend on phase
pressure. In dead-oil systems, the oil is at such a low pressure that it does not contain
any gas or has lost its volatile components (which presumably have escaped from the
reservoir). Neither of the components are therefore dissolved in the other phases at reservoir
conditions. In live-oil systems, gas is dissolved in the oleic phase. When underground,
hydrocarbon molecules are mostly in liquid phase, but gaseous components are liberated
when the oil is pumped to the surface. The solubility of gas in oil is usually modeled through
the pressure-dependent solution gas–oil ratio, Rs = Vgs /Vos , defined as the volume of
gas, measured at standard conditions, that at reservoir conditions is dissolved in a unit of
stock-tank oil. In condensate reservoirs, oil is vaporized in the gaseous phase, so that when
underground, condensate oil is mostly a gas, but condenses into a liquid when pumped to
the surface. The solubility of oil in gas is modeled as the pressure-dependent vaporized
oil–gas ratio Rv , defined as the amount of surface condensate that can be vaporized in a
surface gas at reservoir conditions. Using this notation,2 the black-oil equations for a live-
oil system reads,
    
∂t φ bo So + bg Rv Sg + ∇ · bo vo + bg Rv vg − bo qo + bg Rv qg = 0,
 
∂t φbw Sw + ∇ · bw vw − bw qw = 0,
    
∂t φ bg Sg + bo Rs So + ∇ · bg vg + bo Rs vo − bg qg + bo Rs qo = 0.
To model enhanced oil recovery, we must extend the system with additional components
representing chemical or microbial species that are added to the injected fluids to mobilize
immobile oil by altering wettability and/or to improve sweep efficiency and hence the
overall displacement of mobile oil. One may also model species dissolved in the resident
fluids. Likewise, the black-oil equations may be expanded by an additional solid phase
to account for salts and other minerals that precipitate during hydrocarbon recovery, and
possibly also extended to include an energy equation that accounts for temperature effects.

2 The convention in the petroleum literature is to use B rather than b .

α α

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 248 Mathematical Models for Multiphase Flow

We will return to a more detailed discussion of the general case of three-phase flow
with components that may transfer between phases later in Chapter 11. In the rest of this
chapter, we continue to study the special case of two single-component fluid phases with
no interphase mass transfer.

8.3 Model Reformulations for Immiscible Two-Phase Flow

For the special case of two immiscible fluids, found in a wetting w and a non-wetting n
phase, the general system of flow equations (8.22) simplifies to
∂
(φSw ρw ) + ∇ · (ρw vw ) = ρw qw,
∂t
∂
(φSn ρn ) + ∇ · (ρn vn ) = ρn qn . (8.27)
∂t
In the following, we discuss various choices of primary variables and how these choices
affect the mathematical structure of the resulting coupled system of nonlinear differential
equations.

8.3.1 Pressure Formulation

If we choose the phase pressures pn and pw as primary unknowns, we must express the
saturations Sn and Sw as functions of pressure. To this end, we assume that the capillary
pressure has a unique inverse function Ŝw = Pc−1 (pc ) (see (8.8)) so that we can write
Sw = Ŝw (pn − pw ), Sn = 1 − Ŝw (pn − pw ). (8.28)
Then, we can reformulate (8.27) as
∂  ρ Kk (Ŝ )  
w rw w
(φρw Ŝw ) + ∇ · ∇pw − ρw g∇z = ρw qw,
∂t μw
∂  ρn Kkrn (Ŝw )  
(φρn (1 − Ŝw )) + ∇ · ∇pn − ρn g∇z = ρn qn . (8.29)
∂t μn
This system is unfortunately highly coupled and strongly nonlinear. The strong coupling
comes from the fact that the difference in the primary variables pn − pw enters the com-
putation of Ŝw in the accumulation terms and also appears in the composite functions
krw (Ŝw (·)) and krn (1 − Ŝw (·)) used to evaluate the relative permeabilities. As an exam-
ple of the resulting nonlinearity, we can look at the van Genuchten model for capillary
and relative permeabilities, (8.12) and (8.17). Figure 8.7 shows the capillary pressure and
relative permeabilities as functions of S as well as the inverse of the capillary pressure
and relative permeabilities as functions of pn − pw . Whereas the accumulation function is
nonlinear, this nonlinearity is further accentuated when used inside the nonlinear relative
permeability functions. The pressure formulation was used in the simultaneous solution
scheme originally proposed by Douglas, Peaceman, and Rachford [86] in 1959, but has

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.3 Model Reformulations for Immiscible Two-Phase Flow 249
1
Pc (S)

8 0.8 krw
kro

6 0.6

4 0.4

2 0.2

0
0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8

1 1
Ŝw krw (Ŝw )
0.8 0.8 kro (Ŝw )

0.6 0.6

0.4 0.4

0.2 0.2

0
0 2 4 6 8 10 0 2 4 6 8 10

Figure 8.7 Capillary pressure and relative permeabilities as functions of S for the van Genuchten
model for βg = 1, mg = 2/3, and κ = 1/2. The lower row shows saturation and relative
permeabilities as function of pc = pn − pw .

later been superseded by other formulations that reduce the degree of coupling and improve
the nonlinear nature of the equation.

8.3.2 Fractional-Flow Formulation in Phase Pressure

The strong coupling and much of the nonlinearity seen in the previous formulation can be
eliminated if we instead express the system in terms of one phase pressure and one phase
saturation. A common choice is to use pn and Sw , which gives the following system

∂  ρ Kk  
w rw
(φSw ρw ) + ∇ · ∇pn − ∇Pc (Sw ) − ρw g∇z = ρw qw,
∂t μw
∂  ρ Kk  
n rn
(φ(1 − Sw )ρn ) + ∇ ∇pn − ρn g∇z = ρn qn . (8.30)
∂t μn
To further develop this system of equations, it is common to expand the derivatives to
introduce rock and fluid compressibilities, as discussed for the single-phase flow equation
in Section 4.2. First, however, we will look closer at the special case of incompressible
flow and develop the fractional-flow formulation, which will enable us to further expose
the mathematical structure of the system.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 250 Mathematical Models for Multiphase Flow

Incompressible Flow
For incompressible flow, the porosity φ only varies in space and the fluid densities ρα are
constant. Using these assumptions, we can simplify the mass-balance equations to be on
the form
∂Sα
φ + ∇ · vα = qα . (8.31)
∂t
To derive the fractional-flow formulation, we start by introducing the total Darcy velocity,
which we can express in terms of the pressure for the non-wetting phase,
v = vn + vw = −λn ∇pn − λw ∇pw + (λn ρn + λw ρw )g∇z
= −(λn + λw )∇pn + λw ∇pc + (λn ρn + λw ρw )g∇z. (8.32)
We then add the two continuity equations and use the fact that Sn + Sw = 1 to derive a
pressure equation without temporal derivatives
vn + vw ) = ∇ · v = qn + qw
∇ · ( (8.33)
If we now define the total mobility λ = λn + λw = λK and total source q = qn + qw ,
insert (8.32) into (8.33), and collect all terms that depend on pressure on the left-hand side
and all other terms on the right-hand side, we obtain an elliptic Poisson-type equation
 
− ∇ · (λK∇pn ) = q − ∇ λw ∇pc + (λn ρn + λw ρw )g∇z . (8.34)
This pressure equation (or flow equation) is essentially on the same form as the equation
(4.10) governing single-phase, incompressible flow, except that both the variable coefficient
and the right-hand side now depend on saturation through the relative mobility λ and the
capillary pressure function pc = Pc .
To derive an equation for Sw , we multiply each phase velocity by the relative mobility
of the other phase, subtract the result, and use Darcy’s law to obtain
λn vw − λw vn = λ
vw − λw v
= −λn λw K(∇pw − ρw g∇z) + λw λn K(∇pn − ρn g∇z)

= λw λn K ∇pc + (ρw − ρn )g∇z].
If we now solve for vw and insert into (8.31) for the wetting phase, we obtain what is
commonly referred to as the saturation equation (or transport equation)
∂Sw   
φ + ∇ · fw v + fw λn Kρg∇z = qw − ∇ · fw λn KPc ∇Sw . (8.35)
∂t
Here, ρ = ρw − ρn , and we have introduced the fractional flow function fw = λw /(λn +
λw ), which measures the fraction of the total flow that consists of the wetting fluid.
Equation (8.35) is parabolic and accounts for the balance of three different forces:
viscous advection fw v, gravity segregation (fw λn )Kρg∇z, and capillary forces
fw λn KPc ∇Sw . The first two terms both involve a first-order derivative and hence have
a hyperbolic character, whereas the capillary term contains a second-order derivative

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.3 Model Reformulations for Immiscible Two-Phase Flow 251

0.4
μw = 5μn μw = 5μn
μw = μn μw = μn
0.8
5μw = μn 5μw = μn
0.3

0.6

0.2
0.4

0.1
0.2

0 0
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8

Figure 8.8 Fractional flow function fw (left) and gravity segregation function λn fw for the van
Genuchten–Mualem model (8.17) with mg = 2/3, κ = 1/2, and ρ > 0.

and hence has a parabolic character. The negative sign in front of the capillary term
is misleading; the overall term is positive since Pc < 0 and hence gives a well-posed
problem. Moreover, since λn is zero at the end point, the capillary term disappears at this
point and the parabolic equation thus degenerates to being hyperbolic. This, in turn, implies
that a propagating displacement front – in which a wetting fluid displaces a non-wetting
fluid – has finite speed of propagation, and not the infinite speed of propagation that is
typical for parabolic problems.
Figure 8.8 shows the functional form of the two hyperbolic terms for the van Genuchten–
Mualem model (8.17). The fractional flow function fw is monotonically increasing and has
a characteristic S-shape. If the flow is dominated by viscous terms so that gravity and cap-
illary forces are negligible, the flow will always be cocurrent with both phases flowing in
the direction of v because the derivative of fw is positive. The gravity segregation function,
on the other hand, has a characteristic bell shape and describes the upward movement of
the lighter phase and downward movement of the heavier phase. (The function is turned
upside down if ρw < ρn .) If gravity is sufficiently strong, it may introduce countercurrent
flow, with the fluid phases flowing in opposite directions. On a field scale, capillary forces
are typically small and hence the overall transport equation will have a strong hyperbolic
character. Neglecting the capillary term is generally a good approximation and gives a
purely hyperbolic saturation equation. We will return to this model later and discuss a few
examples in the special case of purely hyperbolic 1D flow.
The coupling between the elliptic pressure equation and the parabolic saturation equa-
tion is much weaker than the coupling between the two continuity equations in the for-
mulation (8.29) with two pressures as primary unknowns. In the pressure equation (8.34),
the coupling to saturation appears explicitly in the effective mobility, which makes up the
variable coefficient in the Poisson problem, and on the right-hand side through the phase
mobilities and the derivative of the capillary function. In (8.35), on the other hand, the
saturation is only indirectly coupled to the pressure through the total Darcy velocity. In

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 252 Mathematical Models for Multiphase Flow

typical waterflooding scenarios, saturation variations are relatively smooth in most of the
domain, except at the interface between the invading and the displaced fluids, and hence the
pressure distribution is only mildly affected by the evolving fluid saturations. In fact, the
pressure and transport equations are completely decoupled in the special case of linear rel-
ative permeabilities, no capillary pressure, and equal fluid densities and viscosities. These
weak couplings are exploited in many efficient numerical methods, which use various forms
of operator splitting to solve pressure and fluid transport in separate steps. This is also our
method of choice for incompressible flow in MRST, and will be discussed in more details
in Chapter 10.
The reformulation discussed above can easily be extended to N immiscible phases, and
will result in one pressure equation and a (coupled) system of N − 1 saturation equations.
It is also possible to develop fractional-flow formulations for multicomponent flow without
mass exchange between the phases.

Compressible Flow
To develop a fractional-flow formulation in the compressible case, we start by expanding
the accumulation term in (8.22) and then use fluid compressibilities to rewrite the derivative
of the phase densities
∂ ∂φ dρα ∂pα ∂Sα
φρα Sα = ρα Sα + φSα + φρα
∂t ∂t dpα ∂t ∂t
∂φ ∂pα ∂Sα
= ρα Sα + φρα Sα cα + φρα .
∂t ∂t ∂t
If we now insert this into (8.22), divide each equation by ρα , and sum the equations, we
obtain
∂φ ∂pn ∂pw 1  1 
+ φcn Sn + φcw Sw + ∇ · ρn vn + ∇ · ρw vw = qt , (8.36)
∂t ∂t ∂t ρn ρw
where qt = qn + qw . To explore the character of this equation, let us for the moment
assume that capillary forces are negligible so that pn = pw = p and that the spatial density
variations are so small that we can set ∇ρα = 0. If we now introduce rock compressibility
cr = d ln(φ)/dp, (8.36) simplifies to
∂p
φc − ∇ · (λK∇p) = q̂. (8.37)
∂t
Here, we have introduced the total compressibility c = (cr + cn Sn + cw Sw ) and a source
function q̂ that accounts both for volumetric sources and pressure variations with depth.
Equation (8.37) is clearly parabolic, but simplifies to the elliptic Poisson equation (8.34)
for incompressible flow (c = 0). Fluid compressibilities tend to decrease with increasing
pressure, and we should therefore expect that the pressure equation has a strong ellip-
tic character at conditions usually found in conventional reservoirs during secondary and
tertiary production, in particular for weakly compressible systems. On the other hand, the
pressure evolution will usually have a more pronounced parabolic character during primary

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.3 Model Reformulations for Immiscible Two-Phase Flow 253

depletion, in gas reservoirs, and in unconventional reservoirs, where K may be in the range
of one micro Darcy or less.
These observations remain true also when capillary forces and spatial density varia-
tions are included. However, including these effects makes the mathematical structure of
the equation more complicated. In particular, the pressure equation becomes a nonlinear
parabolic equation if the spatial variations in density cannot be neglected. To see this, let us
once again neglect capillary forces and consider the second-last term on the left-hand side
of (8.36),
1 1
∇ · (ρn v) = ∇ · vn + vn · ∇ρn
ρn ρn

= ∇ · vn + vn · cn ∇pn .

By using Darcy’s law, we can transform the inner product between the phase flux and the
pressure gradient to involve a quadratic term in ∇pn or vn ,

vn · cn ∇pn = −λn (∇pn − ρn g∇z) · (cn ∇pn )
= cn vn · (−λ−1
n vn + ρn g∇z).

The last term on the left-hand side of (8.36) has a similar form, and hence we obtain
a nonlinear equation if at least one of the fluid phases has significant spatial density
variations.
Going back to (8.36), we see that the equation contains both phase pressures and hence
cannot be used directly alongside with a saturation equation. There are several ways to
formulate a pressure equation so that it only involves a single unknown pressure. We can
either pick one of the phase pressures as the primary variable, or introduce an average
pressure pa = (pn + pw )/2, as suggested in [246]. With pn as the primary unknown, the
full pressure equation reads
∂pn  1  1  
φc − ∇ · ρn λn ∇pn + ∇ · ρw λw ∇pn
∂t ρn ρw
1  ∂Pc
= qn + qw − ∇ · ρw λw ∇Pc + cw Sw
ρw ∂t
1  2 1  2
− ∇ · ρn λn g∇z − ∇ · ρw λw g∇z . (8.38)
ρn ρw
Likewise, with pa as the primary unknown, we get
∂pa  1  1  
φc − ∇ · ρn λn ∇pa + ∇ · ρw λw ∇pa
∂t ρn ρw
1 1  1   ∂Pc
= qt + ∇ · ρn λn − ∇ · ρw λw ∇Pc + 12 cw Sw
2 ρn ρw ∂t
∂P c 1  1 
− 12 cn Sn − ∇ · ρn2 λn g∇z − ∇ · ρw2 λw g∇z . (8.39)
∂t ρn ρw

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 254 Mathematical Models for Multiphase Flow

The same type of pressure equation can be developed for cases with more than two fluid
phases. Also in this case, we can define a total velocity and use this to obtain N −1 transport
equations in fractional-flow form.
The saturation equation for compressible flow is obtained exactly in the same way as in
the incompressible case,
∂   
φρw Sw + ∇ · ρw fw v + λn ρg∇z
∂t

= ρw qw − ∇ · ρw fw λn Pc ∇Sw . (8.40)

However, whereas total velocity was the only quantity coupling the saturation equation to
the fluid pressure in the incompressible case, we now have coupling also through porosity
and density, which may both depend on pressure. This generally makes the compressible
case more nonlinear and challenging to solve than the incompressible case.

8.3.3 Fractional-Flow Formulation in Global Pressure

The formulation in phase pressure discussed above has a relatively strong coupling between
pressure and saturation in the presence of capillary forces. The strong coupling is mainly
caused by the ∇pc term on the right-hand side of (8.34). Let us therefore go back and see if
we can eliminate this term by making a different choice of primary variables. To this end,
we start by looking at the definition of the total velocity, which couples the pressure to the
saturation equation:

v = −λn ∇pn − λw ∇pw + ρn λn + ρw λw )g∇z

= −(λn + λw )∇pn + λw ∇(pn − pw ) + ρn λn + ρw λw )g∇z
 
= −λ ∇pn − fw ∇pc ) + ρn λn + ρw λw )g∇z.

If we now introduce a new pressure variable p, called the global pressure [62], defined
so that ∇p = ∇pn − fw ∇pc , we see that the total velocity can be related to the global
pressure through an equation that looks like Darcy’s law

v = −λ ∇p − (ρw fw + ρn fn )g∇z . (8.41)

By adding the continuity equations (8.31), and using (8.41), we obtain an elliptic Poisson-
type equation for the global pressure
  
− ∇ · λK ∇p − (ρw fw + ρn fn )g∇z = qt . (8.42)

The advantages of this formulation is that it is very simple and highly efficient in the
incompressible case, where pressure often is treated as an immaterial property and one is
mostly interested in obtaining a velocity field for the saturation equation. The disadvantages
are that it is not obvious how one should specify and interpret boundary conditions for the
global pressure and that global-pressure values cannot be used directly in well models (of
Peaceman type). The global-pressure formulation can also be extended to variable densities

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.3 Model Reformulations for Immiscible Two-Phase Flow 255

and to three-phase flow; the interested reader should consult [67, 65] and references therein.
The global formulation is often used by academic researchers because of its simplicity, but
is, to the best of my knowledge, hardly used for practical simulations in the industry.

8.3.4 Fractional-Flow Formulation in Phase Potential

The two fractional-flow formulations discussed above are generally well suited to study
two-phase immiscible flow with small or negligible capillary forces. In such cases, flow is
mainly driven through high-permeable regions, and flow paths are to a large extent deter-
mined by the permeability distribution, the shape of the relative permeability functions,
and density differences (which determine the relative importance of gravity segregation). In
other words, the flow is mainly governed by viscous forces and gravity segregation, which
constitute the hyperbolic part of the transport equation, whereas capillary forces mostly
contribute to adjust the width of the interface between the invading and displaced fluids.
For highly heterogeneous media with strong contrasts in capillary functions, on the other
hand, the capillary forces may have pronounced impact on flow paths by enhancing cross-
flow in stratified media or reducing the efficiency of gravity drainage. Likewise, because
capillary pressure is continuous across geological interfaces, differences in capillary func-
tions between adjacent rocks of different type may introduce (strong) discontinuities in the
saturation distribution.
An alternative fractional-flow formulation was proposed by Hoteit and Firoozabadi
[136] to study heterogeneous media with strong contrasts in capillary functions. The
formulation is similar to the ones discussed above, except that we now work with fluid
potentials rather than fluid pressures. The fluid potential for phase α and the capillary
potential are given by
α = pα − ρα gz, (8.43)

c = n − w = pc + (ρw − ρn )gz. (8.44)

By manipulating the expression for the total velocity, we can write it as a sum of two new
velocities, which each is given by the gradient of w and c , respectively:
   
v = −λn ∇pn − ρn g∇z − λw ∇pw − ρn g∇z
   
= −λn ∇pw + ∇pc − (ρn − ρw + ρw )g∇z − λw ∇pw − ρn g∇z
   
= −(λn + λw ) ∇pw − ρw g∇z − λn ∇pc + (ρw − ρn )g∇z
= −λ∇w − λn ∇c = va + vc .
Here, we observe that
λw
vw = −λw ∇w = − λK∇w, = fw va
λ
so that the velocity a represents the same phase differential as the phase velocity of the
wetting phase. The only exception is that in va , the potential gradient is multiplied by the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 256 Mathematical Models for Multiphase Flow

total mobility, which is a smoother and less varying function than the mobility λw of the
water phase. This means that the saturation dependence of va is less than for the total
velocity v.
Proceeding analogously as above, we can derive the following coupled system for two-
phase, immiscible flow:
 
−∇ · λK∇w = q + ∇ · λn K∇c (8.45)
∂ 
φ + ∇ · fw va = qw . (8.46)
∂t
Compared with the two other fractional-flow formulations, this system is simpler in the
sense that capillary forces are only accounted for in the pressure equation. Equally impor-
tant, the saturation equation is purely hyperbolic and only contains advective flow along
va . This makes the saturation much simpler to solve, since one does not have to resolve
delicate balances involving nonlinear functions for gravity segregation and capillary forces.
One can thus easily employ simple upwind discretizations and highly efficient solvers that
have been developed for purely co-current flow, like streamline methods [79] or methods
based on optimal ordering [220, 196]. On the other hand, the advective velocity va may
vary significantly with time and hence represent a strong coupling between pressure and
transport.

8.3.5 Richards’ Equation

Another special case arises when modeling the vadose or unsaturated zone, which extends
from the ground surface to the groundwater table, i.e., from the top of the earth to the depth
at which the hydrostatic pressure of the groundwater equals one atmosphere. Soil and rock
in the vadose zone will generally contain both air and water in its pores, and the vadose zone
is the main factor that controls the movement of water from the ground surface to aquifers,
i.e., the saturated zone beneath the water table. In the vadose zone, water is retained by a
combination of adhesion and capillary forces, and in fine-grained soil one can find pores
that are fully saturated by water at a pressure less than one atmosphere.
Because of the special conditions in the vadose zone, the general flow equations mod-
eling two immiscible phases can be considerably simplified. First of all, we can expect
that any pressure differences in air will be equilibrated almost instantaneously relative to
water, since air typically is much less viscous than water. (At a temperature of 20 ◦ C, the
air viscosity is approximately 55 times smaller than the water viscosity.) Secondly, air will
in most cases form a continuous phase so that all parts of the pore space in the vadose zone
are connected to the atmosphere. If we neglect variations in the atmospheric pressure, the
pore pressure of air can therefore be assumed to be constant. For simplicity, we can set
the atmospheric pressure to be zero, so that pa = 0 and pc = pa − pw = −pw . This, in
turn, implies that the water saturation (and the water relative permeability) can be defined
as functions of water pressure using one of the models for capillary pressure presented in
Section 8.1.3. Inserting all of this into (8.27), we obtain

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.3 Model Reformulations for Immiscible Two-Phase Flow 257
∂   
φρw Sw (pw ) + ∇ · ρw λrw (pw )K ∇pw − ρw g∇z = 0.
∂t
If we further expand the accumulation term, neglect spatial gradients of the water density,
we can divide the above equation by ρw to obtain the so-called generalized Richards’
equation
∂pw  
+ ∇ · λrw (pw )K∇(pw − ρgz) = 0,
Cwp (pw ) (8.47)
∂t
where Cwp = cw θw + d(θw )/dpw is a storage coefficient and θw = φSw is the water
content. In hydrology, it is common to write the equation in terms of the pressure head
∂hw  
+ ∇ κw krw (hw )∇(hw − z) = 0,
Cwh (hw ) (8.48)
∂t
where Cwh = ρw gCwp and κw is the hydraulic conductivity of water (see Section 4.1 on
page 113). If we further neglect water and rock compressibility, we obtain the pressure
form of the classical Richards’ equation [264]
∂hw   dθw
Cch (hw ) + ∇ κw krw (hw )∇(hw − z) = 0, Cch = . (8.49)
∂t dhw
The corresponding equation for the water flux
vw = −κs krw (hw )∇(hw − z) (8.50)
was suggested earlier by Buckingham [55] and is often called the Darcy–Buckingham
equation. Richards’ equation can also be expressed in two alternative forms,
∂θw  
+ ∇ · κw krw (hw )∇(hw − z) = 0,
∂t
∂θw  
+ ∇ D(θw )∇θw = κw krw (θw )∇z,
∂t
where D(θw ) = dh w
dθw κw krw (θw ) is the hydraulic diffusivity tensor. The latter, called the
water-content form, is generally easier to solve numerically than the other forms, but
becomes infinite for fully saturated conditions.
Various forms of Richards’ equation are widely used to simulate water flow in the vadose
zone. However, one should be aware that this class of models has important limitations.
First of all, because air pressure is assumed to be constant, these models cannot describe
air flow, and also assume that the air phase in the whole pore space is fully connected to
the atmosphere. This may not be the case if the porous medium contains heterogeneities
blocking the air flow, e.g., in the form of layers that are almost impermeable to air. Likewise,
the derivation of the equations assumed that the mobility of air is infinite compared with
water, which may not be the case if the relative permeability of air is much smaller than that
of water. If this is not the case, one should expect significant discrepancies in simulation
results obtained from Richards’ equations and the general two-phase flow equations.
The official release of MRST does not yet have any implementation of Richards’ equa-
tions, but work is underway to develop a new module [299].

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 258 Mathematical Models for Multiphase Flow

8.4 The Buckley–Leverett Theory of 1D Displacements

To shed more light into the physical behavior of two-phase immiscible flow systems, we
derive analytical solutions in a few simple cases. As part of this, you will also get a
glimpse of the (wave) theory of hyperbolic conservation laws. This theory plays a very
important role in the mathematical and numerical analysis of multiphase flow models,
but is not described in great detail herein for brevity. If you want to learn more of this
interesting topic, there are many good mathematical textbooks you can consult, e.g., [76,
134, 183, 293, 294]. Chapter 9 also introduces a few central concepts and briefly discusses
development of specialized numerical methods.

8.4.1 Horizontal Displacement

As a first example, we consider incompressible displacement in a 1D homogeneous and
horizontal medium, x ∈ [0,∞), with inflow at x = 0. In the absence of capillary forces,
the two phase pressures are equal and coincide with the global pressure p. The pressure
equation simplifies to
v  (x) = q, v(x) = −λ(x)p (x).
If we assume that there are no volumetric source terms, but a constant inflow rate at x = 0, it
follows by integration that v(x) is constant in the domain [0,∞). Under these assumptions,
the saturation equation (8.35) simplifies to,
∂S v ∂f (S)
+ = 0. (8.51)
∂t φ ∂x
Let us use this equation to study the propagation of a constant saturation value. To this end,
we consider the total differential of S(x,t),
∂S ∂S
dt +
0 = dS = dx,
∂t ∂x
which we can use to eliminate ∂S/∂t from (8.51),
∂S  dx  v df ∂S
−  + = 0.
∂x dt dS=0 φ dS ∂x
This gives us the following equation for the path of a value of constant saturation
 dx 
 v df (S)
 = . (8.52)
dt dS=0 φ dS
If we assume that f is a function of S only, it follows that states of constant saturation will
propagate along straight lines, i.e., along paths given by
v df (S) 
x(t) = x0 (t0 ) + t − t0 ).
φ dS
To make a well-posed problem, we must pose initial and boundary conditions for the satu-
ration equation (8.51) in the form S(x,0) = S0,x (x) and S(0,t) = S0,t (t). For the special

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.4 The Buckley–Leverett Theory of 1D Displacements 259

case of a linear fractional flow function f (S) = S, which corresponds to a displacement

with linear relative permeabilities and equal viscotities, the analytical solution is given on
closed form

S0,x (x − at), x ≥ at,
S(x,t) =
S0,t (at − x), x < at,
where a = v/φ. Loosely speaking, the interpretation of this solution is that the saturation
equation takes the initial and boundary data and transports them unchanged along the x-axis
with a speed a or a speed equal f  (S)v/φ in the general case. Notice that to get a classical
solution, the initial and boundary data must be continuous and matching so that S0,x (0) =
S0,t (0). This way of constructing a solution is called the method of characteristics and
has been used a lot throughout the literature of porous media to study various kinds of
displacement scenarios.
Next, let us consider the case of one immiscible fluid displacing another, assuming
constant initial saturation S0 and a constant injected saturation Si , where Si generally is
different from S0 . Without lack of generality, we set v = φ = 1 or alternatively introduce
new time t ∗ = t v/φ to rescale the saturation equation so that v and φ disappear. This gives
the classical Buckley–Leverett problem
∂S ∂f (S)
+ = 0, S(x,0) = S0, S(0,t) = Si , (8.53)
∂t ∂S
named after the authors [56] who first developed this type of analysis. Here, we see that
all changes in saturation must originate from the jump in values from S0 to Si at the point
(x,t) = (0,0) and that these changes, which we henceforth refer to as waves, will propagate
along straight lines. In (8.52), we saw that a wave corresponding to a constant S-value will
propagate with a speed proportional to the derivative of the fractional flow function f .
In most cases, f (S) has a characteristic S-shape with one inflection point, i.e., a point at
which f  (S) has an isolated extremum. In essence, this means that saturation values near
the extremum should travel faster than saturation values on both sides of the extremum.
Thus, if we apply the characteristic equation (8.52) naively, we end up with a multivalued
solution as illustrated in Figure 8.9, which is clearly not physical. Indeed, if several waves
emanate from the same location in space and time, the speeds of these waves have to be non-
decreasing in the direction of flow so that the faster waves move ahead of the slower waves.
We can avoid the unphysical behavior if we simply replace the multivalued solution by
a discontinuity, as shown in Figure 8.9. To ensure that mass is conserved, the discontinuity
can be constructed graphically so that the two shaded areas are of equal size. This, however,
means that the solution cannot be interpreted in the classical sense, since derivatives are
not guaranteed to exist pointwise. Instead, the solution must be defined by multiplying the
saturation equation by a smooth function ϕ of compact support and integrating the result
over time and space
 
∂S ∂f (S) 
+ ϕ(x,t) dxdt = 0.
∂t ∂x

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 260 Mathematical Models for Multiphase Flow

Figure 8.9 Construction of the self-similar Buckley–Leverett solution S(x/t). The left plot shows
f  (S), whereas the right plot shows how the multivalued function given by a naive application of the
characteristic equation (8.52) plotted as a dashed line is turned into a discontinuous solution plotted
as a solid line by requiring that the shaded areas are of equal size.

We now integrate by parts and change the order of integration to transfer the derivatives to
the smooth function ϕ,
 
∂ϕ ∂ϕ 
S + f (S) dxdt = 0. (8.54)
∂t ∂x
A weak solution is then defined as any solution satisfying (8.54) for all smooth and com-
pactly supported test functions ϕ.
In general, propagating discontinuities like the one in Figure 8.9 must satisfy certain
conditions. To see this, let xd (t) denote the path of a propagating discontinuity, and pick
two points x1 and x2 such that x1 < xd (t) < x2 . We first use the integral form of the
saturation equation to write
 x2  x2
∂f (S) d
f (S1 ) − f (S2 ) = − dx = S(x,t) dx, (8.55)
x1 ∂x dt x1
and then decompose the last integral as follows:
 x2  xd (t)−  x2
S(x,t) dx = lim S(x,t) dx + lim S(x,t) dx.
x1 →0+ x1 →0+ xd (t)+

Since the integrand is continuous in each of the integrals on the right-hand side of the
equation, we can use the Leibniz rule to write
 xd (t)−  xd (t)−
d ∂S dxd (t)
S(x,t) dx = dx + S(xd (t) − ,t) .
dt x1 x1 ∂t dt
Here, the first term will vanish in the limit x1 → xd (t)−. By collecting terms, substituting
back into (8.55), and taking appropriate limits, we obtain what is known as the Rankine–
Hugoniot condition
σ (S + − S − ) = f (S + ) − f (S − ). (8.56)
Here, S±
denote the saturation values immediately to the left and right of the discontinuity
and σ = dxd /dt. Equation (8.56) describes a necessary relation between states on opposite

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.4 The Buckley–Leverett Theory of 1D Displacements 261

Figure 8.10 Illustration of self-sharpening (left) and spreading waves (right). The upper plots show
the characteristics, with the characteristics belonging to the Riemann fan marked in blue. The lower
plots show the initial solution (solid line) and the solution after some time (dashed line).

sides of a discontinuity, but does not provide sufficient conditions to guarantee that the
resulting discontinuity is physically admissible. To this end, one must use a so-called
entropy condition to pick out the physically correct solution. If the flux function is strictly
convex with f  (S) > 0, or strictly concave with f  (S) < 0, a sufficient condition is
provided by the Lax entropy condition,

f  (S − ) > σ > f  (S + ). (8.57)

This condition basically states that the discontinuity must be a self-sharpening wave in the
sense that saturation values in the interval between S − and S + tend to come closer together
upon propagation. Such a wave is commonly referred to as a shock, based on an analogy
from gas dynamics. In the opposite case of a spreading wave, commonly referred to as
a rarefaction wave, nearby states become more distant upon propagation, and for these
continuous waves, the characteristic speeds are increasing in the flow direction. In Figure
8.9 this corresponds to the continuous part of the solution that decays towards the shock.
For cases with convex or concave flux function, waves arising from discontinuous initial-
boundary data will either be shocks or rarefactions, as illustrated in Figure 8.10. If the flux
function is linear, or has linear sections, we can also have linearly degenerate waves for
which the characteristic speeds are equal on both sides of the wave.
For cases where the flux function is neither strictly convex nor concave, one must use
the more general Oleinik entropy condition to single out admissible discontinuities. This
condition states that
f (S) − f (S − ) f (S) − f (S + )
−
>σ > (8.58)
S−S S − S+
for all values S between S − and S + . For the classical displacement case with S0 = 0 and
Si = 1, this condition implies that the correct saturation S ∗ at the shock front is given by

f (S ∗ )/S ∗ = f  (S ∗ ), (8.59)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 262 Mathematical Models for Multiphase Flow

i.e., the point on the fractional-flow curve at which the chord between the points (0,0) and
(S ∗,f (S ∗ )) coincides with the tangent at the latter point.
If you are well acquainted with the theory of hyperbolic conservation laws, you will
immediately recognize (8.53) as an example of a Riemann problem, i.e., a conservation law
with a constant left and right state. Let these states be denoted SL and SR . If SL > SR , the
solution of the Riemann problem is a self-similar function given by
⎧
⎪
⎪ L, x/t < fc (SL ),
⎨S  −1
S(x,t) = fc (x/t), fc (SL ) < x/t < fc (SR ), (8.60)
⎪
⎪
⎩S , 
x/t ≥ f (S ),
L c R

which is often referred to as the Riemann fan. Here, fc (S) is the upper concave envelope of
f over the interval [SR ,SL ] and (fc )−1 the inverse of its derivative. Intuitively, the concave
envelope is constructed by imagining a rubber band attached at SR and SL and stretched
above f in between these two points. When released, the rubber band assumes the shape of
the concave envelope and ensures that fc (S) < 0 for all S ∈ [SR ,SL ]. The case SL < SR is
treated symmetrically with fc denoting the lower convex envelope. The petroleum literature
usually attributes this construction to Welge [309]. Figure 8.10 shows examples of two such
Riemann fans: a single shock in the left plot and a rarefaction fan in the right plot.

Example 8.4.1 Let us first consider the Buckley–Leverett solution for a pure imbibition,
for which the left and right states are SL = 1 and SR = 0. From the discussion above,
we know that the self-similar solution of the associated Riemann problem is found by
computing the upper concave envelope of the fraction flux function, which is linear in the
interval [0,S ∗ ] and then coincides with f (S) in the interval [S ∗,1], where the saturation S ∗
behind the shock front is determined from (8.59). If we assume a simple case with relative
permeabilities given by the Corey model (8.15) on page 241 with nw = no = 2 and
μw /μo = M, (8.59) reads
SM 2M(1 − SM )SM
2 + M(1 − S )2
= 2 .
SM M SM + M(1 − SM )2
2

√
Here, we have used SM to denote the front saturation S ∗ = M/(M + 1) to signify that it
depends on the viscosity ratio M. Figure 8.11 shows the Buckley–Leverett solution for three
different viscosity ratios. In all three cases, the solution consists of a shock followed by a
trailing continuous rarefaction wave. When the water viscosity is five times higher than
oil (M = 5), the mobility of pure oil is five times higher than the mobility of pure water.
Hence, we have a favorable displacement in which the water front acts like a piston that
displaces almost all the oil at once. In the opposite case with oil viscosity five times higher
than water viscosity, we have an unfavorable displacement because the water front is only
able to displace a small fraction of the oil. Here, the more mobile oil will tend to finger
through the oil and create viscous fingers for displacements in higher spatial dimensions.
We will come back to examples that illustrate this later in the book. In the opposite case of

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.4 The Buckley–Leverett Theory of 1D Displacements 263

M=1
M=5
M = .2

Figure 8.11 Buckley–Leverett solutions for a pure imbibition case with Corey relative permeabilities
with exponent nw = n0 = 2 and mobility ratio μw /μo = M. The left plot shows the upper concave
envelopes of the flux functions, while the right plot shows the self similar solution Sw (x/t) with
x/t = 1 shown as a dotted line.

a drainage process, it follows by symmetry that the solution will consist of a shock from 1
to 1 − SM followed by a rarefaction wave from SM to 0.

Instead of an infinite domain, let us place a producer some distance to the right of the
origin. We then see that since the saturation is constant ahead of the shock, the amount
of produced oil will be equal the amount of injected water until the first water reaches
the producer. This means that the cumulative oil production has a linear slope until water
breakthrough. After water has broken through, the well will be producing a mixture of oil
and water and the slope of the cumulative oil production decreases.
Let us try to determine the amount of produced oil at water breakthrough. To this end,
we can assume that water is injected at x = 0 and oil produced at x = L. At water
breakthrough, the saturation at the producer is S(x = L) = S ∗ . The amount of produced
oil equals the amount of water that has displaced it LS̄, where S̄ is the average water
saturation defined as
  ∗
1 L 1  x=L 1 S
S̄ = S dx = xS − x dS
L 0 L x=0 L 1
 S∗ 
t dfc  t ∗ 
= S∗ − dS = S ∗ − f −1 .
L 1 dS L

The second equality follows from the total differential d(xS) = xdS + Sdx, while the
third equality follows from the equation (8.52) that describes the position x(t) of a constant
saturation value at time t. Finally, the water breakthrough occurs at time t = L/f  (S ∗ ), and
hence we have that,

1 − f (S ∗ )
= f  (S ∗ ).
S̄ − S ∗
We interpret this graphically as follows: Sw is the point at which the straight line used to
define the front saturation S ∗ intersects the line f = 1 as shown in Figure 8.12.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 264 Mathematical Models for Multiphase Flow

S∗ S̄

Figure 8.12 Graphical determination of front saturation and average water saturation at water
breakthrough.

0 0.2 0.4 0.6 0.8 1

Figure 8.13 Fractional flow function for 1D gravity segregation for Corey relative permeabilities
(8.15) with nw = 3, n0 = 2, and mobility ratio μw = 5μo for ρw > ρn .

8.4.2 Gravity Segregation

Next, we consider a pure gravity displacement, for which the transport equation reads,
∂S ∂  λ n λw 
+ = 0, (8.61)
∂S ∂z λw + λn
where we, without lack of generality, have scaled away the constant gρ. With a slight
abuse of notation, we call the flux function for g(S). This flux function will have a char-
acteristic bell-shape, e.g., as shown in Figure 8.13, which points downward if ρw > ρn
and upward in the opposite case. We assume that the wetting fluid is most dense and
consider the flow problem with one fluid placed on top of the other so that the fluids are
separated by a sharp interface. If the lightest fluid is on top, we have that SL < SR , and
from the discussion of (8.60) on page 262 we recall that the Riemann solution is found
by computing the lower convex envelope of g(S). Here, gc (S) ≡ 0 and hence we have a
stable situation. In the opposite case, SL > SR and gc (S) is a strictly concave function for

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.4 The Buckley–Leverett Theory of 1D Displacements 265

SR∗ SL∗

Figure 8.14 Buckley–Leverett solution for a gravity column with sealing top and bottom before the
moving fluids have contacted the top/bottom of the domain. (Model: Corey relative permeabilities
(8.15) with nw == n0 = 2, and mobility ratio μw = μo and ρw > ρn .)

which gc (0) > 0 > gc (1). This is an unstable situation and the heavier fluid will start to
move downward (in the positive z-direction) and the lighter fluid upward (in the negative
z-direction) near the interface. Let us consider more details in a specific case:
Example 8.4.2 We consider a gravity column inside a homogeneous sand body confined by
a sealing medium at the top and the bottom. For simplicity, we scale the domain to be [0,1],
assume that the initial fluid interface is at z = 0.5, and use a Corey relative permeability
model with nw = no = 2 and equal viscosities.
This choice of parameters gives a symmetric flux function g(S), as shown in the left plot
of Figure 8.14. Our initial condition with a heavy fluid on top of a lighter fluid is described
by setting SL = 1 in the top half of the domain and SR = 0 in the bottom half of the domain.
As explained above, the Riemann problem is solved by computing the upper concave hull
of g(S). The hull gc (S) is a linear function lying above g(S) between 0 and SR∗ , follows the
graph of g(S) in the interval SR∗ to SL∗ , and is a linear function lying above g(S) between
SL∗ and 1. The point SR∗ is given by the solution to g  (SR∗ )SR∗ = g(SR∗ ). At the boundaries
x = 0,1, we have g(SL ) = g(SR ) = 0, and hence no-flow conditions as required. Initially,
the solution thus consists of three different regions:

• a single-phase region on top that only contains the heavier fluid (SL = 1), which is
bounded above by the lines z = 0 and below by the position of the shock 1 → SL∗ ;
• a two-phase transition region that consists of a mixture of heavier fluid moving downward
and lighter fluid moving upward; this corresponds to a centered rarefaction wave in
which the saturation decays smoothly from SL∗ to SR∗ ;
• a single-phase region at the bottom that only contains the light fluid. This region is
bounded above by a shock SR∗ → 0 and from below z = 1.
As time passes, the two shocks that limit the single-phase regions of heavy and light fluids
will move upward and downward, respectively, so that an increasing portion of the fluid
column is in a two-phase state; see the right plot in Figure 8.14. Eventually, these shocks
will reach the top and bottom of the domain. Because of the symmetry of the problem, it is
sufficient for us to only consider one end of the domain, say z = 0. Since we have assumed

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 266 Mathematical Models for Multiphase Flow

SR∗ SL∗

SL∗∗ SR∗∗

Figure 8.15 Buckley–Leverett solution for a gravity column with sealing top and bottom just after
the light/heavy fluids started to accumulate at the top/bottom of the domain. To ensure no-flow, the
solution is S = 0 at z = 0 and S = 1 at z = 1.

no flow across the top of the domain, the shock 1 → SL∗ cannot continue to propagate
upward. We thus get a Riemann problem, defined by right state SL∗ and a left state SL ,
which could be any value satisfying g(SL ) = 0. The left state cannot be 1, since this would
give back the same impermissible shock 1 → SL∗ . If the left value is SL = 0 instead, i.e., a
single-phase region with only light fluid forming on the top, the Riemann solution will be
given by the concave envelope of g(S) over the interval [0,SL∗ ] and hence will consist fully
of waves with positive speed that propagate downward again. The left plot in Figure 8.15
shows the corresponding convex envelope, which we see gives a rarefaction wave 0 to SL∗∗
and a shock SL∗∗ → SL∗ . As the leading shock propagates downward, it will interact with
the upward-moving part of the initial rarefaction wave and form a shock, whose right state
gradually decays towards S = 0.5. The solution at the bottom of the domain is symmetric,
with accumulation of a single-phase region of the heavy fluid.

8.4.3 Front Tracking: Semi-Analytical Solutions

Whereas it was quite simple to find an analytical solution in closed form for simple Rie-
mann problems like the ones discussed above, it becomes more complicated for general
initial data or when different waves start to interact, as we observed when the two initial
shocks reflected from the top and bottom of the gravity column. Generally, one must
therefore resort to numerical discretizations in the form of a finite element, difference, or
volume method. However, for one-dimensional saturation equations there is a much more
powerful alternative, which we will introduce you to very briefly.
In the semi-analytical front-tracking method, the key idea is to replace the functions
describing initial and boundary values by piecewise constant approximations. For a small
time interval, the solution can then be found by piecing together solution of local Riemann
problems, and consists of a set of constant states separated by shocks and rarefaction waves.
Shocks correspond to the linear parts of the local convex/concave envelopes, whereas
rarefaction waves correspond to the nonlinear parts of envelopes. If we further approximate
the flux function by a piecewise linear function, all convex/concave envelopes are always

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.4 The Buckley–Leverett Theory of 1D Displacements 267
4.5 1

t=4.5
4 0.5

0
3.5
1
3 t=3.0
0.5

2.5 0

2 1
t=1.5
0.5
1.5
0
1
1
t=0.0
0.5
0.5

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Figure 8.16 Solution for a gravity column with heavy fluid on top of a light fluid computed by front
tracking for a model with Corey relative permeabilities with nw = no = 2 and equal viscosities.
The left column shows the wave pattern in the (z,t) plane, where each blue line corresponds to the
propagation of a discontinuous state. The right column shows the solution S(z,t) at four different
times.

piecewise linear, and the solution of Riemann problems will consist only of shocks. In other
words, the solution to the continuous, but approximate PDE problem consists entirely of
constant states separated by discontinuities propagating at a constant speed given by the
Rankine–Hugoniot condition (8.56) on page 260.
Each time two or more discontinuities “collide,” we have a new Riemann problem that
gives rise to a new set of discontinuities emanating out of the collision point. Computing
local convex/concave envelopes and keeping track of propagating discontinuities can be
formulated as numerical algorithms, and we end up with a semi-analytical method, which
is unconditionally stable and formally first-order accurate. In fact, one can prove that for
any finite spatial domain, there will be a finite number of shock collisions, and hence we
can compute an approximate solution up to infinite time in a finite number of steps; see
[134] for more details. In the following example, we use this method to study a few simple
1D cases.

Example 8.4.3 We continue to study gravity segregation of a heavy fluid placed on top
of a lighter fluid inside a sand box with sealing top and bottom. Figure 8.16 shows the
solution computed by approximating the flux function with a piecewise linear function
with 100 equally spaced segments. Initially, the solution contains two constant states
SL = 1 and SR = 0, separated by a composite wave that consists of an upward-moving
shock, a centered rarefaction wave, and a downward moving shock, as explained earlier

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 268 Mathematical Models for Multiphase Flow

in Example 8.4.2. Notice that the approximate solution at time t = 1.5 is of the same
form as shown in Figure 8.14, except that in the approximate front-tracking solution, the
continuous rarefaction wave is replaced by a sequence of small discontinuities.
The right plot in Figure 8.15 illustrated the form of the exact solution just after the two
initial shocks have reflected at the top and bottom of the domain. Figure 8.16 shows the
solution after the reflected waves have propagated some time upward/downward. Here,
we clearly see that these waves consist of a fast shock wave followed by a rarefaction
wave. As the leading shock interacts with the centered rarefaction wave originating
from (0.5,0), the difference between left and right states diminishes, and the shock
speeds decay. Later, the two shocks collide and form a stationary discontinuity that will
eventually become a stable steady state with the lighter fluid on top and the heavier fluid
beneath.
In general, the transport in an inclined reservoir section will be driven by a combination
of gravity segregation and viscous forces coming from pressure differentials. The next
example studies one such case.
Example 8.4.4 (CO2 storage) We consider two highly idealized models of CO2 storage
in an inclined aquifer with a slow drift of brine either in the upslope or in the downslope
direction. If the injection point lies deeper than approximately 800 meters below the sea
level, CO2 will appear in the aquifer as a supercritical liquid phase having much lower
density than the resident brine. The injected fluid, which is commonly referred to as the
CO2 plume, will therefore tend to migrate in the upslope direction. As our initial condition,
we assume (somewhat unrealistically) that CO2 has been injected so that it completely fills
a small section of the aquifer x ∈ [ 14 , 34 ]. To study the migration after injection has ceased,
we can now use a fractional flow function of the form,
S 2 ∓ 4S 2 (1 − S)3
F (S) = ,
S 2 + 15 (1 − S)3
where the minus sign denotes upslope background drift and the plus sign downslope drift.
Figure 8.17 shows the flux functions along with the envelopes corresponding to the two
Riemann problems.
Let us first consider the case with upslope drift. At the tip of the plume (x = 34 ), there will
be a drainage process in which the more buoyant CO2 displaces the resident brine. This
gives a composite wave that consists of a shock Sd∗ → 1 that propagates upslope, followed
by a centered rarefaction wave from Sd∗ to Sd∗∗ having wave speeds both upslope and
downslope. The rarefaction is followed by a trailing shock wave 0 → Sd∗∗ that propagates
in the downslope direction. Likewise, at the trailing edge of the plume, the resident brine
will imbibe into CO2 , giving a wave consisting of a shock Si∗ → 0 followed by a rarefaction
wave, both propagating in the upslope direction. The two Riemann fans can be seen clearly
in Figure 8.18, and will stay separated until t ≈ 0.18, when the shock from the imbibition
process collides with the trailing drainage shock. The result of this wave interaction is a
new and slightly weaker shock that propagates increasingly faster in the upslope direction,

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.4 The Buckley–Leverett Theory of 1D Displacements 269

Upslope background drift Downslope background drift

1 1.5
Si∗ Si∗∗
Si∗
1
0.5

0.5
0 Sd∗∗

Sd∗ Sd∗∗
0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Figure 8.17 Flux functions for the CO2 migration example. The dashed curves are the flux function
F (S), which accounts for a combination of gravity segregation and viscous background drift. The
blue lines show the concave envelopes corresponding to imbibition at the left edge of the plume,
while the red lines are the convex envelopes corresponding to drainage at the right edge the plume.

0 0.2 0.4 0.6 0.8 1

2 1

1.8 0.5
t=1.5
0
1.6

1
1.4

0.5
1.2 t=0.5
0
1
1
0.8
0.5
t=0.1
0.6
0

0.4
1

0.2 0.5
t=0.0

0 0
0 1 2 3 0 1 2 3

Figure 8.18 Solution for a conceptual model of CO2 injected into an inclined 1D aquifer with a
background upslope drift. Upslope direction is to the right. Dashed lines in the left plot signify the
snapshots shown in the right column.

followed by a continuous rarefaction wave, which is an extension of the initial imbibition

rarefaction. At time t ≈ 0.665, the new shock wave has overtaken the tip of the plume,
and the result of this interaction is that the upslope migration of the plume gradually slows
down.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 270 Mathematical Models for Multiphase Flow
0 0.2 0.4 0.6 0.8 1

1.5 1
t=0.8
0.5

1.25
0

1
t=0.35
1
0.5

0
0.75
1
t=0.2
0.5
0.5
0

1
0.25
t=0.1
0.5

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Figure 8.19 Solution for a conceptual model of CO2 injected into an inclined 1D aquifer with a
background downslope drift. Upslope direction is to the left.

If we remove the upslope drift, the initial imbibition process is no longer present, but the
long-term behavior remains the same and consists of a leading shock wave followed by a
rarefaction that slowly eats up and slows down the shock.
Figure 8.19 shows the opposite case with a downslope background drift. Here, the
imbibition process at the left edge of the plume gives a shock 0 → Si∗∗ that propagates
in the downslope direction (to the right in the figure), a centered rarefaction wave from Si∗∗
to Si∗ with waves that propagate both in the downslope and upslope direction, followed by a
shock Si∗ → 1 that propagates upslope. The drainage at the right edge gives a strong shock
Sd∗ → 1 that propagates downslope, followed by a weak rarefaction wave from 0 to Sd∗ . At
time t ≈ 0.25 the downslope imbibition shock has overtaken both the rarefaction wave and
the shock from the drainage process. The resulting wave interaction first forms a stationary
shock with left state Si∗∗ . However, as the positive part of the imbibition rarefaction wave
continues to propagate downslope, the left state gradually increases, and at some point
√
the left state passes the value S = 1/10 for which F (S) = 1. When this happens,
the stationary shock turns into a composite wave consisting of a shock and a trailing
rarefaction wave that both propagate in the upslope direction.
Figure 8.20 shows the two migration cases over a longer time. Injecting CO2 so that it
must migrate upward against a downslope brine drift is clearly advantageous since it will
delay the upslope migration of the gradually thinning plume.
The purpose of the previous example was to introduce you to the type of wave patterns that
arise in hyperbolic saturation equations, and show that these can be relatively complicated
even for simple initial configurations. While front-tracking is probably the best method

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 8.4 The Buckley–Leverett Theory of 1D Displacements 271

Upslope drift Downslope drift

15 15

10 10

5 5

0 0
0 1 2 3 4 −3 −2 −1 0 1

Figure 8.20 Long-time solutions for the two conceptual models of CO2 storage.

you can find to study wave patterns in 1D hyperbolic saturation equations, it would not
be the method of choice to study multiphase flow models in multiple spatial dimensions,
unless we use a streamline method that transform the 3D transport problems to a family of
1D problems. In the next chapter, we therefore continue to discuss finite-volume methods
that are more suitable for multidimensional reservoir simulation. In simulation of real
reservoirs, one cannot hope to resolve the dynamics to the level of detail seen in the last
example. To be able to compute shocks and other types of discontinuities, finite-volume
(and finite-element) methods contain a certain amount of numerical dissipation that tends
to smear any discontinuity. Likewise, strong heterogeneities and radial flow near wells
can introduce orders of magnitude variations in Darcy velocities. This necessitates the
use of small time steps or implicit temporal discretizations, which both tend to smear
discontinuities and make it more difficult to resolve complex wave interactions in full detail.
We end our discussion of 1D solutions with a few remarks about the simulation of
CO2 storage, since this was touched upon conceptually in Example 8.4.4. In more realistic
modeling of CO2 storage, one would need a more detailed model that accounts for the effect
that the lighter CO2 plume will flow to the top of the formation and migrate upslope as a
thin layer under the sealing caprock that bounds the aquifer from above. This can be done
by either using a 3D or a 2D cross-sectional saturation equation, possibly coupled with
a pressure equation. However, the most efficient approach to study large-scale, long-term
CO2 migration is to integrate the pertinent flow equations in the vertical direction to form
a vertically averaged model that accounts for the vertical fluid distribution in an averaged
sense. MRST has a large module, co2lab, that offers a wide variety of vertically averaged
models and other types of reduced models suitable for modeling of large-scale CO2 storage.
A discussion of such models is beyond the scope of this book. I you are interested in more
details, you should instead consult one of the papers describing the methods implemented
in MRST-co2lab [19, 230, 227, 228, 232, 193, 21].

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 9
Discretizing Hyperbolic Transport Equations

So far in the book, we have introduced you to several basic discretizations: Section 4.4
introduced the two-point scheme for elliptic operators of the type ∇ · K( x )∇ and the
upstream scheme for hyperbolic operators ∇ · v. Likewise, Section 7.2 introduced fully
implicit discretizations for compressible single-phase flow, and discussed how to use a
Newton–Raphson method to linearize and solve the resulting system of discrete equations.
With some modifications, these techniques are all we need to describe the most widespread
approach to simulate compressible, multiphase flow.
For incompressible multiphase flow, on the other hand, it is common to write the sys-
tem as a pressure equation and one or more transport equations and use specialized dis-
cretization schemes for each subequations. Chapter 6 discussed consistent schemes for
elliptic pressure equations. This chapter reviews some of the particular challenges that lie
in discretizing transport equations. If you are not interested in this wider perspective and
only wish to know the simplest possible approach to solving saturation equations, you can
jump directly to Section 9.4 on page 286, which discusses the standard upstream mobility-
weighting method implemented for general polyhedral grids in the incomp module of
MRST.

9.1 A New Solution Concept: Entropy-Weak Solutions

In the first part of this chapter, we continue to discuss the homogeneous conservation law
introduced in the previous chapter,
ut + f (u)x = 0, u(x,0) = u0 (x). (9.1)
Here, u is some conserved quantity, which need not necessarily be a fluid saturation, and
f (u) is a generic flux function. Equation (9.1) usually arises from a more fundamental
physical law on integral form,
 x2
d  
u(x,t) dx = f u(x1,t) −f u(x2,t) , (9.2)
dt x1
which states that the rate of change of quantity u within the interval [x1,x2 ] equals the flux
across the ends x = x1 and x = x2 of the interval.

272

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 9.1 A New Solution Concept: Entropy-Weak Solutions 273

In the previous chapter, we saw that solutions to (9.1) may develop discontinuities in
finite time, even for smooth initial data if the flux function f is nonlinear. This means that
the solution of (9.1) is usually understood in the weak sense,
 ∞ 

uϕt + f (u)ϕx dtdx = u0 (x)ϕ(x,0) dx. (9.3)
0 R R

Here, ϕ(x,t) is a continuous and smooth test function that has compact support so that it
vanishes outside a bounded region in the (x,t)-plane.
Solutions defined by the weak form (9.3) are not necessarily unique. The solution con-
cept must therefore be extended to include additional admissibility conditions to single out
the correct solution among several possible candidates satisfying the weak form. A classical
method to obtain uniqueness is to add a regularizing second-order term to (9.1), giving a
parabolic equation

ut + f (u )x = uxx ,

which has unique, smooth solutions. The unique solution of the hyperbolic equation (9.1)
is then defined as the limit of u (x,t) as  tends to zero. In models from fluid dynamics,
such a second-order term can be proportional to the viscosity of the fluid, and the method
is therefore called the vanishing viscosity method. For flow in porous media, the transport
equations will already have a second-order term coming from capillary forces (see (8.35)).
However, capillary forces are often neglected when studying flow dominated by global
pressure gradients, or the capillary function may vary with rock type and this can introduce
discontinuities in the phase saturations at the interface between different rock types.
Since the vanishing viscosity solutions u (x,t) are smooth, it is possible to use tech-
niques from classical analysis to prove the existence, uniqueness, and stability of the solu-
tion to (9.1). This was first done in a seminal work by Kružkov [173], which paved the road
for the modern theory of nonlinear partial differential equation of a hyperbolic-parabolic
type. Working with limits of viscous solutions is not very practical. Instead, we impose
other admissibility conditions. We have already encountered the Lax and Oleinik entropy
conditions, (8.57) and (8.58), which were used to single out permissible discontinuities. An
alternative approach is to introduce a (convex) entropy function η(u) and a corresponding
entropy flux ψ(u), and require that an admissible weak solution u must satisfy the entropy
condition

η(u)t + ψ(u)x ≤ 0, (9.4)

which must be interpreted in the weak sense as

 ∞ 

η(u)ϕt + ψ(u)ϕx dtdx + η(u0 (x))ϕ(x,0) dx ≥ 0. (9.5)
0 R R

The solution u(x,t) is called an entropy weak solution of (9.1) if it satisfies (9.5) for all
k ∈ R and nonnegative test functions ϕ.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 274 Discretizing Hyperbolic Transport Equations

9.2 Conservative Finite-Volume Methods

You have already seen that using a finite-volume method made it simple to formulate a
discretization on general polyhedral grids and ensured that the resulting method obeys the
important property of mass conservation. Approximating the unknown function in terms of
its cell averages has one additional advantage for hyperbolic conservation laws. Because the
differential equation ceases to be pointwise valid in the classical sense once a discontinuity
arises in u(x,t), we should expect that pointwise approximations used in standard finite-
difference methods will break down, since these rely on the assumption that the function is
smooth in a small neighborhood where the discrete difference is taken. Instead of seeking
solutions in a pointwise sense, the finite-volume approach seeks globally defined solutions
of the integral form (9.2), which is more fundamental, represented in the form of discrete
cell averages, as we have already seen in previous chapters.
To develop a finite-volume method for (9.1), we first define the sliding average,
 x+x/2
1
ū(x,t) = u(ξ,t) dξ . (9.6)
x x− 1 x
2
We then associate these sliding averages with grid cells, {[xi−1/2,xi+1/2 ]}, where xi±1/2 =
xi ± 12 xi , and set ui (t) = ū(xi ,t). By inserting ui (t) into the integral form of the
conservation law (9.2), we obtain a semi-discrete version of (9.1),
d ūi 1    
= f u(xi−1/2,t) − f u(xi+1/2,t) . (9.7)
dt xi
Alternatively, we may define uni = ūi (tn ) for a set of discrete times tn and then integrate
(9.7) from tn to tn+1 to derive a fully discrete version of (9.1),
 tn+1  tn+1
n+1 1 1
ui − ui = n
f (u(xi−1/2,t)) dt − f (u(xi+1/2,t)) dt. (9.8)
x tn x tn
While (9.7) and (9.8) tell us how the unknown cell averages ui (t) and uni evolve in time,
they cannot be used directly to compute these cell averages, since we do not know the
point values u(xi±1/2,t) needed to evaluate the integrand. To obtain these, we need to make
additional assumptions and approximations, e.g., as discussed in Section 4.4.3 for the linear
case. Nevertheless, (9.8) suggests that a viable numerical method for (9.1) should be of the
form
 n
un+1
i = uni − rin Fi+1/2 − Fi−1/2
n
, rin = t n /xi , (9.9)
n
where Fi±1/2 approximates the average flux over each cell interface,
 tn+1
1 
Fi±1/2 ≈
n
f u(xi±1/2,t) dt. (9.10)
t n tn
In Section 8.4 we showed that information propagates at a finite speed along so-called
characteristics, which implies that the flux integral (9.10) will only depend on the solution
u(x,tn ) in a local neighborhood of the interface xi±1/2 . This means, in turn, that Fi±1/2 can

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 9.3 Centered versus Upwind Schemes 275

be approximated in terms of a small collection of neighboring cell averages, i.e., Fi+1/2 =

F (uni−p, . . . ,uni+q ), which we alternatively will write as Fi±1/2 = F (un ;i ± 12 ).
Any numerical method written on the form (9.9) will be conservative. To see this, we
multiply (9.9) by xi and sum over i. The flux terms will cancel in pairs and we are left
with

N 
N
 n
un+1
i xi = uni xi − t n FN+1/2 − F−M−1/2
n
.
i=−M i=−M

From this, it follows that the method is conservative on a finite domain, since the accumula-
tion inside any interval balances the sum of the flux across the interval edges. To see that the
same is true in an infinite domain, we must make some additional assumptions. If the initial
solution u0 (x) has bounded support, it follows that the two flux terms will cancel if we
choose M and N sufficiently large. Hence, the scheme (9.9) is conservative in the sense that
  
un+1
i x i = un
i x i = · · · = u0 (x) dx.
i i

9.3 Centered versus Upwind Schemes

To provide important insight into the numerical solution of hyperbolic equation, we briefly
outline two main classes of classical schemes.

9.3.1 Centered Schemes

The simplest approach to reconstruct the point values u(xi±1/2,t) needed in (9.7) and (9.9),
is to assume that u(x,t) = ui (t) inside each grid cell and average the flux values on opposite
sides of the grid interface, i.e., set
 
n
Fi±1/2 = 12 f (uni±1 ) + f (uni ) . (9.11)
This approximation is referred to as a centered approximation, which unfortunately gives a
2
notoriously unstable scheme. We can add an artificial diffusion term, x 2
t ∂x u, to stabilize,
but then it is no longer possible to go back to the semi-discrete form (9.7), since the artificial
diffusion will blow up in the limit t → 0. Discretizing the artificial diffusion through
standard centered differences gives the classical Lax–Friedrichs scheme
1 n  1  
un+1 = u + un
− r f (un
) − f (un
i−1 .
) (9.12)
i
2 i+1 i−1
2 i+1

Alternatively, the scheme can be written in conservative form (9.9) using the numerical flux
1 n 1 
ui − uni+1 +
F (un ;i + 1/2) = f (uni ) + f (uni+1 ) . (9.13)
2r 2
To ensure stability, we have to impose a restriction on the time-step through a CFL con-
dition, named after Courant, Friedrichs, and Lewy, who wrote one of the first papers on

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 276 Discretizing Hyperbolic Transport Equations

finite-difference methods in 1928 [75]. The CFL condition states that the true domain of
dependence for the PDE (9.1) should be contained in the domain of dependence for (9.9).
For the Lax–Friedrichs scheme, this means that
t
max |f  (u)| ≤ 1. (9.14)
x u
Under this condition, the scheme is very robust and will always converge, albeit painstak-
ingly slow in many cases. The robustness is largely due to the added numerical diffusion,
which tends to smear discontinuities. As an illustration, consider a stationary discontinuity
satisfying ∂t u = 0. Here, the approximate solution in each cell is defined as the arithmetic
average of the cell averages in the neighboring cells, i.e., un+1
i = 12 (uni+1 + uni−1 ). From
this, we see that the scheme will smear the discontinuity one cell in each direction per time
step. By using a formal Taylor expansion on a single time step, it follows that the scheme
has a truncation error of order two. In practice, we are more interested in the error at a
fixed time, which we need to use an increasing number of time steps to reach as x → 0
because of (9.14). This means that we must divide with t so that the error of the scheme
is O(x) as x → 0, and we say that the scheme is formally first-order accurate.
We can improve accuracy if we make a more accurate approximation to the integral
n
defining Fi±1/2 . Instead of evaluating the integral at the endpoint tn , we can evaluate it at
the midpoint tn+1/2 = tn + 12 t. One can show that the corresponding point values can be
predicted with acceptable accuracy by the Lax–Friedrichs scheme on a grid with half the
grid spacing. This gives a second-order, predictor-corrector scheme called the (Richtmeyer
two-step) Lax–Wendroff method [178]
n+1/2   
ui+1/2 = 12 uni + uni+1 − 12 r f (uni±1 ) − f (uni ) ,
 n+1/2 n+1/2 
un+1
i = ui − r f (ui+1/2 ) − f (ui−1/2 ) , (9.15)

which is stable under the same CFL condition (9.14) as the Lax–Friedrichs scheme. The
corresponding numerical flux reads
  
F (un ;i + 1/2) = f 12 (uni + uni+1 ) − 12 r f (uni+1 ) − f (uni ) . (9.16)

Both the Lax–Friedrichs and the Lax–Wendroff scheme can either be interpreted as a finite-
difference or as a finite-volume scheme, and by only looking at the formulas written in
terms of uni , it is not possible to determine which formulation is in use. However, the under-
lying principles are fundamentally different. Finite-difference methods evolve a discrete set
of point values by using discrete differences to approximate the differential operators in
(9.1). Finite-volume methods evolve globally defined solutions given by (9.2) and realize
them in terms of a discrete set of cell averages. This latter perspective is the key to modern
so-called high-resolution methods [183, 294], which have proved to be very successful.
Let us now see how we can implement these two schemes compactly in MATLAB.
When computing on a bounded domain, one generally has to modify the stencil of the
scheme in the cells next to the domain boundary to account for boundary conditions.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 9.3 Centered versus Upwind Schemes 277

A widely used trick to avoid this is to pad the domain with a layer of ghost cells, i.e.,
extra cells that are not really part of the domain and whose states can be set manually to
impose the desired boundary conditions.
function u=lxf(u0,cfl,dx,T,flux,df,boundary)
u = u0; t = 0.0;
dt = cfl*dx/max(abs(df(u0)));
i = 2:numel(u0)-1; % do not compute on the ghost cells
while (t<T)
if t+dt>T, dt = T-t; end
t=t+dt;
u = boundary(u); % set values in the ghost cells
f = flux(u);
u(i) = 0.5*(u(i+1)+u(i-1)) - 0.5*dt/dx*(f(i+1)-f(i-1));
dt = cfl*dx/max(abs(df(u)));
end

Here, u0 gives the initial data, cfl and dx are the CFL number and the spatial discretization
parameter, T is the desired end time, and flux, df, and boundary are handles to functions
that compute the flux f and its derivative and set appropriate states in the ghost cells.
Notice, in particular, how we use the vector i to avoid computing in the ghost cells that pad
our domain. To get the Lax–Wendroff scheme, we introduce an auxiliary array U and then
simply replace the second last line by
U(i) = 0.5*(u(i)+u(i+1)) - 0.5*r*(f(i+1)-f(i));
U = boundary(U); f = flux(U);
u(i) = u(i) - r*(f(i)-f(i-1));

9.3.2 Upwind or Godunov Schemes

The centered schemes introduced in the previous subsection are black-box schemes that can
be utilized without any particular knowledge about the flux function apart from an estimate
of its maximum and minimum derivative (or eigenvalues for systems of equations). The
reason for this is that these schemes utilize information from both sides of the grid interface
when computing approximations to the flux integral in (9.10). In many cases we can do
better, since we know more of how the solution behaves at the interface. If f  (u) > 0, for
instance, we know that all characteristics are positive, and hence we can use the solution
from the left side of the grid interface to evaluate the integral in (9.10). With a constant
reconstruction inside each cell, this means that Fi+1/2 = f (uni ) is the exact value of the
integral, provided that the solution at time tn is constant and equal uni in grid cell i. The
resulting scheme,
 
un+1
i = uni − r f (uni ) − f (uni−1 ) , (9.17)
is called an upwind scheme, which you have already encountered in Chapters 4 and 7.
Similarly, if f  (u) ≤ 0, the upwind scheme takes the form

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 278 Discretizing Hyperbolic Transport Equations
 
un+1
i = uni − r f (uni+1 ) − f (uni ) . (9.18)
In either case, the upwind scheme is a two-point scheme that, instead of using a cen-
tered difference, uses a one-sided difference in the upwind direction, i.e., in the direction
from which the waves (or information) is coming. This type of upwind differencing is the
underlying design principle for so-called Godunov schemes, named after the author of the
seminal paper [118]. If we assume a constant reconstruction so that u(x,t n ) ≡ uni inside
grid cell number i, the evolution of u(x,t) can be decomposed into a set of local Riemann
problems

uni, x < xi+1/2,
∂t v + ∂x f (v) = 0, v(x,0) = (9.19)
ui+1, x ≥ xi+1/2,
n

which are of the same type as those we have already accounted in Section 8.4. Each of these
Riemann problems admits a self-similar solution vi+1/2 (x/t), which can be constructed by
introducing a local convex or concave envelope1 fc (u;uni,uni+1 ) of the flux function f
over the interval [uni,uni+1 ], as discussed in (8.60). The cell averages can then be correctly
evolved a time step t forward in time by (9.9) if we use vi±1/2 (0) – or a good approxi-
mation thereof – to evaluate f in the flux integral (9.10). The Riemann fan will generally
expand in each direction at a constant speed, and using vi±1/2 (0) to evaluate the integrand
of (9.10) is only correct until the first wave from the neighboring Riemann problem reaches
the interface. This implies a time-step restriction of the form
t n
max |fc (· ;uni,uni+1 )| ≤ 1, (9.20)
i,n xi
which we recognize as a sharper version of the CFL condition (9.14). The resulting scheme
is formally first-order accurate, but will generally be much less diffusive than the centered
Lax–Friedrichs scheme.
Let us now go back to the general transport equation (8.35), which is hyperbolic if
Pc ≡ 0. Here, the fractional flow function f (S) has a characteristic S-shape and its
derivative is therefore always positive. This means that if gravity forces are negligible,
like in the horizontal displacements we studied in Section 8.4.1, the correct Godunov
scheme would be either (9.17) or (9.18), depending upon the sign of the flux v. For the
gravity segregation problem considered in Section 8.4.2, on the other hand, the flux function
g(S) = λw λn /(λw +λn ) has derivatives of both signs and we would generally have to solve
the Riemann problem (9.19) like we did in Figure 8.14 on page 265 to apply the Godunov
scheme. The same is generally true for flow involving a combination of viscous and gravity
forces, as illustrated in Figure 8.17. Having to solve a Riemann problem each time we need
n
to evaluate the flux Fi±1/2 may seem rather cumbersome, and fortunately there is a rather
elegant and physically motivated fix to this. Instead of upwinding the flux itself, we can use
the phase fluxes to evaluate the relative mobilities in the upstream direction

1 The notation f (· ;a,b) signifies that the functional form of the function f depends on the parameters a and b.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 9.3 Centered versus Upwind Schemes 279

 λα (Sin ), if vα > 0,
λα (S)u x = n ),
(9.21)
i+1/2
λα (Si+1 otherwise.

and then use these upwind values to evaluate both the S-shaped viscous flux fw (S) and the
bell-shaped gravity flux λn (S)fw (S). Here, we have used the phase flux vα , which in the
multidimensional case can be computed as ∇pα −gρα ∇z)· ni , where ni is the normal to the
interface. The resulting method, called the explicit upstream mobility-weighting scheme, is
usually the method of choice for porous media application because of its accuracy and
simplicity. In [49] it was shown that the method also satisfies the mathematical properties
necessary to get a consistent and convergent scheme.

9.3.3 Comparison of Centered and Upwind Schemes

The classical schemes introduced so far in this section all have certain weaknesses and
strengths that are common for many other schemes as well. In short, first-order schemes
tend to smear discontinuous solutions, whereas second-order schemes introduce spurious
oscillations. To highlight these numerical artifacts, we consider two different cases; you
can find complete source codes for both examples in the nummet directory of the book
module. The directory also contains two examples of high-resolution methods, which to a
large extent cure these problems. Very cursorily explained, the idea of such a scheme is to
introduce a nonlinear function that enables the scheme to switch from being high-order on
smooth solutions to being first-order near discontinuities.

Example 9.3.1 Consider linear advection on the unit interval with periodic boundary
conditions

ut + ux = 0, u(x,0) = u0 (x), u(0,t) = u(1,t).

The advantage of using periodic data is that we know that the exact solution at time t = n
for an integer number n equals u0 (x). As initial data u0 (x) we choose a combination of a
smooth squared sine wave and a double step function,
 x − 0.1 
u(x,0) = sin2 π χ[0.1,0.4] (x) + χ[0.6,0.9] (x).
0.3
Using the lxf solver outlined in Section 9.3.1, we can set up the case as follows, where the
function periodic simply sets u(1)=u(end-1) and u(end)=u(2).
dx = 1/100;
x = -.5*dx:dx:1+.5*dx;
u0 = sin((x-.1)*pi/.3).^2.*double(x>=.1 & x<=.4);
u0((x<.9) & (x>.6)) = 1;

f = @(u) u; % flux function

df = @(u) 0*u+1;
uf = lxf(u0, .995, dx, 20, f, df, @periodic);

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 280 Discretizing Hyperbolic Transport Equations
Upwind Lax−Friedrichs Lax−Wendroff Nessyahu−Tadmor

Figure 9.1 Approximate solutions at time t = 20.0 for the linear advection equation ut + ux = 0,
with periodic boundary conditions computed by three classical schemes and a high-resolution scheme
on a grid with 100 cells and CFL number 0.995.

Figure 9.1 shows approximate solutions after 20 periods (t = 20) computed by the upwind,
Lax–Friedrichs, and Lax–Wendroff methods and the high-resolution Nessyahu–Tadmor
scheme [224] on a grid with 100 cells for t = 0.995x. Both the centered schemes
clearly give unacceptable resolution of the solution profile. The first-order Lax–Friedrichs
scheme smears both the smooth and the discontinuous part of the advected profile. The
second-order Lax–Wendroff scheme preserves the smooth profile quite accurately, but
introduces spurious oscillations at the two discontinuities. This behavior is representative
for classical schemes. The upwind method represents a reasonable compromise in the sense
that it does not smear the smooth wave and the discontinuities as much as Lax–Friedrichs,
and does not introduce oscillations like Lax–Wendroff. The Nessyahu–Tadmor scheme
gives accurate resolution of both the smooth and the discontinuous parts of the profile,
thereby combining the best of the low and high-order schemes.
The advection equation can serve as a conceptual model for more complex cases. As
we saw in Figures 8.9 and 8.10, a displacement front will be self-sharpening in the sense
that if the leading discontinuity is smeared, smeared states behind the displacement front
will travel faster than the discontinuity (since they have characteristics pointing into it) and
will hence “catch up.” Likewise, smeared states ahead of the front will travel slower and
be overrun by the front, since these states have characteristics that point into the front. For
linear waves, the characteristics of any smeared state will run parallel to the discontinuity
and hence there will be no self-sharpening to counteract the numerical dissipation inher-
ent in any numerical scheme. Linear waves are therefore more susceptible to numerical
dissipation than nonlinear shock waves.
In water-based EOR methods, for instance, active chemical or biological substances
are added to modify the physical properties of the fluids or/and the porous media at the
interface between oil and water. The resulting displacement processes are governed by
complex interplays between the transport of chemical substances and how these substances
affect the flow by changing the properties of the fluids and the surrounding rock. These

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 9.3 Centered versus Upwind Schemes 281
Upwind Lax−Friedrichs Lax−Wendroff Nessyahu−Tadmor

Figure 9.2 Approximate solutions at time t = 1.0 for the linear advection equation ut + ux = 0,
with periodic boundary conditions computed by three classical schemes on a grid with 100 cells and
time step of 0.5 relative to the stability limit.

property changes are often very nonlinear and highly sensitive to threshold parameters
that determine sharp transitions between regions of very different behavior. The transport
of chemical substances is largely linear and hence more affected by numerical diffusion.
A simulation may therefore fail to resolve the local displacement process if the chemi-
cal fronts are smeared out or contain overshoots. As a result, unresolved simulation can
therefore lead to misleading predictions of injectivity and recovery profiles.
Example 9.3.2 For many systems, the linear waves travel significantly slower than the
leading shock wave, and hence will be computed with a smaller effective CFL number. In
Figure 9.2, we have rerun the experiment assuming that the linear waves travel at half
the speed of a leading shock wave. Even after one period, the first-order schemes have
introduced significantly more smearing, and the oscillations for Lax–Wendroff are much
worse. The high-resolution scheme, on the other hand, maintains its resolution much better.
The two previous examples are admittedly idealized cases compared with what we see in
reservoir simulation, but illustrate very well the potential pitfalls of classical schemes. In
the next example, we consider a more relevant case.
Example 9.3.3 Let us revisit the classical Buckley–Leverett profile studied in Example
8.4.1, i.e., consider the following equation,

 S2  1, x < 0.1,
St + 2 = 0, S(0,t) = 1, S(x,0) =
S + (1 − S)2 x 0, x ≥ 0.1.
Figure 9.3 shows the solution at time t = 0.65 computed by the same four schemes. The
first-order schemes compute qualitatively correct approximations of both the leading shock
wave and the trailing rarefaction wave. As expected, Lax–Friedrichs introduces significant
smearing of both the shock and the kink. The upwind scheme delivers acceptable accuracy.
Lax–Wendroff fails completely to capture the correct structure of the composite: not only
does it introduce oscillations behind the displacement front, but the propagation speed is

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 282 Discretizing Hyperbolic Transport Equations
Upwind Lax−Friedrichs Lax−Wendroff High−resolution

Figure 9.3 Approximate solutions at time t = 0.65 for the Buckley–Leverett problem on a grid with
50 cells computed by three classical schemes.

also incorrect. The reason is that once overshoots are introduced near the leading disconti-
nuity, the only way to maintain a mass-conservative solution is if the numerically computed
wave (with overshoots) travels slower than the true wave. Likewise, the scheme introduces
an overshoot at the kink where the solution has discontinuous derivative. These are serious
deficiencies typical of higher-order classical schemes. The high-resolution scheme com-
putes a qualitatively correct solution and overall delivers the best accuracy. On the other
hand, it is more computationally costly and would be outperformed by the upwind scheme
if we compare accuracy versus computational cost.

Computer exercises

9.3.1 Try to set x = t so that the CFL number equals one exactly for the periodic
advection problem in Example 9.3.1. Can you explain what you observe? Why is
using a CFL number identical to one not that interesting from a practical point of
view?
9.3.2 Implement a Godunov solver that works correctly for strictly convex or strictly
concave flux functions. Hint: try to find a formula that gives the correct self-similar
Riemann solution along the line x/t = 0.
9.3.3 Implement the single-point upstream mobility-weighting scheme (9.21) and use it
to simulate the 1D cases discussed in Examples 8.4.2–8.4.4.
9.3.4 Extend the schemes introduced in this section to the two-dimensional conservation
law ut + f (u)x + g(u)y = 0, and implement a solver that can run a 2D analogue of
the periodic advection problem in Example 9.3.1. Try to make the changes in the
existing solver as small as possible. Hint: In 2D, the Lax–Friedrichs scheme reads,

un+1
ij = 14 uni+1,j + uni−1,j + uni,j +1 + uni,j −1
1   1  
− r f (uni+1,j ) − f (uni−1,j ) − r g(uni,j +1 ) − g(uni,j −1 ) .
2 2

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 9.3 Centered versus Upwind Schemes 283

9.3.5 The nummet folder of the book module also contains another high-resolution
scheme, the semi-discrete central-upwind scheme Kurganov et al. [174]. Use the
supplied scripts to also familiarize yourself with this scheme.

9.3.4 Implicit Schemes

So far, we have only considered schemes based on explicit temporal discretization, since
this is the most common approach in the literature on numerical methods for hyperbolic
conservation laws. In principle, we could also have used implicit temporal discretization, as
we have already encountered in Chapter 7. Going back to (9.8), we could approximate the
integral using point values at tn+1 rather than at time tn , which would give us a numerical
method on the form,
 n+1 n+1
un+1
i + rin Fi+1/2 − Fi−1/2 = uni . (9.22)

To get a specific scheme, we must specify how the cell averages un+1 n+1
i−p , . . . ,ui+q are
n+1
used to evaluate the numerical flux Fi+1/2 . The standard approach in reservoir simulation
is to use the upstream mobility-weighting approximation from (9.21) with S n replaced
by S n+1 .
In the explicit method (9.9), the flux terms were written on the right-hand side to signify
that they are given in terms of the known solution uni . In (9.22), the numerical flux terms
appear on the left-hand side to signify that they depend on the unknown cell averages
un+1
i . Equation (9.22) hence represents a coupled system of nonlinear discrete equations,
which typically must be solved by use of a Newton–Raphson method as discussed in
Section 7.1. Computing a single time step of an implicit method is typically significantly
more costly than computing an explicit update, since the latter involves fewer evaluations
of the flux function and its derivatives. On the other hand, the fully implicit discretiza-
tion (9.22) is unconditionally stable in the sense that there is no stability condition (CFL
condition) limiting the size of the time step. In principle, this means that the additional
computational cost can be offset by using much larger time steps. In practice, however,
viable time-step sizes are limited by numerical errors (as we will see shortly) and by the
convergence of the numerical method used to solve the nonlinear system. For Newton–
Raphson-type methods, in particular, the convergence may be quite challenging if the flux
function contains inflection points/lines separating regions of different convexity; see e.g.,
[147, 306, 213] for a more detailed discussion.
For 1D horizontal displacements, we have already seen that the flux function has a
characteristic S-shape with positive derivatives. Like in the explicit case, we can therefore
use one-sided values to evaluate the flux integrals, giving the following scheme,
   
un+1
i + rin f un+1i − f un+1
i−1 = uni .

The result is a triangular nonlinear system with one nonzero band below the diagonal.
For cases like the Buckley–Leverett displacement considered in Example 9.3.3, we can

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 284 Discretizing Hyperbolic Transport Equations

0.8

0.6

0.4

0.2
Explicit
Implicit, CFL=1
0 Implicit, CFL=10
0 0.2 0.4 0.6 0.8 1

Figure 9.4 Approximate solutions at time t = 0.65 for the Buckley–Leverett problem on a grid
with 100 cells computed by the single-point upstream mobility-weighting scheme with explicit and
implicit time discretization.

therefore solve the nonlinear system very efficiently using a substitution method. That is,
we start at the cell next to the inflow boundary,

un+1
1 + r1n f (un+1
1 ) = u1 + r1 fL,
n n

where fL is the known inflow. This single nonlinear equation can be solved robustly for an
arbitrarily large t by use of a classical bracketing method that tracks the end points of an
interval containing the unknown root. Once un+1 1 has been computed, we can move to the
next cell, and solve a similar scalar equation,

un+1
2 + r2n f (un+1 n+1
2 ) = u1 + r2 f (u1 ),
n n

and so on. The same strategy can, in fact, be applied to the multidimensional case under
certain assumptions guaranteeing that the system has a similar cocurrent flow property; see
[220] for more details.

Example 9.3.4 Let us revisit the setup from Example 9.3.3 and compare the explicit (9.9)
and the implicit (9.22) discretizations. Instead of implementing the 1D implicit scheme,
we rely on the general transport solvers from the incomp module in MRST, which will be
discussed in more detail in the next chapter. Figure 9.4 shows solutions computed by the
explicit scheme with a unit CFL number and by the implicit scheme with CFL numbers 1
and 10. At a unit CFL number, the explicit scheme resolves the displacement front somewhat
sharper than the implicit scheme. When the CFL number is increased ten times for the
implicit scheme, the numerical smearing is significant and will lead to the prediction of a
(much) too early water breakthrough.

To explain the difference in numerical smearing, let us do some simple numerical analy-
sis. However, rather than studying the nonlinear Buckley–Leverett model, we go back to the
simple advection equation, ut +aux = 0, and consider the schemes discussed earlier in this

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 9.3 Centered versus Upwind Schemes 285

section as finite-difference schemes rather than finite-volume schemes. The classical way of
analyzing numerical schemes is to first compute the truncation error, which is obtained by
inserting the exact solution into the numerical scheme, assume a smooth solution, and use
Taylor expansions to express the various terms around a common point (x,t). In addition,
we also use the fact that ut = −aux and that t and x are proportional. Starting with the
explicit scheme,
u(x,t + t) − u(x,t) u(x,t) − u(x − x,t)
0= +a
t x
= ut (x,t) + 2 t utt (x,t) + O(t ) + aux (x,t) − 12 ax uxx (x,t) + O(x 2 )
1 2

= ut + 12 t a 2 uxx + O(t 2 ) + aux − 12 ax uxx + O(x 2 )

 
= ut + aux − 12 a x − at uxx + O(x 2 ).

In other words, the explicit scheme will up to order O(x 2 ) solve a parabolic equation.
This equation is only well-posed if the coefficient in front of the second-order term is
negative, implying that at/x ≤ 1, which is exactly the same requirement as in the
CFL condition. We also notice that the higher the CFL number we use, the less numerical
diffusion we introduce in the approximate solution and in the special case that at = x,
the scheme has no numerical diffusion.
For the implicit scheme, we only need to expand the un+1
i−1 term since we have already
n+1
computed the expansion for ui for the explicit scheme,

un+1
i−1 = u(x − x,t + t)
= u(x − x,t) − atux (x − x,t) + 12 (at)2 uxx (x − x,t) + O(t 3 )
 
= u − xux + 12 x 2 uxx + O(x 3 ) − at∂x u − xux + O(x 2 )
 
+ 12 (at)2 ∂x2 u + O(x) + O(t 3 ).

Collecting terms, we have that

u(x,t + t) − u(x,t) u(x,t) − u(x − x,t)
0= +a
t x
 
= ut + aux − 12 a x + at uxx + O(x 2 ).

This means that the implicit scheme also will solve a parabolic equation with accuracy
O(x 2 ). In this case, the coefficient in front of the uxx term is always negative and the
scheme is therefore stable regardless of the choice of t. However, this also means that the
numerical diffusion increases with increasing time steps.
One can also conduct a similar analysis in the nonlinear case and arrive at similar con-
clusions. This in turn implies an interesting observation if we look back at our discussion
of linear waves on page 280. For an explicit scheme, slow waves are computed with more
numerical diffusion than fast waves. For an implicit scheme, on the other hand, slow waves
are computed with less numerical diffusion than fast waves. Explicit schemes are therefore

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 286 Discretizing Hyperbolic Transport Equations

best suited for systems having relatively small differences in wave speeds, whereas implicit
schemes are better suited for systems with large differences in wave speeds. In a typical
reservoir setting, wave speeds are large in the near-well region, of medium size in high-flow
regions away from the wells, and low near or in stagnant zones. It is therefore reasonable to
expect that implicit schemes are best suited for simulating real reservoir systems. However,
explicit schemes are still used, mainly because of their simplicity and in connection with
higher-resolution discretizations.

Computer exercises
9.3.1 Implement the implicit upwind scheme in 1D and verify against the solver from the
incomp module.
9.3.2 Use the single-point upstream mobility-weighting scheme (9.21) to extend your
implicit solver to also account for countercurrent flow and use it to simulate the 1D
cases discussed in Examples 8.4.2–8.4.4. Compare with the corresponding explicit
scheme.

9.4 Discretization on Unstructured Polyhedral Grids

So far, this chapter has introduced you to various types of numerical methods that can be
used to discretize saturation equations. The methods discussed, and some of their high-
resolution extensions included in the book module, are all reasonably simple to implement
on regular Cartesian and rectilinear grids and on grids consisting entirely of simplices (tri-
angles in 2D and tetrahedrons in 3D). However, as you have seen multiple times throughout
the book, realistic grid models are rarely that simple. For the baseline solvers in MRST we
have therefore chosen to go with the most robust choice of them all, namely the explicit
or implicit version of the single-point upstream mobility weighting scheme. This section
therefore discusses in more detail how to formulate this scheme on unstructured, polyhedral
grids for a general transport equation
∂S
φ + ∇ · H (S) = 0. (9.23)
∂t
Here, the flux function H includes viscous flow driven by pressure gradients, as well as
gravity segregation and capillary forces,
λw λ w λn K 
H (S) = v + ρ g + ∇Pc (S)
λ w + λn λ w + λn

= Hf (S) + Hg (S) + Hc (S).
 (9.24)
To discretize this equation, we start from a multi-dimensional analogue of the integral form
(9.8) of the conservation law,
 
1  tn+1 
n+1
Si − Si = n
H S(
x,t) · ni,k ds dt, (9.25)
φi |i | tn ikk

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 9.4 Discretization on Unstructured Polyhedral Grids 287

i,k

pi
pk
i ci,k
ni,k
g k

Figure 9.5 Two cells used to define the finite-volume discretization of the general two-phase transport
equation (9.23). Here, the interface i,k between the two cells i and k has normal vector ni,k and
area Ai,k .

defined over a general grid cell i as illustrated in Figure 9.5. We do not want to develop a
higher-order discretization, and hence it is sufficient to evaluate the time integral at one of
its end-point, so that (9.25) simplifies to

t 
n+1
Si − Si = n
H S(x,tm ) · ni,k ds, m = n,n + 1. (9.26)
φi |i | ik
Before we describe the discretization of the three flux functions Hf , Hg , and Hc , let us
quickly recap how we discretize the viscous flux v = −K∇p in the corresponding pressure
equation. Referring to Figure 4.10 on page 133, the flux reads
(pi − πi,k ) ci,k
vi,k ≈ Ai,k Ki · ni,k = Ti,k (pi − πi,k )
|ci,k | 2
 −1 −1 −1
= Ti,k + Tk,i (pi − pk ),
where the first line gives the one-sided formulation computed entirely from quantities
associated with cell i, and the second line gives the formulation that also couples to the
neighboring cell k. From this expression, it follows naturally that the capillary term should
be discretized as follows:
 −1 −1 −1  
Ai,k ∇Pc (S) · ni,k ≈ Ti,k + Tk,i Pc (Si ) − Pc (Sk ) = Pi,k (S). (9.27)
Likewise, we can define a “gravity flux” gik that is independent of saturation,
 −1 −1 −1 gi,k = (ρ)|i (Ki g) · ni,k ,
gik = gi,k + gk,i , (9.28)
gk,i = (ρ)|k (Kk g) · nk,i .
With this, all we need to do is to find the correct upstream value λuα for each of the two
phases, and then the overall approximation of the interface flux reads:
λuw λuw λun  
Hik = vik + gik + Pik . (9.29)
λuw + λun λuw + λun

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 288 Discretizing Hyperbolic Transport Equations

To determine the upstream directions, we need to compare the Darcy flux vik and the gravity
flux gik . (If there are capillary forces, these are henceforth assumed to be added to the
gravity flux, i.e., gik + Pik → gik .) If vik and gik have the same signs, we know that the
correct choice of the wetting mobility is from the left (right) of the interface if the fluxes
are positive (negative). Likewise, when the two fluxes have different signs, we know that
the correct choice of upstream value for the non-wetting fluid is independent of the actual
mobility values. These cases can be summarized in the following tabular:
sign(vik ) sign(gik ) λuw sign(vik ) sign(gik ) λun
≥0 ≥0 λw (SL ) ≥0 ≤0 λn (SL )
≤0 ≤0 λw (SR ) ≤0 ≥0 λn (SR )
If neither of these cases are fulfilled, we need to check the sign of the phase fluxes to
determine the correct upstream direction. In practice, we do not check the phase fluxes
themselves, but rather the phase fluxes divided by the fractional flow functions as these
quantities are less expensive to compute. That is, we check the sign of vik + gik λn and
vik − gik λw and use this to pick the correct upstream value.
For most cases in reservoir simulation, we also have source terms that drive flow, so that
(9.23) is replaced by an inhomogeneous equation
∂S
+ ∇ · H (S) = max(q,0) + min(q,0)f (S).
φ
∂t
In the next chapter, we describe how the resulting transport solvers are implemented in
MRST, and discuss how to combine them with incompTPFA, or one of its consistent
alternative, in a sequential solution procedure to solve incompressible, multiphase flow
on general polygonal and polyhedral grids.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10
Solvers for Incompressible Immiscible Flow

The multiphase flow equations introduced in Chapter 8 can describe very different flow
behavior depending upon what are the dominant physical effects. During the formation of
petroleum reservoirs, fluid movement is primarily driven by buoyancy and capillary forces,
which govern how hydrocarbons migrate upward and enter new layers of consolidated
sediments. The same effects are thought to dominate long-term geological carbon storage
if the buoyant CO2 phase continues to migrate upward in the formation long after injection
has ceased. In recovery of conventional hydrocarbon resources, on the other hand, the
predominant force is viscous advection caused by pressure differentials. Here, pressure
disturbances will in most cases propagate much faster through the porous medium than
the material waves that transport fluid phases and chemical components. This is one of the
main reasons why solving multiphase flow equations turns out to be relatively complicated.
In addition, we have all the other difficulties already encountered for single-phase flow in
Chapters 5–7. The variable coefficients entering the flow equations are highly heteroge-
neous with orders of magnitude variation and complex spatial patterns involving a wide
range of correlation lengths. The grids used to describe real geological media tend to
be highly complex, having unstructured topologies, irregular cell geometries, and orders
of magnitude aspect ratios. Flow in injection and production wells takes place on small
scales relative to the reservoir and hence needs to be modeled using approximate analytical
expressions, and so on.
This chapter will teach you how to solve multiphase flow equations in the special case
of incompressible rock and immiscible and incompressible fluids. As we saw in Chapter 8,
the system of PDEs can then be reformulated so that it consists of an elliptic equation for
fluid pressure and one or more transport equations. These transport equations are generally
parabolic, but have a strong hyperbolic character (see Section 8.3). Since the pressure and
saturations equations have very different mathematical characteristics, it is natural to solve
them in consecutive substeps. Examples of such methods include the classical IMPES
method [276, 284] (see also [73] and references therein), the adaptive-implicit method
(AIM) [280], and the sequentially implicit method [307]. Operator splitting is also used
in streamline simulation [79] and in recent multiscale methods [195].

289

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 290 Solvers for Incompressible Immiscible Flow

Incompressible models are best suited for fluid systems consisting mainly of liquid
phases such as in waterflooding of oil reservoirs that either do not contain gas com-
ponents or are well below the bubble point. Such systems often have weak coupling
between pressure and fluid transport, which can be exploited by sequential solution
procedures. For systems with a high gas content, significant compressibility effects,
strong coupling between different types of flow mechanisms, or large differences in
time constants, one generally has to use compressible flow models and fully implicit
solvers. Nonetheless, also in this case the mixed elliptic-hyperbolic character of the model
equations plays a key role in developing efficient preconditioning strategies [304, 305].
The combination of incompressible models and sequential solution procedures is very
popular in academia and for research purposes, since it provides a simple means to
develop more clean-cut model equations that still have many of the salient features for
multiphase flow.

10.1 Fluid Objects for Multiphase Flow

In Chapter 5, we discussed the basic data objects entering a flow simulation. When going
from a single-phase to a multiphase flow model, the most prominent changes take place in
the fluid model. It is this model that generally will tell your solver how many phases are
present and how these phases affect each other when flowing together in the same porous
medium. We therefore start by briefly outlining a few fluid objects that implement the basic
fluid behavior discussed in Chapter 8.
To describe an incompressible flow model, we need to know the viscosity and the
constant density of each fluid phase, as well as the relative permeabilities of the fluid phases.
If the fluid model includes capillary forces, we also need one or more functions that specify
the capillary pressure as function of saturation. The most basic multiphase fluid object in
MRST implements a simplified version of the Corey model (8.15)
fluid = initSimpleFluid('mu' , [ 1, 10]*centi*poise , ...
'rho', [1014, 859]*kilogram/meter^3, ...
'n' , [ 2, 2]);

Here, the residual saturations Swr and

nw Snr are assumed
 to be
nnzero and the end-points are
scaled to unity, so that krw = Sw and Srn = 1 − Sw . To recap from Chapter 5,
the fluid object offers the following interface to evaluate the petrophysical properties of the
fluid:
mu = fluid.properties(); % gives mu_w and mu_n
[mu,rho] = fluid.properties(); % .... plus rho_w and rho_n

New to multiphase flow is the relperm function, which takes a single fluid saturation or
an array of fluid saturations as input and outputs the corresponding values of the relative
permeabilities. To plot the relative permeability curves of the fluid object, we can use the
following code:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.1 Fluid Objects for Multiphase Flow 291
1 4.5
9
0.9 4
8
0.8
3.5
7
0.7
3
6
0.6
2.5
5
0.5
2
4
0.4
1.5 3
0.3
1 2
0.2

0.1 0.5 1

0 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Figure 10.1 Corey relative permeabilities (left) and their first and second derivatives (middle and
right) constructed by the initCoreyFluid function.

s=linspace(0,1,20)';
kr = fluid.relperm(s);
plot(s,kr(:,1),'-s',s,kr(:,2),'-o');

The relperm function can also return the first and second derivatives of the relative per-
meability curves when called with two or three output arguments.
The incomp module also implements the general Corey model with end-point scaling
kα0 and nonzero residual saturations Swr and Snr .
fluid = initCoreyFluid('mu' , [ 1, 10]*centi*poise , ...
'rho', [1014, 859]*kilogram/meter^3, ...
'n' , [ 3, 2.5] , ...
'sr' , [ 0.2, .15] , ...
'kwm', [ 1, .85]);

Figure 10.1 shows the relative permeabilities and their first and second derivatives for this
particular model.
Whether the flow equations incorporate capillary pressure is specified by the fluid object.
The incomp module implements two different capillary-pressure models, a simple linear
relationship of the form Pc (S) = C(1 − S) and the Leverett J -function scaling (8.9).
Both models take the same input as initSimpleFluid and are generated the following
functions
fluid = initSimpleFluidPc( .., 'pc_scale', 2*barsa);
fluid = initSimpleFluidPc( .., 'rock', rock,
'surf_tension',10*barsa*sqrt(100*milli*darcy/0.1));

fluid =
properties: @(varargin)properties(opt,varargin{:})
saturation: @(x,varargin)x.s
relperm: @(s,varargin)relperm(s,opt,varargin{:})
pc: @(state)pc_funct(state,opt)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 292 Solvers for Incompressible Immiscible Flow

You may notice that the capillary function pc is evaluated using a state object and not a
saturation. This may seem awkward, but provides a simpler interface to Pc functions that
depend on rock properties, like Leverett-J .

Computer exercises

10.1.1 Modify the Corey model so that it also can output the residual saturations and the
end-point scaling values.
10.1.2 Implement the Brooks–Corey (8.16) and the van Genuchten models (8.17) and
(8.18).
10.1.3 Extend the models to also include the capillary functions (8.11) and (8.12).

10.2 Sequential Solution Procedures

To solve the two-phase, incompressible model, we rely entirely on the fractional-flow
formulation developed in Section 8.3.2. As you may recall, in this formulation, the flow
equations consists of an elliptic pressure equation

∇ · v = q, v = −λ ∇pn − fw ∇Pc − (ρw fw + ρn fn )g∇z (10.1)

and a parabolic transport equation

∂Sw   
φ + ∇ · fw v + λn (ρg∇z + ∇Pc ) = qw . (10.2)
∂t
Here, the capillary pressure pc = pw − pn is assumed to be a known function Pc
of the wetting saturation Sw , and the transport equation becomes hyperbolic whenever Pc
is zero.
In the standard sequential solution procedure, the system (10.1)–(10.2) is evolved in
time using a set of discrete time steps ti . Let us assume that p, v, and Sw are all known at
time t and that we want to evolve the solution to time t + t. At the beginning of the time
step, we first assume that the saturation Sw is fixed. This means that the parameters λ, fw ,
and fn in (10.1) become functions of the spatial variable x only. We then use the resulting
Poisson-type equation to update pressure pn and Darcy velocity v. Next, we hold v and
pn fixed while (10.2) is evolved a time step t to define an updated saturation Sw ( x,t +
t). This saturation is then held fixed when we update pn and v in the next time step,
and so on.
Some authors refer to this solution procedure as an operator splitting method, since
the solution procedure effectively splits the overall solution operator of the flow model
into two parts that are evolved in consecutive substeps. Likewise, some authors refer to
the sequential solution procedure as IMPES, which is short-hand for implicit pressure,
explicit saturation. Strictly speaking, using the name IMPES is only correct if the saturation
evolution is approximated by a single time step of an explicit transport solver. The size of
the splitting step t is then restricted by the CFL condition of the explicit scheme. In many

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.2 Sequential Solution Procedures 293

cases, the overall flow system does not have stability requirements that necessitate such
a restriction on t. Indeed, by writing the flow model in the fractional-flow formulation,
we have isolated the parts of the system that have stability restrictions to the hyperbolic
saturation equation. The elliptic pressure equation, on the other hand, describes (smooth)
solutions resulting from the instant redistribution of pressure in a system with infinite speed
of propagation. For this equation, we can therefore in principle use as large time step as we
want.
As long as we ensure that the evolving discontinuities and sharp transitions are propa-
gated in a stable manner in the saturation equation, our only concern when choosing the
size of the splitting step should be to control or minimize the splitting error introduced by
accounting for pressure and transport in separate substeps. The fractional-flow formulation
underlying our operator splitting was developed to minimize the coupling between satura-
tion and pressure. For incompressible flow models, the effect that dynamic changes in the
saturation field have on pressure is governed entirely by the total mobility λ(S), which in
many cases is a function that locally has small and relatively smooth variation in time. For
this reason, you can typically use splitting steps that are significantly larger than the CFL
restriction from the hyperbolic part of the saturation equation and still accurately resolve
the coupling between pressure and saturation. In other words, for each pressure update, the
saturation can be updated by an explicit solver using multiple saturation substeps, or by an
implicit solver using either a single or multiple saturation substeps. If necessary, you can
also iterate on the splitting steps.

10.2.1 Pressure Solvers

The pressure equation (10.1) for incompressible, multiphase flow is time dependent. This
time dependence comes as the result of three factors:

• K/μ is replaced by the total mobility λ(Sw ), which depends on time through the satura-
tion Sw (
x,t),
• the constant density ρ is replaced by a saturation-dependent quantity ρw fw (Sw )
+ ρn fn (Sw ), and
• the new source term q − ∇λw (Sw )∇Pc (Sw ) depends on saturation.
Nevertheless, once Sw is held fixed in time, all three quantities become functions of x
only, and we hence end up again with an elliptic Poisson-type equation having the same
spatial variation as in (4.10) on page 117. Hence, we can either use the two-point scheme
introduced in Section 4.4.1 or one of the consistent discretization methods from Chapter 6,
mutatis mutandis. The incompTPFA solver discussed in Chapter 5 and the incompMimetic
and incompMPFA solvers from Chapter 6 are implemented so that they solve the pressure
equation for a general system of m incompressible phases. Whether this system has one or
more phases is determined by the fluid object and the reservoir state introduced in Section
5.1.2. We will therefore not discuss the pressure solvers in more detail.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 294 Solvers for Incompressible Immiscible Flow

10.2.2 Saturation Solvers

Apart from the time loop, which we have already encountered in Chapter 7, the only
remaining part we need is a solver for the transport equation (10.2) that implements the
discretizations we introduced in Section 9.4. Summarized, this can be written in the fol-
lowing residual form, for each cell i
 
t 
Fi (s,r) = si − ri + Hik (s) − max(qi ,0) − min(qi ,0)f (Si ) . (10.3)
φi |i |
k
Here, s and r are cell-averaged quantities and subscript i refers to the cell the average is
evaluated in. The sum of the interface fluxes for cell i
 λuw (si ,sk )  
Hi (s) = vik + λun (si ,sk )(gik + Pik ) . (10.4)
λw (si ,sk ) + λn (si ,sk )
u u
k
is computed using the single-point, upstream mobility-weighting scheme discussed on
page 287, whereas the fractional flow function f in the source term is evaluated from the
cell average of S in cell i . The explicit scheme is given as S n+1 = S n − F(S n,S n ) and
the implicit scheme follows as a coupled system of discrete nonlinear equations if we set
F(S n+1,S n ) = 0. In the following, we discuss the inner workings of these solvers in more
detail.

Explicit Solver
The incomp module offers the following explicit transport solver
state = explicitTransport(state, G, tf, rock, fluid, 'mech1', obj1, ..)

which evolves the saturation given in the state object a step tf forward in time. The func-
tion requires a complete and compatible model description consisting of a grid structure G,
petrophysical properties rock, and a fluid model fluid. For the solver to be functional, the
state object must contain the correct number of saturations per cell and an incompressible
flux field that is consistent with the global drive mechanisms given by the 'mech' argu-
ment ('src', 'bc', and/or 'wells') accompanied by correctly specified objects obj, as
discussed in Sections 5.1.3–5.1.5.
In practice, this means that the input value of state must be the output value of a previ-
ous call to an incompressible solver like incompTPFA, incompMPFA, or incompMimetic.
In addition, the function takes a number of optional parameters that determine whether
the time steps are prescribed by the user or to be automatically computed by the solver.
The solver can also ignore the Darcy flux and work as a pure gravity segregation solver if
the optional parameter onlygrav is set to true. Finally, the solver will issue a warning if
the updated saturation value is more than satwarn outside the interval [0,1] of physically
meaningful values (default value: sqrt(eps)).
The explicit solver involves many of the same operations and formulas used for the
spatial discretizations as the implicit solver. To avoid duplication of code we have therefore
introduced a private helper function

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.2 Sequential Solution Procedures 295

[F,Jac] = twophaseJacobian(G, state, rock, fluid, 'pn1', pv1, ..)

that implements the residual form (10.3) and its Jacobian matrix J = dF, returned as two
function handles, F and Jac. With this, the key lines of the explicit saturation solver read
F = twophaseJacobian(G, state, rock, fluid, 'wells', opt.wells, ..);
s = state.s(:,1);
t = 0;
while t < tf,
dt = min(tf-t, getdt(state));
s(:) = s - F(state, state, dt);
t = t + dt;
s = correct_saturations(s, opt.satwarn);
state.s = [s, 1-s];
end

Here, the function getdt implements a CFL restriction on the time step by estimating
the maximum derivative of each function used to assemble interface fluxes and source
terms (you can find details in the code). The function correct_saturations ensures
that the computed saturations stay inside the interval of physically valid states. If this
function issues a warning, it is highly likely that your time step exceeds the stability limit,
or something is wrong with your fluxes or setup of the model.

Implicit Solver
The implicit solver has the same user interface and parameter requirement as the explicit
solver
state = implicitTransport(state, G, tf, rock, fluid, 'mech1', obj1, ..)

In addition, there are optional parameters controlling the Newton–Raphson method used to
solve for S n+1 . To describe this method, we start by writing the residual equations (10.3)
for all cells in vector form
t  
F (s) = s − S + H (s) − Q+ − Q− f (s) = 0. (10.5)
φ||

Here, s is the unknown state at time tf and S is the known state at the start of the time step.
As you may recall from Section 7.1, the Newton–Raphson linearization of an equation like
(10.5) can be written as

0 = F (s 0 + δs) ≈ F (s 0 ) + ∇F (s 0 )δs,

which naturally suggests an iterative scheme in which the approximate solution s +1 in the
( + 1)-th iteration is obtained from

J (s  ) δs +1 = −F (s  ), s +1 ← s  + δs +1 . (10.6)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 296 Solvers for Incompressible Immiscible Flow

Here, δs +1 is called the Newton update and J is the Jacobian matrix. The incomp module
was implemented before automatic differentiation was introduced in MRST, and hence the
Jacobian is computed analytically through the following expansion
dF t  dH df 
J (s) = (s) = 1 + (s) − Q− (s) ,
ds φ|| ds ds
dH dH dλw dH dλn dP
= + + f w λn ,
ds dλw ds dλn ds ds
dH f   dH f  
= w v + λn (g + P ) , = − w v − λw (g + P ) .
dλw λ dλn λ

In general, we are not guaranteed that the resulting values in the vector s +1 lie in the
interval [0,1]. To ensure physically meaningful saturation values, we can introduce a line-
search method, which uses the Newton update to define a search direction p  = δs +1
and tries to find the value α that minimizes h(α) = F (s  + α p  ). We may now either
solve h (α) = 0 exactly, or use an inexact line-search method that only asks for a sufficient
decrease in h. In the implicit solver discussed herein, we have chosen the latter approach
and use an unsophisticated method that reduces α in a geometric sequence. The following
code should give you the idea:
function [state, res, alph, fail] = linesearch(state, ds, target, F, ni)
capSat = @(sat) min(max(0, sat),0);
[alph,i,fail] = deal(0,0,true);
sn = state;
while fail && (i < ni),
sn.s(:,1) = capSat(state.s(:,1) + pow2(ds, alph));
res = F(sn);
alph = alph - 1; i = i + 1;
fail = (norm(res, inf) < target);
end
alph = pow2(alph + 1); state.s = sn.s;

Here, F is a function handle to the residual function F . The number of trials ni in the line-
search method is set through the optional parameter 'lstrails', whereas the target value
is set as the parameter 'resred' times the residual error upon entry. Default values for
'lstrails' and 'resred' are 20 and 0.99, respectively.
The implicit discretization is stable in the sense that there exists a solution S n+1 for
an arbitrarily large time increment t. Unfortunately, there is no guarantee that we will
be able to find this solution using the line-search method described previously. If the time
step is too large, the Newton method may simply compute search directions that do not
point us toward the correct solution. To compensate for this, we also need a mechanism
that reduces the time step if the iteration does not converge and then uses a sequence of
shorter time step to reach the prescribed time tf. First of all, we need to define what we
mean by convergence. This is defined by the optional 'nltol' parameter, which sets the
absolute tolerance  (default value 10−6 ) on the residual F(S n+1,S n )∞ ≤ . In addition,

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.3 Simulation Examples 297

we use a parameter 'maxnewt' that gives the maximum number of iteration steps (default
value 25) the method can take to reach a converged solution. The following code gives
the essence of the overall algorithm of the iterative solver, as implemented in the helper
function newtonRaphson2ph
mints = pow2(tf, -opt.tsref);
[t, dt] = deal(0.0, tf);
while t < tf && dt >= mints,
dt = min(dt, tf - t);
redo_newton = true;
while redo_newton,
sn_0 = resSol; sn = resSol; sn.s(:) = min(1,sn.s+0.05);
res = F(sn, sn_0, dt);
err = norm(res(:), inf);
[nwtfail, linfail, it] = deal(err>opt.nltol,false,0);
while nwtfail && linfail && it < opt.maxnewt,
J = Jac(sn, sn_0, dt);
ds = -reshape(opt.LinSolve(J, reshape(res', [], 1)), ns, [])';
[sn, res, alph, linfail] = update(sn, sn_0, ds, dt, err);
it = it + 1;
err = norm(res(:), inf);
nwtfail = err > opt.nltol;
end
if nwtfail,
% Chop time step in two, or use previous successful dt
else
redo_newton = false;
t = t + dt;
% If five successful steps, increase dt by 50%
end
end
resSol = sn;
end

The algorithm has two optional parameters: 'tsref' with default value 12 gives the num-
ber of times we can halve the time step, whereas 'LinSolve' is the linear solver, which
defaults to mldivide. Beyond this, the best way to find more details about the solver is to
read the code.

10.3 Simulation Examples

You have now been introduced to all the functionality you need to solve incompressible,
multiphase flow problems. It is therefore time to start looking into the qualitative behavior
of such systems and the typical multiphase phenomena you may encounter in practice.
The examples presented in the following are designed to highlight individual effects, or
combinations of effects, and may not always be fully realistic in terms of physical scales,
magnitude of the parameters and effects involved, etc. We will also briefly look at the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 298 Solvers for Incompressible Immiscible Flow

structure of the discrete systems arising in the implicit transport solver. The last section of
the chapter discusses numerical errors resulting from specific choices of discretizations and
solution strategies.
We have already seen that there are essentially three effects that determine the direction
of a single-phase flow field. The first is heterogeneity (i.e., spatial variations in permeabil-
ity) that affects the local magnitude and direction of the flow. The second effect is intro-
duced by drive mechanisms, such as wells and boundary conditions, that determine where
fluids flow to and from. However, the further you are from the location of a well or a bound-
ary condition, the less effect it will have on the flow direction. The last effect is gravity.
Multiphase flow is more complicated, since the fluid dynamics is now also affected by
the viscosity and density ratios of the fluids present, as well as by relative permeability and
capillary pressure. These effects will introduce several challenges. If the displacing fluid
is more mobile than the resident fluid, it will tend to move rapidly into this fluid, giving
weak shock fronts and long rarefaction waves. For a homogeneous reservoir, this will result
in early breakthrough of the displacing fluid, and slow and incremental recovery of the
resident fluid. In a heterogeneous reservoir, you may also observe viscous fingering, which
essentially means that the displacing fluid moves unevenly into the resident fluid. This is a
self-reinforcing effect that causes the fingers to move farther into the resident fluid.
Gravity segregation, on the other hand, will force fluids having different density to
segregate, and lead to phenomena such as gravity override, in which a lighter fluid moves
quickly on top of a denser fluid. This is a problem in many recovery methods relying on
gas injection and for geological storage of CO2 . Finally, we have capillary effects, which
tend to spread out the interface between the invading and displaced fluids. When combined
with heterogeneity, these effects are generally difficult to predict without detailed simula-
tions. Sometimes, they work in the same direction to aggravate sweep and displacement
efficiency, but can also counteract each other to cancel undesired behavior. Gravity and
capillary forces, for instance, may both reduce viscous fingering that would otherwise give
undesired early breakthrough.
This section discusses most of the phenomena just outlined in more detail. In most cases,
we only consider a single effect at the time. Throughout the examples, you will also learn
how to set up various types of simulations using MRST. As a rule, for brevity we will not
discuss complete codes, and when reading the material you should therefore take the time
to also read the accompanying codes found in the in2ph directory of the book module.
You will gain much more insight if you run these codes and try to modify them to study the
effect of different parameters and algorithmic choices. I also strongly encourage you to do
as many of the computer exercises as possible.

10.3.1 Buckley–Leverett Displacement

As a first example, let us revisit the 1D horizontal setup from Example 9.3.4 on page 284, in
which we compared explicit and implicit transport solvers for the classic Buckley–Leverett

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.3 Simulation Examples 299

displacement profile arising when pure water is injected into pure oil. The following code
sets up a slightly rescaled version of the problem, computes the pressure solution, and then
uses the explicit transport solver to evolve the saturations forward in time
G = computeGeometry(cartGrid([100,1]));
rock = makeRock(G, 100*milli*darcy, 0.2);
fluid = initSimpleFluid('mu' , [1, 1].*centi*poise, ...
'rho', [1000, 1000].*kilogram/meter^3, 'n', [2,2]);
bc = fluxside([], G, 'Left', 1, 'sat', [1 0]);
bc = fluxside(bc, G, 'Right', -1, 'sat', [0 1]);
hT = computeTrans(G, rock);
rSol = initState(G, [], 0, [0 1]);
rSol = incompTPFA(rSol, G, hT, fluid, 'bc', bc);
rSole = explicitTransport(rSol, G, 10, rock, fluid, 'bc', bc, 'verbose',true);

The explicit solver uses 199 time steps to reach time 10. Let us try to see if we can do this
in one step with the implicit solver:
[rSoli, report] = ...
implicitTransport(rSol, G, 10, rock, fluid, 'bc', bc, 'Verbose', true);

This corresponds to running the solver with a CFL number of approximately 200. From
the output in Figure 10.2 we see that this is not a big success. With an attempted time
step of t = 10, the solver only manages to reduce the residual by a factor 2.5 within the
allowed 25 iterations. Likewise, when the time step is halved to t = 5, the solver still
only manages to reduce the residual one order of magnitude within the 25 iterations. When
the time step is halved once more, the solver converges in 20 iterations in the first step and
then in 9 iterations in the next 3 substeps. This time-step chopping is not very efficient:
Altogether, more than half of the iterations (50 out of 97) were wasted trying to compute
time steps that would not converge. The explicit solver avoids this problem since the time
step is restricted by a CFL condition, but requires significantly more time steps.
To overcome the problem with wasted iterations, we can explicitly subdivide the pres-
sure step into multiple saturation steps:
rSolt = rSol;
for i=1:n
rSolt = implicitTransport(rSolt, G, 10/n, rock, fluid, 'bc', bc);
end

Figure 10.3 reports the approximate solutions and the overall number of iterations used
by the implicit solver with n equally spaced substeps. The solver typically needs more
iterations during the first time steps when the displacement front is relatively sharp. As
the simulation progresses, the shock is smeared across multiple cells, which contributes to
weaken the nonlinearity of the discrete system and hence reduce the number of required
iterations. This explains why the reported number of iterations is not an exact multiple
of the number of time steps. We also see that to get comparable accuracy as the explicit

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 300 Solvers for Incompressible Immiscible Flow
--------------------------------------------------------------
Time interval (s) iter relax residual rate
--------------------------------------------------------------
10
[0.0e+00, 1.0e+01]: 1 1.00 7.82670e+00 NaN
[0.0e+00, 1.0e+01]: 2 0.12 6.81721e+00 0.07
: : : : : :
[0.0e+00, 1.0e+01]: 25 0.03 3.17993e+00 1.00 20
-------------------- Reducing step -------------------
[0.0e+00, 5.0e+00]: 1 1.00 7.03045e+00 NaN
: : : : : : 30
[0.0e+00, 5.0e+00]: 25 0.25 6.34074e-01 0.74
-------------------- Reducing step -------------------
[0.0e+00, 2.5e+00]: 1 1.00 5.64024e+00 NaN 40
: : : : : :
[0.0e+00, 2.5e+00]: 20 1.00 1.23535e-08 1.91
-------------------- Next substep -----------------------
50
[2.5e+00, 5.0e+00]: 1 0.25 3.59382e-01 NaN
: : : : : :
[2.5e+00, 5.0e+00]: 9 1.00 1.71654e-08 1.86
-------------------- Next substep ----------------------- 60
[5.0e+00, 7.5e+00]: 1 0.25 2.67914e-01 NaN
: : : : : :
[5.0e+00, 7.5e+00]: 9 1.00 8.49299e-09 1.94 70
-------------------- Next substep -----------------------
[7.5e+00, 1.0e+01]: 1 0.25 2.57462e-01 NaN
: : : : : : 80
[7.5e+00, 1.0e+01]: 9 1.00 8.74041e-11 2.01

Iterations : 47 Wasted iterations : 50 90

Sub steps : 4 Failed steps : 2
Final residual : 8.74e-11 Convergence rate : 1.9
-------------------------------------------------------------- 10
−10
10
−5 0
10

Figure 10.2 Results from running the implicitTransport solver on a 1D Buckley–Leverett

displacement problem with CFL number 200. Left: screen output, where several lines have been
deleted for brevity. Right: convergence history for the residual, with cumulative iteration number
increasing from top to bottom.

1
Expl: 199 steps
0.9 n= 4: 47 its
n= 10: 65 its
0.8 n= 20: 102 its
n= 40: 161 its
0.7 n=100: 301 its
n=200: 407 its
0.6

0.5

0.4

0.3

0.2

0.1

0
0 10 20 30 40 50 60 70 80 90 100

Figure 10.3 Approximate solutions computed by the explicit transport solver and the implicit
transport solver with n time steps.

transport solver, the implicit solver needs to use at least 40 time steps, which amounts to
more than 160 iterations. In this case, there is thus a clear advantage of using the explicit
transport solver if we want to maximize accuracy versus computational cost.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.3 Simulation Examples 301

10.3.2 Inverted Gravity Column

In the next example, we revisit the inverted gravity column from Example 8.4.3 on page 267
with a light fluid at the bottom and a heavier fluid at the top. We change the setup slightly
so that the fluids are representative for supercritical CO2 and brine found at conditions that
would be plausible when storing CO2 in a deep saline aquifer. The following is the essence
of the simulator (plotting commands are not included for brevity):
gravity reset on
G = computeGeometry(cartGrid([1, 1, 40], [1, 1, 10]));
rock = makeRock(G, 0.1*darcy, 1);
fluid = initCoreyFluid('mu' , [0.30860, 0.056641]*centi*poise, ...
'rho', [975.86,686.54]*kilogram/meter^3, ...
'n', [2,2], 'sr', [.1,.2], 'kwm',[.2142,.85]);
hT = computeTrans(G, rock);
xr = initResSol(G, 100.0*barsa, 1.0); xr.s(end/2+1:end) = 0.0;
xr = incompTPFA(xr, G, hT, fluid);
dt = 5*day; t=0;
for i=1:150
xr = explicitTransport(xr, G, dt, rock, fluid, 'onlygrav', true);
t = t+dt;
xr = incompTPFA(xr, G, hT, fluid);
end

In Example 8.4.3 the fluids had the same viscosity and hence moved equally fast upward
and downward. Here, supercritical CO2 is much more mobile than brine and will move
faster to the top of the column than brine moves downward. Hence, whereas the CO2
reaches the top of the column after 250 days, it takes more than 400 days before the
first brine has sunk to the bottom. After approximately two years, the fluids are clearly
segregated and separated by a sharp interface; see Figure 10.4.
In the simulation, we used relatively small splitting steps (150 steps of 5 days each)
to march the transient solution towards steady state. Looking at Figure 10.5, which shows
the vertical pressure distribution every fiftieth day (i.e., for every tenth time step), we see
that the pressure behaves relatively smoothly compared with the saturation distribution.
It is therefore reasonable to expect that we could get away with using a smaller num-
ber of splitting steps in this particular case. How many splitting steps do you think we
need?
Before leaving the problem, let us inspect the discrete nonlinear system from the implicit
transport solver in some detail. Figure 10.6 contrasts the sparsity pattern at two instances
in time to that of the 1D horizontal Buckley–Leverett problem. From the discussion in
Section 9.3.4, we know that the latter only has a single nonzero band below the diagonal
and hence can be solved more robustly if we instead of using Newton’s method, use
a nonlinear substitution method with a bracketing method for each single-cell problem
[220]. Since the lighter fluid moves upward and the heavier fluid moves downward during
gravity segregation, there will be nonzero elements above and below the diagonal in the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 302 Solvers for Incompressible Immiscible Flow

Figure 10.4 Simulation of an inverted gravity column where pure CO2 initially fills the bottom half
and brine the upper half of the volume. The upper plot shows S(z,t) with yellow color signifying
pure CO2 and blue color signifying pure brine. At equilibrium, the CO2 at the top contains some
irreducible water and the brine at the bottom contains residual CO2 , which thus can be considered as
safely trapped within the immobile brine.

4
x 10
8

−1
0 1 2 3 4 5 6 7 8 9 10

Figure 10.5 Pressure distribution at every tenth time step as function of depth from the top of the
gravity column, with blue color and dotted markers indicating initial time and red color and cross
markers indicating end of simulation.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.3 Simulation Examples 303
0 0 0

10
5 5
20
10 10
30
15 15
40

50 20 20

60 25 25

70
30 30
80
35 35
90

100 40 40
0 10 20 30 40 50 60 70 80 90 100 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
nz = 199 nz = 78 nz = 118

Figure 10.6 Sparsity pattern for the 1D Buckley–Leverett problem (left) and the inverted gravity
column after 125 days (middle) and 390 days (right)

two-phase region and at the interface between the two phases. In the figure, we also see
how the sparsity pattern changes as the two-phase region expands upward and downward
from the initial interface. Since the nonlinear system is no longer triangular, a substitution
method cannot be used, but each linearized system can be solved efficiently using the
Thomas algorithm, which is a special O(n) Gaussian elimination method for tridiagonal
systems. Knowing the sparsity pattern of the problem is a key to efficient solvers. The
mldivide solver (i.e., A\b) in MATLAB performs an analysis of the linear system and
picks efficient solvers for triangular, tridiagonal, and other special systems. To confirm that
the optimal solver is indeed used, you can type spparms('spumoni',2) before you run the
example.

10.3.3 Homogeneous Quarter Five-Spot

To gain more insight into the simulation of multiphase displacement processes, we consider
the classical confined quarter five-spot test case discretized on a 128×128 grid. As you may
recall from Section 5.4.1, this test case consists of one quarter of a symmetric pattern of four
injectors surrounding a producer (or vice versa), repeated to infinity in each direction. We
neglect capillary and gravity forces and assume a simplified Corey model with exponent
2.0 and zero residual saturation, i.e., krw = S 2 and kro = (1 − S)2 . We start by setting
the viscosity to 1 cP for both fluid phases, giving a unit mobility ratio. The injector and
producer operate at fixed bottom-hole pressure, giving a total pressure drop of 100 bar
across the reservoir. Since we have assumed incompressible flow, equal amounts of fluid
must be produced from the reservoir so that injection and production rates sum to zero. The
total time is set such that 1.2 pore volumes of fluid will be injected if the initial injection is
maintained throughout the whole simulation. However, the actual injection rate will depend
on the total resistance to flow offered by the reservoir, and hence vary with time when the
total mobility varies throughout the reservoir as a result of fluid movement. (Remember
that the total mobility is less in all cells containing two fluid phases.) The simulation code

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 304 Solvers for Incompressible Immiscible Flow

t=0.20 PVI t=0.40 PVI t=0.60 PVI t=0.80 PVI Sw

1
0.8
0.6
0.4
0.2

Figure 10.7 Evolution of a two-phase displacement front for a homogeneous quarter five-spot√ case
with water injected into oil. The single line shown in each plot is a contour at value t/(2 2 − 2) PVI
for the time-of-flight field computed from the corresponding single-phase problem (i.e., with λ ≡ 1).

follows the same principles as outlined in the two examples just discussed; you can find
details in the file quaterFiveSpot2D.m.
Figure 10.7 shows how the displacement profile resulting from injection of water into
the oil-filled reservoir expands circularly near the injector. As the displacement front prop-
agates into the reservoir, it gradually elongates along the diagonal and forms a finger that
extends towards the producer. As a result, water breaks through in the producer long time
before the displacement front has managed to sweep the stagnant regions near the northwest
and southeast corners. The evolution of the saturation profile is the result of two different
multiphase effects.
To better understand these effects, it is instructive to transform our 3D transport equation
into streamline coordinates. Since the flow field is incompressible, we can use (4.44) to
write v · ∇ = φ ∂τ ∂
so that (10.2) transforms to a family of 1D transport equations, one
along each streamline,
∂S ∂fw (S) qw
+ = . (10.7)
∂t ∂τ φ
Hence, the first flow effect is exactly the same Buckley–Leverett displacement as we saw
in Section 10.3.1, except that it now acts along streamlines rather than along the axial
directions. It is therefore tempting to suggest that to get a good idea of how a multiphase dis-
placement will evolve, we can solve a single-phase pressure equation for the initial oil-filled
reservoir, compute the resulting time-of-flight field, and then map the 1D Buckley–Leverett
profile (8.60) computed from (8.53) onto time-of-flight. How accurate this approximation
is depends on the coupling between the saturation and pressure equations.
With linear relative permeabilities and unit mobility ratio, there is no coupling between
pressure and transport, and mapping 1D solutions onto time-of-flight thus produces the
correct solution. In other cases, changes in total mobility will modify the total Darcy
velocity and hence the time-of-flight. To illustrate this, let us compare the propagation of
the leading shock predicted by the full multiphase simulation and our simplified streamline
analysis. For fluids with a viscosity ratio μw /μn = M, it follows by solving f  (S) =
f (S)/S that√ the leading shock of the Buckley–Leverett displacement profile moves at a
speed M/(2 M + 1 − 2) relative to the Darcy velocity, shown as a single black line

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.3 Simulation Examples 305

for each snapshot in Figure 10.7. Compared with our simplified streamline analysis, the
movement of the injected water is retarded by the reduced mobility in the two-phase region
behind the leading displacement front.
In most simulations, the primary interest is to predict well responses. To extract these,
we introduce the following function:
function wellSol = getWellSol(W, x, fluid)
mu = fluid.properties();
wellSol(numel(W))=struct;
for i=1:numel(W)
out = min(x.wellSol(i).flux,0); iout = out<0; % find producers
in = max(x.wellSol(i).flux,0); iin = in>0; % find injectors
lamc = fluid.relperm(x.s(W(i).cells,:))./mu; % mob in completed cell
fc = lamc(:,1)./sum(lamc,2); %
lamw = fluid.relperm(W(i).compi)./mu; % mob inside wellbore
fw = lamw(:,1)./sum(lamw,2); %
wellSol(i).name = W(i).name;
wellSol(i).bhp = x.wellSol(i).pressure;
wellSol(i).wcut = iout.*fc + iin.*fw;
wellSol(i).Sw = iout.*x.s(W(i).cells,1) + iin.*W(i).compi(1);
wellSol(i).qWs = sum(out.*fc) + sum(in.*fw);
wellSol(i).qOs = sum(out.*(1-fc)) + sum(in.*(1-fw));
end

Inside the simulation loop, this function is called as follows

[wellSols, oip] = deal(cell(N,1), zeros(N,1));
for n=1:N
x = incompTPFA(x, G, hT, fluid, 'wells', W);
x = explicitTransport(x, G, dT, rock, fluid, 'wells', W);
wellSols{n} = getWellSol(W, x, fluid);
oip(n) = sum(x.s(:,2).*pv);
end

The loop also computes the oil in place at each time step. Storing well responses in a cell
array may seem unnecessarily complicated and requires a somewhat awkward construction
to plot the result
t = cumsum(dT);
plot(t, cellfun(@(x) x(2).qOs, wellSols));

The call to cellfun passes elements from the cell array wellSols to an anonymous
function that extracts the desired field qOs containing the surface oil rate of the second
well (the producer). Each element represents an individual time step. The reason for using
cell arrays is to provide compatibility with the infrastructure developed for compressible
models of industry-standard complexity. Here, use of a cell array provides the flexibility
needed to process much more complicated well output. As a direct benefit, we can use a
GUI developed for visualizing well responses:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 306 Solvers for Incompressible Immiscible Flow

Figure 10.8 Graphical user interface plotWellSols from the ad-core module for plotting com-
puted well responses. The GUI does not work in GNU Octave.

mrstModule add ad-core

plotWellSols(wellSols,cumsum(dt))

This brings up a plotting window as shown in Figure 10.8, where we have chosen to
visualize the surface water rate for injector and producer.
To better understand the deviation between the ideal and the actual recovery, we can
look at the well responses and total mass balance in more detail, as shown in Figure
10.9. We see that the oil rate drops immediately as water enters the system in the first
time step and then decays slowly until water breaks through in the producer around time t
= 0.7. Since this well now produces a mixture of water and oil, the oil rate decays rapidly
as the smoothed displacement front enters the well, and then decays more slowly when
the inflow of water is determined by the trailing rarefaction wave. The left plot shows
the cumulative oil production computed in two different ways: (i) using the oil rate from
the well solution, and (ii) measured as the difference between initial and present oil in
place. Up to water breakthrough, the two estimates coincide. After water breakthrough, the
production estimated from the well solution will be slightly off, since it for each time step
uses a simplified approximation that multiplies the size of the time step with the total flow
rate computed at the start of the time step and the fractional flow in the completed cell at
the end of the time step.

Computer exercises
10.3.1 Repeat the experiment with wells controlled by rate instead of pressure. Do you
observe any differences and can you explain them?
10.3.2 Repeat the experiment with different mobility ratios and Corey exponents.
10.3.3 Can you correct the computation of oil/water rates?

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.3 Simulation Examples 307

Figure 10.9 Well responses computed for the homogeneous quarter five-spot test. The left plot shows
the cumulative oil production computed from the well solution shown as a thin line compared with the
amount of extracted oil derived from a mass-balance computation (initial oil in place minus current
oil in place) shown as a thick dashed line. The lower-right plot shows oil rate and the upper-right plot
shows water saturation and water cut (fractional flow) in the well perforation.

10.3.4 Heterogeneous Quarter Five-Spot: Viscous Fingering

In the previous example we studied imbibition in a homogeneous medium, which resulted
in symmetric displacement profiles. However, when a displacement front propagates
through a porous medium, the combination of viscosity differences and permeability het-
erogeneity may introduce viscous fingering effects. In general, the term viscous fingering
refers to the onset and evolution of instabilities at the interface between the displacing and
displaced fluid phases. Fingering can arise because of viscosity differences between two
phases or as a result of viscosity variations within a single phase that, for instance, contains
solutes. In the laboratory, viscous fingering is usually studied in so-called Hele–Shaw cells,
which consist of two flat plates separated by a tiny gap. The plates can be completely
parallel, or contain small-scale variations (rugosity) to emulate a porous medium. When
a viscous fluid confined in the space between the two plates is driven out by injecting a
less viscous fluid (e.g., dyed water injected into glycerin), beautiful and complex fingering
patterns can be observed. I recommend a search for “Hele–Shaw cell” on YouTube.
Figure 10.10 shows results of three different simulations of a quarter five-spot on a
square domain represented on a uniform 60 × 120 grid with petrophysical properties sam-
pled from the topmost layer of the SPE 10 data set. The wells operate at a fixed rate,
corresponding to the injection of one pore volume over a period of 20 years. To reach a
final time, we use 200 pressure steps and the implicit transport solver. As in the previous
example, the fluid is assumed to obey a simple Corey fluid model with quadratic relative
permeabilities and no residual saturations. (Complete code for the example is found in
viscousFingeringQ5.m.)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 308 Solvers for Incompressible Immiscible Flow

Figure 10.10 Quarter five-spot solutions on a subsample from the first layer of the SPE 10 model for
three different viscosity ratios μw : μn .

If the viscosity of the injected fluid is (significantly) less than that of the resident fluid,
we get an unfavorable displacement in which the displacing phase forms a weak shock front
that will “finger” rapidly through the less mobile phase that initially fills the medium. Once
a finger develops, it will create a preferential flow direction for the injected phase, which
causes the finger to extend towards the producer, following the path of highest permeability.
In the opposite case of a (significantly) more viscous fluid being injected into a less viscous
fluid, one obtains a strong front that acts almost like a piston and creates a very favorable
and stable displacement with a leading front that has much fewer buckles than in the
unfavorable case. Not only does this front have better local displacement efficiency (i.e.,
it can push out more oil), but the areal sweep is also better. The unit viscosity case is
somewhere in between the two, having a much better local displacement efficiency than
the unfavorable case, but almost the same areal sweep at the end of the simulation.
Figure 10.11 reports well responses for the three simulations. Because water is injected
at a fixed rate, the oil rate will remain constant until water breaks through in the producer.
This happens after 1,825 days in the unfavorable case, after 4,050 days for equal viscosities,
and after 6,300 days in the favorable case. As discussed in the previous example, the decay
in oil rate depends on the strength of the displacement front and will hence be much more
abrupt in the favorable mobility case, which has an almost piston-like displacement front.
On the other hand, by the time the favorable case breaks through, the unfavorable case has
reached a water cut of 82%. Water handling is generally expensive and in the worst case
the unfavorable case might shut down before reaching the end of the 20-year production
period.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.3 Simulation Examples 309
cumulative oil production
18,000 Water cut
initial oil in place 0.8
16,000
0.6

14,000 0.4

12,000 0.2

0
10,000 2,000 4,000 6,000
Time [Days]
8,000
Oil surface rate

6,000 2
P (Ratio 1:10) 1.5
4,000 P (Ratio 1:1)
1
P (Ratio 10:1)
2,000
0.5

0 0
1,000 2,000 3,000 4,000 5,000 6,000 7,000 2,000 4,000 6,000
Time [days]
Time (days)

Figure 10.11 Well responses computed for the heterogeneous quarter five-spot with different
mobility ratios μw : μn .

Next, we look at the sparsity of the transport equation. In the homogeneous case, all
fluxes point in the positive axial directions and the Jacobian matrix is thus lower triangular.
With heterogeneous permeability, or another well pattern, this unidirectional flow property
is no longer present and the Jacobian matrix will have elements above and below the diag-
onal; see Figure 10.12. However, if we look at the transformation to streamline coordinates
(10.7), it is clear that we still have unidirectional flow along streamlines. This means that
the Jacobian matrix can be permuted to triangular form by performing a topological sort on
the flux graph derived from the total Darcy velocity. In MATLAB, you can try to permute
the matrix to upper-triangular form by use of the Dulmage–Mendelsohn decomposition:
[p,q] = dmperm(J); Js = J(p,q);

The result is shown in the middle plot in Figure 10.12. This permutation is similar to what
is done inside MATLAB’s linear solver mldivide. Since we can permute the Jacobian
matrix to triangular form, we can also do the same for the nonlinear system, and hence
apply a highly efficient nonlinear substitution method [220] as discussed for 1D Buckley–
Leverett problems in Section 9.3.4. The same applies also for 3D cases as long as long
as capillary forces are neglected and we have purely co-current flow. Countercurrent flow
can be introduced by gravity segregation, as we saw in Section 10.3.2, or if the flow field
is computed by one of the consistent discretization schemes from Chapter 6, which are
generally not monotone.
It is also possible to permute the discretized system to triangular form by performing a
potential ordering [175], as shown to the right in Figure 10.12:
[,i] = sort(x.pressure); Jp = J(i,i);

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 310 Solvers for Incompressible Immiscible Flow

Figure 10.12 The left plot shows the sparsity structure for the Jacobian matrix for the heterogeneous
quarter five-spot with viscosity ratio 1:10. The middle plot shows the sparsity structure after a
topological sort, whereas the right plot shows the sparsity structure after potential ordering.

Implementing this type of nonlinear substitution methods is unfortunately not very efficient
in MATLAB and should be done using a compiled language. In MRST, we therefore
mainly rely on the intelligence built into mldivide to give us the required computational
performance.
A remark at the end: in a real case, injectivity would obviously be a decisive factor,
i.e., how high pressure is required to ensure a desired injection rate without fracturing the
formation, or vice versa, which rate one would obtain for a given injection pressure below
the fracturing pressure. This is not accounted for in our earlier discussion; for illustration
purposes we tacitly assumed that the desired injection rate could be maintained.

Computer exercises
10.3.1 Repeat the experiments with wells controlled by pressure, fixed water viscosity,
and varying oil viscosity. Can you explain the differences you observe?
10.3.2 Run a systematic study that repeats the quarter five-spot simulation from the pre-
vious exercise for each of the 85 layers of the SPE 10 model. Plot and compare
the resulting production curves. (You can run the experiment for a single mobility
ratio to save computational time.)
10.3.3 Run the same type of study with 100 random permeability fields, e.g., as generated
by the simplified gaussianField routine from Section 2.5.2.
Alternatively, you can use any kind of drawing pro-
gram to generate a bitmap and generate a channelized
permeability as follows
K = ones(G.cartDims)*darcy;
I = imread('test.pbm');
I = flipud(I(:,:,1))';
K(I) = milli*darcy;

This can easily be combined with different random fields for the foreground and
background permeability.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.3 Simulation Examples 311

Figure 10.13 Illustration of the sloping sandbox used for the buoyancy example and how it is
simulated by rotating the gravity vector. (Color: Gaussian porosity field.)

10.3.5 Buoyant Migration of CO2 in a Sloping Sandbox

In Section 10.3.2, we considered the buoyant migration of supercritical CO2 inside a ver-
tical column. Here, we extend the problem to three spatial dimensions and simulate the
upward movement of CO2 inside a sloping sandbox with sealing boundaries. The rectangu-
lar sandbox has dimensions 100 × 10 × 200 m3 , and we consider two different petrophysi-
cal models: homogeneous properties or Gaussian porosity with isotropic permeability given
from a Carman–Kozeny transformation similar to (2.6). The sandbox is rotated around the
y-axis so that the top surface makes an angle of inclination θ with the horizontal plane.
Instead of rotating the grid so that it aligns with the aquifer geometry, we will rotate the
coordinate system by rotating the gravity vector an angle θ around the y-axis; see Figure
10.13. The rotation is introduced as follows:
R = makehgtform('yrotate',-pi*theta/180);
gravity reset on
gravity( R(1:3,1:3)*gravity().' );

MRST defines the gravity vector as a persistent, global variable, which by default (for his-
 The second line sets g to the standard value (pointing downward
torical reasons) equals 0.
vertically) before we perform the rotation.
To initialize the problem, we assume that CO2 , which is lighter than the resident brine,
fills up the model from the bottom and to a prescribed height,
xr = initResSol(G, 1*barsa, 1);
d = gravity() ./ norm(gravity);
dc = G.cells.centroids * d.';
xr.s(dc>max(dc)-height) = 0;

For accuracy and stability, the time step is ramped up gradually as follows,
dT = [.5, .5, 1, 1, 1, 2, 2, 2, 5, 5, 10, 10, 15, 20, ...
repmat(25,[1,97])].*day;

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 312 Solvers for Incompressible Immiscible Flow
Initial 20 days 100 days 250 days 500 days 1000 days 1500 days 2500 days 1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

Figure 10.14 Buoyant migration of CO2 in a sandbox with sealing boundaries. The upper plots show
the homogeneous case and the lower plots the case with Gaussian heterogeneity. In the plots, only
cells containing some CO2 are colored.

to reach a final simulation time of 2,500 days. The remaining code is similar to what was
discussed in earlier examples; details can be found in buoyancyExample.m.
For the homogeneous case, the buoyant CO2 plume will initially form a cone shape
as it migrates upward and gradually drains the resident brine. After approximately 175
days, the plume starts to accumulate as a thin layer of pure CO2 under the sloping east
face of the box. This layer will migrate quickly up towards the topmost northeast corner
of the box, which is reached after approximately 400 days, This corner forms a structural
trap that will gradually be filled as more CO2 migrates upward. The trapped CO2 forms a
diffused and curved interface (see the plots at 500 and 1,000 days), but as time passes, the
interface becomes sharper and flatter. During the same period, brine will imbibe into the
trailing edge of the CO2 plume and gradually form a layer of pure brine at the bottom. After
approximately 1,000 days, the only CO2 left below the interface in the northeast corner is
found at small saturation values and will therefore migrate very slowly upward.
The heterogeneous case follows much of the same pattern, except that the leading
drainage front will finger into high-permeability regions of the sandbox. Low-permeability
cells, on the other hand, will retard the plume migration. Altogether, we see a significant
delay in the buoyant migration compared with the homogeneous case. Since the permeabil-
ity is isotropic, the plume will still mainly migrate upward. This should not be expected
in general. Many rocks have significantly lower permeability in the vertical direction or
consist of strongly layered sandstones containing mud drapes or other thin deposits that
inhibit vertical movement between layers. For such cases, one can expect a much larger
degree of lateral movement.

Computer exercises
10.3.1 To gain more insight into the flow physics of a buoyant phase, you should experi-
ment more with the buoyancyExample script. A few examples:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.3 Simulation Examples 313

• Set θ = 88 and initial height to 10 meters for the heterogeneous case.

◦
• Set θ = 60 and impose an anisotropic permeability with ratio 0.1:1:5 to mimic
a case with strong layering.
• In the experiments so far, we used an unrealistic fluid model without residual
saturations. Replace the fluid model by a more general Corey model having
residual saturations (typical values could be 0.1 or 0.2) and possibly also end-
point scaling. How does this affect the upward plume migration? (Hint: in addi-
tion to the structural trapping at the top of the formation, you will now have
residual trapping.)

10.3.2 Rerun the experiment with capillary forces included, e.g., by using initSimple
FluidJfunc with parameters set as in its documentation.
10.3.3 Replace the initial layer of CO2 at the bottom of the reservoir by an injector that
injects CO2 at constant rate before it is shut in. For this, you should increase the
spatial and temporal scales of the problem.
10.3.4 Gravity will introduce circular currents that destroy the unidirectional flow
property we discussed in Section 10.3.4. To investigate the sparsity of the
discretized transport equations, you can set a breakpoint inside the private
function newtonRaphson2ph used by the implicit transport solver and use the
plotReorder script from the book module to permute the Jacobian matrix
to block-triangular form. Try to stop the simulation in multiple time steps to
investigate how the sparsity structure and degree of countercurrent flow change
throughout the simulation.

10.3.6 Water Coning and Gravity Override

Water coning is a production problem in which water (from a bottom drive) is sucked
up in a conical shape towards a producer. This is highly undesirable since it reduces the
hydrocarbon production. As an example, we consider a production setup on a sector model
consisting of two different rock types separated by a fully conductive, inclined fault as
shown in Figure 10.15.
A vertical injector is placed in the low-permeability stone (K= 50 md and φ = 0.1)
to the east of the fault, whereas a horizontal producer is perforated along the top of the
more permeable rock (K= 500 md and φ = 0.2) to the west of the fault. The injector
operates at a fixed bottom-hole pressure of 700 bar and the producer operates at a fixed
bottom-hole pressure of 100 bar. To clearly illustrate the water coning, we consider oil
with somewhat contrived properties: density 100 kg/m3 and viscosity 10 cP. The injected
water has density 1,000 kg/m3 and viscosity 1 cP. Both fluids have quadratic relative
permeabilities. The large density difference was chosen to ensure a bottom water drive,
while the high viscosity was chosen to enhance the coning effect. Complete source code
can be found in coningExample.m in the book module.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 314 Solvers for Incompressible Immiscible Flow

Figure 10.15 Sector model used to demonstrate water coning. Blue color indicates low permeability
and red color high permeability.

The idea of using a vertical injector is that the lower completions will set up a bottom
water drive in the good rock to the west of the fault, whereas the upper completions will
provide volumetric sweep of the low-quality rock to the east of the fault. The water front
from the lowest perforations penetrates through to the better zone west of the fault after
approximately 40 days and then gradually builds up a water tongue that moves westward
more rapidly along the bottom of the reservoir as seen in the two plots in the upper row
of Figure 10.16. The advancing water front reaches the far west side of the reservoir after
approximately 670 days and forms a cone that extends upward towards the horizontal pro-
ducer. After 810 days, the water front breaks through in the midsection of the producer, and
after 1,020 days, the whole well is engulfed by water. If the well had been instrumented with
intelligent inflow control devices, the operator could have reduced the flow rate through the
midsection of the well to try to delay water breakthrough. As the water is sucked up to the
producer, it gradually forms a highly conductive pathway from the injector to the producer
as seen in the snapshot from time 1,500 days shown at the bottom-left of Figure 10.16. A
significant fraction of the injected water will therefore cycle through the water zone without
contributing significantly to sweep any unproduced oil. Cycling water like this contributes
to significantly increase the energy consumption and the operational costs of the production
operation and is generally not a good production strategy.
Looking at the well responses in Figure 10.17, we first of all observe that the initial
injection rate is very low because of the high viscosity of the resident oil. Thus, we need
high injection pressure to push the first water into the reservoir. Once this is done, the
injectivity increases steadily as more water contacts and displace a fraction of the oil.
Because the reservoir rock and the two fluids are incompressible, increased injection rates
give an equal increase in oil production rates until water breaks through after approximately
800 days. Since there is no heterogeneity to create pockets of bypassed oil, and residual
saturations are zero, we will eventually be able to displace all oil by continuing to flush the
reservoir with water. However, the oil rate drops rapidly after breakthrough, and increasing
amounts of water need to be cycled through the reservoir to wash out the last parts of the
remaining oil. By 4,500 days, the recovery factor is 73% and the total injected and produced
water amount to approximately 2.6 and 1.8 pore volumes, respectively.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.3 Simulation Examples 315

Figure 10.16 Evolution of the displacement profile for the water-coning case.

Another related problem is that of gravity override, in which a less dense and more
mobile fluid flows preferentially above a denser and less mobile fluid. To illustrate this
multiphase flow phenomenon, we consider a reservoir consisting of two horizontal zones
placed on top of each other. Light fluid of high mobility is injected into the lower zone by
a vertical well placed near the east side. Fluids are produced from a well placed near the
west side, perforated in the lower zone only; see Figure 10.18. Densities of the injected and
resident fluids are assumed to be 700 kg/m3 and 1,000 kg/m3 , respectively. To accentuate
the phenomenon, we assume that both fluids have quadratic relative permeabilities, zero
residual saturations, and viscosities 0.1 cP and 1.0 cP, respectively. In the displacement
scenario, 0.8 pore volumes of the light fluid are injected at constant rate. For simplicity, we
refer to this fluid as water and the resident fluid as oil.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 316 Solvers for Incompressible Immiscible Flow

Oil surface rate [m3/s] Water surface rate [m3/day]

0.3
0.25

0.2
0.2
0.15

0.1
0.1 P
0.05 I

0 0
1,000 2,000 3,000 4,000 1,000 2,000 3,000 4,000
Time [days] Time [Days]

7 Cumulative oil production [m3] 7

3
Cumulative water injection/production [m ]
x 10 x 10
5
Initial oil in place 10
4
8
3
6 Initial oil in place
2
4
1 2

0 0
0 1,000 2,000 3,000 4,000 1,000 2,000 3,000 4,000
Time [Days] Time [Days]

Figure 10.17 Well responses for the simulation of water coning. The top plots show rates in each
perforation of the horizontal producer; the red lines indicate water breakthrough. The middle plots
show total surface rates, whereas the lower plots show cumulative oil production and cumulative
water injection and production.

We consider two different scenarios: one with high permeability in the upper zone and
low permeability in the lower zone, and one with low permeability in the upper zone and
high permeability in the lower zone. To simplify the comparison, the injection rate is the
same in both cases. Figure 10.19 compares the evolving displacement profiles for the two
cases. In both cases, buoyancy quickly causes the lighter injected fluid to migrate into
the upper zone. With high permeability at the top, the injected fluid accumulates under
the sealing top and flows fast towards the west boundary in the upper layers, as seen

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.3 Simulation Examples 317

Figure 10.18 Setup for the case used to illustrate gravity override. Blue and red color indicate
different permeabilities. Perforated cells are colored green.

Figure 10.19 Evolution of the displacement profiles for the gravity-override setup. The reservoir is
viewed from a position below and to the southeast of the reservoir. Cells containing only the resident
fluid are not plotted.

in the upper-left plot of Figure 10.19. Looking at Figure 10.20, you may observe that
since the production well is pressure-controlled and perforated in the low-permeability
zone, the perforation rates will increase toward the top, i.e., the closer the perforation lies to
the high-permeability zone above. As the leading displacement front reaches the producer

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 318 Solvers for Incompressible Immiscible Flow

6,000 6,000

4,000 4,000

2,000 2,000

0 0

1 1

2 2

3 3
2,160 2,160
2,180 2,180
2,200 2,200
4 2,220 4 2,220
2,240 2,240

´104 ´104

2 2

1.5 1.5

1 1
2,160 2,160
0.5 0.5
2,180 2,180
0 2,200 0 2,200
1 2,220 1 2,220
2 2
3 3
4 2,240 4 2,240

´104 ´104

2 2

1.5 1.5

1 1

0.5 0.5

2,160 4 2,160 4
2,180 3 2,180 3
2,200 2 2,200 2
2,220 1 2,220 1
2,240 2,240

Figure 10.20 Perforation rates for the gravity-override case: oil (top), water (middle), and total
rate (bottom). For the oil rate, the red lines indicate peak production, while they indicate water
breakthrough in the plots of water and total rate. The left column reports the case with a high-
permeability upper zone, and the right column the opposite case.

near the west boundary, it is sucked down toward the open perforations and engulfs them
almost instantly, as seen after 360 days in the middle-left plot of Figure 10.19 and the red
line in the lower-left plot of Figure 10.20. This causes a significant drop in oil rate and a
corresponding increase in water rates towards the top of the well, which lies closer to the
flooded high-permeability zone. The oil rate is also reduced in the lowest perforations, but

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.3 Simulation Examples 319
7 3
x 10 Cumulative oil production [m ]

Low permeability on top

2.5
High permeability on top

2
0.35

1.5 0.3

0.25

0.2
1
0.15

0.1
0.5
0.05
200 400 600 800 1000 1200 1400

0
200 400 600 800 1,000 1,200 1,400
Time [Days]

Figure 10.21 Cumulative oil production for the gravity-override case. The inset shows surface oil
rates.

because these grid layers do not produce much of the injected fluid, the drop in oil rate is an
effect of the internal adjustment of total perforation rates along the well, because mobility
is so much higher near the top of the well. Indeed, some time after water breakthrough the
oil rate is higher in the lower than in the upper perforations.
Also with a low-permeability zone on top, buoyancy will cause the injected fluid to
migrate relatively quickly into the upper zone. However, because this case has better direct
connection between the injector and producer through the high-permeability lower zone,
the displacement front will move relatively uniformly through all layers of the top zone and
the upper layer of the lower zone. The leading part of the displacement front is now both
higher and wider and has swept a larger part of the upper zone by the time it breaks through
in the producer. Unlike in the first case, the displacing fluid will not engulf the whole well,
but only break through in the three topmost perforations. However, after breakthrough,
most of the production comes from the topmost perforation, and the total rate in the three
lowest perforations quickly drops below 20% of the rate in topmost perforation.
Finally, looking at cumulative oil production and surface oil rates reported in Figure
10.21, we see that low permeability on top not only delays the water breakthrough, but also
enables us to maintain a higher oil rate for a longer period. This case is more economically
beneficial than with high permeability on the top. To increase the recovery of the latter case,
one possibility would be to look for a substance to inject with the displacing fluid to reduce
its mobility in the upper zone.

10.3.7 The Effect of Capillary Forces – Capillary Fringe

As you may recall from Section 8.1.3, the pressure in a non-wetting fluid is always greater
than the pressure in the wetting phase. In a reservoir simulation model, the capillary pres-
sure – defined as pc = pn − pw in a two-phase system – has the macroscale effect
of determining the local saturation distribution at the interface between the wetting and

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 320 Solvers for Incompressible Immiscible Flow

non-wetting fluid, or in other words, there is a relation pc = Pc (S) between capillary

pressure and saturation. This is seen in two ways: In a system that is initially in hydrostatic
equilibrium, capillary forces enforce a smooth, vertical transition in saturation upward from
a (horizontal) fluid contact. This transition is often referred to as the capillary fringe,
and will be discussed in more detail in this section. The second effect is that capillary
forces will redistribute fluids slightly near a dynamic displacement front so that this is
not a pure discontinuity, as assumed in the hyperbolic models discussed in Chapter 9, but
rather a smooth wave. For many field and sector models, the characteristic width of the
transition zone is small compared to the typical grid size, and hence capillary forces can be
safely neglected. In other cases, capillary forces have a significant dampening effect on the
tendency for viscous fingering and are therefore crucial to include in the simulation model.
Returning to the formation of a capillary fringe near a fluid interface, let zi denote the
depth of the contact between the wetting and non-wetting fluid, and let pn,i and pw,i denote
the phase pressures at this depth. The phase pressures and the capillary pressure are then
given by

pw (z) = pw,i + gρw (z − zi ), pn (z) = pn,i + gρn (z − zi )

(10.8)
pc (z) = pc,i + gρ(z − zi ),
where ρ = ρn − ρw is density difference and pc,i is capillary pressure at zi . This pressure
is the capillary pressure necessary to initiate displacement of the wetting fluid by the non-
wetting fluid and is called the entry pressure. It follows from (10.8), that the total height of
the capillary fringe is given by pc,n /gρ. Figure 10.22 illustrates the concept of a capillary
fringe for a case with zero residual saturations.
If we know the phase contact zi , we can find the saturation directly by first computing
the capillary force as function of depth using (10.8) and then using the capillary pressure

Figure 10.22 Diagrams showing phase and capillary pressures (left) and saturation (right) as function
of depth. Here, zi is the depth of the contact between the non-wetting and the wetting fluid, and zn is
the depth of pure non-wetting fluid.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.3 Simulation Examples 321

function Pc (S) to invert for saturation. Alternatively, if we only know the volume of the
fluids, we can use the incompressible solvers to determine the hydrostatic fluid distribution.
To illustrate, we consider a 100 × 100 m2 vertical cross-section represented on a 20 × 40
grid. We assume a fluid system with CO2 and brine having the same basic properties as
in Section 10.3.2. With a density difference of approximately 290 kg/m3 , the capillary
fringe corresponding to a capillary pressure of 1 bar has a height of 35 m. The relationship
between saturation and capillary pressure depends on permeability and porosity, and to
model this, we use the Leverett J -function (8.9). The initSimpleFluidJfunc imple-
ments a simplified Corey-type fluid in which J (S) = 1 − S. This gives
-
φ
Pc (S) = σ (1 − S).
K
The surface tension σ is usually specific to the fluid. To illustrate capillary raise, we choose
σ such that median rock properties give a capillary pressure of one bar,
fluid = initSimpleFluidJfunc('mu' , [0.30860, 0.056641]*centi*poise, ...
'rho', [ 975.86, 686.54]*kilogram/meter^3, ...
'n' , [ 2, 2], ...
'surf_tension',1*barsa/sqrt(mean(rock.poro)/(mean(rock.perm))),...
'rock',rock);

We set initial data such that the rock is filled with half a pore volume of wetting fluid at
the bottom and half a pore volume of non-wetting fluid at the top, and then simulate the
system forward in time until steady-state is reached. The time-loop is set up with a gradual
ramp-up of the time step to increase the stability of the solution procedure (we will come
back to the choice of time step in Section 10.4.1),
dt = dT*[1 1 2 2 3 3 4 4 repmat(5,[1,m])]*year;
dt = [dt(1).*2.^[-5 -5:1:-1], dt(2:end)];
s = xr.s(:,1);
for k = 1 : numel(dt),
xr = incompTPFA(xr, G, hT, fluid);
xr = implicitTransport(xr, G, dt(k), rock, fluid);
t = t+dt(k);
if norm(xr.s(:,1)-s,inf)<1e-4, break, end;
end

We consider two different permeability setups. In the first case, the permeability field
varies linearly from 50 md in the west to 400 md in the east. The porosity is assumed to be
constant. When the system is released from the artificial initial state with a sharp interface
at z = 50, the non-wetting fluid starts draining downward into the wetting phase, whereas
the wetting phase starts imbibing upward into the non-wetting phase. After approximately
3.5 years, the system reaches the steady state shown in the middle plot in the upper row
of Figure 10.23. Here, we say that steady-state is reached when the saturation difference
between two consecutive time steps is less than 10−4 measured in the L∞ norm. Because

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 322 Solvers for Incompressible Immiscible Flow
2

1.8

1.6

1.4

1.2

0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1

2.5

1.5

0.5

0
0 0.2 0.4 0.6 0.8 1

Figure 10.23 Capillary fringe for two different permeability fields: permeability increasing linearly
from west to east (top), and lognormal permeability (bottom).

permeability is homogeneous in the vertical direction, the steady-state saturation decreases

linearly upward from pure wetting to pure non-wetting fluid. The height of the capillary
fringe
√ is much higher in west where the permeability is low, since the capillary scales as
1/ K, and lower in the east where the permeability is high. In the plot of capillary pressure
versus saturation to the upper-right in Figure 10.23, you may also be able to identify the
twenty different lines corresponding to the twenty columns of homogeneous permeability
in the grid.
The second case has random petrophysical parameters with a permeability field that is
related to a Gaussian porosity field through a Carman–Kozeny relationship; see the lower-
left plot in Figure 10.23. The permeability values span three orders of magnitude, from 1
md to 1 darcy, which in turn gives a wider span in time constants than in the case with
linear permeability. Because of the heterogeneities, the imbibing wetting phase will be
sucked higher up and sideways into regions of lower permeability. If you run the script
capillaryColumn yourself, you will see that the high-permeability regions surrounding
the initial sharp interface reach equilibrium within a few years, whereas the rise is much
slower in the low-permeability regions. The column next to the east boundary, in particular,
is the last to reach steady state after approximately 41 years. At steady state, the fringe
extends above the top of the reservoir section in the east-most column. Moreover, because
permeability is heterogeneous, the saturation at steady state is no longer monotone in the
vertical (and horizontal) direction.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.3 Simulation Examples 323

Figure 10.24 The Norne test case with grid and petrophysical data from the simulation model of the
real field. The well pattern is artificial and has nothing to do with how the real field is operated.

10.3.8 Norne: Simplified Simulation of a Real-Field Model

Having worked with highly idealized models so far in this chapter, it is now time to look at
a more realistic model. For this, we will use the grid geometry and petrophysical properties
from a simulation model of the Norne field from the Norwegian Sea. More details about
Norne and the reservoir geometry were given in Section 3.3.1. Figure 10.24 shows the
petrophysical properties as well as a well-pattern that was chosen somewhat haphazardly
for illustration purposes. We notice that the permeability is anisotropic and heterogeneous,
with a clear layered structure. This layered structure is also reflected in the histograms,
which show several modes. (Such histograms are discussed in more detail in Sections
2.5.3 and 2.5.5 for the SPE 10 and SAIGUP models.) The lateral permeability has four
orders of magnitude variations, whereas the vertical permeability is up to two orders lower
and has five orders of magnitude variations. In addition, the vertical communication is
further reduced by a multiplier field (MULTZ keyword), which contains large regions having
values close to zero in the middle layers of the reservoir. The porosities span the interval
[0.094, 0.347], but since the model has a net-to-gross field to model that a portion of
the cells may consist of impermeable shale, the effective porosity is much smaller in
some of the cells. For a model like this, we thus cannot expect to be able to use the
explicit transport solver and must instead rely on the implicit solver. (Note also that the
incomp module contains a similar example with synthetic petrophysical properties; see
incompExampleNorne2ph.)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 324 Solvers for Incompressible Immiscible Flow

4
x 10
6 18,000
Oil rate [stb/day]
16,000 Water rate [stb/day]
5
14,000

4 12,000
10,000
3 8,000
6,000
2
4,000
1 2,000
0
500 1,000 1,500 2,000 2,500 3,000 3,500 4,000 500 1,000 1,500 2,000 2,500 3,000 3,500 4,000
Time [Days] Time [Days]
P1 P2 P3 P4 P5 P1 P2 P3 P4 P5

Figure 10.25 Incompressible two-phase simulation for the Norne model. The upper plots show
snapshots of the solution (after 6 and 12 years, respectively) and the lower plots show oil and water
surface rates for all producers.

Setting up the model proceeds as discussed previously; you can consult the runNorne
Simple script in the book module for full details. The only difference is how to account
for the transmissibility multipliers. This is done as follows:
hT = computeTrans(G, rock, 'Verbose', true);
tmult = computeTranMult(G, grdecl);
hT = hT.*tmult;

Here, the Eclipse input structure grdecl contains data for the MULTZ keyword. The second
call takes the MULTZ values associated with cells and assigns a corresponding reduction
value between 0 and 1 to all half-faces. We then multiply the half-transmissibilities hT by
tmult to get the reduced transmissibility. It is important that these multipliers are assigned
before computing the intercell transmissibilities. Altogether, approximately 6.5% of the
half-faces have reduced transmissibility.
Figure 10.25 shows results of a simulation of a scenario in which water is injected
into oil. The oil has a five times higher viscosity and hence the injected water will form
an unstable displacement with a weak displacement front similar to the case discussed in
Section 10.3.4. The main displacement takes place in the region that involves injectors I1

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.3 Simulation Examples 325

to I5 and producers P1–P3. The last two producers are located in regions that are poorly
connected to the rest of the reservoir where the injectors are placed, and hence contribute
less to the overall production. This is particularly true for producer P4, which most likely is
completely misplaced. As we have seen in previous examples, once water breaks through
in a well (primarily in P1 and P2), the oil rate decays significantly.
The main purpose of the example lies in the actual code and not in the results it produces.
Discussing the simulation results beyond this point is therefore somewhat futile since
the fluid system and the well pattern have limited relevance for the real reservoir, which
contains a three-phase oil–gas–water system that is modeled by the compressible black-
oil equations that will be discussed in the next chapter. The main takeaway message is
that the incompressible solvers can be applied to models that have the geometrical and
petrophysical complexity seen in real reservoir models.

Computer exercises

10.3.1 To get more acquainted with the multiphase incompressible solvers and see their
versatility, you should return to a few examples presented earlier in the book and
try to set them up as multiphase test cases:

• Consider the reservoir in Exercise 3.1.3 and place one injector to the south
and two producers symmetrically along the northern perimeter. Simulate the
injection of one pore volume of water into an oil.
• Consider the test case with non-rectangular reservoir geometry in Figure 5.6
on page 164 and set up a simulation that injects one pore volume from the
flux boundary. How would you compute the flux out of the pressure-controlled
boundary?
• Pick any of the faulted grids generated by the simpleGrdecl routine as shown
in Figure 3.31 on page 97 and place an injector in one fault block and a producer
in the other and simulate the injection of half a pore volume of water.

10.3.2 Consider a rectangular reservoir with two wells (see Figure 3.37) and compare
solutions computed with three different grids: a uniform coarse grid, a uniform
fine grid, and a coarse grid with radial well refinement.
10.3.3 Try to study the Norne model in more detail.

• Are the multipliers important for the simulation result?

• Is gravity important or can it be neglected?
• Can you come up with a better recovery strategy, i.e., improved placement and
control strategy for wells?
• Do you get very different solutions if you use a consistent solver? (Hint:
Although multipliers can be incorporated into these solvers, as described in
[231], this is not part of the public implementation and for this comparison you
should therefore neglect the MULTZ keyword.)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 326 Solvers for Incompressible Immiscible Flow

10.4 Numerical Errors

There are several errors involved in the computations in the previous section. First of all,
numerical discretization errors obviously arise when approximating a continuous differen-
tial equation by a set of discrete finite-volume equations. For single-phase, incompressible
flow, errors were purely spatial. These errors will decrease with decreasing size of the
grid as long as the spatial discretization is consistent. However, as we saw in Chapter 6,
the standard two-point scheme is not consistent unless the grid is strictly K-orthogonal,
and the incompTPFA pressure solver can in general be expected to produce errors for
anisotropic permeabilities and skewed grids. When using a sequential method to solve
multiphase flow equations, there will also be temporal errors arising from three different
factors: discretization errors arising when temporal derivatives in the transport equations
are discretized by finite differences, amplifications of spatial errors with time, and errors
introduced by the operator splitting underlying the sequential solution procedure. In this
section, we briefly discuss the two last error types in more detail.

10.4.1 Splitting Errors

When using a sequential solution procedure, the total velocity is computed from the fluid
distribution at the start of each time step. This means that the effect of mobility on the
flow paths is frozen in time, and for each time step appears as if we solved a single-phase
flow problem with reduced permeability in all parts of the domain that contain more than
one fluid phase. Within a single splitting step the transport solver will thus only resolve
the dynamic effect of mobility along each flow path, but will not account for the fact that
mobility changes reduce the effective permeability along each flow path or move the flow
paths themselves. This introduces a time lag in the simulation, which may lead to signif-
icant errors in the propagation of displacement fronts if the splitting steps are chosen too
large.

Homogeneous Quarter Five-Spot

To illustrate this, we can revisit the homogeneous quarter five-spot from Section 10.3.3
and study the self-convergence of approximate solutions defined on a fixed grid as the
number of splitting steps increases. Figure 10.26 shows approximate solutions at time t =
0.6 PVI (pore volumes injected) computed with 4 steps for  = 0, . . . ,3. With a single
pressure step, the displacement front coincides with the time line from the single-phase
flow field, since the pressure computation only sees the initial oil saturation. The only
exception is a certain smearing introduced by the spatial and temporal discretizations of the
explicit scheme used to compute the transport step. The retardation effect that oil has on the
invading water is better accounted for as the number of splitting steps increases, and hence
the splitting solution gradually approaches the correct solution. Table 10.1 reports the self-
convergence towards a reference solution computed on the same grid with 256 splitting
steps. To estimate the convergence rate we assume that the error scales like O(t r ). If the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.4 Numerical Errors 327
Table 10.1 Self-convergence for the homogeneous quarter
five-spot computed on a 128 × 128 grid with n splitting steps
relative to a reference solution computed with 256 time steps after
0.6 PVI. The errors are reported in relative norms.

saturation pressure
n L1 -error rate L2 -error rate

1 5.574e−02 – 4.950e−03 –
2 4.368e−02 0.35 5.140e−04 3.27
4 2.778e−02 0.65 1.554e−04 1.73
8 1.445e−02 0.94 4.524e−05 1.78
16 6.389e−03 1.18 1.226e−05 1.88
32 2.394e−03 1.42 2.998e−06 2.03
64 7.869e−04 1.61 5.990e−07 2.32

1 steps 4 steps 16 steps 64 steps

Figure 10.26 Quarter five-spot solution at time 0.6 PVI computed on a uniform 128 × 128 grid with
the explicit transport solver and different number of splitting steps. The solid lines are time lines at
t = 0.6a for a single-phase displacement.

solution having error E2 is computed using twice as many time steps as the solution having
error E1 , the corresponding convergence rate is

r = log(E1 /E2 )/ log(2).

Pressure is smooth and will therefore converge faster than saturation, which is a discontin-
uous quantity and hence will have much larger errors. The convergence for high n values is
exaggerated since we are measuring self-convergence toward a solution computed with the
same method, but with a larger number of steps. The code necessary to run this experiment
is found in splittingErrorQ5hom.m in the in2p directory of the book module.

Heterogeneous Quarter Five-Spot

As pointed out earlier, the coupling between the pressure and transport equations depends
on the variation in total mobility λ(S) throughout the simulation. If variations are small and
smooth, the two equations will remain loosely coupled, and relatively large time steps can

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 328 Solvers for Incompressible Immiscible Flow
1

0.9
1:10
0.8 1:1
10:1
0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.2 0.4 0.6 0.8 1

Figure 10.27 Total mobility for three fluid models with different viscosity ratios.

2 steps 4 steps 8 steps 16 steps 32 steps 64 steps 256 steps

Figure 10.28 Self convergence with an increasing number of equally spaced splitting steps to reach
time 0.5 PVI for a quarter five-spot setup on a subsample from the first layer of the SPE 10 model
for three different viscosity ratios M = μw /μn (top: M = 1/10, middle: M = 1, bottom: M = 10).
Saturation profiles are shown at a time scaled by ν(1)/ν(M), where ν(M) is the characteristic wave
speed of the displacement front with viscosity ratio M.

be allowed without seriously decaying solution accuracy. On the other hand, when λ has
large variations over the interval [0,1], pressure and transport are more tightly coupled, and
we cannot expect to be able to use large splitting steps. To illustrate this, we revisit the setup
with three different fluid models used to study viscous fingering in Section 10.3.4. From
the plot of total mobilities in Figure 10.27 it is obvious that both the unfavorable (1:10) and
the favorable (10:1) mobility cases have stronger coupling between saturation and pressure
than the case with equal viscosities.
Figure 10.28 shows the self convergence of the saturation profiles with respect to the
number of equally spaced time steps used to reach time 0.5 PVI. To isolate the effect of
splitting errors and avoid introducing excessive numerical smearing in the solutions with
few splitting steps, we have subdivided the transport steps into multiple steps so that the
implicit solver uses the same step length and hence introduces the same magnitude of

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.4 Numerical Errors 329
0.9 0.9
Water cut for M=0.1 Water cut for M=1.0
0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5
P (3) P (3)
0.4 0.4
P (12) P (12)
0.3 P (48) 0.3 P (48)
0.2 P (192) 0.2 P (192)
P (768) P (768)
0.1 0.1
P (rampup) P (rampup)
0 0
0.2 0.4 0.6 0.8 1 1.2 1.4 0.2 0.4 0.6 0.8 1 1.2 1.4

0.9
Water cut for M=10.0
0.8

0.7

0.6

0.5
P (3)
0.4 P (12)
0.3 P (48)
P (192)
0.2
P (768)
0.1
P (rampup)
0
0.2 0.4 0.6 0.8 1 1.2 1.4

Figure 10.29 Water cut in the producer for various number of splitting steps for the heterogeneous
quarter five-spot setup from Figure 10.28.

numerical smearing in all simulations. From the figure it is clear that even with very few
splitting steps, the sequential solution method manages to capture the qualitatively correct
behavior for all viscosity ratios. As expected, the discrepancies in solutions with few and
many time steps are larger for the favorable and unfavorable cases than for unit viscosity
ratio.
To investigate how the size of the splitting steps affects the quantitative behavior of
the approximate solutions, we reran the same experiments up to 1.5 PVI. Figure 10.29
reports water cut in the producer for all three fluid models. Starting with the unfavorable
case, we see that the water cut has a dent. This is a result of the secondary finger that
initially extends along the western edge, making contact with the main finger and hence
contributing to a more rapid incline in water production. All curves, except the one using
only three time steps, follow the same basic trend. The main reason is that small variations
in the saturation profile will not have a large effect on the water cut since the displacement
front is so weak. With three splitting steps only, the main dent is a result of the second
pressure update. For unit viscosity ratio, the production curves are still close, but here we
notice a significant incline in the curve computed with 12 steps after the pressure update
at 0.75 PVI. The 12-step and 48-step curves also show non-monotone behavior at 0.625
PVI and 0.69 PVI, respectively. Similar behavior can be seen for the favorable case, but
since this case has an almost piston-like displacement front, the lack of monotonicity
is significantly amplified. For comparison, we have also included a simulation with 96
splitting steps, in which the first two first steps have been replaced by ten smaller splitting

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 330 Solvers for Incompressible Immiscible Flow

steps that gradually ramp up to the constant time step. These profiles, and similar profiles
run with 48 steps, are monotone, which suggests that the inaccuracies in the evolving sat-
uration profiles are introduced early in the simulation when the profile is rapidly expanded
by high fluid velocities in the near-well region. In our experience, using such a ramp-
up is generally advisable to get more well-behaved saturation profiles. You can find the
code for this experiment in splittingErrorQ5het.m in the in2p directory of the book
module.

Capillary-dominated flow
The sequential solution procedure discussed in this chapter is as a general rule reasonably
well-behaved for two-phase scenarios where the fluid displacement is dominated by the
hyperbolic parts of the transport equation, i.e., by viscous forces (pressure gradients) and/or
gravity segregation. If the parabolic part of the solution dominates, on the other hand,
a sequential solution procedure will struggle more, in particular when computing fluid
equilibrium govern by a delicate balance between gravity and capillary forces. To illustrate
this, we revisit the computation of capillary fringe from Section 10.3.7. If you look carefully
in the accompanying code, you will see that we compute the two cases with a time step that
is ten times larger for the Gaussian case than for the case with linear permeability. The time
steps (and the initial ramp-up sequence) were chosen by trial and error and are close to what
appears to be the stability limit. If one, for instance, increases the final time steps by 150%
for the case with linear permeability, the simulation will not converge but instead ends up
predicting an oscillatory interface, as illustrated in Figure 10.30. Similar problems may
arise, e.g., when simulating structural trapping of CO2 using vertical equilibrium models,
in which gravity gives rise to a parabolic term that plays the same role as capillary pressure
in the upscaled flow equations; see e.g., [227] for more details.

10.4.2 Grid Orientation Errors

As you may recall from the discussion on page 287, the saturation-dependent mobility at
the interface between two grid blocks is usually approximated by single-point, upstream
mobility weighting. Like the TPFA method, the resulting scheme only accounts for infor-
mation on opposite sides of a cell interface and does not take any transverse transport
effects into account. It is therefore well known that this method also suffers from grid
orientation errors, especially when applied to unfavorable displacements, as we will see
in the following example. In passing, we also note that to evaluate gravity and capillary-
pressure terms, the transport solvers use two-point approximations similar to what is used
in the TPFA method. This may introduce additional errors, but we will not discuss these in
detail herein.

Homogeneous Quarter Five-Spot

When the single-point transport solver is combined with the classical TPFA pressure solver,
computed displacement fronts tend to preferentially move along the axial directions of the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.4 Numerical Errors 331

0
10

−1
10

−2
10

−3
10

−4
10
Δ t = 36.50 days
Δ t = 54.75 days
−5
10
0 1 2 3 4 5 6 7 8 9

Figure 10.30 Unstable solution computed by a sequential solution with too large splitting step.
The lower plot shows difference in L∞ norm between consecutive time steps – which is used as
convergence criterion by the solver – for a convergent solution with t = 36.5 days and for a
divergent solution with t = 54.75 days. The upper plot shows three consecutive solutions for the
divergent simulation.

grid, i.e., in the direction of the normal vectors of the cell faces. This will lead to grid
orientation effects like those discussed earlier in Chapter 6, even if the resulting grids are
K-orthogonal. To illustrate this, we compare and contrast solutions of the standard quarter
five-spot setup with a rotated setup in which the grid is aligned with the directions between
injectors and producers, as illustrated in Figure 10.31. This test problem was first suggested
by Todd et al. [290] and has later been used by many other authors to study grid orientation
errors in miscible displacements [321, 255, 279], which are particularly susceptible to this
type of truncation error. To avoid introducing too much diffusion when using few time
steps, we use the explicit transport solver. (See runQ5DiagParal for complete setup of the
two cases and gridOrientationQ5 for the following experiments.)
For the standard setup, the combination of a single-point transport solver and a two-point
pressure solver overestimates the movement into the stagnant regions along the x and y
axes and underestimates the diagonal movement in the high-flow direction between injector
and producer along the diagonal. In the rotated setup, the grid axes follow the directions
between the wells and the solvers will hence tend to overestimate flow in the high-flow
zone and underestimate flow toward the stagnant zones. The upper row in Figure 10.32
shows that this effect is pronounced for the unfavorable displacement (M = 0.1), evident
with equal viscosities (M = 1), and hardly discernible for the favorable displacement
(M = 10). For equal viscosities, the difference between the two grids can be almost

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 332 Solvers for Incompressible Immiscible Flow

Producer

Injector

Quarter five-spot

Rotated five-spot

Figure 10.31 Well setup for the quarter five-spot comparison. Displacement fronts have preferential
movement parallel to the axial directions and hence the rotated setup will predict earlier breakthrough
than the original setup.

Mobility ratio 1:10. Time: 0.3 PVI Mobility ratio 1:1. Time: 0.6 PVI Mobility ratio 10:1. Time: 0.7 PVI

Grid: 32 x 32. Steps: 1 Grid: 32 x 32. Steps: 8 Grid: 32 x 32. Steps: 64

Grid: 16 x 16. Steps: 16 Grid: 32 x 32. Steps: 16 Grid: 64 x 64. Steps: 16

Figure 10.32 Quarter five-spot solutions computed on the rotated (colors) and original (solid lines)
geometry for a 32 × 32 grid. The top row shows solutions computed with 16 steps for different
mobility ratios. The middle plot shows convergence for mobility ratio 1:10 with respect to time step
and the lower row with respect to grid resolution.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.4 Numerical Errors 333

eliminated by increasing the number of splitting steps for a fixed x and/or by increasing
the grid resolution provided that a certain number of time steps are used. In the unfavorable
case, the plots in the middle row of Figure 10.32 show that increasing the number of
time steps changes both solutions in the same direction, but does not necessarily reduce
their difference. However, since both the original and the rotated grid are regular, the grid
orientation effects will diminish if we increase the grid resolution.
A possible remedy to the behavior we just observed is to replace the single-point
scheme by a multidimensional upwind scheme (see e.g., [155]) or a modern high-resolution
scheme. Such methods are not yet part of the public MRST release, but work is currently
in progress [190].

Symmetric Well Pattern on a Skew Grid

We have already seen several times that since the TPFA scheme cannot approximate
transverse fluxes that are parallel to grid interfaces, the incompTPFA solver will introduce
grid orientation errors for anisotropic permeabilities and grids that are not K-orthogonal.
In Chapter 6, we showed that we can significantly reduce, but not completely eliminate,
these errors if we replace TPFA by a consistent scheme. In this example, we will revisit
one of the cases discussed in Chapter 6 to investigate if we see the same reduction for
two-phase flow.
The computational setup consists of a horizontal 400 × 200 m2 sector model. Water
is injected from at the midpoint of the northern perimeter and fluids are produced from
two wells located 50 m from the southeast and southwest corners, respectively. Since the
well pattern is symmetric within a confined domain and the petrophysical parameters are
homogeneous and isotropic, the true displacement profile will also be symmetric. The
grid, however, is skewed and compressed towards the southeast corner. This will induce
a preferential flow direction towards the southeast producer. Seeing that grid orientation
effects are more pronounced for unstable displacements, we use the same configuration
as in the previous example with a viscosity ratio M = 10, which gives a very mobile,
weak displacement front that will tend to finger rapidly into the resident oil. Figure 10.33
shows a snapshot of the pressure and saturation profile after 250 days along with water
cuts and cumulative oil production over the whole 1,200-day simulation period. The flow
field computed with the incompTPFA pressure solver exhibits the same lack of symme-
try as seen in Figure 6.5 on page 187. The result is a premature breakthrough in the
southeast producer and delayed breakthrough in the southwest producer. With a mimetic
finite-difference (MFD) solver, the water-cut curves are much closer and less affected by
grid orientation errors. Moreover, since all wells operate under pressure control, we see
that the total oil production predicted by TPFA is significantly less than for the MFD
solver.
As a second example, we use the same grid to describe a vertical cross-section, in which
we inject water from two horizontal injectors at the bottom of the reservoir and produce
fluids from a horizontal producer at the top of the reservoir. Producers and injectors operate
under the same pressure control as for the horizontal reservoir section. Figure 10.34 shows

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 334 Solvers for Incompressible Immiscible Flow
TPFA: 250 days Mimetic: 250 days
100 100 200

50 50 150

0 0 100
0 100 200 300 400 0 100 200 300 400

100 100 1

50 50 0.5

0 0 0
0 100 200 300 400 0 100 200 300 400

0.9
Water cut
0.8 Cumulative oil production [m3]
2000
0.7

0.6
1500
0.5

0.4
1000
0.3 P1 (TPFA) P1 (TPFA)
0.2 P2 (TPFA) P2 (TPFA)
500
P1 (MFD) P1 (MFD)
0.1
P2 (MFD) P2 (MFD)
0 0
200 400 600 800 1,000 1,200 200 400 600 800 1,000 1,200
Time [Days] Time [Days]

Figure 10.33 Simulation of a symmetric flow problem in a horizontal, homogeneous domain

represented on a skew grid.

simulation results with the TPFA and MFD solvers. The density difference between the
injected and resident fluids is 150 kg/m3 , and hence gravity tends to oppose the imbibing
water front in a way that accentuates the grid orientation effects for both solvers. For com-
parison, Figure 10.35 reports the simulations performed on a regular Cartesian grid. Both
schemes produce symmetric, but slightly different displacement profiles and the match in
production profiles is largely improved compared with the skew grid.
Altogether, the two examples presented in this section hopefully show you that you not
only need to take care when designing your grid, but should also be skeptic to simulations
performed by a single method or a single choice of time steps. A good piece of advice is
to conduct simulations with more than one scheme, different time-step selection, and, to
the extent possible, different grid types to get an idea of how numerical errors influence
your results. To further investigate grid orientation effects, consider any of the following
computer exercises.

Computer exercises
10.4.1 Repeat the test case with the original/rotated quarter five-spot using one of the
consistent solvers from Chapter 6 to compute the pressure. Do you see any differ-
ences?

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 10.4 Numerical Errors 335

TPFA: 500 days Mimetic: 500 days

100 100 200

50 50 150

0 0 100
0 100 200 300 400 0 100 200 300 400

100 100 1

50 50 0.5

0 0 0
0 100 200 300 400 0 100 200 300 400

0.8
Water cut 3,500 3
Cumulative oil production [m ]
0.7
3,000
0.6
2,500
0.5

0.4 2,000

0.3 1,500

0.2 1,000
P (TPFA) P (TPFA)
0.1 500
P (MFD) P (MFD)
0 0
200 400 600 800 1,000 1,200 200 400 600 800 1,000 1,200
Time [Days] Time [Days]

Figure 10.34 Simulation of a symmetric flow problem in a vertical, homogeneous domain repre-
sented on a skew grid with two horizontal injectors at the bottom and a horizontal producer at the
top.

TPFA: 500 days Mimetic: 500 days

100 100 1

50 50 0.5

0 0 0
0 100 200 300 400 0 100 200 300 400

0.8
Water cut 3500
0.7 Cumulative oil production [m3]
3000
0.6
2500
0.5
2000
0.4
1500
0.3

0.2 1000
P (TPFA) P (TPFA)
0.1 500
P (MFD) P (MFD)
0 0
200 400 600 800 1,000 1,200 200 400 600 800 1,000 1,200
Time [Days] Time [Days]

Figure 10.35 Simulation of a symmetric flow problem in a vertical, homogeneous domain computed
on a regular Cartesian grid.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 336 Solvers for Incompressible Immiscible Flow

10.4.2 Set up a flow problem to compare solutions computed on the extruded Delaunay
or Voronoi grids shown in Figure 3.32 on page 97.
10.4.3 Set up a quarter five-spot test using the grids shown in Figure 3.44 on page 108.
10.4.4 The CaseB4 test case shown in Figure 3.20 on page 80 represents the same geology
with two different grid formats, using either a deviated pillar grid or a stair-stepped
grid. Both grids are sampled at two different resolutions. Set up a flow problem
with four wells, two injectors and two producers, one in each corner of the reser-
voir. Run simulations on all four grids and compare production curves. You can
either use homogeneous permeability or a layered permeability with homogeneous
properties within each layer.
10.4.5 Pick any of the models in the bedModels1 or bedModel2 data sets and run two-
phase simulations injecting one pore volume from south to north and from west to
east. Investigate splitting and grid orientation effects.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11
Compressible Multiphase Flow

The black-oil equations constitute the industry-standard approach to describe compressible

three-phase flow. Black-oil models generally have stronger coupling between fluid pressure
and the transport of phases/components than the two-phase, incompressible flow models
discussed in the previous chapter. For this reason it is common to use a fully coupled
solution strategy, in which the whole system of equations is discretized implicitly and
all primary unknowns are solved for simultaneously. This chapter introduces you to the
underlying physics and describes the various rock-fluid and PVT properties that enter these
models. We also explain the basics of how these models are discretized and implemented
in MRST. Our implementation will rely heavily on the discrete operators discussed earlier
in the book. As a gentle introduction to the numerics of the black-oil equations, we pick up
again the simple AD solver from Chapter 7 and extend it to compressible two-phase flow
without mass transfer between the phases. The code is applicable to general unstructured
grids in 2D and 3D, but it is still a far stretch from a full-fledged reservoir simulator capable
of simulating real assets, as discussed in Chapter 12.

11.1 Industry-Standard Simulation

A commercial reservoir simulator is a complex software that involves a wide variety of
flow models, computational algorithms, and a lot of subtle tricks to ensure robust behavior.
Covering all of these features, as well as the many different scenarios for which a simulator
can be used, is well beyond the scope of this chapter. Instead, I will try to provide you with
a minimum of insight into the key principles and show how to set up your first black-oil
simulation in MRST using the AD-OO framework (this framework is described in more
detail in Chapter 12). Let us start by a brief overview before we discuss the different
constituents in more detail.

Flow Models and Phase Behavior

Reservoir fluids are complex mixtures of many chemical components that distribute dif-
ferently among the fluid phases present in the reservoir depending upon temperature and
pressure. The composition and characteristics of hydrocarbon fluids vary largely from one
reservoir to another, but may also have large variations across reservoir compartments.
As we have seen already, the black-oil model groups the chemical components into two
337

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 338 Compressible Multiphase Flow

pseudo-components reflecting the phase they belong to at surface conditions. At reservoir

conditions, each component can be present in both a liquid oleic and a gaseous phase. We
motivate and briefly explain the phase diagrams used to describe the phase distribution of
the pseudo-components, but an in-depth discussion of the underlying physics and the many
different forms these diagrams can take is outside the scope of the book.
The black-oil equations can also model processes involving other chemical species. In
enhanced oil recovery, various forms of chemical and biological substances are added to the
injected fluid to increase sweep efficiency and improve local displacement. You can model
such systems by adding extra fluid components to the three phases. MRST offers an EOR
module with basic models for polymer and surfactant flow; see [30] for more details, as
well as a module for simulating solvents. We have already discussed CO2 storage multiple
times. Here, the buoyant and supercritical phase formed when CO2 is injected sufficiently
deep into the subsurface can dissolve into the resident brine. The behavior of the CO2 –brine
system can thus be modeled as a hydrocarbon system in which brine plays the role as oil
and CO2 the role as gas. Extensive support has been developed in MRST to describe the
long-term migration of buoyant CO2 in large-scale aquifer systems, in which fully implicit
solution methods developed for the black-oil equations play a key role. You can find more
details in the thesis by Andersen [19].
Grouping the hydrocarbons into two pseudo-components does not describe all recovery
processes with sufficient accuracy. Compositional simulation models the reservoir fluids as
a system with M components and describes their phase behavior by an equation of state.
The basic conservation equations for such systems was given in Section 8.2.2 and typical
examples of equations of state were briefly reviewed as part of our discussion of single-
phase flow in Section 4.2. As of release 2017b, MRST offers a module for compositional
simulation. We will not discuss this at all herein; you can find preliminary details in the
thesis by Møyner [212]. Likewise, a new module for geochemistry was also added in
the same release, which, in particular, offers non-isothermal aqueous speciation, surface
chemistry, redox chemistry, as well as solid and gas phase equilibrium.

Wells and Production Control

In previous chapters, wells have always been completed in a single cell or along straight
lines following the axial directions, and have either been controlled by constant pressure or
constant fluid rate. Real wells are significantly more complex and generally follow highly
deviated trajectories and possibly consist of multiple branches. Many subtleties go into
well modeling. In particular, special care is needed in the solution algorithm to handle
cross-flow, in which fluids may flow into one part of the wellbore and out of another.
To ensure that each well operates under realistic engineering constraints, reservoir sim-
ulators offer a comprehensive set of control options that can be applied to fluid rates and
bottom-hole or tubing-head pressures. A typical control strategy consists of setting a target
rate and constraints on the pressure the well can sustain before the rate must be reduced.
Well control may thus switch depending upon the computed pressures and flow rates once
the well fails to meet one of its constraints. In so-called history mode, observed fluid rates

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.1 Industry-Standard Simulation 339

at reservoir conditions are set as controls to reproduce production history as closely as

possible. Controls also include economic constraints, which force wells to be completely
shut in or recompleted if their production falls below a lower economic limit, and gas lift
or other mechanisms introduced to help lift the well fluids to the surface when the oil or
liquid rates fall below a specified limit. This will strongly affect the bottom-hole pressure
in individual wells.
Control strategies can also be imposed on groups of wells, or on the field as a whole.
Typical strategies here either involve guide controls, in which total rates are distributed
to individual rates according to their production potential, or prioritization schemes that
turn wells on in decreasing order of priority. Group and field controls may also involve
various restrictions in terms of total pump and lift capacities, reinjection of produced
formation gas or water, limitations in surface facilities, etc. Altogether, the resulting control
strategies may involve a lot of quite intricate logic. MRST contains data structures for
representing surface facilities, but only a limited number of models and control options
have so far been implemented in the public version. I therefore consider a more detailed
discussion of control strategies to be outside the scope of this book. Examples of third-party
implementations with MRST are discussed in two master theses [141, 115].
Many wells also have advanced completions and devices for inflow control. In multiseg-
ment well models, the well trajectory is discretized into a series of segments that can be
linked together into branches. Each branch can, in turn, contain sub-branches. A segment
describes how the flow rate between two points (nodes) relates to the pressure drop between
the same two points. Nodes can be located at different positions along the wellbore or on
opposite sides of devices that restrict the flow, such as chokes and valves. Each node has
an associated volume, and mass-conservation equations are solved between the nodes to
describe pressure drops and wellbore storage effects. As of release 2017a, MRST contains
a general multisegment model that allows branches of segments to be defined in the tubing
and in the void space (annulus) between the inner and outer tubing. The branches can
be joined at multiple points in a flexible manner to form a general network. Altogether,
this enables accurate description of inflow devices and the near-well region, as well as
complex cross-flow patterns that may arise in multilateral wells. An in-depth discussion of
multisegment well modeling is outside the scope of this book, but as is always the case with
MRST, you are free to consult the code to understand the necessary details.

Data Input
To model real assets, a simulator needs comprehensive and flexible input functionality
to specify the setup of all ingredients that constitute the simulation model. The industry-
standard approach is to define an input language consisting of a number of keywords the
user can utilize to construct a model of a computation (i.e., a state machine). The detailed
syntax will vary from one simulator to another, but the general setup is conceptually the
same. You have already encountered the ECLIPSE input format [270, 33] multiple times
throughout the book. Over the years, this software has amassed options for simulating a
wide variety of physical processes, and the input language currently has several thousand

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 340 Compressible Multiphase Flow

keywords, many of which have multiple available options. There are also similar file for-
mats that can be used to store simulation results so that these can be reloaded and used to
continue a previous simulation.
So far, we have only discussed how ECLIPSE input describes grids and petrophysical
properties, and even for this purpose there are a large number of keywords we have not
discussed. In general, an ECLIPSE input deck consists of eight different sections, some of
which are optional:

RUNSPEC – required section, includes a description of the simulation, including parameters

such as name of the simulation case, grid dimensions, phases and components
present, number of wells, etc.
GRID – required section, describes basic grid geometry/topology and petrophysical prop-
erties (porosity, permeability, net-to-gross).
EDIT – optional section describing user-defined changes that are applied to the grid data
after these have been processed to modify pore volume, cell centers, transmissi-
bilities, local grid refinements, etc.
PROPS – required section describing fluid and rock properties such as relative permeabil-
ities, capillary pressure, and PVT specification as a function of fluid pressures,
saturations, and compositions.
REGIONS – optional section defining how the computational grid can be split into multi-
ple regions for calculating PVT properties, relative permeabilities and capillary
pressures, initial conditions, and fluids in place.
SOLUTION – required section specifying how the model is to be initialized: calculated from
specified depths of fluid contacts, read from a restart file (output) from a previous
run, or specified for each grid cell.
SUMMARY – optional section that enables the user to specify the reservoir responses (well
curves, field-average pressure, etc.) to be written to a summary file after each time
step.
SCHEDULE – required section, defines wells and how they are to be operated, specifies
time-step selection and tolerances for the solvers, controls the output of quantities
defined over the grid (pressures, saturations, fluxes, etc.).

The sections must come in this prescribed order in the input file, but the keywords within
each section can be given in arbitrary order. Altogether, the input data file can be seen as a
kind of scripting language that not only enables you to input data to the simulator, but also
provides control of the time dependency and how the simulator will run a prescribed set of
time steps.
In the following, we borrow the naming convention from ECLIPSE and refer to the way
a simulation is to be run as a schedule. In the simplest case, a schedule consists of a number
of prescribed time steps and one control per well. However, schedules representing real
assets tend to be much more complicated, as can be inferred from the previous discussion
of wells.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.1 Industry-Standard Simulation 341

Let me also make two additional remarks: Since several of the keywords, and in partic-
ular those in the GRID section, may contain huge amounts of data, the input format offers
an INCLUDE keyword allowing input data to be separated into multiple files. Moreover,
although MRST aims to support the most widely used keywords in the required sections,
there are many keywords that are not supported. In particular, no effort has been put into
supporting the flexible, yet complex, options for user-specified postprocessing of geological
models in the EDIT section.
The following text will give some details about ECLIPSE input [270, 33], but only
when this is prudent to understand how you can set up the various part of a black-oil fluid
description in MRST. If you are interested in a more thorough introduction, I recommend
the lecture notes by Pettersen [250], which are much to the point, freely available, and
complementary to the discussion in my book. An early version of the ECLIPSE 100 User
Course [269] can also be found online.

More Advanced Solution Methods

As evidenced by our discussion, discrete flow models found in reservoir simulators are
significantly more complex than the incompressible, two-phase flow equations consid-
ered in previous chapters. For small models with relatively simple phase behavior and
solution trajectories, it is straightforward to linearize the flow equations and solve them
with standard linear algebra, much in the same way as we did for single-phase flow in
Section 7.2. However, for more complex phase behavior, there are many features compli-
cating the solution process, like the need to switch primary unknowns from one cell to
another when phases appear or disappear. Likewise, you may need extra postprocessing
steps to record historic solution trajectories to model hysteretic behavior of relative perme-
abilities and capillary pressure, and so on. As the complexity increases for each of the three
constituents of the reservoir model (geology model, fluid model, well model), one really
sees the benefit of using automatic differentiation to perform the necessary linearizations.
All these model complications translate to increased nonlinearity in the discretized
system, and a simple nonlinear Newton–Raphson solver cannot be expected to always con-
verge unless we instrument it with inexact line-search methods and methods for time-step
control, as briefly introduced in Section 10.2.2. Likewise, there is a need for correct scaling
of the residual equations, as well as more advanced linear solvers, e.g., a preconditioned
Krylov subspace method like the flexible general minimal residual (GMRES) method,
combined with specialized preconditioning strategies aimed at lessening the impact of the
largely different mathematical characteristics of the various sub-equations. Classic choices
of single-stage preconditioners include the incomplete lower/upper (ILU) family or the
nested factorization method [24], which is still the default method in ECLIPSE. Another
popular choice is the two-stage constrained-pressure residual (CPR) method [304, 305].
Here, the first step computes an approximate solution of a reduced-pressure matrix that
is appropriately scaled and selected so that one can use, e.g., a very efficient algebraic
multigrid solver [285, 119]. The second step then applies a broadband smoother like ILU
to the entire system. Section 12.3.4 briefly outlines such preconditioners and the exact
solution strategy used in MRST to obtain robust simulation of real cases. Full details can,
as always, be found in the code.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 342 Compressible Multiphase Flow

Details of the MRST Implementation

To a large degree, MRST has many of the model complexities discussed thus far imple-
mented in its black-oil type solvers. Our first successful attempt at implementing black-oil
simulators relied on discrete operators for discretization and automatic differentiation and
used the type of procedural programming discussed so far in the book. As more complex
behavior was introduced, the code gradually became more unwieldy, and at a certain stage,
we realized that a major restructuring was required to make the code easier to extend and
maintain. The result was the object-oriented AD-OO framework [170, 212, 30], which will
be discussed in more detail in the next chapter. The key idea of the AD-OO framework
is to separate the implementation of physical models (i.e., conservation equations and
constitutive relationships), discrete operators, nonlinear solvers and time-stepping, and
assembly and solution of the linear system. By using a combination of classes, structures,
and functions for generic access, details from within each of these contexts can be hidden
from the user/developer and only exposed if really needed. As a result, it is now much
simpler to implement generic algorithmic components or extend existing models with more
comprehensive flow physics.
This introduces a pronounced divide in MRST between simulators written for incom-
pressible and compressible flow. It also represents a pedagogical challenge. I have solved
it by relying on the simplified codes written in a procedural form to introduce key concepts
that are also found in the object-oriented framework, albeit implemented in a somewhat
different form. Once the basic concepts have been introduced, I switch to using the AD-
OO framework when presenting simulation cases with black-oil models. This means, in
particular, that the codes used to run black-oil simulations will not be presented at the same
level of detail as in previous chapters of the book, but all examples have complete source
codes in the ad-blackoil or book modules.

11.2 Two-Phase Flow without Mass Transfer

For a system consisting of two immiscible phases, each with a fixed chemical composition
and no mass transfer between the phases, the basic flow model consists of the conservation
equations (8.22) and the standard multiphase extension of Darcy’s law, (8.23). This if often
referred to as a dead oil model, and is representative for reservoirs with relatively thick or
residue oil that has lost its light and volatile hydrocarbon components, or for oils containing
a constant amount of dissolved gas because pressure is so high that gas cannot appear as a
separate phase at reservoir conditions (more about this later).
If we for simplicity neglect capillary pressure and introduce discrete approximations for
all derivatives, the discrete flow equations read,
(φS α ρ α )n+1 − (φS α ρ α )n
+ div(ρv)n+1
α = (ρq)n+1
α , (11.1a)
t n
rα 
Kk n+1 
v n+1
α = − n+1
grad(p n+1 ) − gρ n+1
α grad(z) . (11.1b)
μα

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.2 Two-Phase Flow without Mass Transfer 343

There are only two issues that are conceptually different from the corresponding single-
phase model (7.2) we studied in Chapter 7: first of all, we now have two phases and must
ensure that all saturation-dependent quantities are averaged correctly, as discussed in the
previous chapter. Secondly, wells must be treated slightly differently, since the wellbore
now may contain a mixture of two fluid phases. We shall put this issue off for now, and
return to it later.

A Simple Compressible Two-Phase Simulator

To develop our first simulator, we consider a simple rectangular 200 × 200 × 50 m3
reservoir, initially filled with oil with oil in the upper 2/3 and by water in the bot-
tom 1/3 of the volume. Water is injected in the lower southwest corner and fluids
are produced from the upper northeast corner. To mimic the effect of injection and
production wells, we manipulate the flow equations to ensure reasonable inflow/outflow
for the perforated cells. The resulting code, called twoPhaseAD.m, can be found in
the book module and follows essentially the same steps as the single-phase code in
Section 7.2. Setup of the geological model is essentially the same and we skip the
details.
For the fluid model, we assume that the water phase is slightly compressible with
quadratic relative permeabilities, whereas the oil phase is lighter, more compressible, and
has cubic relative permeabilities
[muW, muO ] = deal(1*centi*poise, 5*centi*poise);
[cw, co ] = deal(2e-6/psia, 1e-5/psia);
[krW, krO ] = deal(@(S) S.^2, @(S) S.^3);

The reference densities at 200 bar are 1014 kg/m3 for water and 850 kg/m3 for oil. Since the
capillary pressure is assumed to be zero, we can compute the initial pressure distribution
exactly as for single-phase flow on page 206. Transmissibilities and discrete operators are
independent of the flow equation and are defined as in Section 7.2.2, except that we now
also have to define the upwind operator:
upw = @(flag, x) flag.*x(N(:, 1)) + flag.*x(N(:, 2));

The setup of the simulation loop follows along the same path as in Section 7.2.4. We
start by initializing the unknown variables and introduce indices so that we that later can
easily extract the individual variables from the global vector of unknowns,
[p, sW] = initVariablesADI(p0, sW0);
[pIx, sIx] = deal(1:nc, (nc+1):(2*nc));

Next, we initialize the necessary iteration variables and create a simulation schedule con-
taining a gradual ramp-up of the time step. Then, the main loop is more or less identical to
the one shown on page 210. The only difference is how we compute the discrete equations.
We start by computing all cell-based properties

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 344 Compressible Multiphase Flow

% Densities and pore volumes

[rW,rW0,rO,rO0] = deal(rhoW(p), rhoW(p0), rhoO(p), rhoO(p0));
[vol, vol0] = deal(pv(p), pv(p0));

% Mobility: Relative permeability over constant viscosity

mobW = krW(sW)./muW;
mobO = krO(1-sW)./muO;

In the flux terms, we approximate the product of density and mobility at the interface
between cells using upwind-mobility weighting. To this end, we first define the difference
in phase pressures across each cell interface,
dp = grad(p);
dpW = dp-g*avg(rW).*gradz;
dpO = dp-g*avg(rO).*gradz;

which we then use to determine the upstream direction for each phase and pick cell values
from the upstream side
vW = -upw(double(dpW) <= 0, rW.*mobW).*T.*dpW;
vO = -upw(double(dpO) <= 0, rO.*mobO).*T.*dpO;

We now have all terms we need to compute the residual equations in each cell,
water = (1/dt(n)).*(vol.*rW.*sW - vol0.*rW0.*sW0) + div(vW);
oil = (1/dt(n)).*(vol.*rO.*(1-sW) - vol0.*rO0.*(1-sW0)) + div(vO);

The injector is set to inject 0.1 pore volume of water over a period of 365 days. We model
the injection by a source term having constant mass rate equal the total volumetric rate
times the surface density of water
water(inIx) = water(inIx) - inRate.*rhoWS;

For the producer, we just fake plausible behavior; that is, we replace the oil equation by
an equation that equates the cell pressure to the specified bottom-hole pressure and set the
residual equation for water equal to Sw ,
oil(outIx) = p(outIx) - outPres;
water(outIx) = sW(outIx);

These conditions will ensure that fluids flow into the perforated cell.

Structure of the Linearized System

As before, the linearized system can be assembled into a global system matrix by concate-
nating the two ADI vectors,
eqs = {oil, water}; eq = combineEquations(eqs{:}); % i.e., cat(eqs{:});

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.2 Two-Phase Flow without Mass Transfer 345

dmperm

∂O ∂O

po
ten
∂p ∂Sw

tia
lo
rde
rin
g
∂W ∂W
∂p ∂Sw

Figure 11.1 Linearized equations for the compressible, two-phase system without mass transfer
between phases. The discrete equations have a 2 × 2 block structure corresponding to the derivatives
of the water (W) and oil (O) equations with respect to the two primary variables p and Sw .

Time: 0.3125 days Time: 100 days Time: 360 days

Figure 11.2 Saturation profiles for the simple two-phase problem without mass transfer between the
phases. The plots only show cells in which Sw > 10−3 .

Let us inspect this system in some detail on a small 5 × 5 × 3 grid. A large number of the
coefficients in the residual equations are zero in the first iteration steps since the pressure
differential induced by the wells has not yet started to cause significant movement of fluids.
We therefore run five of the ramp-up steps, bringing us to time five days, before we inspect
the linearized system.
In the linearized system in Figure 11.1 we recognize the sparsity pattern of a standard
seven-point scheme in matrix blocks that have derivatives with respect to p. Blocks with
Sw -derivatives can be reordered to obtain the triangular structure we discussed in Section
10.3.4, which is typical for hyperbolic problems. Indeed, we have already seen that we can
rewrite this compressible system as a parabolic pressure equation and a hyperbolic transport
equation in the absence of capillary forces.
Figure 11.2 clearly shows that the suggested production strategy might not be optimal.
Already in the first time step, water is sucked up from the bottom layer and towards the
producer. However, by injecting in the water layer, rather than in the oil zone, we ensure
that oil is not pushed down into the water zone. Maybe you can think of a better design to

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 346 Compressible Multiphase Flow

produce this reservoir? You can for instance try to move the injector to the upper southwest
corner and repeat the simulation.
We leave the setup of wells and instead investigate what effect compressibility has on the
system by comparing solutions in the compressible and the incompressible case. If we reset
all the compressibilities to zero and rerun the code, we get the following error message:
Warning: Matrix is close to singular or badly scaled.
Results may be inaccurate.
RCOND = 2.977850e-23.

This is usually reason to worry, and we will get back to this issue shortly. However, let
us first look at the computed results in terms of oil production, pressure in cell perforated
by the injector, and field-average pressure. Oil production can be measured from a simple
mass-balance calculation. To this end, we need the initial oil in place, which we compute
as part of the initialization.
ioip = sum(pv(p0).*(1-sW0).*rhoO(p0));

To extract injector pressure, average field pressure, and oil in place at the end of each time
step we postprocess the array of structures used to store the computed solutions,
t = arrayfun(@(x) x.time, sol)/day;
ipr = arrayfun(@(x) x.pressure(inIx), sol)/barsa;
fpr = arrayfun(@(x) sum(x.pressure)/G.cells.num, sol)/barsa;
oip = arrayfun(@(x) sum(pv(x.pressure).*(1-x.s).*rhoO(x.pressure)), sol);

Notice here the combination of arrayfun, which enables us to operate on each element
of an array, and anonymous functions used to extract and process the fields belonging to
each array element. Figure 11.3 reports these reservoir responses. Both the pressure in
the perforated injector cell and the field-average pressure shown in the left plot exhibit

240
Injector pressure (ipr)
Field-average pressure (fpr) 4
x 10
230 Incompressible ipr 2
Incompressible fpr

220
1.5
oil production [m ]
3
pressure [bar]

210

200 1

190
0.5
180 Compressible
Incompressible
170
0
0 50 100 150 200 250 300 350 0 50 100 150 200 250 300 350
time [days] time [days]

Figure 11.3 Reservoir responses for the first two-phase example without mass transfer. The left plot
shows the pressure in the perforated cell and field-average pressure (FPR) and the right plot shows
total field oil production (FOPT). (The dotted line indicates initial FPR.)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.2 Two-Phase Flow without Mass Transfer 347

significantly different behavior in the compressible and incompressible models. In the

compressible case, we see that the injector pressure1 is significantly higher since some of
the injected energy is used to compress the oil and expand the pore space rather than only
pushing fluids, as in the incompressible case. At the same time, the reduced pressure at the
producer will cause the compressible oil to expand. Hence, we also observe a somewhat
higher total field-oil production. Looking at the FPR, we see that this is high initially in
the compressible case, but decays gradually as more fluids are produced. We also observe
the same trend in the incompressible case, but here we have an initial transient period
during which the pressure first decreases and then increases slightly before we observe the
expected gradual decay. Given that we do not have a real well model, it is somewhat futile
to put too much attempt into interpreting this behavior.

Numerical Challenges
Let us return to the numerical difficulties of the problem. Figure 11.4 reports the number
of iterations observed along with a plot of how the reciprocal condition number changes
with time for the two simulations. The more well-conditioned the system is, the closer
the reciprocal condition number should be to unity. All numbers are very small in our
case, indicating quite ill-conditioned systems. In particular, during the first four iterations
steps, the reciprocal condition numbers for the incompressible case are of the order 10−37
to 10−24 , which means that this system is effectively singular. Nevertheless, the direct
solver manages to compute a solution that can be used to continue the nonlinear iteration
toward convergence. This is a well-known problem in reservoir simulation; fully implicit
formulations are not suited for solving incompressible flow problems. A common trick to
avoid singular matrices is therefore to add a little compressibility to the problem to make it
less ill-posed. In our example, we could, for instance, assume incompressible fluids but set
the rock compressibility to 10% of the value used in the compressible simulation without
visibly affecting the pressure and production curves reported in Figure 11.3.
Also in the compressible case, the condition numbers are not very good. From a prag-
matic point of view, you may think that this is not a serious problem since the direct solver
in MATLAB has no problem to solve the system. Nevertheless, the problem will get worse
if the reservoir has strong heterogeneity. Moreover, when moving to larger systems, we
would eventually have to utilize an iterative solver, which in most cases needs a suitable
preconditioner to deliver satisfactory convergence, as we will discuss later.
To summarize: This simple example has introduced you to the typical structure of a
compressible, multiphase problem. You have seen that the elliptic and hyperbolic character-
istics discussed in previous chapters indeed can be observed in the discrete system. We have
also demonstrated that there are significant differences in compressible and incompressible
systems and that these generally require different solution strategies; hence the divide in
MRST between incompressible and compressible solvers. We will return to more numerical

1 Referring back to the Peaceman well model (4.35) on page 128, this pressure corresponds to p and will be lower than the
R
bottom-hole pressure pbh inside the wellbore.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 348 Compressible Multiphase Flow
Compressible
10 10
8 8
6 6
4 4
2 2
0 0
0 5 10 15 20 25 30 35 40
Incompressible
10 10
8 8
6 6
4 4
2 2
0 0
0 5 10 15 20 25 30 35 40
−10
x 10
5 0
10

4 10
−20

−40
3 10

Compressible
1 Incompressible

0
0 5 10 15 20 25 30 35 40

Figure 11.4 Numerical stability for the two-phase problem. In the upper plot, the bars show the
number of iterations to convergence for each of the 41 time steps; time-step lengths (in units days)
are shown as a line plot. The lower plot shows estimates of the variation in reciprocal condition
numbers for the linearized system.

examples later in the chapter, but now it is time for a more in-depth discussion of the fluid
and phase behavior of three-phase models.

11.3 Three-Phase Relative Permeabilities

When extending our two-phase simulator to three phases, we have to determine how to
model saturation-dependent quantities like capillary pressure and relative permeabilities. In
principle, one could imagine that all three phases could flow simultaneously in all parts of
the pore volume and that relative permeability and capillary pressure generally were func-
tions of two independent saturations. It is more common to assume that the flow is essen-
tially two-phase, with the third phase acting as a passive background phase that restricts
the flow volume and provides resistance to flow but does not contribute actively to the flow.
Effectively, this means that the relative permeabilities of gas and water are functions of gas
and water saturation, respectively, as we saw in Section 8.1.4. (The same applies to capillary

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.3 Three-Phase Relative Permeabilities 349

pressure.) A commercial simulator like ECLIPSE offers several different options [271]
for combining two-phase models to obtain three-phase relative permeabilities, including
both the Stone I [284] and II [283] models. Common for all options is that they rely
on interpolation of tabulated input. The various options are classified into families that
either use one-dimensional or two-dimensional input. This offers great flexibility, but may
also slow down the simulation if you are not careful when you specify the input tables.
Fully implicit simulators implemented with the AD-OO framework in the ad-blackoil
module utilize fluid objects, as discussed in Sections 5.1.1 and 10.1 to evaluate relative
permeabilities and capillary pressures as general functions of phase saturations. It is up to
you whether these functions should be based on analytic expressions or tabulated input,
and the default dependence on saturation only can easily be overridden should you need to
include other variable dependencies.
For convenience, the ad-props module offers a generic interface for constructing fluid
objects based on input in ECLIPSE format. So far, we have only implemented the most
common family of input parameters and the default option for interpolating three-phase oil
relative permeability from two-phase curves. Nonetheless, it should not be a very difficult
exercise to extend this implementation to include other families of input options. The rest
of this section describes the default model and its implementation in the ad-props module
in more detail.

11.3.1 Relative Permeability Models from ECLIPSE 100

The default model for calculating three-phase oil relative permeability in ECLIPSE 100
is based on the same key idea as the Stone I model (8.19)–(8.20) on page 243, but where
the Stone I model assumes a nonlinear relationship to interpolate the two-phase curves,
the default model employs a simpler linear relationship, which according to [271] avoids
problems of negative values and poor conditioning. The model makes the key assumption
that water and gas are completely segregated within each cell, giving the fluid distribution
illustrated in Figure 11.5. This means that each cell is divided into two zones (here, Sw , So ,
and Sg denote the average saturations):

• in the upper gas zone, only oil and gas flow, and the oil saturation is So , the water
saturation is Swc , and the gas saturation is Sg + Sw − Swc ;
• in the lower water zone, only oil and water flow, and the oil saturation is So , the water
saturation is Sg + Sw , and the gas saturation is 0.
Then, the oil relative permeability is interpolated linearly from two-phase relative perme-
abilities
Sg krog (So ) (Sw − Swc ) krow (So )
kro (Sg ,Sw ) = + , (11.2)
Sg + Sw − Swc Sg + Sw − Swc
with So = 1 − Sw − Sg . Here, krog is the oil relative permeability for a system with oil, gas,
and connate water, and krow is the oil relative permeability for a system with oil and gas

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 350 Compressible Multiphase Flow
Swc 1 − So − Swc So
1

Sg
Sg +Sw −Swc
gas

oil

Sw −Swc
Sg +Sw −Swc water

0
1 − So So

Figure 11.5 Fluid distribution inside each cell assumed by the default method for computing three-
phase oil relative permeability in MRST; adapted from [271].

only, both tabulated as a function of oil saturation. This model is implemented as default
within MRST, too.
To create a fluid object, we must read and process data from the PROPS section of
ECLIPSE input files. For this, we use the deckformat module, which offers general
functionality for reading and processing input decks. The most important function is the
reader,
deck = readEclipseDeck(fn);

which reads the input file fn and constructs a structure deck with one field for each
of the sections in the input file, except EDIT and SUMMARY. In passing, we mention that
the deckformat module also offers several functions to construct objects for grids, rock
properties, and wells, as well as functionality for reading and processing ECLIPSE output
files. Hence, you can use MRST to visualize output from ECLIPSE simulations, and it is
also possible to use existing ECLIPSE runs to initialize simulation cases in MRST.
To understand the input, we must look in more detail at the input format of the PROPS
section. All data are given as capital letter keywords, followed by numbers and terminated
by a single /. As an example, let us take an excerpt from the file BENCH_SPE1.DATA, which
is supplied with MRST and describes input for the SPE 1 benchmark case [241]
PROPS ===============================================================
-------- THE PROPS SECTION DEFINES THE REL. PERMEABILITIES, CAPILLARY
-------- PRESSURES, AND THE PVT PROPERTIES OF THE RESERVOIR FLUIDS
----------------------------------------------------------------------

-- ROCK COMPRESSIBILITY
-- REF. PRES COMPRESSIBILITY
ROCK
14.7 3.0E-6 /

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.3 Three-Phase Relative Permeabilities 351

Here, all lines starting with a double-dash (--) are interpreted as comments and the only
data contained in this part of the file are the two parameters necessary to describe the
compressibility of the rock, i.e., a reference pressure (in units bar) and a compressibility (in
units 1/bar). For this particular input file, the relative permeabilities are given by the two
keywords SWOF and SGOF:
SWOF
-- SWAT KRW KRO PCOW
0.1200000000 0 1.000000000000 0
0.1210000000 0.000000011363636 1.000000000000 0
0.1400000000 0.000000227272727 0.997000000000 0
0.1700000000 0.000000568181818 0.980000000000 0
0.2400000000 0.000001363636364 0.700000000000 0
0.3199999990 0.000002272727261 0.350000004375 0
0.3700000000 0.000002840909091 0.200000000000 0
0.4200000000 0.000003409090909 0.090000000000 0
0.5200000000 0.000004545454545 0.021000000000 0
0.5700000000 0.000005113636364 0.010000000000 0
0.6200000000 0.000005681818182 0.001000000000 0
0.7200000000 0.000006818181818 0.000100000000 0
0.8200000000 0.000007954545455 0.000000000000 0
1.0000000000 0.000010000000000 0 0
/

The four columns represent water saturation (SWAT), relative permeability of water (KRW),
relative permeability of oil (KRO), and oil–water capillary pressure (PCOW). After process-
ing, the deck.PROPS field will have one subfield containing each of these three data
deck.PROPS =
SWOF: {[14x4 double]}
SGOF: {[15x4 double]}
:
ROCK: [1.0135e+05 4.3511e-10 NaN NaN NaN NaN]

The careful reader may observe that each of the SWOF and SGOF fields represent the
input data as a cell containing an array. The reason is that the relative permeability input
may contain multiple arrays of different length corresponding to different regions of
the reservoir. Each array will then be represented as a cell and the whole function as a
cell array.
We can then create a fluid object by calling initDeckADIFluid from ad-props
fluid = initDeckADIFluid(deck);

The essential part of initDeckADIFluid is a generic loop that traverses the fields of
deck.PROPS and tries to pass each of them onto a constructor function written specifically
for each field
props = deck.PROPS;
fns = fieldnames(props);
for k = 1:numel(fns)
fn = fns{k};
if doAssign(fn)
asgn = str2func(['assign',fn]);
try

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 352 Compressible Multiphase Flow

fluid = asgn(fluid, props.(fn), reg);

catch %#ok
warning(msgid('Assign:Failed'), ...
'Could not assign property ''%s''.', fn)
end
end
end

The second line extracts a list with the names of all fields present, the fifth line checks the
field name against a list of field names that either have not been implemented yet or should
for some reason not be processes. The sixth line makes a function pointer by appending
the field name to the string “assign,” in our case creating function pointers to assignROCK,
assignSWOF, and assignSGOF. These functions are then called in the eight line, with the
fluid object fluid and the appropriate fluid data as input. (The third input argument, reg,
represents region information.) Let us look at the implementation of rock compressibility.
If we for simplicity disregard region information, the model for rock compressibility is set
up as follows:
function f = assignROCK(f, rock, reg)
[cR, pRef] = deal(rock(2),rock(1));
f.cR = cR;
f.pvMultR = @(p)(1 + cR.*(p-pRef));
end

Although rock compressibility, strictly speaking, is not a fluid relationship, it has for his-
toric reasons been bundled with the saturation and pressure-dependent fluid and rock-fluid
relationships since it represents a purely pressure-dependent quantity and will enter the
flow equations in the same way as the other elements of the fluid object. Notice also that
with the generic setup in initDeckADIFluid, all you must do to support a new keyword
MYKEY, is to implement the function assignMYKEY and place it in the right directory of the
ad-props module.
Assigning relative permeabilities is somewhat more involved. If we disregard code lines
having to do with the region information (i.e., that the model can use different relative per-
meability functions in different sub-regions of the grid), the constructor takes the following
form
function f = assignSWOF(f, swof, reg)
f.krW = @(sw, varargin) krW(sw, swof, reg, varargin{:});
f.krOW = @(so, varargin) krOW(so, swof, reg, varargin{:});
f.pcOW = @(sw, varargin) pcOW(sw, swof, reg, varargin{:});
swcon = cellfun(@(x)x(1,1), swof);
f.sWcon = swcon(1);
end

That is, we add three function pointers to the fluid object implementing the relative
permeability of water, the relative permeability from an oil–water displacement, and

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.3 Three-Phase Relative Permeabilities 353

the capillary pressure of between oil and water. Let us take the krOW function as an
example,
function v = krOW(so, swof, reg, varargin)
satinx = getRegMap(so, reg.SATNUM, reg.SATINX, varargin{:});
T = cellfun(@(x) x(:,[1,3]), swof, 'UniformOutput', false);
T = extendTab(T);
v = interpReg(T, 1 - so, satinx);
end

The first line sets up a map that lets us select the correct saturation functions if there
are multiple regions. The next line extracts the first and the third column of the input
data, i.e., the SWAT and KRO columns. Here, we use cellfun in combination with an
anonymous function to extract individual elements in the same way as we used arrayfun
to extract simulation results from the sol array of structures in the previous section. The
saturation values in the relative permeability tables may not necessarily span the whole unit
interval, and in the third line we add a constant extrapolation to ensure that we can use T
to extrapolate for all legal saturation values. The last line calls an interpolation function
that uses region information to pick the correct relative permeability function from the cell
array and then calls an appropriate interpolation function. If T had been a simple array, the
last line would have amounted to
F = griddedInterpolant(T(:,1), T(:,2) , 'linear', 'linear');
v = F(1-so);

The functions krW and pcOW are implemented exactly the same way, with the obvious
modifications, and so are the functions in assignSGOF.

11.3.2 Evaluating Relative Permeabilities in MRST

By default, fluid objects constructed by initDeckADIFluid only contain functions for
evaluating the tabulated two-phase systems found in the input file through functions
krW, krOW, krG, and krOG. For convenience, the ad-props module offers the function
assignRelPerm that implements the three-phase relative permeability model described in
the beginning of the section. You can run assignRelPerm to instrument your fluid object
with a generic interface to evaluate relative permeabilities. However, before this function
is called, you must have run the necessary routines to assign data tables and functions for
interpolating these data to the fluid object f, or alternatively have implemented your own
fluid object that offers the same functions krW, krOW, etc. Notice also that if you later
update any of the two-phase data in the fluid object, assignRelPerm must be run again to
ensure correct evaluation.
The assignRelPerm function inspects the fields that are present in the fluid object and
assigns the correct relative-permeability functions accordingly

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 354 Compressible Multiphase Flow

function f = assignRelPerm(f)
if isfield(f, 'krOG') % two-phase water-oil
f.relPerm = @(sw, varargin)relPermWO(sw, f, varargin{:});
elseif isfield(f, 'krOW') % two-phase oil-gas
f.relPerm = @(sg, varargin)relPermOG(sg, f, varargin{:});
else % three-phase
f.relPerm = @(sw, sg, varargin)relPermWOG(sw, sg, f, varargin{:});
end
end

The logic is quite simple: If there is no oil–gas data (krOG) for a three-phase system
available, we must have a water–oil system. If there is no oil–water data (krOW) for a three-
phase system available, we must have an oil–gas system. Otherwise, all three phases are
present and we must set up a full three-phase model. The relative permeabilities for the
two-phase water–oil system are set up as follows:
function [krW, krO] = relPermWO(sw, f, varargin)
krW = f.krW(sw, varargin{:});
if isfield(f, 'krO')
krO = f.krO(1-sw, varargin{:});
else
krO = f.krOW(1-sw, varargin{:});
end
end

Here, the default is to look for a two-phase table called krO for the oil relative permeability.
If this is not present, the user must have supplied an oil–water table from a three-phase
model, and we use this instead. The oil–gas system is treated analogously. Finally, for the
three-phase model:
function [krW, krO, krG] = relPermWOG(sw, sg, f, varargin)
swcon = f.sWcon;
swcon = min(swcon, double(sw)-1e-5);

d = (sg+sw-swcon);
ww = (sw-swcon)./d;
krW = f.krW(sw, varargin{:});

wg = 1-ww;
krG = f.krG(sg, varargin{:});

so = 1-sw-sg;
krow = f.krOW(so, varargin{:});
krog = f.krOG(so, varargin{:});
krO = wg.*krog + ww.*krow;

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.3 Three-Phase Relative Permeabilities 355

11.3.3 The SPE 1, SPE 3, and SPE 9 Benchmark Cases

In 1981, Aziz S. Odeh [241] launched an initiative for the independent comparison of the
state-of-the-art reservoir simulators for 3D black-oil models. To represent a good cross-
section of solution methods used throughout the industry, seven companies were selected,
each participating with their own simulator. This type of comparison was well received by
both the industry and the academic community. Over the next two decades, the Society of
Petroleum Engineering (SPE) published a series of nine additional studies, which are today
known as the ten SPE comparative solution projects. In each study, a number of oil com-
panies, software providers, research institutes, universities, and consultants participated on
a voluntarily basis, reporting their best attempt to solve a specified problem. Altogether,
these ten problems constitute a rich source of test problems that can be used to verify and
validate various multiphase models and solution strategies. You have already seen data
from the tenth project used multiple times throughout this book.
In this example, we use input decks for the SPE 1 [241], SPE 3 [158], and SPE 9 [159]
comparison projects to demonstrate how you can use MRST to set up and compare different
fluid models. If you have not already done so, you can download the necessary files by use
of MRST’s data set manager:
mrstDatasetGUI;

This will place the data sets on a standard path, so that they can be read and processed using
the following set of commands (with the obvious modifications for SPE 3 and SPE 9):
mrstModule add deckformat ad-core ad-props
pth = getDatasetPath('spe1');
fn = fullfile(pth, 'BENCH_SPE1.DATA');
deck = readEclipseDeck(fn);
deck = convertDeckUnits(deck);
f = initDeckADIFluid(deck);
f = assignRelPerm(f);

This produces a fluid object with a number of fields (here for SPE 1):

krW: @(sw,varargin)krW(sw,swof,reg,varargin{:})
krOW: @(so,varargin)krOW(so,swof,reg,varargin{:})
pcOW: @(sw,varargin)pcOW(sw,swof,reg,varargin{:})
sWcon: 0.1200
krG: @(sg,varargin)ireg(Tkrg,sg,varargin{:})
krOG: @(so,varargin)ireg(Tkro,sgas(so),varargin{:})
pcOG: @(sg,varargin)ireg(Tpcog,sg,varargin{:})
:
relPerm: @(sw,sg,varargin)relPermWOG(sw,sg,f,varargin{:})

Here, we recognize relative permeability of water, oil relative permeability of an oil–water

system, oil–water capillary pressure, connate water injection, etc. The last function is a

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 356 Compressible Multiphase Flow
Water relative permeability curve Gas relative permeability curve
1 1
SPE1
0.9 SPE3 0.9
SPE9
0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
SPE1
0.1 0.1 SPE3
SPE9
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Water saturation Gas saturation

Figure 11.6 Relative permeabilities for water and gas for the fluid models from three of the SPE
Comparative Solution Projects.

generic interface for evaluating relative permeabilities of all phases in the three-phase
system. Let us look at the two-phase relative permeabilities krW and krG, which we evaluate
by a call on the form
s = linspace(0,1,51)';
krW = f.krW(s);

Figure 11.6 shows the resulting curves (extrapolated krG values are not plotted for satura-
tions Sw < Swc ). SPE 1 is represented as a three-phase system in this particular input file,
but it is really a two-phase gas–oil problem that describes gas injection into an undersatu-
rated oil, i.e., an oil that can dissolve more gas. (We discuss this sort of PVT behavior in
the next section.) The water relative permeability is therefore virtually zero over the whole
saturation interval. For the SPE 3 model, both relative permeability curves have monotone
derivatives similar to what we have seen in the analytic models discussed in Section 8.1.4.
This is not the case for the gas relative permeability of SPE 1, which has an S-shape similar
to the fractional flux functions we have seen in two-phase systems. For the SPE 9 model,
the relative permeability curves have significant kinks, and this will contribute to increase
the nonlinearity of the discretized flow equations.
Figure 11.7 shows the corresponding relative permeability curves for two-phase oil–
water and gas–oil systems and the resulting interpolated three-phase oil relative permeabil-
ity curve. The plots were generated as follows:
[sw, sg] = meshgrid(linspace(0,1,201)); so=1-sw-sg;
[, krO, ] = f.relPerm(sw(:),sg(:));
krO = reshape(krO,size(sw)); krO(sw+sg>1)=nan;

[mapx,mapy] = ternaryAxis('names',{'S_w';'S_g';'S_o'});
contourf(mapx(sw,sg,so), mapy(sw,sg,so), krO, 20)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.3 Three-Phase Relative Permeabilities 357

Oil relative permeability curves

So=1
1
Oil/water system
SPE 1:
0.9 Oil/gas system
SPE1
0.8 SPE3 0.2 0.8
SPE9
0.7

0.6 0.4 0.6

0.5

0.4 0.6 0.4

0.3

0.2 0.8 0.2

0.1 Oil relative permeability

0
0 0.2 0.4 0.6 0.8 1 S =1 0.2 0.4 0.6 0.8 S =1
w g
Oil saturation

So=1 So=1
SPE 3: SPE 9:

0.2 0.8 0.2 0.8

0.4 0.6 0.4 0.6

0.6 0.4 0.6 0.4

0.8 0.2 0.8 0.2

Oil relative permeability Oil relative permeability

S =1 0.2 0.4 0.6 0.8 S =1 S =1 0.2 0.4 0.6 0.8 S =1

w g w g

Figure 11.7 Three-phase relative permeabilities of oil for three of the SPE Comparative Solution
Projects generated by linear interpolation of two-phase curves from oil–water and oil–gas displace-
ments (upper left plot).

The krOW and krOG curves representing two-phase displacements coincide for both the
SPE 1 and SPE 3 data sets, and hence the interpolated krO surface is smooth and monotone
along straight lines in phase space. For SPE 9, the two-phase curves are not only signifi-
cantly different, but the krOW curve has a very steep gradient. As a result, we observe kinks
and it is also possible to identify straight lines in the phase diagram along which krO is
non-monotone; see e.g.,
sw = linspace(0,.6,21); sg = linspace(0,.4,21);
[,krO] = f.relPerm(sw',sg');

As is warned about in [270], abrupt changes in relative permeability values over small
saturation intervals can be the source of significant convergence problems when solving
the discretized flow equations and may cause the simulator to slow down. Like ECLIPSE,
MRST honors the input data and does not attempt to overcome convergence problems by
smoothing kinks or retabulating the input data.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 358 Compressible Multiphase Flow
4
3
Oil production [m ] 4 Gas production [m3] Pressure [bar]
x 10 x 10 400
5 14
34%
12 350
4
10
300
3 8
19% 250
2 6
200
4
1 no injection
4.6% 150
2 0.5 PV water
1.5 PV water
0 0 100
0 200 400 600 800 1000 0 200 400 600 800 1000 0 200 400 600 800 1000
time [days] time [days] time [days]

Figure 11.8 Conceptual three-phase simulation with no mass transfer between the phases. The
reservoir is initially filled with gas on top, oil in the middle, and water at the bottom. Fluids are
produced from the middle layer. The plots show results of simulations with no water injection and
with injection of 0.5 and 1.5 pore volumes of water into the bottom layer. The plots to the right show
average pressure in the injector cell (solid lines) and FPR (dashed lines).

11.3.4 A Simple Three-Phase Simulator

Having established how to model three-phase relative permeabilities, it is a simple exercise
to extend the simple simulator from Section 11.2 to three phases under the assumption of no
mass transfer between the phases. This assumption is not very realistic, but is introduced to
isolate and illustrate certain aspects of multiphase flow. To exemplify, we consider a setup
similar to the one discussed in Section 11.2, with the upper 1/3 initially filled with gas
(Sg = 1 − Swc ), the middle 1/3 filled with oil (So = 1 − Swc ), and the bottom 1/3 is filled
with water. Fluids are produced from a cell at the northeast corner of the middle section by
setting the pressure in this cell to be 100 bar, which is approximately 100 bar lower than
the initial average reservoir pressure. We change the fluid model so that water is twice as
viscous as oil, and assume that relative permeabilities follow the SPE 3 fluid model.
When the producer starts depleting the reservoir pressure, the compressed gas will
expand so that a large fraction of it is driven out of the reservoir through the oil layer
below. Gas is more mobile than oil and will flow past the oil rather than displace it. Hence,
only a minor fraction of the oil will be recovered. To increase oil recovery, we inject water
into the water zone at the lower southwest corner. Code to simulate this case is given in
threePhaseAD.m in the book module. We run a comparison of no water injection and
injection of 0.5 and 1.5 pore volumes of water
pvi=0; threePhaseAD; pvi=.5; threePhaseAD; pvi=1.5; threePhaseAD

Figure 11.8 reports observed hydrocarbon production and pressure behavior. Without water
injection, the reservoir pressure is quickly depleted and reaches a steady state after approx-
imately 100 days, and out of the volumes initially in place, we recover 75%, but only 4.6%
of the oil. By injecting 0.5 pore volumes of water, we are able to maintain reservoir pressure
and hence get higher oil production in the period until all the gas is depleted and the injected
water breaks through in the producer. This happens after approximately 740 days. Towards

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.4 PVT Behavior of Petroleum Fluids 359
1.5 PV water injection

10
8
6
4
2
0
0 20 40 60 80 100

Figure 11.9 Number of nonlinear iterations for the conceptual three-phase simulation.

the end of this period, the pressure increases significantly as the water front approaches the
producer and then stabilizes at a plateau. When gas disappears and oil is the most mobile
phase, the oil production increases abruptly, but then decays slowly as a result of gradually
increasing water cut in the producer. Similar behavior can be observed when injecting 1.5
pore volumes of water over the same time period. However, since injection rates are three
times higher, water breakthrough occurs much earlier and we also observe pressures that
are much higher than the initial field pressure.
If we try to repeat the simulation with a total injection of two pore volumes, the sim-
ulator fails to converge within 15 iterations in the 23rd time step, right before the water
front breaks through in the producer. We also observe a similar increase in the number of
iterations towards water breakthrough for the other water injection cases; see Figure 11.9. It
is well known that Newton’s method may experience convergence problems if saturations
(or compositions) change too much during a single time step. We have already discussed
this issue in Section 10.2.2, where we also introduced a reactive strategy for chopping of
the time step.

Computer exercises
11.3.1 Visualize the structure of the linearized three-phase system in the same way as
in Figure 11.1. Try to permute the system so that the numbering in the vector of
unknowns first runs over variables and then over cells.
11.3.2 What happens if you set water to be incompressible?
11.3.3 Extend the simulator in threePhaseAD.m to include the reactive time-step control
discussed in Section 10.2.2 and try to extend it so that it also uses changes in
saturation as a criterion to chop time steps. Does this enable you to run a simulation
with more than two pore volumes injected?

11.4 PVT Behavior of Petroleum Fluids

Our assumptions of no mass transfer between phases and that each fluid phase has constant
compressibility, but otherwise constant fluid properties, represent gross oversimplifications.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 360 Compressible Multiphase Flow

Pressure and temperature will not only affect the compressibility and viscosity of each
fluid phase, but also determine which phases are present and how the different chemical
components in the reservoir fluids distribute among the aqueous, oleic, and gaseous phases
at reservoir and surface conditions. Hydrocarbon fluids can come in a wide variety of forms,
including the bitumen (semisolid rock form of crude oil) of the Athabasca oil province in
Canada and heavy (dead) oils containing almost no light hydrocarbon components. We
will not discuss such cases in any detail but instead focus on hydrocarbon fluids that have
a significant gas component at surface conditions. That is, medium oils containing lighter
components that are liberated as free gas at surface conditions; volatile oils that shrink
significantly when liberating free gas; rich gases that condense liquid at reduced pressure;
wet gases that condensate fluids at lower temperatures; and dry gases that mainly consist
of light hydrocarbons and stay in single-phase vapor form.

11.4.1 Phase Diagrams

We have previously discussed equations of state modeling the relationship among pressure,
volume (density), and temperature. For reservoir fluids it is more common to use phase
diagrams, which in this setting are a type of chart showing pressure and temperature
conditions at which thermodynamically distinct phases occur and coexist at equilibrium.
This section discusses such diagrams for reservoir fluids.

Single-Component Substances
As a starting point, however, let us look at a typical phase diagram for a substance consist-
ing of a single component. To explain such diagrams, we first look at Gibbs’ phase rule
for a system without chemical reactions. This rule states that the number of independent
degrees of freedom F that must be specified to determine the intensive state of a system is
given by

F = 2 + nc − np,

where nc is the number of components and np is the number of phases. For a single-
component system (nc = 1), we have F = 3 − np , and hence there can at most be
three phases. A state with three phases corresponds to zero degrees of freedom and the
phases must therefore coexist in a single point called the triple point, at which pressure
and temperature are uniquely determined. If two phases coexist, there is one degree of
freedom, and pressure can be chosen freely if temperature is fixed, and vice versa. When
the substance is in a single-phase state, both pressure and temperature can be chosen
independently.
Figure 11.10 illustrates the resulting phase diagram. In the diagram, the three single-
phase regions corresponding to solid, liquid, or vapor phase are separated by the subli-
mation, melting, and vaporization curves. Along these curves, two phases coexist in equi-
librium. Across the curves, various properties of the substance (compressibility, viscosity,

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.4 PVT Behavior of Petroleum Fluids 361

Figure 11.10 Conceptual phase diagram for a single-component substance. The dashed blue line
illustrates the melting curve for water, which unlike most other fluids has negative slope.

etc.) change abruptly and have discontinuous derivatives with respect to temperature and
pressure. At the triple point, all phases exist in equilibrium. For temperatures below the
critical point Ttp , the substance is in vapor phase if the pressure is below the sublimation
curve and in solid state for pressures above. For pressures and temperatures along the
sublimation curve, vapor and solid are in equilibrium. For temperatures between the triple
point and the critical point (Ttp < T < Tcp ), the substance is in vapor state if the pressure is
below the vaporization curve, in liquid state if the pressure is above the vaporization curve,
and in solid state if the pressure is above the melting curve.
The liquid–vapor curve terminates at the critical point. The critical temperature Tcp is
the temperature at, and above, which vapor of the substance cannot be liquefied, no matter
how much pressure is applied. Critical temperatures depend on the size of the molecule
and hence vary a lot for hydrocarbons, from low temperatures for small molecules to high
temperatures for large molecules. For water, the critical temperature is 374◦ C, whereas
carbon dioxide has a much lower critical temperature of 31.2◦ C. Likewise, the critical
pressure Pcp is the pressure required to liquefy a gas at its critical temperature, or in other
words, the pressure above which distinct liquid and gaseous phases do not exist. The critical
pressure of water is 217.7 atm and for CO2 it is 73.0 atm. For temperatures above Tcp , the
substance is said to be in a supercritical state, i.e., in a state in which the liquid and gaseous
phases are indistinguishable. This is the state in which CO2 is usually found when stored
in a deep geological formation (approximately below 800 m).
For most substances, the melting curve separating liquid and solid phase has a posi-
tive derivative with respect to temperature; the denser the molecules of the substance are
brought together, the higher temperature is required to break them apart to form a liquid
state. Water is one exception to this rule.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 362 Compressible Multiphase Flow

Figure 11.11 PVT behavior of a single-component hydrocarbon substance. The left figures illustrate
measurement of pressure-volume behavior in a high-pressure cell at fixed temperature. Mercury is
connected to a pump that can inject or extract this fluid to control pressure. Mercury has high surface
tension and will hence be non-wetting for most container surfaces like glass. The right plot shows
the measurements plotted in a PVT diagram, with the bubble-point and vaporization curves projected
onto a pressure-temperature diagram.

For hydrocarbon recovery, we are primarily interested in the vaporization curve. To

better understand the phase behavior for multicomponent systems, we first briefly explain
an experimental setup for determining this curve using a high-pressure cell, as illustrated
in Figure 11.11. Initially, the cell is filled with liquid hydrocarbon at the top and mercury
at the bottom. As mercury is gradually removed, the liquid hydrocarbon expands and the
pressure drops. This continues until the first bubble of gas comes out of the solution as we
reach the bubble point (point 2 in the figure). Beyond this point, the gas phase expands the
volume rapidly while keeping the pressure constant (point 3 in the figure). Eventually, we
reach the , where the last liquid droplet vaporizes (point 4 in the figure). As more mercury
is removed, the gas expands while the pressure decreases. If we repeat the experiments for
multiple temperatures, we get the diagram shown in the right plot. By projecting the locus
of bubble points, which we henceforth refer to as the bubble-point curve, and the locus of
dew points, which is called the dew-point curve, we obtain the vaporization curve in the
pressure-temperature diagram.

Binary Substances
Hydrocarbon reservoir fluids are made up of a wide range of chemical substances, and
hence their phase diagrams will be very complicated. Gibbs’ phase rule obviously gives
us little guidance since the number of actual phases occuring in the reservoir is much less

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.4 PVT Behavior of Petroleum Fluids 363

Figure 11.12 A pressure-temperature diagram for a binary system of fixed composition. Bubble-point
curve is shown in blue and dew-point curve shown in red. The lines inside the two-phase envelope
represent constant percentages of liquid. The diagram is specific to a given chemical composition of
the binary mixture and will change if this composition changes, e.g., if components are extracted at
different rates.

than the maximum number of phases that may exist. In black-oil models, the hydrocarbon
fluids are considered to be a two-component system. (Notice that this distinction between
oil and gas components is only relevant when the reservoir is in a state where both can
coexist.) For binary substances, the number of degrees of freedom for a two-phase system
is F = 2, and hence two phases could in principle exist for arbitrary combinations of
temperature and pressure, but this is not the case in practice. To describe the reservoir
fluids, we utilize a bubble-point/dew-point projection similar to the one we studied for
the single-component fluid above. This projection is defined for a constant mole fraction
of one of the components, and hence the phase diagram will depend on the composition
of the binary mixture. In the projected plot, the bubble-point and dew-point curves no
longer coincide along a single vaporization curve. Once the bubble point is reached, the
expansion is accompanied by a pressure decrease, since the gaseous phase has a different
chemical composition than the oleic phase. The projection hence produces a temperature
and pressure envelope, inside which the binary mixture forms two phases as shown in
Figure 11.12.
At pressures and temperatures above the bubble-point curve, the binary mixture forms
a liquid phase. Likewise, a gaseous phase is formed for all points located below or to the
right of the dew-point curve. The bubble-point and dew-point curves meet in the critical
point, where the phase compositions and all phase properties are identical. Notice that the
critical point does not necessary coincide with the maximum temperature of the two-phase

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 364 Compressible Multiphase Flow

envelope. The maximum temperature is called the cricondentherm and the maximum pres-
sure is called the cricondenbar. The critical point represents a mathematical discontinuity
in the system, and the phase behavior is generally difficult to define near this point, meaning
that it is not obvious whether a (slight) pressure decline will cause the formation of bubbles
of vapor or droplets of liquid.

11.4.2 Reservoir Types and Their Phase Behavior during Recovery

In petroleum literature, the phase diagrams discussed in the previous subsection are some-
times used to classify reservoirs according to the type of phase behavior that takes place
during (primary) production. To motivate this classification, and the choice of physical vari-
ables and closure relationships to be discussed in the next section, we start by a schematic
of a typical production setup for a depleting reservoir as shown in Figure 11.13.
When a well is opened to start production of hydrocarbons, the pressure inside the
reservoir will start to decline. Reservoirs are thermally inert, and unless other fluids having
significantly different temperature are injected into the reservoir, the temperature of the
reservoir fluids will remain (almost) constant at the value given by the geothermal gradi-
ent of the area. Inside the reservoir, the extraction of hydrocarbons therefore takes place
along a vertical line in the pressure-temperature diagram. Also when fluids are injected
to displace the resident hydrocarbons, temperature effects are in many cases so small that
the temperature can be considered as constant (near the producing well) in the reservoir
model. This means that various types of reservoirs can be classified by the location of the

Figure 11.13 Illustration of pressure and temperature changes resulting in liberation and expansion
of gas and shrinkage of oil during production of a hydrocarbon reservoir containing both oil and gas.
Stock-tank conditions vary but will roughly be 1 atm and 20◦ C.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.4 PVT Behavior of Petroleum Fluids 365

initial reservoir pressure and temperature relative to the bubble-point/dew-point envelope.

If the temperature of the reservoir is less than the critical temperature (Tr < Tc ) of the
hydrocarbon mixture, the reservoir is said to be an oil reservoir. Conversely, if the tem-
perature is larger than the critical temperature (Tr > Tc ), the reservoir is said to be a gas
reservoir. In both cases, the reservoir may contain hydrocarbons in both liquid and vapor
form, depending on the type of the reservoir.
As the hydrocarbons leave the reservoir and flow to the surface, they undergo a series
of changes that reduce the pressure and temperature to surface conditions. These changes
generally cause liberation of gas, expansion of the gaseous phase, and shrinkage of the
liquid oleic phase. Pressure reductions come in two types: in flash liberation, gas comes out
of solution as a result of a sudden pressure drop, and then stays confined with the remaining
oil. In differential liberation, gas is liberated from the oil as pressure is gradually decreased
and is continuously removed from the oil. How the reservoir fluids contained in the well
stream are brought from reservoir pressure and temperature to surface condition can have
a large effect on how much oil and gas is recovered from a given quantity of reservoir
fluids. Traditionally, oil has had much higher economic value than gas, which often has
been considered as a waste product. It has therefore been highly important to design a
production system that maximizes the stock-tank oil yield.
There are many mechanisms that induce pressure reduction: friction and gravity forces
acting in the opposite direction of the flow, loss of kinetic energy due to expansion, contrac-
tion in the fluid area, flow through constrictions, etc. At the surface, the fluids pass through
the Christmas tree mounted on top of the well head, whose main purpose is to control the
flow of reservoir fluids out of the well. The Christmas tree has isolation valves to stop
the fluid flow for maintenance or safety purposes, and flow lines and chokes that control
the flow into the downstream production system. Inside the production system, the fluids
can continue into one or more pressure vessels (separators) used to separate the produced
fluids into oil, gas, and water components, and possibly remove gas and/or water. The oil
component may continue to a stock tank for temporary storage, inside which further gas
liberation may take place.
We already know that the pressure drop inside the reservoir and the pressure drop from
the reservoir to the wellbore both depend on flow rate; this follows from Darcy’s law and the
Peaceman well model, respectively. Similar relationships between pressure drop and flow
rate can be developed for the other parts of the flow system from the bottom hole to the
stock tank; that is, pressure loss in the tubing, in the Christmas tree and flow lines, across
the choke, and in the separator units. A few examples are presented in Section 11.7.2. These
relationships are used by production engineers to optimize the hydrocarbon recovery. For a
more detailed discussion, you can consult a textbook on so-called nodal analysis [145].

Gas Reservoirs
All reservoirs in which the initial temperature is higher than the critical temperature are
called gas reservoirs. Methane is by far the dominant hydrocarbon component in gas reser-
voirs. If the initial temperature is higher than the cricondentherm, the hydrocarbon mixture

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 366 Compressible Multiphase Flow

Figure 11.14 Phase diagrams for gas reservoirs. In a single-phase dry/wet gas reservoir (left),
the temperature is above the cricondentherm. In a retrograde gas condensate reservoir (right), the
temperature is above the critical temperature, but below the cricondentherm. During retrograde
depletion below the dew-point curve, the composition of the mixture will change inducing a change
in the diagram indicated by the dotted lines.

will be in a single-phase gaseous state and will continue to be so during depletion. As

the reservoir is depleted, the composition of the produced fluids do not change, but the
hydrocarbon mixture may enter the two-phase envelope on its way to the surface, depending
upon the cooling process, as illustrated in the left plot of Figure 11.14.
If a gas reservoir produces liquid hydrocarbons at the surface (condensate), the reservoir
is said to be a condensate gas, wet gas, or rich gas reservoir. Conversely, if the temperature–
pressure path followed during production does not penetrate the two-phase envelope, all
hydrocarbons produced at the surface will be in gas phase and the reservoir is said to be a
dry gas reservoir. Dry gas reservoirs rarely contain hydrocarbons with carbon number five
or higher and typically consist of C1 , C2 , N2 , CO2 , and H2 S. No hydrocarbon separation is
needed at the surface since the well stream consists of gas only, but separators/dehydrators
may be needed to remove produced/condensed water. Although wet-gas reservoirs mainly
consist of methane, they also contain both intermediate hydrocarbons (ethane to hexane)
as well as a significant fraction of hydrocarbons with carbon number seven or higher. The
condensation of liquid hydrocarbons at the surface is a result of cooling, but whether the
well stream is brought to surface conditions in one big flash or through a series of smaller
flashes does not have a large effect on the relative amounts of gas and liquid observed at
the surface.
Unlike dry/wet gas reservoirs, dew-point or retrograde gas condensate reservoirs con-
tain a significant fraction of heavier hydrocarbon components (C7+ ). The right plot in
Figure 11.14 shows such a hydrocarbon mixture, which is initially in (an indeterminate)
vapor state and will continue to be so as the reservoir is depleted toward point A on
the dew-point curve. During this initial decline, the composition of the produced fluid
remains constant. Below point A, a liquid oleic phase containing heavier hydrocarbons

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.4 PVT Behavior of Petroleum Fluids 367

will condense out of the vapor as a dew or fog and continue to do so until the process
reaches point B of maximum liquid percentage. This process is called retrograde, since
one normally would see vaporization and not condensation when pressure decays.
The condensed liquid is immobile at low saturations and generally less mobile than the
vapor phase. Hence, more light hydrocarbons will be recovered from the reservoir, giving
an increased gas–oil ratio at the surface. This changes the composition of the hydrocarbon
mixture, increasing the fraction of heavier hydrocarbons, which in turn shifts the two-phase
envelope toward the right, as indicated by the dotted lines in the figure. This aggravates the
loss of heavier hydrocarbons, which are retained as immobile droplets inside the rock pores.
If production is continued, the system will sooner or later reach a point of maximum liquid
percentage, below which the condensed liquid will start to vaporize. This improves the
mobility of the liquid components and leads to a decrease in the surface gas–oil ratio.
For a dew-point reservoir, the exact design of how the reservoir fluids are brought to sur-
face conditions can have a large impact on the amount of stock-tank oil and gas. This may
call for a complex network of separators in which separated vapor or liquid is reinjected
into the liquid–vapor mixture to better distribute heavier hydrocarbon components to the
stock-tank oil and lighter components to the final gas stream.
In closing, we remark that the distinction between condensate and oil is somewhat
artificial and has more to do with how the liquid hydrocarbons are recovered than their
chemical composition. Liquid hydrocarbons in the stock tank are called oil if they originate
from a liquid oleic state in the reservoir and condensate when they have been condensed
from a gaseous state.

Bubble-Point Oil Reservoirs

A reservoir having initial temperature below the cricondentherm is characterized as an oil
reservoir. Unlike the dead oil considered in Section 11.2, which either contained a constant
amount of dissolved gas or had lost all its components, the oleic phase in bubble-point
reservoirs consists of a live oil containing dissolved gas, which may be released at lower
pressures inside the reservoir and at the surface. This type of reservoir is also called a
solution-gas reservoir, and will be the topic for most of the simulation examples discussed
later in the chapter. Although somewhat artificial and imprecise, it is common to subdivide
oil reservoirs into two categories: black-oil reservoirs are mainly made up of a variety of
hydrocarbons, including heavy and non-volatile components (C7+ ). Volatile reservoirs con-
tain fewer heavy components and have more intermediate hydrocarbon molecules (ethane
to hexane) and the oleic phase therefore tends to shrink dramatically when gas is liberated.
Black-oil reservoirs tend to have much less shrinkage. The main property that distinguishes
the two categories, however, is the presence of vaporized oil in the gaseous phase, which
is assumed to be negligible in black-oil reservoirs and significant in volatile reservoirs.
This might be a bit confusing since we later will use black-oil equations to model both
types of reservoirs. The reason is most likely historic. Traditionally, the largest and most
profitable reservoirs in the world have predominantly been of black-oil type, and the black-
oil equations were developed to model these. Volatile reservoirs are typically found at

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 368 Compressible Multiphase Flow

Figure 11.15 Phase diagram for bubble-point oil reservoirs. The left plot shows a black-oil reservoir
with undersaturated initial oil and the right plot a volatile oil reservoir.

greater depths, and as exploration goes deeper, more reservoirs of this type have been
discovered and entered production. Volatile reservoirs contain less oil per volume, but since
the reservoir fluids are more mobile because of lower oil viscosity and higher dissolved-
gas content, this type of reservoir can still be more economically attractive than a black-oil
reservoir.
If the pressure is above the bubble-point curve as shown in the left plot of Figure 11.15,
the hydrocarbon mixture of the oil reservoir is in a liquid state usually referred to as
undersaturated, since all the available gas has been dissolved in the oleic phase. However,
when the first liquid is brought to the surface, it liberates free gas. As production continues,
the pressure will decline toward the bubble-point curve. Below the bubble point, gas will
be released from the oleic phase as vapor bubbles that gradually form a free gaseous phase.
Because the oil cannot hold all the available gas, we say that it is in a saturated state. When
the gaseous phase exceeds its residual saturation, it will start to flow toward the producer.
At the surface, the gas thus consists of both expanded free gas and gas liberated from the
oil during production. Oftentimes, surface facilities have to limit the gas rate, which in turn
can reduce the oil rate below an economic level unless secondary mechanisms like water
injection are applied to maintain the reservoir pressure and push the oleic phase toward the
producer (as we saw for the dead oil system in Section 11.3).
The volatile reservoir shown to the right in Figure 11.15 is initially in an saturated state
inside the two-phase envelope. The reservoir fluids thus consist of a liquid oleic phase
which exists in equilibrium with a gaseous phase containing gas condensate. Because the
gaseous phase has a different composition and is lighter than the oleic phase, the gaseous
phase will form a gas zone or gas cap overlying the liquid oil zone and has a phase diagram
that is largely different from that of the oleic phase. Equilibrium between the two phases
implies that the oleic phase must be at its bubble point, whereas the gaseous phase must be
at its dew point and may either exhibit retrograde or non-retrograde behavior, as illustrated

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.4 PVT Behavior of Petroleum Fluids 369

Figure 11.16 Phase diagrams for a reservoir at saturated conditions so that liquid oil is in equilibrium
with a gas cap containing vaporized oil. The gaseous phase will exhibit retrograde behavior in the
diagram to the left and non-retrograde behavior in the diagram to the right.

Figure 11.17 Sketch of phase diagrams for different types of hydrocarbon reservoirs.

in Figure 11.16. Notice also that volatile reservoirs are not very different from retrograde
gas condensate reservoirs for temperatures close to the critical temperature.
We end our discussion about reservoir classification by presenting two comparisons
of the different reservoir types. Figure 11.17 shows a schematic with the different phase
diagrams plotted relative to each other. This sketch is by no means accurate, but illustrates
how the phase diagrams move to the right when the hydrocarbon mixture contains (a larger
fraction of) heavier hydrocarbon components. Table 11.1 reports a number of characteristic
properties for different reservoir types.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 370 Compressible Multiphase Flow
Table 11.1 Petroleum fluids and their characteristics at initial conditions or
bubble-point pressure (adapted from [177]).

Oils Gases
Heavy oils Gas Wet and dry
Characteristic and tars Black oils Volatile oils condensates gases

Mol. weight 210+ 70–210 40–70 23–40 <23

Oil color in stock Black Brown to Greenish to Orange to Clear
tank light green orange clear
Stock-tank ◦ API 5–15 15–45 42–55 45–60 45+
C7 + fraction (mol%) >50 35–50 10–30 1–6 0–1
Reservoir 90–200 100–200 150–300 150–300 150–300
temperature ◦ F
Saturation pressure 0–500 300–5,000 3,000–7,500 1,500–9,000 —
[psia]
Initial Bo [RB/STB] 1.0–1.1 1.1–1.5 1.5–3.0 3.0–20.0 20+
Initial Rs [scf/STB] 0–200 200–900 900–3,500 3,500–30,000 30,000+
Initial Rv 0 0–10 10–200 50–300 0–50
[STB/MMscf† ]
Max vol% liquid‡ 100 100 100 0–45 0

† At bubble-point pressure.
‡ At constant composition expansion of reservoir fluid.

11.4.3 PVT and Fluid Properties in Black-Oil Models

The phase diagrams discussed in the previous section are useful to classify reservoir fluids
and visualize the pressure-temperature path from reservoir to surface conditions. They aid
the reservoir engineer in deciding the depletion process, designing top-side processing
facilities (separators, etc.), and devising appropriate strategies for improved oil recovery
(water injection or reinjection of gas to maintain pressure, etc.). Phase diagrams are also
useful when setting up a fluid sampling program to determine parameters describing PVT
behavior in flow models.
As we have seen already in Section 8.2.3 on page 246, the fluid and PVT descrip-
tion in a black-oil model tries to mimic how the gas and oil pseudo-components at
surface conditions partition into the gaseous and oleic phases at reservoir conditions,
as illustrated in Figure 11.13 and discussed in detail in the previous section. To this
end, the composition and density of the oil and gas components are defined at surface
conditions (where they easily can be measured), and then the following three param-
eters are used to determine the density of the oleic and gaseous phases at reservoir
conditions:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.4 PVT Behavior of Petroleum Fluids 371

Bo – The oil formation-volume factor is defined as the volume of oil in liquid phase that
must be recovered from the reservoir to produce one volume unit of stock-tank
oil; in other word, Bo = Vo /Vos . A reservoir volume V with porosity φ and
oil saturation So thus contains φV So /Bo stock-tank volumes of oil. In SI units,
the formation-volume factor is a dimensionless quantity, but in oilfield units it
is usually reported in units reservoir barrels per stock-tank barrels [bbl/STB],
where 1 bbl equals 0.1589873 m3 . Some authors prefer the reciprocal quantity,
bo = 1/Bo , i.e., the factor by which one reservoir volume of oil shrinks (or
expands) when brought to surface conditions. MRST uses shrinkage factors rater
than formation-volume factors.
Bg – The gas formation-volume factor is defined similarly, as the volume of gaseous phase
that must be recovered from the reservoir to produce a unit volume of gas at the
surface, i.e., Bg = Vg /Vgs . It is common to express Bg in units reservoir barrels
per standard cubic feet [bbl/scf]. MRST uses the reciprocal expansion factor bg
= 1/Bg .
Rs – The solution gas–oil ratio is the volume of gas at standard conditions that dissolves
into a unit stock-tank volume of oil at reservoir conditions and is usually reported
in units standard cubic feet per stock-tank barrel [scf/STB]. A reservoir volume V
with porosity φ and oil saturation So thus contains φV So Rs /Bo surface volumes
of liberated gas.

This description assumes that the reservoir is in instantaneous phase equilibrium. In

addition to these three parameters, we also need the bubble point pb of the oil, a similar
formation-volume factor for water, as well as the dynamic viscosities of the gaseous,
oleic, and aqueous phase as function of pressure (and temperature). To model volatile-oil,
retrograde dew-point, and wet-gas reservoirs, one also need to model the amount of
vaporized oil present in the gaseous phase:

Rv – The vaporized oil–gas ratio (or volatile oil–gas ratio) is defined as the stock-tank vol-
ume of condensate oil per unit volume of free gas at surface conditions produced
from the gaseous phase at reservoir conditions.

The formation-volume factors and gas–oil and oil–gas ratios all depend on reservoir pres-
sure and temperature and are determined by laboratory experiments (or simulations) repro-
ducing the sequence of differential (and flash) liberations that most likely take place during
production of the reservoir. Since reservoirs are almost thermally inert, it is usually suffi-
cient to express these quantities as function of pressure and not temperature.
In the following, we discuss various fluid properties and relationships in some more
detail, and in most cases try to mention what are plausible ranges of their values. Readers
interested in a more in-depth discussion should consult a textbook on reservoir fluids like
e.g., [207, 316, 177] or www.petrowiki.org.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 372 Compressible Multiphase Flow

Equation of State: Fluid Densities

Using the quantities defined earlier, we can write the following equations of state that
relate the densities of the aqueous, oleic, and gaseous phases (denoted by capital letters
W,O,G) at reservoir conditions to the densities of the oil, water, and gas pseudo-
components (denoted by o,w,g) at surface conditions
ρws
ρW = = bw ρws (11.3)
Bw
ρos + Rs ρgs
ρO = = bo (ρos + Rs ρgs ) (11.4)
Bo
ρgs + Rv ρos
ρG = = bg (ρgs + Rv ρos ) (11.5)
Bg
Water density is almost invariant to pressure changes. The oleic phase behaves very dif-
ferently above and below the bubble point. Above the bubble point, it behaves like a
standard fluid in the sense that the density increases with increasing pressure. Below the
bubble point, gas dissolution will in most cases dominate compression effects, so that
density decreases with increasing pressure as a result of dissolved gas that increases the
fraction of lighter hydrocarbon components. The gaseous phase is strongly affected by
fluid compression so that density increases with pressure.

API and Specific Gravity

In our discussion of phase behavior, we repeatedly talked about light and heavy hydrocar-
bon components in the mixture. In petroleum literature, it is common to use the American
Petroleum Institute gravity as a measure of how heavy or light a petroleum liquid is. The
API gravity is defined as
141.5 ρ
γAP I = − 131.5, γ = , (11.6)
γ ρw
where γ is the specific gravity of liquid  to water, i.e., the ratio of the density of liquid
to the density of water. Bubble-point pressure pb , oil viscosity μo , and the solution gas–oil
ratio Rs , all depend strongly on oil gravity. It also has an indirect effect on oil compress-
ibility and oil formation-volume factors, since these depend on Rs .
The API gravity is specified as degrees on a hydrometer (◦ API) and is defined so that any
liquid having API gravity greater than 10◦ is lighter and floats on water, whereas liquids
with values less than 10◦ are heavier and sink. Using this scale, crude oil is commonly
classified as follows:

• light oil: API gravity higher than 31.1◦ , i.e., less than 870 kg/m3
• medium oil: API gravity between 22.3 and 31.1◦ , i.e., 870–920 kg/m3
• heavy oil: API gravity below 22.3◦ , i.e., 920–1,000 kg/m3
• extra heavy oil: API gravity below 10.0◦ , i.e., greater than 1,000 kg/m3
Typical API gravity values are also reported in the overview of fluid characteristics for
different reservoirs in Table 11.1.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.4 PVT Behavior of Petroleum Fluids 373

For gas, the specific gravity is measured relative to air and is defined as
ρg Mg
γg = = ,
ρair Mair
where M denotes the molar mass and Mair = 28.94 kg/kmol. Specific gas gravity is one
of the parameters that determine bubble-point pressure, viscosity, and compressibility of
gas, as well as the solution gas–oil ratio. Typical values range from 0.55 for dry gas to
approximately 1.5 for wet sour gas. Petroleum gases have gravity approximately equal
0.65.
Reservoir water is saline and hence has a specific gravity greater than unity, but has no
impact on compressibility, formation-volume factor, and viscosity.

Compressibility and Formation-Volume Factor of Dry Gas

You may recall from Section 4.2 that the equation of state of an ideal gas is given as
pV = nRT , where n is the number of moles of gas and R is the universal gas constant.
The literature is full of equations of state for real gases, but in petroleum engineering it is
common to use a so-called compressibility equation of state to describe dry gas
Vreal
pV = ZnRT , Z= . (11.7)
Videal
The compressibility factor or gas deviation factor Z measures how much the real gas
volume deviates from that of an ideal gas at the same temperature. The gas compressibility
factor typically varies between 0.8 and 1.2. For constant temperature and moderate pres-
sures, the gas molecules are close enough to exert an attraction on each other that causes
the actual gas volume to be less than that of an ideal gas. As p approaches 0, the molecules
move more apart and the gas behaves like an ideal gas. At higher pressures, repulsive forces
will tend to force modules more apart and the real gas therefore has a higher volume than
an ideal gas. The same type of equation of state can be used also for wet gas and mixtures
if Z is replaced by a pseudo function that depends on the composition of the mixture. For
volatile oils and retrograde-condensate gases it is more common to use a more advanced
equation of state like one of the cubic equations discussed in Section 4.2. For a dry gas, the
formation-volume factor can be derived from (11.7),
ps ZT
Bg = , (11.8)
Zs Ts p
where subscript s refers to standard conditions. This varies throughout the world, and can
be e.g., 14.65 psia and 60◦ F or 1 atm and 25◦ C.
The isothermal compressibility of the gas is defined as the fractional change of volume
with respect to pressure for constant temperature,
1  ∂V  1  ∂Bg 
cg = − =− . (11.9)
V ∂p T B g ∂p T

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 374 Compressible Multiphase Flow

Z(p)

pressure, p

Figure 11.18 Characteristic shape of gas compressibility factor Z as function of pressure for fixed
temperature.

Bg

pressure, p

Figure 11.19 Characteristic shape of gas formation-volume factor as function of pressure for fixed
temperature and chemical composition.

Inserting the compressible equation of state, it follows that

1  nRT  ∂Z  ZnRT  1 1  ∂Z 
cg = − − = − .
V p ∂p T p2 p Z ∂p T

At low pressures, ∂Z/∂p is negative and hence the compressibility of a real gas is larger
than that of an ideal gas. At high pressures, ∂Z/∂p is positive and hence the compressibility
is lower than for an ideal gas.

Gas Viscosity
Dynamic gas viscosities typically range from 0.01 to 0.03 cP and are generally difficult to
determine experimentally. It is therefore common to use correlations to determine values.
At low pressures, gas molecules are far apart and can easily move past each other. As
pressure increases the molecules are forced closer to each other and hence the viscosity of
the gas increases. In MRST, gas viscosities are either given as simple explicit formulas (as
seen so far) and correlations [316], or as tabulated input read from an ECLIPSE simulation
deck (as we will see later).

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.4 PVT Behavior of Petroleum Fluids 375

Rs

Rv

pb pressure, p

Figure 11.20 Characteristic shape of the solution gas–oil ratio and vaporized oil–gas ratio as
functions of pressure for fixed temperature and chemical composition.

Bubble-Point Pressure
Bubble-point pressure is one of the key parameters in the phase behavior of reservoir
fluids since it marks the limit between the liquid and two-phase liquid–vapor region for
hydrocarbon fluids. In the previous section we outlined briefly how bubble-point pressure
can be measured when studying fluid expansion as function of pressure and temperature.
The literature is full of correlations that can be used to determine the bubble-point pressure
all on the form pb = f (T ,γAP I ,γg ,Rs ); see e.g., [177] or www.petrowiki.org for an
overview. None of these are implemented in MRST. Instead, the software follows the
convention from leading commercial simulators and requires tabulated B and Rs factors,
in which pb is only given implicitly.

Solution Gas–Oil Ratio and Vaporized Oil–Gas Ratio

When discussing dissolved gas, we are in reality referring to light hydrocarbon molecules
that appear in gaseous phase at the surface and in liquid oleic phase at reservoir conditions.
For a given temperature, the quantity of light molecules present in the oleic phase is limited
only by the pressure and the number of light molecules present. For pressures below the
bubble point, the oleic phase will release some free gas if the pressure is lowered slightly,
and hence we say that the phase is saturated. Above the bubble point, the oleic phase is able
to dissolve more gas than what is available and is said to be undersaturated. Altogether, this
means that a plot of the quantity of dissolved gas in the oleic phase as function of pressure
increases linearly up to the bubble point and then remains constant; see Figure 11.20. The
vaporized oil–gas ratio is only defined for pressures below the bubble point.

Oil Formation-Volume Factor

Most measures of fluid production has traditionally taken place at the surface. It has there-
fore been customary to introduce the formation volume-factor to relate volumes measured
at surface conditions to volumes at reservoir conditions. The oil formation-volume factor
is affected by three mechanisms: The most important is the liberation of gas when bringing
liquid oil from the reservoir to the surface, which will cause a decrease in the liquid volume,

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 376 Compressible Multiphase Flow

Bt
2

Bo

1
pb pressure, p

Figure 11.21 Characteristic shape of the oil and total formation-volume factor as functions of
pressure for fixed temperature and chemical composition.

so that the Bo factor always assumes values greater or equal 1; the more gas that is dis-
solved, the larger the Bo factor. At reservoir conditions, increased pressure introduces two
additional effects: fluid compression and swelling caused by dissolution of gas. The latter
is dominant and below the bubble point, the formation-volume factor therefore increases
with increasing pressures. Above the bubble point, there is no more gas to be dissolved,
and hence fluid compression causes a gradual decrease in Bo ; see Figure 11.21.
The compressibility of oil is defined in the same way as for gas. For undersaturated oil,
we therefore have that
1 ∂Bo 
co = − , Bo = Bob exp co (pb − p) , p > pb, (11.10)
Bo ∂p
where Bob is the oil formation-volume factor at the bubble point.

Total Formation-Volume Factor

To better understand the behavior for saturated oil, we introduce the total or two-phase
formation-volume factor, defined as the ratio between the total hydrocarbon volume at
reservoir conditions and the volume of oil or gas at standard conditions. For simplicity,
let us first consider the case of a black oil with a surface oil volume Vo , i.e., a system
with dissolved gas but no vaporized oil. At the bubble point, the reservoir fluids inside the
reservoir are found as a single oleic phase having a volume Vo Bob and containing Vo Rsb
volumes of dissolved oil. If we lower the pressure slightly, the system will enter the two-
phase region, so that the oleic phase liberates an amount of gas that at the surface would
amount to a volume Vo (Rsb − Rs ). At reservoir conditions, this corresponds to a volume
Bg (Rsb −Rs ) of free gas. Hence, the total formation-volume factor with respect to oil reads,

Bto = Bo + Bg Rsb − Rs .
Because of the expanding gas, the Bto function can have significantly larger values at lower
pressures than at the bubble point.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.4 PVT Behavior of Petroleum Fluids 377

Figure 11.22 Illustration of the volumes involved in a volatile oil model.

For a retrograde gas condensate reservoir, we can start at the dew point and carry out a
similar argument to obtain

Btg = Bg + Bo Rvd − Rv ,
where Rvd is the vaporized oil–gas ratio at the dew point.
For volatile oils, things get a bit more complicated as Figure 11.22 illustrates. If no water
is present, the pore volume Vφ is filled by the oleic and gaseous phase that together contain
an amount Vo of oil and Vg of gas. Let Vof be the stock-tank volume of free oil and Vgf
the surface volume of free gas. Then, we have
Vφ = Vof Bo + Vgf Bg .
For pressures below the bubble point, oil and gas will appear in both the oleic and gaseous
phase, and hence Vo = Vof + Vgf Rv and Vg = Vgf + Vof Rs . Combining these two and
solving for Vgf gives Vgf = (Vg − Vo Rs )/(1 − Rv Rs ). Inserting into the equation above
and simplifying, we have
 
Bo − Rs Bg Vo + Bg − Rv Bo Vg
Vφ = . (11.11)
1 − Rs Rv
Just above the bubble point, the hydrocarbon is in liquid phase and hence all gas is resolved
in the oil, i.e., Vg = Vo Rsb . By definition, Vφ = Bto Vo . Inserting into the above equation,
dividing by Vo , and reorganizing terms, we obtain the following expression for the two-
phase formation-volume factor
⎧
⎨Bo, p ≥ pb ,
Bto = Bo 1−R Rv +Bg R −Rs (11.12)
⎩ sb sb
, p < pb .
1−Rv Rs

The two-phase formation-volume factor for retrograde gas condensate can be derived
similarly as

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 378 Compressible Multiphase Flow

10−3

10−7
pb pressure, p

Figure 11.23 Isothermal oil compressibility as function of pressure for fixed temperature and
chemical composition.

⎧
⎨Bg , p ≥ pd ,
Btg = Bg 1−R Rs +Bo R −Rv (11.13)
⎩ vd vd
, p < pd .
1−Rv Rs

Isothermal Oil Compressibility

Isothermal oil compressibility is defined as the relative change in oil volume per change
in pressure. It follows from the earlier discussion that oil compressibility has different
behavior above and below the bubble point. For an undersaturated oleic phase above
the bubble point, the compressibility is defined in the same way as for a gas, i.e.,
co = (−1/Bo )(∂Bo /∂p)T , and is a primary drive mechanism. At the bubble point, gas
comes out of solution and causes a sharp increase in compressibility. Below the bubble
point, the compressibility becomes a much stronger function of pressure as the oil volume
now depends on expansion/compression of both liquid oil and dissolved gas
1  ∂Bo  1 Bg − Bo Rv  ∂Rs 
co = − + . (11.14)
B o ∂p T B o 1 − Rs Rv ∂p T
As a drive mechanism, however, oil compressibility is dominated by the compressibility of
gas that has come out of solution.

Oil Viscosity
We have already seen in Section 7.3 that the viscosity of a fluid is affected by both tempera-
ture and pressure, with temperature being the most dominant for a fluid of constant compo-
sition. Viscosity is also related directly to the type and size of molecules the fluid is made
up of. Above the bubble point, the oleic phase has a constant chemical composition and
the lower the pressure is, the easier the molecules can move past each other, since they are
further apart. Hence, the viscosity increases almost linearly with pressure. Below the bubble
point, the oleic phase changes chemical composition as more lighter molecules are removed
from the liquid when gas is liberated. The fraction of larger molecules with complex shapes
will gradually become larger, causing the viscosity to increase. One order of magnitude

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.5 Phase Behavior in ECLIPSE Input Decks 379

μo (p)

pb pressure, p

Figure 11.24 Oil viscosity as function of pressure for fixed temperature and chemical composition.

increase in viscosity from bubble point to low reservoir pressure is not uncommon, and this
largely reduces the mobility of the remaining oil.

Water Properties
The aqueous phase is usually a brine that contains a number of different dissolved salts and
hence has specific density greater than unity. Water may also contain certain amounts of
dissolved gas, but this is usually ignored in standard simulation models since the solubility
factor in water is orders of magnitude lower than in the liquid oleic phase.
Water viscosity decreases with increasing temperature and increases with increasing
pressure, but is generally a very weak function of pressure. Typical values at reservoir con-
ditions range from 0.5 to 1 cP. Water is only slightly compressible at reservoir conditions,
with values in the range from 1 × 10−6 to 9 × 10−6 psi−1 , which implies that the water
formation-volume factor is usually close to unity.

11.5 Phase Behavior in ECLIPSE Input Decks

Use of tabulated data from the PROPS section of ECLIPSE input is the primary input mech-
anism to set up simulation models for the simulators implemented in the ad-blackoil,
ad-eor, and other derived modules. As we saw earlier on page 350, functionality for
reading these input files, parsing and processing them to create specific fluid objects is
offered by the deckformat and ad-props modules.
To make the book as self-contained as possible, this section outlines the necessary
ECLIPSE keywords and show a few examples of specific phase behavior for the first, third,
and ninth SPE benchmark cases. For a more comprehensive discussion of the input format,
you should consult the lecture notes of Pettersen [250], or the documentation of ECLIPSE
[270, 269] or OPM Flow [33]. The following presentation is inspired by these sources.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 380 Compressible Multiphase Flow

Dead Oil
Dead oils are assumed to always stay above the bubble point and hence behave like a
standard fluid that is compressed and becomes more viscous at higher pressures. Most
dead oils contain a constant amount of dissolved gas that is liberated as the oleic phase is
brought to the surface. There are two different ways to describe dead oil in an ECLIPSE
input deck:

PVDO – specifies the properties of an undersaturated oil above the bubble point in terms of a
table of formation-volume factors and viscosities versus monotonically increasing
pressure values. The table is attached to specific (groups of) cells, and the keyword
can contain multiple tables if the reservoir contains several zones with distinct oils.
The number of tables is therefore given by the NTPVT element of the TABDIMS
keyword from the RUNSPEC section. The constant amount of dissolved oil and the
associated bubble-point pressure should be specified using the RSCONST keyword.
The following example is taken from [250]:
PVDO
-- Po Bo mu_o
27.2 1.012 1.160
81.6 1.004 1.164
136.0 0.996 1.167
190.5 0.988 1.172
245.0 0.9802 1.177
300.0 0.9724 1.181
353.0 0.9646 1.185 /
RSCONST
-- Rs Pbp
180 230 /

Here, the pressures Po and Pbp are given in units bar, the formation-volume
factor Bo and the gas–oil ratio Rs are dimensionless, and the viscosity mu_o has
unit cP.
PVCDO – describes an oleic phase with constant compressibility by specifying the five
constants in the formulas

Bo (p) = Bor e−co (p−pr ), (Bo μo )(p) = Bor μor e−(co −cμ )(p−pr ) .

Here, pr is reference pressure, Bor reference formation-volume factor, co com-

pressibility, μor reference viscosity, and cμ “viscosibility.” In addition, one must
specify RCONST and a maximum pressure PMAX. The following example is taken
from [269] (with pressures given in units psi):
PVCDO
-- P Bo Co muO Cmu
2500 1.260 6E-6 0.5 1E-6 /
RSCONST
-- GOR Pb
0.656 2500 /
PMAX
4500 /

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.5 Phase Behavior in ECLIPSE Input Decks 381

Live Oil
Undersaturated oil has a well-defined bubble point that can be determined from the temper-
ature and the oil’s chemical composition. Saturated oils, on the other hand, change chemical
composition as gas is dissolved or liberated. To model the phase behavior of a live oil, you
need to describe both how the bubble point changes with changing composition of the oil,
as well as the properties in the undersaturated region for each unique composition. In a
black-oil model, the fluid composition of the oleic phase is represented in terms of the
solution gas–oil ratio, and all we need to discriminate saturated from undersaturated states
is a relationship between Rs and the bubble point pb . In ECLIPSE input, this is represented
by tabulating pb (Rs ). For each value of pb (Rs ), the dead-oil behavior in the undersaturated
region is prescribed in the same way as described for dead-oil models. Hence, the input
language offers two keywords for describing live oils:

PVTO – consists of NTPVT tables that each give the properties above and below the bubble
point. Each table contains from two to NRPVT records that specify PVT data for
given Rs values. The records consist of at least four numbers specifying Rs ,
bubble point pressure pb , formation-volume factor Bob for saturated oil at the
bubble point, and viscosity μob at the bubble point. However, each record can
also be composed of from two to NPPVT rows of data organized in three columns
that describe oil pressure (increasing values), formation-volume factor (decreasing
values), and viscosity for undersaturated oil for the specified value of Rs . These
extra data must be specified for the maximum Rs value inside each table but can
optionally be given for any Rs value. We present a specific example of this input
format for the SPE 9 model on page 383.
Notice that ECLIPSE interprets the input data by linear interpolation of the
reciprocal values of Bo and (μo Bo ) between data points.

PVCO – specifies properties above and below the bubble point in compressibility form;
that is, instead of specifying the undersaturated behavior in terms of a tabular,
it is described as for the PVCDO keyword. The data thus consists of NTPV tables that
each may have up to NRPVT rows. The following example is
from [269]:
PMAX
6000 /
PVCO
--Pb Rs Bo Vo Co Cvo
1214.7 0.137 1.17200 1.97000 1E-5 0
1414.7 0.195 1.20000 1.55600 1* 0
1614.7 0.241 1.22100 1.39700 1* 0
1814.7 0.288 1.24200 1.28000 1* 0
2214.7 0.375 1.27800 1.09500 1* 0
2614.7 0.465 1.32000 0.96700 1* 0
3014.7 0.558 1.36000 0.84800 1* 0
3414.7 0.661 1.40200 0.76200 1* 0
3814.7 0.770 1.44700 0.69100 1* 0
/

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 382 Compressible Multiphase Flow

Notice, in particular the use of 1* to indicate a default value that should be com-
puted by linear interpolation. Since the table only gives one value in the first row,
this value should be copied to all the next rows.

Dry Gas
Dry gases are found either below the dew point or (well) to the left of the critical point.
Their phase behavior is equivalent to that of dead oils and is described as a tabular of
formation-volume factor and viscosity as function of pressure, or by a Z-factor and viscos-
ity as function of pressure:
PVDG – specifies NTPVT tables that each consists of three columns and up to NPPVT rows
giving gas pressure (increasing monotonically downwards), gas formation factor,
and gas viscosity. A specific example of this input format is presented for the
SPE 9 model on the next page.
PVZG – specifies properties using Z-factors. The data consists of NTPV tables that each
consists of two records. The first gives the reference temperature used to convert
Z-values into formation-volume factors (see [11.8] on page 373), and the second
consists of from 2 up to NPPVT rows giving gas pressure, Z-factor, and gas vis-
cosity. Pressures and Z-values should increase monotonically down the table. The
lowest pressure value should correspond to the dew point. The gas may contain a
constant amount of vaporized oil, specified by the RVCONST keyword, which also
gives the dew-point pressure.

Wet Gas
Wet or rich gases with vaporized (condensate) oil are similar to live oils in the sense that the
phase behavior is different for saturated and undersaturated states. The composition of the
wet gas is represented by vaporized oil–gas ratio Rv . To discriminate between saturated and
undersaturated states, you must tabulate the maximum amount of vaporized oil as function
of pressure, i.e., Rv (pg ). The corresponding keyword is:
PVTG – describes properties of wet gas with vaporized oil. The format is similar to the PVTO
keyword used to describe live oil, and consists of from two to NRPVT records that
specify PVT data (Rv , Bg , and μg ) at the saturated state for given p values. Each
record can also be composed of from two to NPPVT rows of three-column data
that specify Rv , Bg , and μg for undersaturated states. These extra data must be
specified for the highest pressure data in each of the NTPVT tables.
We shown an example of this input format on page 387 when discussing the PVT behavior
of the SPE 3 benchmark.

Miscellaneous Properties
Water properties are described with the keyword PVTW, which follows the same format as
the PVCDO keyword for dead oil. In addition to the PVT behavior, you also need to prescribe

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.5 Phase Behavior in ECLIPSE Input Decks 383

the density or specific gravity of each fluid at surface condition, which is defined by the
keywords DENSITY or GRAVITY. The PROPS section of the input deck should also describe
how the pore volume depends on pressure, either by the ROCK keyword that specifies a
reference pressure and rock compressibility, or by the ROCKTAB keyword that tabulates
porosity (and permeability) multipliers as function of pressure.

Example: SPE 1 and SPE 9 – Live Oil With Dry Gas

To exemplify live oil and dry gas models, we revisit the SPE 1 and SPE 9 benchmarks
discussed on page 355. The SPE 9 benchmark describes a live oil with dry gas specified as
follows (the setup for SPE 1 is similar):
-- LIVE OIL (WITH DISSOLVED GAS) -- DRY GAS (NO VAPORIZED OIL)
PVTO PVDG
-- Rs Pbub Bo Vo -- Pg Bg Vg
.0 14.7 1.0000 1.20 / 14.7 178.08 .0125
.165 400. 1.0120 1.17 / 400. 5.4777 .0130
.335 800. 1.0255 1.14 / 800. 2.7392 .0135
.500 1200. 1.0380 1.11 / 1200. 1.8198 .0140
.665 1600. 1.0510 1.08 / 1600. 1.3648 .0145
.828 2000. 1.0630 1.06 / 2000. 1.0957 .0150
.985 2400. 1.0750 1.03 / 2400. 0.9099 .0155
1.130 2800. 1.0870 1.00 / 2800. 0.7799 .0160
1.270 3200. 1.0985 .98 / 3200. 0.6871 .0165
1.390 3600. 1.1100 .95 / 3600. 0.6035 .0170
1.500 4000. 1.1200 .94 4000. 0.5432 .0175 /
5000. 1.1189 .94 /
/ DENSITY
-- Oil Water Gas
PVTW 44.98 63.01 0.0702 /
--Pref Bw Comp Vw Cv
3600. 1.0034 1.0E-6 0.96 0.0 /

The model is specified in field units, so that densities are given as lb/ft3 and not kg/m3 and
pressure values are given in unit psi. For the live oil, we see that undersaturated behavior
is described for the largest Rs value. We do not repeat the statements required to load
these model, but simply assume that we have created a processed deck structure with the
necessary input data and created a fluid object f. Once the units have been converted to
the SI system and the input data have been processed to construct a fluid object, the phase
behavior of the oil is represented by the following functions and constants:
bO: @(po,rs,flag,varargin) bO(po,rs,pvto,flag,reg,varargin{:})
muO: @(po,rs,flag,varargin) muO(po,rs,pvto,flag,reg,varargin{:})
rsSat: @(po,varargin) rsSat(po,pvto,reg,varargin{:})
rhoOS: 720.5105

MRST has functions for computing the reciprocal expansion/shrinkage factors rather than
the usual formation-volume factors. We also note that Rs is called rsSat rather than Rs.
The reason is that dissolved gas will be an unknown in our system of flow equations and
hence we change the name of the parameter Rs that describes the maximum dissolved gas
amount for a given pressure to avoid confusion.
Looking at the functions in more detail, we see that the bO function takes three primary
arguments: oil pressure, the amount of dissolved oil, and a flag that indicates whether

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 384 Compressible Multiphase Flow

a given element in po and rs represents a saturated or an undersaturated state. In the

underlying functions, pvto is an object that holds the data of the PVTO keywords, whereas
reg represents region information that picks the right table, should the model have multiple
oils. The implementation of the BO, bO, muO, and rSat functions can be found in the file
assignPVTO.m and consists of calls to two utility functions that extract region information,
which is subsequently used to interpolate the correct data tables. For saturated oil, we use a
simple linear interpolation between the data points, whereas the interpolation for undersat-
urated oil is set up to preserve compressibility and viscosibility if only one undersaturated
data curve is given, or to interpolate compressibility and viscosibility if multiple curves are
given. I will not go through the interpolation functions in detail; these have been optimized
for speed and are only easy to understand if you are a black-belt MATLAB programmer.
Instead, we can plot the various relationships. (You can find the necessary code in the
showSPEfluids script in the book module.) We start by extracting the input data from the
deck structure:
pvto = deck.PROPS.PVTO{1};
rsd = pvto.key([1:end end]);
pbp = pvto.data([pvto.pos(1:end-1)' end],1);
Bod = pvto.data([pvto.pos(1:end-1)' end],2);
muOd = pvto.data([pvto.pos(1:end-1)' end],3);

Here, we have only extracted the data points representing saturated states as well as the
undersaturated data for the maximum Rs value, which is the minimal data requirement and
coincides with what is given in the input file in our case. To simplify subsequent interpo-
lation, the input processing in MRST adds an extra data point inside the undersaturated
region for each data point, but we ignore these in our simplified data extraction.
Let us take the SPE 9 model as an example: The input data has eleven data records for Rs
values 0, 0.165, 0.335,. . . ,1.5. The last record also contains formation-volume factor and
viscosity at pressure 5,000 psi inside the undersaturated region. The three pairs of pressure,
Bo , and μo values specified for Rs = 1.5 define values for compressibility and viscosibility.
Assuming that these are constant inside the whole undersaturated region, the input routine
computes one additional value at pressure values p = pbp + 1,000 psi, so that pvto.data
becomes a 22 × 3 array. If we need Rs values for all data points represented in the deck
structure, including extra points that have been added in the undersaturated region, we can
use the following call:
rsd = pvto.key(rldecode((1:numel(pvto.key))',diff(pvto.pos),1));

To evaluate the fluid model, we make a mesh of pressure and dissolved gas states
and then determine whether each state represents a saturated or undersaturated state by
comparing with the interpolated Rs value for the corresponding pressure:
[RsMax,pMax] = deal(max(rsd), max(pbp));
[rs,p] = meshgrid(linspace(10,RsMax-10,M), linspace(0,pMax,N));
Rs = reshape(f.rsSat(p(:)), N, M);

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.5 Phase Behavior in ECLIPSE Input Decks 385

isSat = rs >= Rs;

rs(isSat) = Rs(isSat);
Bo = reshape( f.BO (p(:), rs(:), isSat(:)), N,M);
muO = reshape( f.muO(p(:), rs(:), isSat(:)), N,M);

Figure 11.25 shows plots of Rs as function of pressure as well as plots of Bo and μo

as function of pressure and amount of dissolved gas. For comparison, we have overlain
the plots for SPE 1 and SPE 9, which describe a gas-injection scenario and a case with

Oil formation-volume factor [−] Maximum solution gas−oil ratio [−]

300

1.8
250

1.7
200

1.6
150

1.5
100

1.4
50 SPE 1
SPE 9
1.3
0
0 100 200 300 400 500 600
Pressure [bar]
1.2
Gas formation−volume factor [−]
0
10
1.1
SPE 1
SPE 9
1
0 100 200 300 400 500 600 10
−1

Pressure [bar]

Oil viscosity [cP]

1.8 −2
10

1.6

−3
10
0 100 200 300 400 500 600
1.4 Pressure [bar]

Gas viscosity [cP]

1.2 0.05

0.045

0.04
1
0.035

0.03
0.8
0.025

0.02

0.6 0.015

0.01
SPE 1
0 100 200 300 400 500 600 0.005 SPE 9
Pressure [bar]
0
0 100 200 300 400 500 600
Pressure [bar]

Figure 11.25 Description of phase behavior for the live-oil and dry-gas models in the SPE 1 and
SPE 9 benchmarks, shown as blue and green lines, respectively. In the Bo and μo plots, the red lines
correspond to undersaturated states.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 386 Compressible Multiphase Flow

water injection, respectively. There is a pronounced difference in the phase properties of

the two models: Whereas the SPE 1 oil shows strong compressibility and viscosibility
effects both for the saturated and undersaturated states, the undersaturated SPE 9 oil is
almost incompressible and has constant viscosity. In part, this is also indicated by the
surface densities of oil and gas, which are 786.5 kg/m3 and 0.9698 kg/m3 for SPE 1
and 720.5 kg/m3 and 1.1245 kg/m3 for SPE 9. The SPE 1 model has a larger fraction of
both lighter and heavier hydrocarbon components and therefore exhibits larger variations
in formation-volume factor and viscosity than SPE 9.
For completeness, let us also look at the gaseous and aqueous phases. The behavior
of dry gas is represented by the following functions and constants (here with values for
SPE 9):

bG: @(pg,varargin) ireg(TbG,pg,varargin{:})

muG: @(pg,varargin) ireg(TmuG,pg,varargin{:})
rhoGS: 1.1245

The constant arguments TBG, TbG, and TmuG to the functions are data objects containing
the tabulated formation-volume factor, its reciprocal expansion factor, and the viscosity
from the PVDG keyword, respectively. The function name ireg is a shorthand for the
interpReg function we encountered when discussing computation of three-phase relative
permeabilities on page 353.
The aqueous phase is represented as a weakly compressible fluid (here for SPE 9),

cW: 1.4504e-10
muWr: 9.6000e-04
bW: @(pw,varargin) bW(pw,pvtw,reg,varargin{:})
muW: @(pw,varargin) muW(pw,pvtw,reg,varargin{:})
rhoWS: 1.0093e+03

To be consistent with ECLIPSE, we use a Taylor expansion to approximate the weakly

compressible functions

Bwr
Bw (p) = Bwr e−cw (p−pr ) = Bwr e−ξ ≈ .
1 + ξ + 12 ξ 2

At this point, I encourage you to consult the tutorial fluidInspectionExample from the
ad-core module, which introduces a graphical interface for fluid models from the AD-OO
framework that enables you to interactively produce many of the same plots as shown here
and in Section 11.3. Creating models and launching the GUI only requires three lines of
code:
[G, rock, fluid, deck] = setupSPE9(); % or setupSPE1(), setupSPE3()
spe9 = selectModelFromDeck(G, rock, fluid, deck);
inspectFluidModel(spe9)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.5 Phase Behavior in ECLIPSE Input Decks 387

Example: SPE 3 – Rich Gas With Retrograde Condensation

The SPE 3 benchmark was designed to compare the ability of compositional simulators
to match PVT data for a problem with gas cycling in a rich-gas reservoir with retrograde
condensation. The input deck distributed with MRST is a translation of the problem into a
black-oil framework, with which one can only hope to describe the pertinent flow physics
to a limited extent. The resulting model consists of a dead oil, described with the PVDO
keyword, and a rich gas described by the PVTG keyword. Here, we only show excerpts of
the specification of the rich gas:

-- ’Gas Pressure’ ’Gas OGR’ ’Gas FVF’ ’Gas Visc’

-- Units: psia stb /Mscf rb /Mscf cp
PVTG
1214.7000 0.0013130 2.2799 0.0149
0 2.2815 0.01488/
1814.7000 0.00353 1.4401 0.01791
0.001313 1.4429 0.01782
0 1.4445 0.01735 /
2414.7000 0.01102 1.0438 0.02328
:
:
3449.3322 0.0642 0.7783 0.0403
0.0612 0.7777 0.0395
0.0567 0.7769 0.03892
0.0454 0.7747 0.03748
0.0331 0.7723 0.03594
0.01102 0.7681 0.03325
0.00353 0.7666 0.03236
0.001313 0.7662 0.0321
0 0.7660 0.03194 /
/

The maximum amount of vaporized oil increases as function of pressure, and the input
table uses an increasing number of data points within each record, starting with two points
for the first pressure value and ending with nine values for the highest pressure value.
The data points are quite unevenly scattered in (pg ,Rv ) space, as you can see from the
plot of Rv and μg versus pg in Figure 11.26. You should therefore be wary of potential
numerical inaccuracies in the numerical interpolation. (The showSPEfluids script in the
book module gives the full code used to generate the plots in the figure.)
Looking at the qualitative behavior of the gas properties shown in Figure 11.26, we see
that as pressure increases, more oil can be vaporized into the gas and this will increase
the viscosity of the gaseous phase. The data points along the vertical lines traverse the
phase envelope for fixed pressure, from the blue line corresponding to a saturated state
with full vaporization (i.e., maximal supplied Rv value) to the red line representing the
dew point, at which the vapor phase contains no vaporized oil (Rv = 0). Vaporization
of oil also affects the gas formation-volume factor, but since the amount of vaporized
oil is so small, the resulting variations in volume across the phase envelope for a fixed
pressure are so small that they are only visible when zooming in on the plot. Interestingly,
we observe that Bg decreases with increasing vaporization for pressures below 220 bar, but
increases with increasing vaporization for pressures above 220 bar. However, the density of

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 388 Compressible Multiphase Flow
−4
x 10 Maximum vaporized oil−gas ratio [−]
Gas viscosity [cP]
0.045

0.04
3
0.035

0.03
2

0.025

1 0.02

0.015

0 0.01
80 100 120 140 160 180 200 220 240 80 100 120 140 160 180 200 220 240
Pressure [bar] Pressure [bar]
x 10
−3 Gas formation-volume factor [−] Density of gaseous phase [kg/m3]
13 350
−3
x 10
12
4.7 300
11
4.6

10 4.5 250
4.4
9
4.3 200
8
4.2
205 210 215 220 225 230 235 240

7 150

6
100
5

4 50
80 100 120 140 160 180 200 220 240 80 100 120 140 160 180 200 220 240
Pressure [bar] Pressure [bar]

Figure 11.26 Description of phase behavior for the rich-gas system of the SPE 3 benchmark problem.
Vertical lines represent transition across the phase envelope for fixed pressure, with the blue lines
representing max Rv and red lines Rv = 0.

the gaseous phase increases monotonically with increasing oil vaporization, as one would
expect physically.
The figure does not report properties of the dead oil, but these behave as expected, with
Bo decreasing and μo increasing with increasing pressure.

11.6 The Black-Oil Equations

In this section, we discuss the flow equations for black-oil reservoir model, which we
presented in Section 8.2.3, in more detail. We go through the equations, component by
component, and discuss the various special cases and present how to discretize the various
terms using discrete differentiation and averaging operators. As described in Section 11.4,
the family of black-oil models assume isothermal conditions inside the reservoir, so that
the PVT behavior is modeled as a function of pressure only. The resulting models work
best when operating in single-phase hydrocarbon regions far from the critical point, but
can also be used in two-phase hydrocarbon regions as long as the variation in composition
is relatively small.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.6 The Black-Oil Equations 389

11.6.1 The Water Component

The discussion in this chapter has primarily focused on the gas and oil components con-
tained in the gaseous and oleic hydrocarbon phase. Reservoirs will nonetheless always con-
tain aqueous fluids present as connate water within the hydrocarbon-bearing rock layers.
The reason is that the reservoir rock was fully or partially saturated with water before hydro-
carbons migrated upwards from a deeper source rock. Since most rocks are fully or partially
water-wet, the invading hydrocarbons would not be able to fully displace the resident
water. Oftentimes, hydrocarbon-bearing rocks will have extensive underlying aquifer that
contribute pressure support and fluid inflow that should be accounted for in the reservoir
model.
As a rule, aqueous reservoir fluids are saline. The most common dissolved salt is sodium
chloride, but the aqueous phase typically contains a wide range of compositions and con-
centrations of other inorganic chemical species. These species are not accounted for in basic
black-oil type models, which simply assume that the resident brine and injected water is
a single, chemically inert component that only is present in the aqueous phase. (You can
consult Trangenstein and Bell [295] for an alternative model, in which gas is allowed to
dissolve into the aqueous phase.) The corresponding flow equations read
 
∂t φbw Sw + ∇ · bw vw − bw qw = 0, (11.15)

with Darcy’s law for the phase velocity

Kkrw
vw = − (∇pw − gρw ∇z).
μw
These equations are discretized exactly in the same way as (11.1) on page 342, except
for the obvious modifications resulting from ρw having been replaced by bw . Let us now
look at how you could implement the discrete flow equations in MRST. We have already
discussed most of the tools you need. As before, we assume that the fluid properties
are collected in a fluid object called fluid. For convenience, the ad-core module col-
lects all the discrete differentiation and averaging operators in a special structure, which
we here will refer to by the short-hand s. We start by computing the necessary fluid
properties:
bW = fluid.bW(pO);
rhoW = bW.*fluid.rhoWS;
rhoWf = s.faceAvg(rhoW);
mobW = krW./fluid.muW(pO);
pvMultR = fluid.pvMultR(pO);

Here, faceAvg is the same as the avg operator we have used previously, but has been
renamed to specify explicitly that this is an averaging operator that should be used to
compute average values at cell interfaces. Notice that we evaluate the pressure-dependent
properties using oil pressure and not water pressure. In general, there are several ways you
can choose the primary unknowns in the black-oil models. Herein, we have chosen to follow

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 390 Compressible Multiphase Flow

the conventions from ECLIPSE and use oil pressure po and gas and water saturations, Sw
and Sg , as our primary unknowns. (We will get back to the choice of primary variables later
in this section.)
The water flux across cell faces will, as before, be computed using single-point
upstream-mobility weighting. To this end, we must first compute the water pressure in
each grid cell,
pW = pO
if isfield(fluid, 'pcOW') && isempty(sW)
pW = pW - fluid.pcOW(sW);
end

Then, we can compute the water flux from Darcy’s law

dpW = s.Grad(pW) - rhoWf.*gdz;
upcw = (double(dpW)<=0);
vW = -s.faceUpstr(upcw, mobW).*T.*dpW;
bWvW = s.faceUpstr(upcw, bW).*vW;

The function faceUpstr implements the upstream operator we previously denoted by upw.
With this, we have all quantities we need to compute the homogeneous water equation on
residual form
water = (s.pv/dt).*( pvMult.*bW.*sW - pvMult0.*bW0.*sW0 ) + s.Div(bWvW);

Here, quantities ending with 0 refer to variables evaluated at the beginning of the time
step and the vector s.pv contains pore volumes in each cell at reference pressure. The
code lines in this subsection are in essence what is implemented in the flow solvers of the
ad-blackoil module. However, if you look at the code in e.g., equationsBlackOil, you
will see that it is more complicated since we in general need to also include (multisegment)
well models, source terms, and boundary conditions. Likewise, we must account for various
user-defined multipliers that modify transmissibilities and mobilities, enable solution of
adjoint equations for computing sensitivities, and store computed results; more about this
in the next chapter.

11.6.2 The Oil Component

In the basic black-oil model, the oil component is only present in the liquid oleic phase,
and we hence get the following flow equation:
 
∂t φbo So + ∇ · bo vo − bo qo = 0, (11.16)

with Darcy’s law for the phase velocity

Kkrw
vo = − (∇po − gρo ∇z).
μo

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.6 The Black-Oil Equations 391

As for the water component, we start by computing the necessary fluid properties. How
we compute density, shrinkage factor, and viscosity depends upon whether the oil phase
contains dissolved gas or not. If the system contains no dissolved gas, the parameters only
depend on oil pressure and can be computed as follows:
bO = fluid.bO(pO);
if isfield(fluid, 'BOxmuO')
muO = fluid.BOxmuO(pO).*bO;
else
muO = fluid.muO(pO);
end
rhoO = bO.*fluid.rhoOS;

As pointed out on page 384, the convention in ECLIPSE 100 input is that tabulated
formation-volume factors and viscosity for oil should not be interpolated independently,
but instead be interpolated as their reciprocal product, 1/(μo Bo ). MRST signifies input data
that require this particular interpolation by a special function BOxmuO in the fluid object. If
this function is not present, the viscosity and the shrinkage factors are computed indepen-
dently. If oil also contains dissolved gas, the same properties are evaluated as follows:
bO = fluid.bO(pO, rs, isSat);
muO = fluid.muO(pO, rs, isSat);
rhoO = bO.*(rs*fluid.rhoGS + fluid.rhoOS);

Here, rs is a vector holding the gas–oil ratio and isSat is a Boolean vector that tells
whether each element of pO and rs represents a saturated or an undersaturated state. We
will get back to how rs is computed in Section 11.6.4. With density, shrinkage factor, and
viscosity given, we can proceed to compute the oleic phase flux bOvO using upstream-
mobility weighting as described for the aqueous phase in Section 11.6.1. Altogether, this
gives us the homogeneous residual equation
oil = (s.pv/dt).*( pvMult.*bO.*sO - pvMult0.*bO0.*sO0 ) + s.Div(bOvO);

If oil is allowed to vaporize into the gaseous phase, the conservation equation for the oil
component reads
    
∂t φ bo So + bg Rv Sg + ∇ · bo vo + bg Rv vg − bo qo + bg Rv qg = 0. (11.17)
We therefore need to compute the properties and the phase flux for the gaseous phase before
we can compute the homogeneous residual oil equation:
oil = (s.pv/dt).*( pvMult.* (bO.* sO + rv.* bG.* sG) - ...
pvMult0.*(bO0.*sO0 + rv0.*bG0.*sG0) ) + s.Div(bOvO + rvbGvG);

Notice, in particular, that we use single-point upstream-mobility weighting to evaluate the

vaporized oil–gas ratio in the term representing flux of oil vaporized in the gas phase
rvbGvG. This is completely analogous to how we evaluate the shrinkage factors for oil
and water at the interface between two neighboring cells.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 392 Compressible Multiphase Flow

11.6.3 The Gas Component

The gas component can generally be present both in the gaseous and the liquid oleic phases.
The corresponding conservation equation reads
    
∂t φ bg Sg + bo Rs So + ∇ · bg vg + bo Rs vo − bg qg + bo Rs qo = 0 (11.18)
with Darcy’s law for the phase velocity
Kkrg
vg = − (∇pg − gρg ∇z).
μg
Like for oil, the computation of the gas properties differs for systems with and without
vaporized oil:
if vapoil
bG = fluid.bG(pO, rv, isSat);
muG = fluid.muG(pO, rv, isSat);
rhoG = bG.*(rv*fluid.rhoOS + fluid.rhoGS);
else
bG = fluid.bG(pO);
muG = fluid.muG(pO);
rhoG = bG.fluid.rhoGS;
end

We also need to compute the phase pressure to determine the phase velocity:
pG = pO;
if isfield(fluid, 'pcOG') && isempty(sG)
pG = pG + fluid.pcOG(sG);
end

With these quantities given, we can proceed as above to compute the remaining terms nec-
essary to evaluate the accumulation and flux terms in the homogeneous residual equation
gas = (s.pv/dt).*( pvMult.* (bG.* sG + rs.* bO.* sO) - ...
pvMult0.*(bG0.*sG0 + rs0.*bO0.*sO0 ) ) + s.Div(bGvG + rsbOvO);

11.6.4 Appearance and Disappearance of Phases

With all the details presented thus far, it should be a surmountable task to extend the
simple two-phase, dead-oil implementation discussed in Section 11.2 to a similar simplified
simulator for black-oil models. In particular, dissolved gas in the liquid oleic phase and/or
vaporized oil in the gaseous phase is not too difficult to include as long as each hydrocarbon
phase either stays in a saturated state or an unsaturated state throughout the whole simula-
tion. For saturated phases, it is common to use one phase pressure (po in MRST) and the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.6 The Black-Oil Equations 393

water and gas saturation (Sw and Sg ) as independent variables since both the solution gas–
oil ratio Rs and the vaporized oil–gas ratio Rv can be computed as a function of pressure.
If oil is in an undersaturated state, there is no free gas available and the flow problem
is essentially a two-phase flow problem (assuming that water is present). However, such
cases can also be simulated as a three-phase problem if we set Rs as a primary variable.
Likewise, we can choose Rv as a primary variable for cases with undersaturated gas.
Somewhat more intricate details are needed when the hydrocarbon phases switch
between saturated and undersaturated states so that phases may appear or disappear during
the simulation. As you have seen already, MRST introduces the solution gas–oil ratio
rs as an extra variable alongside Sg for cases with dissolved gas. In addition, we need
a status flag in each cell that tells which of the two is the primary unknown. When oil
is in a saturated state, Sg is the primary variable and rs is a secondary variable that can
be determined from the pressure. If the computed gas saturation in a cell is zero or falls
below, we know that the cell has switched from a saturated to an undersaturated state. In
cells containing undersaturated oil, there is no free gas available and we solve for rs as a
primary variable to track the amount of gas flowing in and out of each cell with the oil.
Once a new rs value is computed, we need to check whether rs>fluid.rsSat(pO),
which means that we have a transition from undersaturated to saturated state and hence
need to increase Sg to the correct amount of liberated gas.
Switching of variables may unfortunately introduce instabilities in the nonlinear solution
process, and in practice we do not allow states to cross from a saturated/undersaturated state
and well into the undersaturated/saturated region. We rather stop the update at the interface,
i.e., set rs to rsMax=rsSat(p0) and Sg=0, and chop the update to improve robustness of
the Newton iteration loop. Cases with transitions between saturated and undersaturated gas
are treated analogously. The following code summarizes the logic just explained to choose
the primary variable x representing gas:
% Assumption: rs = rsSat for saturated cells, likewise for rv
% status 1 (oil, no gas): x = rs, sg = 0, rv = rvMax
% status 2 (gas, no oil): x = rv, sg = 1-sw, rs = rsMax
% status 3 (oil and gas): x = sg, rs = rsMax, rv = rvMax
watOnly = sW > 1- sqrt(eps);

if vapoil, oilPresent = true; else, oilPresent = or(sO > 0, watOnly); end

if disgas, gasPresent = true; else, gasPresent = or(sG > 0, watOnly); end

status = oilPresent + 2*gasPresent;

if disgas, st1 = false; else, st1 = status==1; end
if vapoil, st2 = false; else, st2 = status==2; end
st3 = status == 3;

x = st1.*rs + st2.*rv + st3.*sG;

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 394 Compressible Multiphase Flow

11.7 Well Models

Earlier in the book, we have discussed well models for single-phase flow. Section 4.3.2
introduced the basic inflow-performance relationship, which states that the flow rate in or
out from a well is proportional to the pressure difference between the wellbore and the
reservoir, q = J (pR − pwb ). This section describes well models for multiphase flow in
more detail. In particular, we discuss wells with multiple connections and explain how to
handle cross flow and introduce a more comprehensive family of models that use a network
of 1D flow models to describe flow in horizontal and multilateral wells.

11.7.1 Inflow-Performance Relationships

The inflow-performance relationship for black-oil models is written in terms of volumetric
production rates for each fluid phase α at stock-tank conditions

qα,i = Ji λα,i pi − pbh − Hwi . (11.19)

Here, Ji is the productivity index (sometimes called connection transmissibility factor or

R-factor) for connection number i, λα,i is phase mobility at the connection, pi is pressure
in the perforated grid block, pbh is bottom-hole pressure, and Hwi is pressure drop from
the well’s datum point to connection i.
The standard Peaceman-type expression for the productivity index of a vertical well
penetrating a Cartesian grid block reads
θ Kh
J = .
ln(re /rw ) + S
Here, Kh is effective permeability times net thickness of the connection, rw is wellbore
radius, and re is the equivalent radius. We have previously discussed expressions for K and
re for various special cases in Section 4.3.2. The parameter θ gives the angle the wellbore
is connected to the reservoir block, with θ = 2π when the well is perforated in the center
of the cell, θ = π when the well is perforated along a cell face, and θ = π/2 when
the well is perforated along a corner edge; see Figure 11.27. The skin factor S accounts
for the difference between ideal pressure drawdown predicted by Darcy’s law and actual
drawdown resulting from formation changes in the near-well region; S also includes the
constant 3/4 from pseudo-steady flow. The typical range of skin factors is −6 < S <

θ = 2π θ=π θ = π/2

Figure 11.27 The parameter θ measures the angular section of the wellbore that is connected to the
reservoir and depends on where the well is placed inside a grid block.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.7 Well Models 395

Figure 11.28 Illustration of crossflow in a well; blue and green colors represent different fluid phases.

100. Negative values represent wells stimulated by hydraulic fracturing or other means to
improve the permeability in the near-well zone, whereas large positive values represent
wells suffering from severe formation damage caused, e.g., by various forms of clogging
induced by drilling fluids, fines migration, precipitation of soluble salts, etc.
The mobility λα,i is computed differently for injection and production wells. For pro-
ducing connections, the mobility of a free phase of gas, oil, or water, is given as
(bα krα )i
λα,i = , (11.20)
μα,i
where the relative permeability krα , the shrinkage/expansion factor bα , and the phase vis-
cosity μα are computed using pressure and saturation values from the grid block perforated
by connection number i. If the produced gaseous or oleic phases contain vaporized oil or
dissolved gas, respectively, the mobilities account for this as follows
λO = λo + Rv λg , λG = λg + Rs λo . (11.21)
For injecting connections, MRST computes the mobility as the sum over all phases
 krβ
λα,i = yα bα . (11.22)
μβ i
β

where yα is the volume fraction of phase α in the mixture inside the wellbore. This means
that the injectivity is a function of the total mobility of the connected grid block.
Wells connected to multiple grid blocks may sometimes experience crossflow as illus-
trated in Figure 11.28. This means that fluids flow into some connections and out of other
connections. Injectivities must therefore be modified so that they represent the mixture
of the fluids inside the wellbore. To this end, we first compute inflow from producing
connections at reference (surface/stock-tank) conditions and use this to compute the aver-
age wellbore mixture at reference conditions. Then, we compute the average volumetric
mixture at the conditions of each injection connection and use this to define the correct
mobility. A similar average wellbore mixture is used to calculate the pressure drop between
the well’s datum point and each connection.

11.7.2 Multisegment Wells

Many wells are designed so that the reservoir fluids do not flow directly from the sandface
and into the wellbore. Oftentimes, the wellbore (production casing) lies inside another cas-
ing that has been cemented in place inside the drilled hole. The void annular space between

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 396 Compressible Multiphase Flow

Open hole Screen Casing Annulus ICD

Figure 11.29 Conceptual illustration of different types of well completions.

the casing and the wellbore is limited by sealing packers, which limit the annular space
into sections, and sleeves that provide flow paths between the annulus and the wellbore.
One can also have various types of screens that aim to reduce/eliminate the inflow of solid
particles that would cause erosion inside the wellbore. Simple versions of sand screens are
made up of ribs, rods, or wire filters, possibly filled with gravel, whereas more modern
ones consist of high-permeable, resin-coated gravel packed between two layers of wire-
wrap filter media (gravel pack). In addition, wells may have various components that can
partially or fully choke the inflow to prevent unwanted fluids to enter the wellbore. Inflow
control devices (ICDs) are passive or autonomous devices installed along the completion
interval to optimize inflow along the length of the wellbore. These devices are designed to
exert higher back pressure at high flow rates and can be used to equalize inflow from high
and low-permeable zones, or delay water or gas breakthrough across selected intervals of
horizontal wells. Inflow control valves serve the same purpose and are controlled from the
surface through electric or hydraulic signals.
To further increase complexity, modern wells typically have directional profiles
designed to intersect multiple targets with a single wellbore. Such wells can have multiple
production zones, or consist of multiple deviated or horizontal branches (multilaterals).
To increase the productivity of a well, it is common to install downhole pumps that
lower the wellbore flowing pressure so that fluids can be drawn from the reservoir at a
higher rate. If the downhole pressure is too low, it may be necessary to apply artificial
gas lift, in which gas is injected downhole to lower the density of the flowing fluids
and use gas expansion to help the produced fluids to the surface. Most oil wells require
artificial lift at some point to ensure that they keep flowing when their bottom-hole pressure
decreases. Wells may also incorporate equipment for downhole compression and separation
of fluid phases, which may be used to separate gas (or water) so that it can be reinjected
downhole. Downhole processing can also be introduced to reduce the size and weight
of surface or seafloor facilities or to overcome bottlenecks in existing surface processing
systems.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.7 Well Models 397

All these different components will induce pressure drawdown and/or changes in the
composition of the flowing fluids. To model this complex behavior, one may use a so-called
multisegment well model, which can be thought of as a flow network consisting of a set of
nodes connected by a set of segments (edges). Each node n has an associated volume and
can store fluids, and fluids can flow along each segment connecting two nodes as described
by a set of one-dimensional flow equations.

Continuous Flow Equations

The one-dimensional flow equations can either be formulated using phase velocities vα and
volume fractions yα , or component mixture velocities vcm and mass fractions xc for the gas,
oil, and water components. Capillary pressure is negligible inside the wellbore, hence we
can write the conservation equations for volumetric quantities as follows:
 
∂t bw yw + ∂x bw vw − bw qw = 0,
 
∂t bo yo + bg Rv yg + ∂x bo vo + bg Rv vg − bo qo − bg Rv qg = 0, (11.23)
 
∂t bg yg + bo Rs yo + ∂x bg vg + bo Rs vo − bg qg − bo Rs qo = 0.
The phase velocities do not relate to pressure gradients through Darcy’s law, but rather
through some nonlinear relation modeling friction, acceleration, inflow devices, valves,
and so on. We can write this relationship as

∇p − gρ∇z − G(v,ρ,μ) = 0, (11.24)

where ρ, v, and μ are the mixture density, velocity, and viscosity, respectively, and G
represents the (nonlinear) part of the pressure drop that cannot be explained by a Darcy
relationship. The equivalent set of equations on mass form reads,

∂t xc ρ + ∇ · vcm − qcm = 0, c ∈ {g,o,w}. (11.25)

Here, qcm is the mass source term of component c.

Discrete Well Equations

To define the discrete multisegment well models, we utilize a vector notation analogous
to the discrete operators we introduced in Section 4.4.2 for flow equations in the porous
medium. Where we previously had two mappings F between cells and their bounding
cell faces and C1,C2 between faces and adjacent cells, we now have similar mappings S
from nodes to all connected segments and N1,N2 from segments to the adjacent nodes; see
Figure 11.30. Given these mappings, the discrete gradient operator reads

grad(v)[s] = p[N2 (s)] − p[N1 (s)],

and similarly the discrete divergence operator reads

 
div(v)[n] = v[s] 1{n=N1 (s)} − v[s] 1{n=N2 (s)} .
s∈S(n) s∈S(n)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 398 Compressible Multiphase Flow

Figure 11.30 Example of well topology and corresponding discrete operators.

These operators can both be represented as sparse matrices, almost like the D matrix.
However, the top segment of a well needs special treatment, since the top node corresponds
to the reference depth at which the well’s bottom-hole pressure is defined. Hence, instead
of calculating a pressure drop across the top segment, we impose a boundary condition on
the top node corresponding to the correct control of the well. This means that we write
the operators as grad(p) = Mp and div(v) = −N Tv, where M equals N except that
the first row is removed. Discrete averaging operators are defined analogously. Given these
operators, the discrete flow equations are defined as follows: If we use the mass form, the
conservation equation in each node reads
V  
x c ρ − (x c ρ)0 + div(v m
c ) − q c = 0,
m
(11.26)
t
where superscript 0 refers to values at the previous time step. The equation for the pressure
drop reads,

grad(p) − g upw(ρ)grad(z) − G v m,upw(ρ),upw(μ) = 0. (11.27)

Here, grad(z) is the discrete gradient of the node depth, i.e., the depth difference between
neighboring nodes. Note that (11.27) is not defined for the topmost segment and we need
to specify a boundary or control equation at the top to close the coupled system (11.26)–
(11.27).

Pressure-Drop Relations
The function G in (11.24) can be used to impose almost any pressure drop. For general
network topologies, however, we assume that ∂G/∂v m is nonzero. As an example, the
pressure drop over a simple valve can be modeled as
ρu2c
p = , (11.28)
2Cv2
where uc is the velocity in the constriction and Cv is the discharge coefficient. Another
example is the frictional pressure drop along the wellbore, which may have a significant

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.8 Black-Oil Simulation with MRST 399

effect on the behavior of (long) horizontal wells. In ECLIPSE 100 [271], this pressure loss
is given by
2f L 2
p =
ρu . (11.29)
D
Here, f is the Fanning friction factor, L is the length of the segment, D is the corresponding
tubing diameter, ρ is the mixture density, and u is the mixture velocity. For a mixture mass
flux v m or mixture volume flux v, the velocity u is given by
vm v
u= = . (11.30)
ρπ(D/2)2 π(D/2)2
For annular pressure drop, (D/2)2 is replaced by (Do /2)2 − (Di /2)2 , where Do and Di
are the diameters of the outer and inner tubing, respectively. The Fanning friction factor
depends on the Reynolds number Re = ρuD/μ and is calculated as
.  6.9  e 10/9 
f −1 = −3.6 log10 + , (11.31)
Re 3.7D
for Re > 4,000, and f = 16/Re in the laminar region for Re < 2,000. In between, one
interpolates between the corresponding end values for Re equal 2,000 and 4,000. MRST
uses the same approach, but also allows use of (11.31) for all velocities. The coefficient e
gives roughness of the tubing.

11.8 Black-Oil Simulation with MRST

Altogether, we have discussed a great number of details that must be brought together to
make a simulator. In addition, there are still a number of other important topics like pre-
conditioning and details of boundary conditions, time chopping, nonlinear solution strategy,
and so on, that I have not discussed in detail. As stated in the introduction of the chapter,
I consider many of these details to be outside the scope of the book and I hence feel that
time is right to switch the presentation to the AD-OO framework, which implements all the
necessary ingredients entering industry-standard reservoir simulators. I therefore end the
chapter by first briefly outlining how you can use the AD-OO framework to simulate the
SPE 1 benchmark model and then discuss some limitations and potential pitfalls that you,
as a user of MRST and other reservoir simulators, should be aware of. The next chapter will
give a more systematic overview of the AD-OO framework and discuss more simulation
examples.

11.8.1 Simulating the SPE 1 Benchmark Case

You have already encountered the fluid model from the SPE 1 benchmark [241] multiple
times in this chapter. The benchmark describes a problem with gas injection in a small
10 × 10 × 3 reservoir with a producer and an injector placed in diagonally opposite corners.
The porosity is uniform and equals 0.3, whereas the permeability is isotropic with values

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 400 Compressible Multiphase Flow
25

20

Time step [days]

15

10

0
0 200 400 600 800 1000 1200
Time [days]

Figure 11.31 Horizontal permeability and well positions for the SPE 1 benchmark case. The injector
is perforated in the upper and the producer in the lower layer. The right plot shows the time steps in
the simulation schedule.

500, 50, and 200 md in the three layers with thickness 20, 30, and 50 ft; see Figure 11.31.
The reservoir is initially undersaturated with constant pressure in each layer and a uniform
mixture of water (Sw = 0.12) and oil (So = 0.88) with no initial free gas (Sg = 0.0) and a
constant solution gas–oil ratio (Rs ≈ 226.2) throughout the model. Although the problem
is really a two-phase gas–oil problem describing gas injection into an undersaturated oil,
the input file describes a three-phase problem with almost stationary water. The original
problem was posed to study ten years of production, but herein we only report and compare
solutions for the first 1,216 days.
We have already discussed the steps to load the ECLIPSE input data and create the
necessary objects representing grid, petrophysical and fluid properties in detail earlier in
the book. For completeness, however, we include them all here:
gravity reset on
deck = readEclipseDeck(fullfile(getDatasetPath('SPE1'), 'BENCH_SPE1.DATA'));
deck = convertDeckUnits(deck);
G = computeGeometry(initEclipseGrid(deck));
rock = compressRock(initEclipseRock(deck), G.cells.indexMap);
fluid = initDeckADIFluid(deck);

In addition, we must set up an initial state. The SOLUTION section in the input file specifies
how the reservoir should be initialized. Here, we hard-code the fluid equilibrium state for
simplicity:
[k, k] = ind2sub([prod(G.cartDims(1:2)), G.cartDims(3)], G.cells.indexMap);
p0 = [329.7832774859256 ; ... % Top layer
330.2313357125603 ; ... % Middle layer
330.9483500720813 ]; % Bottom layer
p0 = convertFrom(p0(k), barsa);
s0 = repmat([ 0.12, 0.88, 0.0 ], [G.cells.num, 1]);
rs0 = repmat( 226.1966570852417 , [G.cells.num, 1]);
rv0 = 0; % dry gas

state = struct('s', s0, 'rs', rs0, 'rv', rv0, 'pressure', p0);

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.8 Black-Oil Simulation with MRST 401

The first statement creates a partition vector containing the layer index, which we use to
lookup in the vector of equilibrium pressures values. All the statements just presented are
collected in a utility function setupSPE1 in the ad-blackoil module.
The next step is to create an object representing the whole simulation model with
description of reservoir geometry, petrophysical and fluid parameters, drive mechanisms,
and parameters determining the correct flow equations
model = selectModelFromDeck(G, rock, fluid, deck);

Technically, the model object is an instance of a class defined using classdef and
includes both properties and methods that can operate on input parameters and the data
contained within the object. Classes can also define events and enumerations, but these
are not used in AD-OO. If you are not familiar with the concept of class objects, you
can think of model as a standard struct containing both data members and function
pointers. In this particular case, the input data will create a three-phase, black-oil model
with dissolved gas but no vaporization of oil. The following is a list of a few of the data
members:

ThreePhaseBlackOilModel with properties:

disgas: 1
vapoil: 0
fluid: [1x1 struct]
rock: [1x1 struct]
water: 1
gas: 1
oil: 1
gravity: [0 0 9.8066]
FacilityModel: [1x1 FacilityModel]
operators: [1x1 struct]
G: [1x1 struct]

The fields G, rock, and fluid are the standard MRST structures we have used earlier
to represent grid, petrophysics, and fluid properties. The operators structure contains
transmissibilities, pore volume, and discrete operators and is the same structure we referred
to as s in the previous section. FacilityModel is an instance of another class designed to
describe models of wells and surface facilities. Last but not least, the object contains five
flags telling which of the three phases are present and whether gas is allowed to dissolve in
oil and oil is allowed to vaporize into gas.
We have already encountered the flag vapoil in the code that computed gas prop-
erties on page 392. Likewise, disgas and vapoil were used to determine the correct
primary unknown for the gas phase on page 393. These, and all the other code pieces
for computing flow equations on residual form presented in Section 11.6, are excerpts
from the function equationsBlackOil and a number of helper functions and have been
simplified and reorganized for pedagogical purposes. The AD-OO framework never calls
equationsBlackOil directly, but instead accesses it indirectly through a generic member
function called getEquations. This way, the parts of the framework that implement the
time loop and the nonlinear solver does not need to know any specifics of the model as
long as we know that the model exists and that calling getEquations will compute the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 402 Compressible Multiphase Flow

corresponding flow equations on residual form with automatic differentiation activated for
all primary unknowns. In addition to the parameters just outlined, the model contains a
number of other parameters that control the nonlinear solver, chopping of time steps, and
output of results. More details about different model classes, time loops, nonlinear and
linear solvers, are given in Chapter 12.
The next thing we must do is create a simulation schedule, which essentially consists of
a set of time steps and a specification of the drive mechanisms that will be present during
each of these time steps. These are described in the SCHEDULE section of the input deck,
which for SPE 1 declares two different wells:
-- WELL SPECIFICATION DATA -- COMPLETION SPECIFICATION DATA
-- --
-- WELL GROUP LOCATION BHP PI -- WELL -LOCATION- OPEN/ SAT CONN WELL
-- NAME NAME I J DEPTH DEFN -- NAME I J K1 K2 SHUT TAB FACT DIAM
WELSPECS COMPDAT
’P’ ’G’ 10 10 8400 ’OIL’ / ’P’ 10 10 3 3 ’OPEN’ 0 -1 0.5 /
’I’ ’G’ 1 1 8335 ’GAS’ / ’I’ 1 1 1 1 ’OPEN’ 1 -1 0.5 /
/ /

The WELSPECS keyword shown to the left introduces the wells, defines their name, the
position of the well head, the reference depth for the bottom holes, and the preferred phase.
The keyword COMPDAT shown to the right, specifies how each well is completed, i.e., in
which cells each well can inject/produce fluids. The next we need to define is how the
wells are controlled during the simulation:

-- PRODUCTION WELL CONTROLS

-- WELL OPEN/ CNTL OIL WATER GAS LIQU RES BHP
-- NAME SHUT MODE RATE RATE RATE RATE RATE
WCONPROD
’P’ ’OPEN’ ’ORAT’ 20000 4* 1000 /
/

-- INJECTION WELL CONTROLS

-- WELL INJ OPEN/ CNTL FLOW
-- NAME TYPE SHUT MODE RATE
WCONINJE
’I’ ’GAS’ ’OPEN’ ’RATE’ 100000 100000 50000/
/

The specification says that the producer should be controlled by oil rate, whereas the
injector should inject gas and be controlled by rate. To turn this input into a data object
describing the simulation schedule in a form that MRST understands, we issue the follow-
ing command:
schedule = convertDeckScheduleToMRST(model, deck);

which parses the deck structure and creates the following data structure:

schedule = schedule.step = schedule.control =

step:[1x1 struct] control:[120x1 double] W:[2x1 struct]
control:[1x1 struct] val:[120x1 double] bc:[]
src:[]

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.8 Black-Oil Simulation with MRST 403

The schedule has two data members: a list of steps and a list of controls. The controls
specify a set of targeted2 drive mechanisms (wells, boundary conditions, and source terms)
that each is unique and will be applied during at least one step. Each step consists of a
time-step length (val) and the number (control) of the corresponding control. In our
case, there is only one control consisting of the injection and production wells described
previously (see Figure 11.31), which will be active during each of the 120 time steps. The
control steps are given by the keyword TSTEP, which can be repeated to specify the full
schedule (see Figure 11.31):
TSTEP -- INITIAL STEP: 1 DAY, MAXIMUM: 6 MONTHS
1.0 2*2.0 2*5.0 5*10.0 11*25.0 TUNING
/ 1 182.5 /
TSTEP 1.0 0.5 1.0E-6 /
25.0 /
/

The TUNING keyword gives parameters for time-step control, time truncation and con-
vergence control (here: mass-balance error ≤10−6 ), and control of Newton and linear
iterations (not given here). This keyword is ignored by MRST.
To run the schedule, we must select a nonlinear solver, which is implemented as a class
in the AD-OO framework:
nls = NonLinearSolver('useLinesearch', true);

The NonLinearSolver class implements an abstract framework for nonlinear solvers,

which uses the Newton linearization (or some other mechanism like a fix-point step) pro-
vided by the model to iterate towards a solution. It is capable of time-step selection and
chopping based on convergence rates and can be extended via subclassing with alter-
native linear solvers and time-step classes. Convergence is handled by the model. The
NonLinearSolver simply responds based on what the model reports in terms of con-
vergence to ensure some level of encapsulation. In the current example, we initialize the
solver class to use line-search, as discussed in Section 10.2.2. The excerpts from the nls
object
NonLinearSolver with properties:
maxIterations: 25
maxTimestepCuts: 6
LinearSolver: [1x1 BackslashSolverAD]
timeStepSelector: [1x1 SimpleTimeStepSelector]
useLinesearch: 1
verbose: 0
:

show that the nonlinear solver will at most use 25 iterations, be allowed to cut the time
step six times before giving up. The solver will use the BackslashSolverAD linear solver
and SimpleTimeStepSelector to select time steps, use line-search as specified, run in
non-verbose mode so that no details about convergence are reported to screen, and so on.

2 This is an important distinction: the schedule may for instance request that a given well is operated according to a prescribed
rate (rate control). However, if this rate cannot be achieved without violating physical limits on the bottom-hole pressure, the
simulation will switch to using a bottom-hole control instead.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 404 Compressible Multiphase Flow

Figure 11.32 Graphical user interfaces for plotting reservoir data and well responses. The left plot
shows the saturation of free gas after 20 time steps. The right plot shows the variation in bottom-hole
pressure in the injector and producer throughout the simulation.

We are now ready to run the full simulation. We specify that we wish output of well
solutions, reservoir states, and a report with details about time steps, convergence, etc.:
[wellSols, states, report] = ...
simulateScheduleAD(state, model, schedule, 'nonlinearsolver', nls);

Solving timestep 001/120: -> 1 Day

Solving timestep 002/120: 1 Day -> 3 Days
: : :
Solving timestep 120/120: 3 Years, 111 Days -> 3 Years, 121 Days
*** Simulation complete. Solved 120 control steps in 103 Seconds .. ***

To inspect the solutions, we issue the following commands, which will bring up the two
graphical user interfaces shown in Figure 11.32. First, the plot of well responses
plotWellSols(wellSols, report.ReservoirTime)

and then the plot of reservoir states

figure;
plotToolbar(G, states)
plotWell(G, schedule.control(1).W,'radius',.5)
axis tight, view(-10, 60)

Let us study the dynamics of the reservoir in some detail. Figure 11.33 shows how
the bottom-hole pressure in the injector and the solution gas–oil ratio inside the reservoir
evolve as functions of time. Before production starts, the reservoir is in an undersaturated
state, shown as a green marker in the plots of rs as function of pO. Initially, the injector
needs a relatively high bottom-hole pressure (bhp) to force the prescribed gas rate into the
reservoir. Because the reservoir is undersaturated, the injected gas will dissolve in the oil
phase, so that the viscosity decreases. A lower bhp is thus required to maintain the injection
rate. As the injection continues, cells near the well will gradually become saturated so that
free gas is liberated and starts accumulating at the top of the reservoir. Similarly, some free

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.8 Black-Oil Simulation with MRST 405

Figure 11.33 Reservoir dynamics of the SPE 1 case during the first three years. The lower-right plot
shows the bottom-hole pressure of the injector as function of time. The other plots show the solution
gas–oil ratio as function of pressure for all cells in the reservoir at four different instances in time.

gas will be liberated near the producer because of the pressure drawdown. As more gas
is injected into the reservoir, more cells transition from an undersaturated to a saturated
state, giving an increasing plume of free gas that expands out from the injector along the
top of the reservoir (see the left plot in Figure 11.32). The total injection rate is higher
than the production rate, and hence the pressure in the reservoir will increase, and after
approximately 200 days, we see that the injector bhp starts to increase again to maintain
the specified rate.

11.8.2 Comparison against a Commercial Simulator

To build confidence in our simulator, we can verify the simulation in the previous subsec-
tion against ECLIPSE. To compare the two, we first extract surface rates and bottom-hole
pressures from MRST as plain arrays:
[qWs, qOs, qGs, bhp] = wellSolToVector(wellSols);

We also extract time steps and determine indices for injector and producer:
T = convertTo(cumsum(schedule.step.val), year);
inj = find([wellSols{1}.sign] == 1);
prod = find([wellSols{1}.sign] == -1);

Output from ECLIPSE is usually stored in so-called summary/restart files, which can be
used to reinitialize a simulation so that it can be continued. The deckformat module in
MRST offers useful functionality for reading and interpreting such (binary) output files (see

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 406 Compressible Multiphase Flow
Gas rate / Oil rate Bottom-hole pressure (producer )
260 320

255 300
MRST
250 MRST 280
Eclipse
245 Eclipse 260

240 240

bar
235 220

230 200
225 180
220 160
215 140
0 0.5 1 1.5 2 2.5 3 3.5 0 0.5 1 1.5 2 2.5 3 3.5
Time [Years] Time [Years]

Figure 11.34 Comparison of well responses for the SPE 1 benchmark simulated by MRST and
ECLIPSE 100.

readEclipseSummaryFmt and readEclipseSummaryUnFmt). To simplify the compari-

son for the SPE 1 benchmark, the data set supplied with MRST includes a *.mat file with
selected well responses obtained by simulating the exact same input deck with ECLIPSE
100. These data are stored as a MATLAB structure, which also contains a few functions for
extracting the raw data. We load the summary
load SPE1_smry
ind = 2:size(smry.data,2);
Te = smry.get(':+:+:+:+', 'YEARS', ind);

Here, we have ignored the first entry, which contains zero values. Figure 11.34 shows a
comparison of producer gas–oil ratio and bottom-hole pressure in the producer. These are
extracted as follows:
eGOR = convertFrom(smry.get('PRODUCER', 'WGOR', ind), 1000*ft^3/stb)';
eBHPp = convertFrom(smry.get('PRODUCER', 'WBHP', ind), psia)';
eBHPi = convertFrom(smry.get('INJECTOR', 'WBHP', ind), psia)';

mGOR = qGs(:,prod)./qOs(:,prod);
mBHPp = bhp(:,prod);
mBHPi = bhp(:,inj);

As we see from the figure, there is (almost) perfect match between MRST and ECLIPSE,
which is what we should expect, given that the two simulators implement (almost) exactly
the same numerical methods.

11.8.3 Limitations and Potential Pitfalls

Numerical flow models based on the black-oil equations and their like are complex crea-
tures and should be treated cautiously to ensure that the results they produce make sense.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.8 Black-Oil Simulation with MRST 407

A commercial simulator usually has a large number of precautions built into it, so that it
on one hand will refuse to simulate unphysical situations, but on the other hand manages
to give an answer even though the user gives incomplete or inconsistent data. (You may
also recall our discussion of time-step chops implemented to ensure convergence when
faced with severe nonlinearity, which potentially could be the result of an inadequate model
or incorrect parameters.) Being aware of limitations and potential pitfalls is important, in
particular when using a prototyping tool like MRST that has very few precautions built into
it, but also when using mature commercial tools, whose alluring robustness may sometimes
be deceptive. In this section, I will try to review a few potential issues and comment on
possible precautions.

When Can Black-Oil Models be Applied?

Any model is no better than its assumptions, and it is important to know what the limitations
of a model are so that it is not used for the wrong purposes. Black-oil models can, in
principle, be applied to simulate all the five types of reservoirs outlined in Section 11.4.2,
as long as these are produced by depletion or waterflooding. As a general rule of thumb, you
should be more precautious when applying black-oil models to gas injection processes that
involve significant changes in the composition of the oil and gas pseudo-components during
the displacement. Vaporization and retrograde condensation in rich gases are examples of
such processes, and here you should be careful to check the validity of black-oil models
against more comprehensive compositional models. To ensure that the black-oil description
is sufficiently accurate, the recovery process should satisfy the following characteristics
[269]:

• The recovery process should be isothermal.

• The path taken by the reservoir fluids should be far from the critical point.
• The amount of liberated gas or condensate dropout should be a small part of the hydro-
carbon in place.
• The remaining hydrocarbon composition should not change significantly when gas is
liberated or vaporized condensate drops out.
If these conditions are not satisfied, you should consider using a compositional simulator.

Consistent PVT Parameters

PVT parameters used in black-oil models were originally derived from laboratory analysis
of (recombined) reservoir fluids, but in recent years it has become more common to derive
the necessary pressure-dependent parameters by simulating differential liberation experi-
ments by use of compositional models with appropriate equations of state. Regardless of
their source, the derived PVT model must satisfy certain physical properties; see [73]. First
of all, the total compressibility of the system must be non-negative, i.e.,

ct = Sw cw + So co + Sg cg + cr ≥ 0,

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 408 Compressible Multiphase Flow

where cr is the rock compressibility and cα is the isothermal compressibility of phase α,

Bw B (Bg − Bo Rv )Rs
cw = − , co = − o + ,
Bw Bo Bo (1 − Rs Rv )
Bg (Bo − Bg Rs )Rv
cg = − + .
Bg Bg (1 − Rs Rv )
Since this property has to be valid for all saturations, it follows that each of the isothermal
compressibilities must also be nonnegative, which gives the following three consistency
conditions:
Bw ≤ 0,
  
1 − Rs Rv Bo ≤ Bg − Rv Bo Rs ,
 
1 − Rs Rv Bg ≤ Bo − Rs Bb Rv . (11.32)
To ensure that the partial volumes in (11.11) are also positive, the following additional
conditions must be satisfied:
Rs Rv < 1, Bw > 0, Bg − Rv Bo > 0, Bo − Rs Bg > 0. (11.33)
As we have seen in Section 11.4, Bo , Rs , and Rv are normally increasing functions with
positive derivatives, whereas Bg and Bw are decaying functions with negative derivatives.

Use of Tabulated Data

Use of tabulated data, as discussed earlier in Section 11.5, is very common in industry-
grade simulation models. This means that when the simulator requires values between
data points, these need to be interpolated. Likewise, values requested outside the range
of tabulated data need to be extrapolated. Algorithmic choices of how data are interpolated
and extrapolated are specific to the individual simulator, but can have a significant impact
on both the accuracy and stability of the simulator, and the user should therefore be aware
of what methods are used.
The default choice in ECLIPSE is to use linear interpolation and zeroth order extrap-
olation for rock-fluid properties, so that the simulator picks the value of the first or last
tabulated point rather than using trend interpolation based on the closest points. This means
that derivatives of tabulated quantities are piecewise constants and effectively are com-
puted using finite-differences between tabulated points. PVT properties are also interpo-
lated within their range of definition, but properties like Rs and Rv may also be extrapo-
lated. In MRST, we have chosen to follow the interpolation conventions from ECLIPSE
in the ad-props module to be able to produce simulation results that are as identical as
possible when using the same discretizations and solvers. This raises several issues you
should be aware of:

• First of all, you should make sure that the data points cover the value range of the
simulation. For PVT properties, this usually means that data should be defined for higher
pressures than what are initially found in the reservoir, since pressure can increase during

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.8 Black-Oil Simulation with MRST 409

the simulation as a result of fluid injection. In ECLIPSE 100, you can use the keyword
EXTRAPMS to force the simulator to issue a warning whenever it extrapolates outside of
tabulated data. At the time of writing, such a safeguard is not yet implemented in MRST.
• To reduce numerical errors from the numerical interpolation, make sure that data points
are sufficiently dense and distributed so that interpolation errors do not vary too much in
magnitude throughout the parameter range.
• Try to avoid discontinuities, strong kinks, and other abrupt changes in tabulated proper-
ties unless these are dictated by physics. (See, e.g., the relative permeabilities for SPE 9 in
Figures 11.6 and 11.7.) If necessary, introduce extra data points near strong nonlinearities
and local extremal points to avoid introducing unphysical oscillations and too-large val-
ues for derivatives. This will generally improve the convergence of the nonlinear solver
and reduce the number of (unnecessary) time-step chops.
• On the other hand, introducing too many points will make table lookup more expensive
and contribute to slow down your simulation. For MRST, you should be particularly
beware of multiple regions. Property evaluation is vectorized over all cells to make it
as efficient as possible, but with multiple relperm or PVT regions, each region must
be processed sequentially in a for loop, and this can add a significant computational
overhead if the number of regions is large.
• In certain cases, ECLIPSE interpolates functions of the input data or functional combina-
tions of multiple properties, rather than the tabulated values. The primary example is the
PVT data for oil. Here, ECLIPSE 100 interpolates reciprocals of Bo and μo Bo , whereas
ECLIPSE 300 interpolates reciprocals of Bo , but interpolates μo directly. Using different
interpolations may have a significant effect on simulation results.
To give you an idea of how things can go wrong, let us revisit two of the SPE benchmark
models, starting with SPE 3. To sample the model, we make a 101 × 21 regular mesh
of (p,Rv ) values covering the rectangular domain [p0,p1 ] × [0,Rv (p1 )] for p0 = 80
and p1 = 280 bar. In the input data, properties are only defined for pressures in the
interval [83.75,237.82] bar, and hence we need to use both interpolation and extrapolation
to compute values. Figure 11.35 reports the computed gas viscosities for values toward the
upper end of the pressure interval. Here, the left plot clearly shows that the viscosity is not
necessarily a monotone function of pressure for fixed Rv . More important, the right plot
shows that we get incorrect behavior with highly unrealistic viscosities for Rv values less
than Rv (p) once the pressure exceeds maximum data values by more than a few percent.
Figure 11.36 shows that the PVT model in SPE 1 P violates the consistency condition
Bo − Rs Bg > 0 in (11.33) for high pressure values, as observed previously by Coats [73].
By convention, the formation-volume factor Bo of oil is computed by interpolating the
shrinkage factor bo rather than the tabulated Bo values. In the tabulated data, bo declines
with increasing Rs values and will hence cross zero for sufficiently high values. This leads
to blowup and negative values for Bo . This behavior is only observed for (p,Rs ) values that
are larger than what is to be expected in practical computations, but it is still interesting to
note this artifact.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 410 Compressible Multiphase Flow
−4
x 10 0.08
0.04
3.5
0.07
3 0.038
0.06
2.5 0.036
0.05
2 0.034

1.5 0.04
0.032
1 0.03
0.03
0.5
0.028 0.02
0
200 210 220 230 240 160 180 200 220 240 260 280
Pressure [bar] Pressure [bar]

Figure 11.35 Interpolated and extrapolated gas viscosity for the SPE 3 fluid model. The left plot
shows contours of μg over parts of the (p,Rv ) domain, with data points shown as filled circles. The
right plot shows μg as function of pressure only, with dashed lines corresponding to constant Rv
values.

10
0.8
1000 1000

0.6 5
800 800
Gas−oil ratio

Gas−oil ratio

0.4
600 600 0

0.2
400 400
−5
0
200 200
−10
200 400 600 800 1000 1200 200 400 600 800 1000 1200
Pressure [bar] Pressure [bar]

Figure 11.36 Inconsistency in the PVT data of the SPE 1 benchmark giving negative partial volumes.
The left plot shows the interpolated shrinkage factor bo , whereas the right plot shows the resulting
values of Bo − Rs Bg . To emphasize the blowup and the negative values, the color scale has been
cropped to [−10,10], but the actual values are up to two orders larger.

Initialization and Equilibration

In all examples studied so far in the book, the initial saturation has either been a constant,
or the fluid phases have been uniformly segregated throughout the reservoir. In a real
reservoir, the initial fluid distribution will usually be more complicated, as we saw for the
heterogeneous capillary fringe in Section 10.3.7. Setting the correct initial state is essential
to obtain accurate estimates of fluids in place and predict how the reservoir will behave
under production. ECLIPSE input offers two ways of setting up the initial state:

• you can either input individual cell values of the initial state using the keywords
PRESSURE, SWAT, SGAS, and RS or PBUB in an ECLIPSE input file, or
• you can calculate the initial state equilibration, i.e., determine it on the basis of hydro-
static equilibrium.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 11.8 Black-Oil Simulation with MRST 411

At the time of writing, MRST’s support of initialization is somewhat unfinished. All cells in
a grid can be assigned specific pressure and saturation values given in a deck by the function
initStateDeck(model,deck). The function also performs a full equilibration initialization,
proceeding in two steps
regions = getInitializationRegionsDeck(model, deck);
state0 = initStateBlackOilAD(model, regions);

The first call extracts the different subregions from the REGIONS section of the input deck,
whereas the second call computes vertical equilibrium inside each region. This part of the
software will be unified in the near future.
Let us discuss this initialization in some more detail. In the previous chapter we saw
how variations in rock properties can cause variations in capillary pressures. Observed
fluid contacts (gas–oil and oil–water) generally show vertical variation and will not always
be planar. To determine the free fluid contacts, we must know the phase pressures inside
the reservoir. The level of free gas is then defined as the depth at which the gas and oil
pressures coincide, i.e., the depth at which the oil–gas capillary pressure is zero. The level
of free water is defined analogously.
Typical input data to an equilibration calculation consist of pressure at a datum depth,
gas–oil and oil–water contacts, capillary pressure at each contact, and gas–oil ratio (Rs )
or bubble-point pressure (pb ) as a function of depth. Given these data, it is, in principle, a
straightforward task to determine the fluid distribution, which generally will consist of five
different zones:
1. A gas cap, in which gas is the only continuous phase (Sg = 1 − Swi )
dpg
= gρg , po = pg − Pcog (Sg ), pw = po − Pcow (Swi ).
dz
2. An oil–gas transition zone, in which both gas and oil are continuous:
dpg dpo
= gρg , = gρo, pw = po − Pcow (Swi ).
dz dz
The saturations are Sw = Swi , Sg = Pcog−1 (p − p ), and S = 1 − S − S .
g o o g wi
3. An oil zone, in which oil is the only continuous phase (So = 1 − Swi ):
dpo
= gρo, pg = po + Pcog (0), pw = po − Pcow (Swi )
dz
4. An oil–water transition zone, in which both oil and water are continuous:
dpo dpw
= gρo, = gρw, pg = po + Pcog (0).
dz dz
−1 (p − p ), and S = 1 − S .
The saturations are Sg = 0, Sw = Pcow o w o w
5. A water zone, in which only water is present:
dpw
= gρw, po = pw + Pcow (1), pg = po − Pcog (0).
dz

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 412 Compressible Multiphase Flow

In practice, however, the equilibration procedure is complicated by the fact that the fluid
contacts may vary between different compartments in the reservoir and that capillary pres-
sure curves may vary from one cell to the next. Notice also that numerical artifacts and
inaccuracies may arise if the grid is not aligned with the fluid contacts; see e.g., [251].
Once your model is initialized, it is a good precaution to run a simulation with no
external drive mechanisms applied to see if the reservoir contains any movable fluids, i.e.,
check whether the initial phase and pressure distributions remain invariant in time.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12
The AD-OO Framework for Reservoir Simulation

In Chapter 7 we showed that combining a fully implicit formulation with automatic differ-
entiation makes it simple to extend basic flow models with new constitutive relationships,
extra conservation equations, new functional dependencies, and so on. By using numerical
routines and vectorization from MATLAB, combined with discrete differential and aver-
aging operators from MRST, these equations can be implemented in a very compact form
close to the mathematical formulation. Writing a simulator as a single script, like we did for
single-phase flow in Chapter 7 or for two-phase flow without phase transfer in Section 11.2,
has the advantage that the code is self-contained, quick to implement, and easy to modify
as long as you work with relatively simple flow physics. However, as the complexity of the
flow models and numerical methods increases, and more bells and whistles are added to
the simulator, the underlying code will inevitably become cluttered and unwieldy.
When you research new computational methods for reservoir simulation, it is important
to have a flexible research tool that enables you to quickly test new ideas and verify
their performance on a large variety problems, from idealized and conceptual cases to full
simulation setups. The AD-OO framework was introduced to structure our fully implicit
simulators and enable us to rapidly implement new proof-of-concept codes also for more
complex flow physics. Altogether, the AD-OO framework offers many of the features
found in commercial simulators. Understanding its design and all the nitty-gritty details
of the actual implementation will obviously provide you with a lot of valuable insight
into practical reservoir simulation. If you are not interested in code details, you can read
Section 12.1 and then jump directly to Section 12.4, which presents a few simulation
examples so as to outline the functionality and technical details you need to set up your
own black-oil simulation both with and without the use of ECLIPSE input. The sections
in between give a somewhat detailed discussion of the most central classes and member
functions to demonstrate the general philosophy behind the framework, but do not discuss
in detail how boundary conditions, source terms, and wells are implemented. The purpose
of these sections is to provide you with sufficient detail to start developing new simulators
with new numerics or other types of flow physics.

413

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 414 The AD-OO Framework for Reservoir Simulation

12.1 Overview of the Simulator Framework

Looking back at the procedural implementations presented so far in the book, we have
introduced a number of data objects to keep track of all the different entities that make up
a simulation model:

• a state object holding the unknown pressures, saturations, concentrations, inter-cell

fluxes, and unknowns associated with wells;
• the grid structure G giving the geometry and topology of the grid;
• a structure rock representing the petrophysical data: primarily porosity and permeability,
but net-to-gross as well;
• a sparse transmissibility vector (or inner-product matrix), possibly including multipliers
that limit the flow between neighboring cells;
• a structure fluid representing the fluid model, having a collection of function handles
that can be queried to give fluid densities and viscosities, evaluate relative permeabilities,
formation volume factors, etc.;
• additional structures that contain the global drive mechanisms: wells, volumetric source
terms, boundary conditions;
• and an optional structure that contains discrete operators for differentiation and
averaging.
In a procedural code, these data objects are passed as arguments to solvers or functions
performing subtasks like computing residual equations. As an example, all incompressible
flow solvers in the incomp family require the same set of input parameters: reservoir
state, grid and transmissibilities, a fluid object, and optional parameters that describe the
drive mechanisms. The transport solvers, on the other hand, require state, grid, fluid, rock
properties, and drive mechanisms. Solvers for the fully implicit equation system require the
union of these quantities to set up the residual equations. An obvious simplification would
be to collect all data describing the reservoir and its fluid behavior in a model object,
which also implements utility functions to access model behavior. Given a reservoir state,
for instance, we can then query values for physical variables with a syntax of the form:
[p, sW, sG, rs, rv] = ...
model.getProps(state, 'pressure', 'water', 'gas', 'rs', 'rv');

As we saw in the previous chapter, the criteria by which we choose sg , rs , and rv as

primary reservoir variable will vary from cell to cell depending on the fluid phases present.
Ideally, the model object should keep track of this and present us with the correct vector of
unknowns for all cells. It is also natural that the model object should be able to compute
residual flow equations (and their Jacobian) by a call like
eqs = model.getEquations(oldState, newState, dT);

With a generic interface like this, you can develop nonlinear solvers that require no specific
or very limited knowledge of the physical model. A next step would be to separate the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.1 Overview of the Simulator Framework 415

time-stepping strategy and the linear solver from the nonlinear solvers, so that we eas-
ily can switch between different methods or replace a generic method by a tailor-made
method. Moreover, a critical evaluation of the many simulation examples developed with
the procedural approach shows that large fractions of the scripts are devoted to input and
output of data, plotting results, etc. Likewise, implementing variations of the same model
either results in code duplication or a large number of conditionals that tend to clutter
the code.
A standard approach to overcome these difficulties is to use object orientation. The
object-oriented, automatic-differentiation (AD-OO) framework in MRST is tailor-made to
support rapid prototyping of new reservoir simulators based on fully or sequentially implicit
formulations. An essential idea of the framework is to develop general model and simulator
classes, and use inheritance to develop these into simulators for specific equations. The
framework splits

• physical models describing the porous medium and the fluids flowing in it,
• nonlinear solvers and time-stepping methods,
• linearization of discrete equations formulated using generic discretization, averaging,
and interpolation operators as discussed previously, and
• linear solvers for the solution of the linear system
into different numerical contexts. This way, we only expose needed details and enable more
reuse of functionality that has already been developed.
Figure 12.1 illustrates how these contexts may appear in the main loop of a simulator.
The left part of the figure shows an unsophisticated Newton loop of the type used for
single-phase flow in Chapter 7, which consists of an outer loop running the time steps, and
an inner loop performing the Newton iteration for each time step. The right part shows a
more advanced nonlinear solver that utilizes some kind of time-control mechanism to run

Figure 12.1 Time loop of a simulator for a simple Newton solver (left) and for a more sophisticated
solver with stabilization and time-step control (right). For the latter, the iteration has been divided
into various numerical contexts.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 416 The AD-OO Framework for Reservoir Simulation

the targeted control steps of a simulation schedule. (Usually, the simulator should report
the solution after each such step.)
In this setup, the physical model is responsible for initializing the reservoir state. The
nonlinear solver implements the mechanism that selects and adjusts the time steps and
performs a nonlinear iteration for this time step. Inside this nonlinear iteration, the physical
model computes residual equations defined over all cells and possibly also cell faces and
discrete entities in the well representation (connections or nodes and segments). When
the residual equations are evaluated, the corresponding Jacobian matrix containing the
linearized residual equations is computed implicitly by the AD library. This produces
a collection of matrix blocks that each represents the derivative of a specific residual
equation with respect to a primary variable. The linearization context assembles the matrix
blocks into an overall Jacobian matrix for the whole system, and if needed, eliminates
certain variables to produce a reduced linear system. The linear solver sets up appropriate
preconditioners and solves the linear system for iteration increments, while the nonlinear
solver performs necessary stabilization, e.g., in the form of an inexact line-search algo-
rithm. Finally, the physical model updates reservoir states with computed increments and
recomputes residuals to be checked against prescribed tolerances by the nonlinear solver.
Figure 12.2 shows how a single nonlinear solve with time-step estimation is realized in
terms of classes and structures in the AD-OO framework. The ad-core module utilizes
MRST’s core functionality to implement a generic nonlinear solver, a set of generic time-
step selectors, and a set of linear solvers. The nonlinear solver does not know any specific
details of the physical model, except that it is able to perform the generic operations
we just outlined; we will come back to model classes in the next section. The time-step
selectors can implement simple heuristic algorithms that limit how much solutions can
change during an iterative step, e.g., as in the Appleyard and modified Appleyard chop
[271] used in commercial simulators. The linear solver class provides an interface for

Figure 12.2 Data diagram for a single nonlinear solve with time-step estimation in the AD-OO
framework.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.1 Overview of the Simulator Framework 417

various solvers including MATLAB’s default backslash solver as well as a state-of-the-

art, constrained pressure residual (CPR) preconditioner [304, 305, 119], which can be
combined with efficient algebraic multigrid solvers such as the aggregation-based method
found in [238], or the more recent AMGCL library [81]. Assembly of the linearized system
is relegated to a special class that stores meta-information about the primary variables and
the residual equations, i.e., whether they are reservoir, well, or control equations. The
linear solver class employs this information to help setting up preconditioning strategies
that exploit special structures in the problem.

Example 12.1.1 (1D Buckley–Leverett) To illustrate the use of the nonlinear solver, let us
revisit the classical Buckley–Leverett problem from Example 9.3.4 on page 284 and Section
10.3.1. For completeness, we show all code lines necessary to set up this problem in the
AD-OO framework (which you also can find in adBuckleyLeverett1D.m in the ad-core
module). First, we construct the reservoir and water–oil fluid data with constant shrinkage
factors:
mrstModule add ad-core ad-props ad-blackoil
G = computeGeometry(cartGrid([50, 1, 1], [1000, 10, 10]*meter));
rock = makeRock(G, 1*darcy, .3);
fluid = initSimpleADIFluid('phases', 'WO', 'n', [2 2]);

We then create a two-phase black-oil model (from ad-blackoil) and set up the initial
state:
model = TwoPhaseOilWaterModel(G, rock, fluid);
state0 = initResSol(G, 50*barsa, [0, 1]);
state0.wellSol = initWellSolAD([], model, state0);

Notice that the water–oil model assumes that the state object contains a subfield holding
well solutions. Since there are no wells in our case, we instantiate this field to be an empty
structure having the correct subfields. Finally, we set up the correct drive mechanisms:
constant rate at the left and constant pressure at the right end.
injR = sum(poreVolume(G,rock))/(500*day);
bc = fluxside([], G, 'xmin', injR, 'sat', [1, 0]);
bc = pside(bc, G, 'xmax', 0*barsa, 'sat', [0, 1]);

Having set up the problem, the next step is to instantiate the nonlinear solver:
solver = NonLinearSolver();

By default, the solvers employs MATLAB’s standard backslash operator to solve the linear
problems. The time-step selector employs a simplified version of the time-chop strategy
outlined in Section 10.2.2 to run the specified control steps: time steps are chopped in two
if the Newton solver requires more than 25 iterations, but no attempt is made to increase
successful substeps during a control step. The setup is the same as for the SPE 1 test
case 11.8.1 on page 399, except that we then used inexact line search to stabilize the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 418 The AD-OO Framework for Reservoir Simulation

Newton updates. Line search is not needed here since the fluid behavior of the Buckley–
Leverett problem is much simpler and does not involve transitions between undersaturated
and saturated states.
There are several ways we can use the nonlinear solver class. One alternative is to
manually set up the time loop and only use the solveTimestep function of the nonlinear
solver to perform one iteration:
[dT, n] = deal(20*day, 25);
states = cell(n+1, 1); states{1} = state0;
solver.verbose = true;
for i = 1:n
states{i+1} = solver.solveTimestep(states{i}, dT, model, 'bc', bc);
end

The results are stored in a generic format and can be plotted with the plotToolbar
GUI discussed in the previous chapter. Alternatively, we can construct a simple simulation
schedule and call the generic simulation loop implemented in ad-core:
schedule = simpleSchedule(repmat(dT,1,25), 'bc', bc);
[,sstates] = simulateScheduleAD(state0, model, schedule);

This loop has an optional input argument that sets up dynamic plotting after each step in
the simulation, e.g., to monitor how the well curves progress during the simulation, or to
print out extra information to the command window. Here, we use this hook to set up a
simple graphical user interface that visualizes the progress of the simulation and enables
you to modify the time steps interactively, dump the solution to workspace, and stop the
simulation and continue running it in debug mode. The GUI may in turn call any or both of
the two GUIs shown in Figure 11.32 to plot reservoir states and well solutions. We request
1D plotting of the water saturation, and that we turn off plotting of wells, since the example
does not have any:
fn = getPlotAfterStep(state0, model, schedule, ...
'plotWell', false, 'plotReservoir', true, 'field', 's:1', ...
'lockCaxis',true, 'plot1d', true);

[,sstates,report] = ...
simulateScheduleAD(state0, model, schedule,'afterStepFn', fn);

Figure 12.3 shows a snapshot of the GUI and the saturation profile just before the displace-
ment front reaches the right end of the domain. The first control step requires 7 iterations,
whereas the next 15 steps only use 4 iterations, giving a total of 67 iterations, or an average
of 4.19 iterations for the first 16 control steps. The GUI also reports the total runtime
so far and uses the iteration history to estimate the total runtime. A total runtime of 7
seconds for such a small model may seem high and can mainly be attributed to the GUI,
plotting, and computational overhead introduced by the AD-OO framework. This over-
head is significant for small models, but decreases in importance as the size of the model
increases.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.1 Overview of the Simulator Framework 419

Figure 12.3 The default graphical user interface in the generic simulation loop reports simulation
progress and enables simple computational steering.

For large models it may not be feasible to run interactive sessions and store all simu-
lation results in memory. The solver also has a hook that enables you to control how the
computed states are stored. To control the storage, we use a so-called result handler, which
behaves almost like a cell array, whose content can either be stored in memory, on disk,
or both places. This is very useful if you have long simulations you want to run in batch
mode, or if you expect that the simulator may fail or be interrupted before the simulation is
finished and you want to store all completed time steps. Let us set up a handler for storing
computed states to disk
handler = ResultHandler('writeToDisk', true, 'dataFolder', 'ad-bl1d');

We can then pass the handler to the simulator as an optional argument

simulateScheduleAD(state0, model, schedule, 'outputHandler', handler);

The simulator will now store each state as a separate MATLAB-formatted binary file in a
subdirectory ad-bl1d somewhere on a standard path relative to where you have installed
MRST. By default, the files are called state1.mat, and so on. The name can be changed
through the option 'dataPrefix', whereas the parameter 'dataDirectory' lets you
change the base directory relative to which the files are stored. To maintain results in
memory, you need to set the option 'storeInMemory'.
You can now use the result handler as a standard cell array, and you need not really
know where the results are stored. You can for instance plot the results through the same
GUI we used in the previous chapter
plotToolbar(G, handler, 'field', 's:1','lockCaxis', true, 'plot1d', true);

If you at a later time wish to retrieve the data, you simply create a new handler and will
then have access to all data stored on disk. The following example shows how to extract
every second time step:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 420 The AD-OO Framework for Reservoir Simulation

handler = ResultHandler('dataFolder', 'ad-bl1d');

m = handler.numelData();
states = cell(m, 1);
for i = 1:2:m
states{i} = handler{i};
end

You can delete the stored data through the call handler.resetData().

To make the AD-OO simulator framework as versatile as possible, we have tried to

develop a modular design, so that any simulator consists of a number of individual com-
ponents that can easily be modified or replaced if necessary. Figure 12.4 shows how vari-
ous classes, structures, and functions work together in the generic simulateScheduleAD
function. The function takes physical model, initial state, and schedule as input, and then
loops through all the control steps, updating time steps and drive mechanisms, calling
the nonlinear solver, and performing any updates necessary after the nonlinear solver has
converged (to track hysteretic behavior, etc.). In the next two sections, we discuss the main
elements shown in Figure 12.4.

12.2 Model Hierarchy

The mrst-core module implements two basic model classes: PhysicalModel and
ReservoirModel. Their main purpose is to provide generic templates for a wide vari-
ety of specific flow models. As such, these two generic model classes do not offer
any flow equations and hence cannot be used directly for simulation. Instead, they
provide generic quantities and access mechanisms, which we will discuss in more
detail shortly. We can then develop specific models by expanding the generic models
with specific flow equations and fluid and rock properties. The ad-core module also
contains classes implementing well models: either simple Peaceman wells of the type we
discussed in Sections 4.3.2, 5.1.5, and 11.7.1, or advanced multisegment well models from
Section 11.7.2 that give more accurate representation of deviated, horizontal, and multilat-
eral wells with advanced completions and inflow control devices.
The ad-blackoil module implements black-oil models, as discussed in Chapter 11,
utilizing functionality from the deckformat module to read ECLIPSE input decks and
from ad-props to construct models for rock-fluid and PVT properties; see Figure 12.5. The
resulting simulators can simulate industry-standard black-oil models and compute adjoint
gradients and sensitivities.
Sequential solvers, like those discussed in Chapter 10, can often be significantly faster
than fully implicit solvers, especially for where the total velocity changes slowly during the
simulation. Such solvers are also a natural starting point when developing, e.g., multiscale
methods [215, 216]. The blackoil-sequential module implements sequential (fully
implicit) solvers for the same set of equations as in ad-blackoil, based on a fractional

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Figure 12.4 Overview of how components in the AD-OO framework are organized to implement a generic simulator loop (here shown
for black-oil equations).
Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 422 The AD-OO Framework for Reservoir Simulation
MRST core mrst-gui deckformat
Basic functions/data structures: Graphical interfaces for interactive Input of ECLIPSE simulation
grid, petrophysics, wells, boundary visualization of reservoir states and decks: read, convert to SI units,
conditions, I/O, grid processing, petrophysical data and construct MRST objects for
AD library, plotting, . . . grids, fluid and rock properties,
wells and simulation schedules

ad-core ad-blackoil ad-props

General simulation framework: General 3-phase black-oil simulator Initialization of fluid models from
abstract model classes, time-step with dissolution and vaporization, ECLIPSE input decks
and iteration control, linearizations, specialized 1- and 2-phase models,
linear solvers, hooks for I/O and CPR preconditioning
plotting, . . .

core functionality
blackoil-sequential ad-eor
Sequential implicit solvers for the Fully implicit simulators for water- utility module
black-oil equations, solving flow based EOR methods: polymer and
and transport in separate steps surfactant AD-OO module

Figure 12.5 Modules used to implement black-oil models in the AD-OO framework.

flow formulation wherein pressure and transport are solved as separate steps and thus are
represented as separate models.
The ad-eor module [30] implements extensions of the black-oil equations to water-
based EOR methods like polymer and surfactant flooding and the co2lab module [19]
uses inheritance from the black-oil module to implement specialized simulators for CO2
sequestration. The black-oil family also includes modules for solvents and routines for
solving optimal control problems based on the AD-OO framework. In addition, MRST
contains other AD-OO modules for compositional simulators [212], geomechanics, and
coupled geomechanics–flow simulations that do inherit from ad-blackoil, and more
modules are in the making. These are all (unfortunately) outside the scope of this book.

12.2.1 PhysicalModel – Generic Physical Models

At the lowest level of the model hierarchy, we have a generic physical model that defines
basic properties to be inherited by other models and model classes. All models in the
AD-OO framework are interpreted in a discrete sense and are assumed to consist of a
set of model equations on residual form. The primary properties of a physical model
are a structure holding discrete operators used to define the model equations and a non-
linear tolerance that defines how close the residual values must be to zero before the
model is fulfilled. Actual operators are not implemented in the PhysicalModel class and
must be specified in derived classes or explicitly set by the user for concrete instances
of PhysicalModel. These operators are typically defined over a grid, and the ad-core
module offers the function setupOperatorsTPFA for setting up the two-point operators
we have discussed in previous chapters. In addition to the optional grid property, the class
contains a flag signifying if the model equations are linear (and hence can be solved by a

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.2 Model Hierarchy 423

single matrix inversion), and a flag determining how verbose the member functions should
be in their output
classdef PhysicalModel
properties
operators
nonlinearTolerance
G
verbose
stepFunctionIsLinear
end
methods
function model = PhysicalModel(G, varargin)
model.nonlinearTolerance = 1e-6;
model.verbose = mrstVerbose();
model = merge_options(model, varargin{:});
model.G = G;
model.stepFunctionIsLinear = false;
end
end

Notice that all class names in MRST start with a capital letter to distinguish them from
functions, which always start with a lowercase letter.

Model Equations and Physical Properties

The basic class has several member functions that implement generic tasks associated with
a model. The first, and most important, is to evaluate residual equations and Jacobians,
which is done by the following function:
[problem, state] = model.getEquations(state0, state, dt, forces, vararg)

Here, problem is an instance of the LinearizedProblem class that implements what we

referred to as the linearization numerical context in the simulator loop in Figure 12.1. This
class has data structures to represent the linearized system for given states and member
functions that correctly assemble the full linear system from residuals and Jacobian block
matrices stored for the individual sub-equations making up the flow model; we discuss this
in more detail in Section 12.3. Input argument forces represents driving forces that are
currently active. These may vary from one step to the next and are specified by the control
part of the schedule; see e.g., Figure 12.4. Active driving forces are set by
forces = model.getDrivingForces(control);

The function creates a cell array that lists the forces on the same form as for the incom-
pressible solvers. What are the valid forces for a model (with reasonable defaults) can be
queried
validForces = model.getDrivingForces()

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 424 The AD-OO Framework for Reservoir Simulation

In addition, the class has utility functions for interacting with the state structure in a con-
sistent manner. This includes querying values for one or more specific variables/properties,
setting or capping specific values, or correctly incrementing values
p = model.getProp (state, 'pressure');
[p,s] = model.getProps(state, 'pressure', 's');
state = model.setProp (state, 'pressure', 5);
state = capProperty (state, 's', 0, 1)
state = model.incrementProp(state, 'pressure', 1);

These operations are generic and hence implemented in the PhysicalModel class. How-
ever, the specific function arguments will only work for instances of ReservoirModel
or derived classes that specify the variables pressure and s. You can check whether a
variable is defined or not as follows
[fn, index] = model.getVariableField(name)

which produces the field name and column index that will extract the correct variable
from the state object. The function produces an error when called with any name for a
PhysicalModel, since this class does not implement any variables.

Linear Updates to States

The second purpose of the basic model class is to correctly update states as part of a linear
or nonlinear solution process. First of all, the function
[state, report] = model.stepFunction(state, state0, dt, forces, ...
linsolver, nonlinsolver, itno, varargin)

implements the generic behavior of a single nonlinear update; that is, it executes the steps
of the inner while loop of the right-most box in Figure 12.1. Here, nonlinsolver is an
instance of the NonLinearSolver class, which we briefly introduced on on page 399, and
linsolver is an instance of a similar linear solver class; more detail are given in Section
12.3. The second output, report, is a standardized step report that provides information
such as the CPU time used by the linear solver, setup of the stabilization step, convergence
status, size of residual, etc. There are also functions to extract the computed update, update
the state, check whether it is converged or not, check whether the controls have changed, as
well as a hook enabling a state to be updated after it has converged, e.g., to model hysteretic
behavior.
Let us in particular look at the function for updating the state with a given increment.
The following code updates the pressure value from 10 to 110:
state = struct('pressure', 10);
state = model.updateStateFromIncrement(state, 100, problem, 'pressure')

We can also restrict the maximum relative changes, e.g., to be at most 10%
state = model.updateStateFromIncrement(state, 100, problem, 'pressure', .1)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.2 Model Hierarchy 425

which sets pressure to 11. Relative limits such as these are important when working with
tabulated and nonsmooth properties in a Newton-type loop, as the initial updates may be
far outside the reasonable region of linearization for a complex problem. On the other
hand, limiting the relative updates can delay convergence for smooth problems with ana-
lytic properties and will, in particular, prevent zero states from being updated, so use
with care.

Quality Assurance
A model must also be able to check its on validity. The function
model.validateModel(forces)

where forces is an optional argument, validates that a model is suitable for simulation.
If missing or inconsistent parameters can be fixed automatically, an updated model is
returned. Otherwise, an error should occur. This function does not have meaning for a
PhysicalModel, but should be implemented in derived classes. Likewise, there is a func-
tion to validate the state for use with the model
model.validateState(state)

The function should check that required fields are present and of the right dimensions,
and if missing fields can be assigned default values, return state with the required fields
added. If reasonable default values cannot be assigned, a descriptive error should be thrown
telling the user what is missing or wrong (and ideally how to fix it). Also this function must
be implemented in derived classes. The only function whose behavior is implemented in
PhysicalModel is
model.checkProperty(state, 'pressure', [nc, 1], [1, 2]);

here called for a model of the derived ReservoirModel class to check that the pressure
field has nc elements along its first dimension and one element along the
second dimension.

Adjoint Equations
One of the reasons for introducing the AD-OO framework was to simplify the solution of
so-called adjoint equations to compute parameter gradients and sensitivities. (How this is
done is outside the scope of this book.) The base class therefore contains an interface for
computing adjoint equations:
[problem, state] = model.getAdjointEquations(state0,state,dt,forces,vararg)

Solving adjoint equations boils down to running a backward simulation with a set
of linearized equations. PhysicalModel also offers a generic update function
solveAdjoint that computes a single step for the corresponding adjoint equations.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 426 The AD-OO Framework for Reservoir Simulation

12.2.2 ReservoirModel – Basic Reservoir Models

The next class in the hierarchy has properties to represent entities found in most reservoir
models: fluid model, petrophysical properties, indicators for each of the basic three phases
(aqueous, gaseous, and oleic), variables representing phase saturations and chemical com-
ponents, and a model of production facilities

classdef ReservoirModel < PhysicalModel

properties
fluid, rock, gravity, FacilityModel
water, gas, oil

% Iteration parameters
dpMaxRel, dpMaxAbs, dsMaxAbs, maximumPressure, minimumPressure,
useCNVConvergence, toleranceCNV, toleranceMB
:

Notice in particular that the reservoir model declares absolute and relative tolerances on
how much pressures and saturations can change during an iteration step as well as lower and
upper bounds on the pressure. In addition, the class declares that convergence of nonlinear
iterations can be measured in two different ways, either by mass balance (MB) or by scaled
residuals (CNV); more details will follow in Section 12.3.
To construct an instance of the class, we first construct an instance of a PhysicalModel
and then proceed to declare the additional properties that are part of a reservoir model. The
following is a somewhat simplified excerpt:

function model = ReservoirModel(G, varargin)

model = model@PhysicalModel(G);
model.rock = varargin{1}; model.fluid = varargin{2};
[model.water, model.gas, model.oil] = deal(false);
model.dpMaxRel = inf; model.dpMaxAbs = inf;
model.dsMaxAbs = .2;
:
model.operators = setupOperatorsTPFA(G, model.rock, ..)
end % call for construction: ReservoirModel(G, rock, fluid, ...)

By default, the reservoir model class does not contain any active phases; these must be
specified in derived classes. However, we do declare that discrete operators should be of
the two-point type. We also declare limits on how much pressures and saturations may
change from one iteration step to the next: the default is that changes in pressure can be
arbitrarily large, whereas saturations are only allowed an absolute change of 0.2.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.2 Model Hierarchy 427

Physical Variables and Active Phases

Physical variables known to a particular class are declared by the following function:
function [fn,ix] = getVariableField(model, name)
switch(lower(name))
case {'pressure' , 'p'}
ix = 1; fn = 'pressure' ;
case {'s', 'sat', 'saturation'}
ix = ':'; fn = 's' ;
case {'so' , 'oil'}
ix = find(strcmpi(model.getSaturationVarNames, name));
fn = 's' ;
:

The purpose of this function is to provide a convenient and uniform way to access physical
variables without knowing their actual storage in the state object. Let us consider saturation
as an example. Saturation is stored in the field state.s, but this vector may have one, two,
or three entries per cell, depending upon the number of phases present. To access individual
phase saturations, the class implements two additional utility functions:
function vars = getSaturationVarNames(model)
vars = {'sw', 'so', 'sg'};
vars = vars(model.getActivePhases());
end
function isActive = getActivePhases(model)
isActive = [model.water, model.oil, model.gas];
end

Here, we see that oil saturation is stored in the second column of state.s for models
containing water, but in the first column for a model without water. Oil saturation can now
be extracted as follows:
[fn, ix] = model.getVariableField('so');
so = state.(fn)(:, ix);

This may seem unnecessary complicated, but enables access to variables in a uniform way
for all models so that the code continues to work if we later decide to change the name
of the saturation variable, or implement models with more than three phases to represent
precipitated solids, emulsified phases, etc.
Altogether, the reservoir model declares the possible presence of variables for pressure;
saturations for the aqueous, oleic, and gaseous phases; temperature; and well solutions, but
does not declare variables for dissolved gas–oil ratio and vaporized oil–gas ratios, as these
are specific to black-oil models. There are also a number of other functions for getting
the active phases and components and correctly storing saturations, mobilities, face fluxes,
densities, shrinkage factors, and upstream indices in the state object.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 428 The AD-OO Framework for Reservoir Simulation

Fluid Behavior
The ReservoirModel class implements a generic method for evaluating relative perme-
abilities:

function varargout = evaluateRelPerm(model, sat, varargin)

active = model.getActivePhases();
nph = sum(active);
varargout = cell(1, nph);
names = model.getPhaseNames();

if nph > 1
fn = ['relPerm', names];
[varargout{:}] = model.(fn)(sat{:}, model.fluid, varargin{:});
elseif nph == 1
varargout{1} = model.fluid.(['kr', names])(sat{:}, varargin{:});
end

The function is a general interface to the relPermWO, relPermOG, relPermWG, and

relPermWOG functions discussed in Section 11.3, which in turn are interfaces to the more
basic krW, krO, krG, krOG, and krOW functions of fluid objects in MRST. The generic
reservoir model neither implements nor offers any interface for PVT behavior, since this
is very different in black-oil models and models relying on an equation of state. However,
there is a function for querying surface densities of all phases.

Global Driving Forces

The basic physical model was aware of driving forces but did not implement any.
ReservoirModel has gravity as a property and can evaluate the gravity vector and
compute a gravity gradient. In addition, the class specifies the potential presence of
boundary conditions, source terms, and wells

function forces = getValidDrivingForces(model)

forces = getValidDrivingForces@PhysicalModel(model);
forces.W = [];
forces.bc = [];
forces.src = [];
end

Wells are special because each well is represented by another model, i.e., an instance
of a derived class of PhysicalModel. Instances of such models are stored in the
FacilityModel property of the reservoir model, which in principle can contain models
of all production facilities used to bring hydrocarbons from the reservoir to the surface.
New well instances are created by the following function whenever the well controls
change

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.2 Model Hierarchy 429

function [model, state] = updateForChangedControls(model, state, forces)

model.FacilityModel = model.FacilityModel.setupWells(forces.W);
state.wellSol = initWellSolAD(forces.W, model, state);
[model,state] = updateForChangedControls@PhysicalModel(model, state, forces);
end

The generic operation of adding the effect of wells to a system of equations by adding the
corresponding source terms and augmenting the system with additional model equations
for the wells is done in the insertWellEquations function. The same holds for source
terms and boundary conditions. If you are more interested, you should consult the code for
further details.

Updating States and Models

The computation of linearized increments in a reservoir model is inherited from the
PhysicalModel class without any special adaptions or extensions, since it follows the
exact same algorithm as in a generic Newton loop. On the other hand, the way these
increments are used to update reservoir states differs from other types of models and hence
requires special implementation. The function updateState splits state variables into four
different categories: well variables, saturation variables, pressure, and optional remaining
variables. Updating well variables is quite involved and will not be discussed herein; as
always, you can find all necessary details in the code.
Saturation variables are characterized by the fact that they should sum to unity so that
pore space is completely filled by fluid phases. If our model contains nph phases, the
linearized increments will contain nph − 1 saturation increments. The nph ’th increment
is set as the negative sum of all the other increments; i.e., values added to the first
nph − 1 variables must be subtracted from the saturation of the last phase so that
the total increment over the nph phases is zero. The increments are then passed to the
updateStateFromIncrement function to compute new saturations satisfying limits on
absolute and relative updates. To ensure physically correct states, the updated saturations
are then cropped to the unit interval, and then we renormalize saturations in any cells

containing capped values, so that Sα ← Sα / α Sα .
We update pressure by first updating according to relative/absolute changes and then
capping values to fulfill limits on minimum and maximum pressure. Any remaining vari-
ables are updated with no limits on changes or capping.

Validating Models and States

You may recall that PhysicalModel defined functions for validating models and states
but did not implement any specific behavior. The reservoir model checks that the model
instance contains a facility model, and if not, instantiates one such object with default
values:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 430 The AD-OO Framework for Reservoir Simulation

function model = validateModel(model, varargin)

if isempty(model.FacilityModel)
model.FacilityModel = FacilityModel(model);
end
if nargin > 1
W = varargin{1}.W;
model.FacilityModel = model.FacilityModel.setupWells(W);
end
model = validateModel@PhysicalModel(model, varargin{:});

The validation function can also be called with a structure containing drive mechanisms as
argument. In this case, we extract the well description and pass it on to the facility model so
that this model can initialize itself properly. The function then calls the validation function
inherited from its parent class (PhysicalModel), which in the current implementation
does nothing. Because the reservoir model does not contain any concrete equations and
parameters, further consistency checks of model equations and petrophysical, fluid, and
PVT properties must be performed in derived classes or implemented explicitly by the
user. However, the model should be able to check a reservoir state and make sure that this
state is compatible with the model:
function state = validateState(model, state)
% Check parent class
state = validateState@PhysicalModel(model, state);
active = model.getActivePhases();
nPh = nnz(active);
nc = model.G.cells.num;
model.checkProperty(state, 'Pressure', [nc, 1], [1, 2]);
if nPh > 1
model.checkProperty(state, 'Saturation', [nc, nPh], [1, 2]);
end
state = model.FacilityModel.validateState(state);

To be consistent, the state should contain one unknown pressure and the correct number
of phase saturations in each cell. Once we have ensured that the reservoir states meets the
requirements, the facility model should check that the reservoir state contains the correct
data structures for representing well states. Neither of these operations check whether the
actual values represented in the reservoir state are physically meaningful.

12.2.3 Black-Oil Models

The ThreePhaseBlackOilModel class is the base class in the ad-blackoil module
and is derived from the ReservoirModel class. Unlike the two previous abstract classes,
this base class is a concrete class designed to represent a general compressible, three-
phase, black-oil model with dissolved gas and vaporized oil. The model therefore declares
two new properties, disgas and vapoil, to signify the presence of dissolved gas and

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.2 Model Hierarchy 431

vaporized oil, respectively. The corresponding data fields rs and rv are defined in the
getVariableField function. We also extend validateState to check that the reservoir
state contains a rs and/or a rv field if dissolution and/or vaporization effects are active in a
specific model instance. The class also defines the maximum absolute/relative increments
allowed for rs and rv . By default, the constructor sets up a dead-oil model containing the
aqueous, gaseous, and oleic phases, but with no dissolution and vaporization effects.
Unlike the generic classes from ad-core, ThreePhaseBlackOilModel implements
concrete equations. The member function getEquations calls the
equationsBlackOil, which generates linearized equations according to how the general
black-oil model is configured. We have already discussed the essential code lines contained
in this function, but as stated in Chapter 11, the actual code is structured somewhat
differently and contains more conditionals and safeguards to ensure that discrete residuals
are computed correctly for all combinations of drive mechanisms and possible special
cases of the general model.
In addition to implementing flow equations, the class extends updateState to imple-
ment variable switching when the reservoir state changes between saturated and undersat-
urated states. There is also a new member function that computes correct scaling factors to
be used by CPR-type preconditioners, which we outline on page 444.
The ad-blackoil module also implements three special cases of the general black-
oil model: a single-phase water model and two-phase oil–water and oil–gas models. Since
most functionality is inherited from the general model, these special cases are implemented
quite compactly, for instance:

classdef TwoPhaseOilWaterModel < ThreePhaseBlackOilModel

properties
end
methods
function model = TwoPhaseOilWaterModel(G, rock, fluid, varargin)
model = model@ThreePhaseBlackOilModel(G, rock, fluid, varargin{:});
model.oil = true;
model.gas = false;
model.water = true;
model = merge_options(model);
end

Residual equations are evaluated by the function equationsOilWater, which is a sim-

plified version of the equationsBlackOil function in which all effects relating to gas,
dissolution, and vaporization have been removed.

function [problem, state] = ...

getEquations(model, state0, state, dt, drivingForces, varargin)
[problem, state] = ...
equationsOilWater(state0, state, model, dt, drivingForces, varargin{:});
end

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 432 The AD-OO Framework for Reservoir Simulation

For completeness, let us quickly go through the essential parts of this function, disregarding
source terms, boundary conditions, and computation of adjoints. We start by getting driving
forces and discrete operators and setting names of our primary variables,
W = drivingForces.W;
s = model.operators;
primaryVars = {'pressure', 'sW', wellVarNames{:}};

Then, we can use the reservoir and facility models to extract values for the primary
unknowns:
[p, sW, wellSol] = model.getProps(state, 'pressure', 'water', 'wellsol');
[p0, sW0, wellSol0] = model.getProps(state0, 'pressure', 'water', 'wellSol');
[wellVars, wellVarNames, wellMap] = ...
model.FacilityModel.getAllPrimaryVariables(wellSol);

Once this is done, we compute the residual equations more or less exactly the same way as
discussed in Section 11.6. That is, we first compute fluid and PVT properties in all cells.
Then, we compute gradient of phase potentials, apply suitable multipliers, and average
properties correctly at the interface to compute intercell fluxes.
sO = 1 - sW;
[krW, krO] = model.evaluateRelPerm({sW, sO});
[pvMult, transMult, mobMult, pvMult0] = getMultipliers(model.fluid, p, p0);
krW = mobMult.*krW; krO = mobMult.*krO;
T = s.T.*transMult;
:

We now have all we need to compute the residual equations

water = (s.pv/dt).*( pvMult.*bW.*sW - pvMult0.*bW0.*sW0 ) + s.Div(bWvW);
oil = (s.pv/dt).*( pvMult.*bO.*sO - pvMult0.*bO0.*sO0 ) + s.Div(bOvO);

In all codes discussed so far, we have included wells by explicitly evaluating the well
equations and adding the corresponding source terms to the residual equations in individual
cells before concatenating all the residual equations. Here, however, we first concatenate
the reservoir equations and then use a generic member function from the ReservoirModel
class to insert the necessary well equations
eqs = {water, oil};
names = {'water', 'oil'};
types = {'cell', 'cell'};
[eqs, names, types, state.wellSol] = ...
model.insertWellEquations(eqs, names, types, wellSol0, wellSol,
wellVars, wellMap, p, mob, rho, {}, {}, dt, opt);

Finally, we construct an instance of the linearized problem class

problem = LinearizedProblem(eqs, types, names, primaryVars, state, dt);

We will discuss this class in more detail in Section 12.3.1.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.2 Model Hierarchy 433

12.2.4 Models of Wells and Production Facilities

The AD-OO framework has primarily been developed to study hydrocarbon recovery, CO2
storage, geothermal energy, and other applications in which fluids are produced from a
porous rock formation and brought up to the surface, or brought down from the surface and
injected into a rock formation. By default, all reservoir models are therefore assumed to
have an entity representing this flow communication between the surface and the subsur-
face. If you only want to describe flow in the porous medium, you can set the well model
to be void, but it must still be present.

Facility Models
The well models discussed earlier in the book apply to a single well (see Section 11.7).
Wells are sometimes operated in groups that are subject to an overall control. The corre-
sponding well models must hence be coupled. As an example, wells can be coupled so that
gas or water produced from one well is injected into another well as displacing fluid. Wells
from different reservoirs may sometimes also produce into the same flowline or surface
network and hence be subject to overall constraints that couple the reservoir models. Hence,
there is a need to represent a wider class of production and injection facilities.
Instead of talking of a well model, the AD-OO framework therefore introduces a general
class of facility models that can represent different kinds of facilities used to enhance and
regulate the flow of reservoir fluids from the sandface to the well head or stock tank. The
FacilityModel class is derived from the PhysicalModel class and is a general container
class that holds a collection of submodels representing individual wells. In principle, the
FacilityModel container class can also incorporate models for various types of topside
production facilities, but at the time of writing, no such models are implemented in MRST.

Peaceman-Type Well Models

The inflow performance relation, q = J (pR −pwb ), discussed in Sections 4.3.2 and 11.7.1,
gives an unsophisticated description of injection and production wells, and is generally
best suited for simple vertical wells, in which fluids inside the wellbore can be assumed to
be in hydrostatic equilibrium. Apart from the well index (and possibly a skin factor), no
attempt is made to model the complex flow physics that may take place in various types
of well completions. The flow in and out of the wellbore is regulated by controlling either
the surface flow rate or the bottom-hole pressure (bhp). In the AD-OO framework, this
type of well model is implemented in the SimpleWell class, which is derived from the
PhysicalModel class.

Multisegment Well Models

The MultisegmentWell class is also derived from the PhysicalModel class and imple-
ments the more general concept of multisegment wells, which we briefly introduced in
Section 11.7.2. In these models, the flow of fluids from the sandface to the well head is

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 434 The AD-OO Framework for Reservoir Simulation

described in terms of a set of 1D flow equations on a general flow network. In ECLIPSE

[270], the network must assume the form of a tree (i.e., be a directed graph). The imple-
mentation in MRST is more general and allows for complex network topologies involving
circular dependencies among the nodes.

A Word of Caution about the Implementation

Implementing the correct behavior of wells and facilities involves quite complicated logic
and a lot of intricate details that quickly make the code difficult to understand. A detailed
discussion of the facility and well classes is therefore outside the scope of this book. As
usual, you can find all necessary details in the actual code, but let me add a small warning:
To retain the same computational efficiency as the rest of the simulator classes, the well
classes have been optimized for computational efficiency. The reason is that well and
facility models typically have orders-of-magnitude fewer unknowns than the number of
states defined over the reservoir grid. The vectorized AD library in MRST is primarily
designed to perform well for long vectors and large sparse matrices and is not necessarily
efficient for scalars and short vectors. To keep the computational efficiency when working
with individual wells, it has proved necessary to meticulously distinguish between variables
that need automatic differentiation and variables that do not. As a result, the code contains
explicit casting of AD variables to standard variables, which in my opinion has reduced
readability.
In other words: please go ahead and read the code, but be warned that it may appear less
comprehensible than other parts of the AD-OO framework.

12.3 Solving the Discrete Model Equations

In this section, we go through the three different types of classes the AD-OO framework
uses to solve the discrete equations, motivate their design, and briefly describe their key
functionality. If you are interested in developing your own solution algorithms and solvers,
the classes define generic information about the discrete equations and interfaces to the rest
of the simulator you can use as a starting point.

12.3.1 Assembly of Linearized Systems

The main purpose of the LinearizedProblem class is to assemble the individual matrix
blocks computed by the AD library invoked within a model class into a sparse matrix that
represents the linearization of the whole reservoir model. To this end, the class contains the
residual equations evaluated for a given specific state along with meta-information about
the equations and the primary variables they are differentiated with respect to. We refer to
this as a linearized problem. This description can be transformed into a linear system and
solved using a subclass of the LinearSolverAD class to be discussed in Section 12.3.4,
provided that the number of equations matches the number of primary variables. The class
also contains the following data members:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.3 Solving the Discrete Model Equations 435

classdef LinearizedProblem
properties
equations, types, equationNames
primaryVariables
A, b
state
dt, iterationNo, drivingForces
end

The cell array equations contains the residuals evaluated at the reservoir state repre-
sented in the state object. The residuals can be real numbers, but are most typically AD
objects. The cell array types has one string per equation indicating its type. (Common
types: 'cell' for cell variables, 'well' for well equations, etc.) These types are available
to any linear solver and can be used to construct appropriate preconditioners and solver
strategies. The cell arrays equationNames and primaryVariables contain the names of
equations and primary variables; these are used, e.g., by the linear solvers and for reporting
convergence. The sparse matrix A represents the linear system and the vector b holds the
right-hand side. For completeness, we also store the time step, the iteration number, and
the driving forces, which may not be relevant for all problems.
The linear system corresponding to a linearized problem can be extracted by the member
function
[A, b] = problem.getLinearSystem();

The function will check whether the linear system already exists, and if not, it will assemble
it using the function:
function problem = assembleSystem(problem)
if isempty(problem.A)
iseq = cellfun(@(x) isempty(x), problem.equations);
eqs = combineEquations(problem.equations{iseq});
if isa(eqs, 'ADI')
problem.A = -eqs.jac{1};
problem.b = eqs.val;
else
problem.b = eqs;
end
end
end

In our previous AD solvers, the equations were always AD objects, and we could assemble
the linearized system simply by concatenating the AD variables vertically. In a general
model, we cannot guarantee that all residual equations are AD objects, and hence the
vertical concatenation is implemented as a separate function we can overload by specialized
implementations. If the residual equations are given as real numbers, this function simply
reads

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 436 The AD-OO Framework for Reservoir Simulation

function h = combineEquations(varargin)
h = vertcat(varargin{:});
end

In this case, we cannot construct a linear system since no partial derivatives are available. If
the residual equations are represented as AD objects, the linear system is, as before, assem-
bled from the Jacobian block matrices stored for each individual (continuous) equation. The
AD library from mrst-core implements an overloaded version of combineEquations,
which ensures that AD objects are concatenated correctly and also enables concatenation
of AD objects and standard doubles.
How the linear equations are solved, will obviously depend on the linear solver. The
simplest approach is to solve all equations at once. However, as we will come back to
in Section 12.3.4, it may sometimes be better to eliminate some of the variables and
only solve for a subset of the primary variables. To this end, the LinearizedProblem
class has member functions that use a block-Gaussian method to eliminate individual
variables

[problem, eliminatedEquation] = problem.eliminateVariable(name)

where name corresponds to one of the entries in problem.equationNames. Likewise, you

can eliminate all variables that are not of a specified type

[problem, eliminated] = reduceToSingleVariableType(problem, type)

Using this function, you can for instance eliminate all equations that are not posed on
grid cells. There is also a function that enables you to recover the increments in primary
variables corresponding to equations that have previously been eliminated. In addition, the
LinearizedProblem class contains member functions for appending/prepending addi-
tional equations, reordering the equations, computing the norm of each residual equation,
querying indices and the number of equations, querying the number of equations of a
particular type, as well as a number of utility functions for sanity checks, clearing the
linear system, and so on.

12.3.2 Nonlinear Solvers

The NonLinearSolver class is based on a standard Newton–Raphson formulation and
is capable of selecting time steps and cutting them if the nonlinear solver convergences
too slowly. To modify how the solver works, you can either develop your own subclass
and/or combine the existing solver with modular linear solvers and classes for time-step
selection. You have already seen some of the properties of the class in Section 11.8.1
on page 399. The following is an almost complete declaration of the class and its data
members

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.3 Solving the Discrete Model Equations 437

classdef NonLinearSolver < handle

properties
identifier
maxIterations, minIterations
LinearSolver, timeStepSelector, maxTimestepCuts
useRelaxation, relaxationParameter, relaxationType, relaxationIncrement
minRelaxation, maxRelaxation
useLinesearch, linesearchReductionFn
linesearchReductionFactor, linesearchDecreaseFactor, linesearchMaxIterations
linesearchConvergenceNames, linesearchResidualScaling
enforceResidualDecrease, stagnateTol
errorOnFailure, continueOnFailure

The nonlinear solver must call a linear solver and hence has a LinearSolver class object;
the default is the standard backslash solver. Likewise, we need time-step selection, which
in the default implementation does not offer any advanced heuristics for computing optimal
time steps. Furthermore, there are limits on the number of times the control steps can be
subdivided and the maximum/minimum number of iterations within each subdivided step.
If useRelaxation is true, the computed Newton increments should be relaxed by a factor
between 0 and 1. The relaxationParameter is modified dynamically if useRelaxation
is true and you should in general not modify this parameter unless you really know what
you are doing. Relaxation can be any of the following three choices
x_new = x_old + dx; % relaxationType = 'none'
x_new = x_old + dx*w; % relaxationType = 'dampen'
x_new = x_old + dx*w + dx_prev*(1-w); % relaxationType = 'sor'

where dx is the Newton increment and w is the relaxation factor. Similarly, there are various
parameters for invoking and controlling line search.
When enforceResidualDecrease is invoked, the solver will abort if the residual does
not decay more than the stagnation tolerance during relaxation. If errorOnFailure is not
enabled, the solver will return even though it did not converge. Obviously, you should never
rely on non-converged results, but this behavior may be useful for debugging purposes.

Use of Handle Classes

You may have observed that NonLinearSolver is implemented as a subclass of the
handle class. All classes defined so far have been so-called value classes, meaning that
whenever you copy an object to another variable or pass it to a function, MATLAB creates
an independent copy of the object and all the data the object contains. This means that
if you want to change data values of a class object inside a function, you need to return
the class as an output variable. Classes representing nonlinear solvers do a lot of internal
bookkeeping inside a simulator and hence need to be passed as argument to many functions.
By making this class a handle class, the class object is passed by reference to functions,

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 438 The AD-OO Framework for Reservoir Simulation

so that any changes occuring inside the function will also take effect outside the function
without having to explicitly return the class object as output argument.

Computing a Control Step

We have already encountered the primary member function
[state, report, ministates] = solver.solveTimestep(state0, dT, model)

whose purpose is to solve a specified control step. Recall from our discussion of nonlinear
Newton-solvers in Section 10.2.2 that each control step may be subdivided into several
substeps (ministeps) to ensure stable computations and proper convergence. Each ministep
will involve of one or more nonlinear iterations. Each of these may consist of one matrix
inversion if we use a direct linear solver or one or more linear iterations if we use an
iterative solver.
Whether the solver will move forwards using a single time step or multiple substeps
depends on the convergence rate and what the time-step selector predicts to be a feasible
time step for the given state. For the solveTimestep function to work, the model object
must contain a valid implementation of the stepFunction member function declared in
the PhysicalModel class. Driving forces that are active for the particular model must be
passed as optional arguments in the same way as for the incompressible solvers.
Upon completion, the function will return the reservoir state after the time step and
a report structure containing standard information like iteration count, convergence
status, and any other information passed on by model.stepFunction. The cell array
ministates contains all ministeps used to advance the solution a total time step dT.
The class also has member functions for computing these ministeps, applying line search,
stabilizing Newton increments, and implementing checks for stagnation or oscillating
residuals. You have already been exposed to the essential ideas of these operations;
the implementation in the NonLinearSolver class contains more safeguards against
possible errors, has more flexible hooks for pluggable linear solvers and time-step selection
schemes, as well as more bells and whistles for monitoring and reporting the progress of the
solver.

Measures of Nonlinear Convergence

Whether the nonlinear solution process has converged or not is determined by the
PhysicalModel class; the NonLinearSolver class simply responds based on what
the model reports in terms of convergence. There are several different ways convergence
can be measured. The AD-OO framework currently implements the same two measures of
convergence as in ECLIPSE [271].
The first measure is mass balance. When summing residual equations over all cells in
the grid, all intercell fluxes cancel because the flux out of a cell across an internal face has
the same magnitude but opposite sign of the corresponding flux into the neighboring cell.
The sum of residuals therefore equals the difference between the net mass accumulation of
the phase minus the net influx from wells, source terms, and open boundaries, which is the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.3 Solving the Discrete Model Equations 439

definition of the mass-balance error. By convention, these errors are scaled to make them
problem-independent,
 / 
MBα = t B̄α i Rα,i i i , (12.1)

where B̄α is the formation-volume factor of phase α evaluated at average pressure, Rα,i is
the residual equation evaluated in cell i, and i is the pore volume of that cell. The result
can be interpreted as mapping from masses to saturations at field conditions. We say the
solution process is converged when the all scaled residuals are less than the toleranceMB
tolerance declared in the ReservoirModel class.
The second test computes the maximum normalized residual defined as
R 
 α,i 
CNVα = t B̄α max  (12.2)
i i
and says that the solution is converged when this quantity is less than toleranceCNV for all
three phases. Default values for the MB and CNV tolerances are the same as in ECLIPSE
100, i.e., 10−7 and 10−3 , respectively.

12.3.3 Selection of Time-Steps

As indicated in the previous subsection, appropriate selection of time step is crucial to
ensure low runtimes and good convergence in the nonlinear solver. We have already
encountered several selection schemes that use experience from previous time steps to
select optimal time steps, e.g., by halving the time step if the nonlinear solver fails to
converge, or increasing the time step back once we have managed to successfully run a
certain number of chopped steps. Questions to be asked:

• How difficult was it to converge the previous time step? As a simple rule of thumb, a
single iteration is usually considered as too easy, two to three iterations is easy, whereas
more than ten iterations indicate problems with time step or model.
• How does a specified control step compare with the last successful step?
• For a ministep: how far are we away from the end of the control step?
• How does the required number of iterations for the last successful step compare with any
iteration targets?
• Which mechanism determined the size of the previous time step?
Heuristics based on these and similar questions are implemented in the time-selector classes
of MRST, which like NonLinearSolver are handle classes:
classdef SimpleTimeStepSelector < handle
properties
history, maxHistoryLength
isFirstStep, isStartOfCtrlStep
previousControl, controlsChanged, resetOnControlsChanged
firstRampupStep, firstRampupStepRelative
maxTimestep, minTimestep
maxRelativeAdjustment, minRelativeAdjustment
stepLimitedByHardLimits

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 440 The AD-OO Framework for Reservoir Simulation

The class implements a generic member function pickTimestep,

dt = selector.pickTimestep( dtPrev, dt, model, solver, statePrev, state)

which is called by the NonLinearSolver class to determine the next time step. The logic
of this function is as follows: We start by checking whether this is the first step, or if the
well controls have changed and we want to induce a gradual ramp-up

if selector.controlsChanged && ...

(selector.resetOnControlsChanged || selector.isFirstStep);
dt = min(dt, selector.firstRampupStepRelative*dt);
dt = min(dt, selector.firstRampupStep);
selector.stepLimitedByHardLimits = true;
end

The default behavior is to not use such a ramp-up. This means that the default value of
the relative adjustment is set to unity, the lower limit on the first ministep attempted after
controls have changed is set to infinity, and the reset flag is disabled. The time step is then
passed on to the function

dt = selector.computeTimestep( dt, dtPrev, model, solver, statePrev, state)

which will try to compute an optimal adjustment, using some sort of heuristics based
on the stored time-step and iteration history. (By default, this consists of 50 entries.) The
base class offers no heuristics and returns an unchanged time step. The
IterationCountTimeStepSelector subclass tries to select a time step that will
maintain the number of nonlinear iterations as close as possible to a prescribed target,
whereas StateChangeTimeStepSelector attempts to ensure that certain properties of
the state change at prescribed target rates during the simulation. This can often be a good
way of controlling time steps and minimizing numerical error if good estimates of the error
are known; see [58] for more details.
The last step of the selection algorithm is to ensure that the computed time step does
not change too much compared with the previous time step, unless we are at the start of a
new control step. The default choice of tolerances is to allow relative changes of at most
a factor two. Likewise, all time steps are required to be within certain lower and upper
bounds (which by default are 0 and ∞, respectively). Alternative heuristics for choosing
time steps can be introduced by implementing a new subclass of the simple selector base
class or one of its derived classes.

12.3.4 Linear Solvers

The linear solver classes in AD-OO implement methods for solving linearized problems.
All concrete solvers are subclasses of the following handle class:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.3 Solving the Discrete Model Equations 441

classdef LinearSolverAD < handle

properties
tolerance, maxIterations
extraReport, verbose
replaceNaN, replacementNaN, replaceInf, replacementInf
end

The intention of the class is to provide a general interface to all types of linear solvers, both
direct and iterative. Hence, the class declares both a tolerance for the linear solver (default:
10−8 ) and an upper limit on the number of linear iterations (default: 25), even though these
do not make sense for direct solvers. The flag extraReport determines the amount of
reporting; turning it on may consume a lot of memory for problems with many unknowns.
The main interface to linear solvers is through the following member function, which
assembles and solves a linear system:
function [dx, result, report] = solveLinearProblem(solver, problem, model)
problem = problem.assembleSystem();
:
timer = tic();
[result, report] = solver.solveLinearSystem(problem.A, problem.b);
[result, report] = problem.processResultAfterSolve(result, report);
report.SolverTime = toc(timer);
:
dx = solver.storeIncrements(problem, result);

The only place this function is called is within the generic stepFunction implemented
in the PhysicalModel class. The code excerpts above show the main steps of solving a
linear problem: first, the linear system is assembled by the problem object, which is an
instance of the LinearizedProblem class discussed in Section 12.3.1. Next, we call the
main member function solveLinearSystem, which is nonfunctional in the base class and
must be implemented in concrete subclasses. The linear solution is then passed back to the
LinearizedProblem class for potential postprocessing. Postprocessing is not necessary
for the models discussed herein, but the hook is provided for generality. Linear solvers
may sometimes produce infinite values or not-a-number, which the linear solver class can
replace by other values if the replaceInf and/or replaceNaN flags are set; this is done
by two simple if statements. Finally, we extract the results from the solution vector and
store them in the cell array dx, which has one increment entry per primary variable in the
linearized problem.
In addition to the two member functions for solving linearized problems and linear
systems, the LinearSolverAD class implements various utility functions that may be
utilized by concrete linear solvers. This includes interfaces for setup and cleanup of linear
solvers as well as a generic function that can reduce the problem by eliminating some of the
variables and a similar generic function for recovering increments for eliminated variables
from the linear solution. There is also a special function that reduces a problem to cell

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 442 The AD-OO Framework for Reservoir Simulation

variables only, solves it, and then recovers the eliminated variables afterwards. Last, but
not least, the solveAdjointProblem function solves backward problems for computing
adjoints.

MATLAB’s Standard Direct Solver

By far, the simplest approach is to use the standard direct solver. Use of this solver is
implemented very compactly:
classdef BackslashSolverAD < LinearSolverAD
methods
function [result, report] = solveLinearSystem(solver, A, b) %#ok
result = A\b;
report = struct();
end

The resulting solver is very robust and should be able to solve almost any kind of linear
system that has a solution. On the other hand, black-oil models have several unknowns per
cell and can have many nonzero matrix entries because of unstructured grid topologies.
Sparse direct solvers do generally not scale very well with the number of unknowns, which
means that both memory consumption and runtime increase rapidly with model sizes.
MATLAB’s direct solvers should therefore only be used for models having at most a few
tens of thousands of unknowns.

GMRES with ILU Preconditioner

To get a more scalable solver, you need to use an iterative method. Iterative methods used
today to solve large-scale, sparse systems are usually of the Krylov subspace type that seek
to find an approximation to the solution of the linear system Ax = b in the Krylov space
K formed by repeatedly applying the matrix A to the residual r 0 = b − Ax 0 , where x 0 is
some initial guess:

Kν = Kν (A,r 0 ) = span{r 0,Ar 0,A2 r 0, . . . ,Aν−1 r 0 }.

One widely used example is the generalized minimal residual (GMRES) method, which
seeks the approximate solution as the vector x ν that minimizes the norm of the ν’th residual
r ν  = b − Ax ν . Krylov vectors are often linearly dependent, and hence GMRES
employs a modified Gram–Schmidt orthogonalization, called Arnoldi iteration, to find an
orthonormal basis for Kν .
The GMRES iteration is generally robust and works well for nonsymmetric systems of
linear equations. Iterative solvers are nonetheless only efficient if the condition number of
the matrix is not too high. Black-oil models often give linear systems that have quite bad
condition numbers. To remedy this, it is common to use a preconditioner. That is, instead
of solving Ax = b, we first construct a matrix B that is inexpensive to invert. Expanding,
we have b = AB −1 Bx = (AB −1 )y, which is less expensive to solve if AB −1 has better
condition number than A.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.3 Solving the Discrete Model Equations 443

Herein, we apply an incomplete LU factorization (ILU) method. The true LU factors

for a typical sparse matrix can be much less sparse than the original matrix, and in an
incomplete factorization, one seeks triangular matrices L and U such that LU ≈ A. When
used as a preconditioner, it is common to construct L and U so that these matrices preserve
the sparsity structure of the original matrix, i.e., that we have no fill-in of new nonzero
elements. The result is called ILU(0).
The GMRES_ILUSolverAD class implements a GMRES solver with ILU(0) precondi-
tioning by use of the built-in gmres and ilu solvers from MATLAB. As for the direct
solver, the implementation is quite compact:
function [result, report] = solveLinearSystem(solver, A, b)
nel = size(A, 1);
if solver.reorderEquations
[A, b] = reorderForILU(A, b);
end
[L, U] = ilu(A, solver.getOptsILU());
prec = @(x) U\(L\x);
[result, flag, res, its] = ...
gmres(A, b, [], solver.tolerance, min(solver.maxIterations, nel), prec);
report = struct('GMRESFlag', flag, 'residual', res, 'iterations', its);
end

Notice that we permute the linear system to ensure nonzero diagonals to simplify the con-
struction of the ILU preconditioner. Notably, this utility is useful whenever well equations
are added, since these may not have derivatives with respect to all well controls.
Well equations generally consist of a mixture of different variables and can hence be
badly scaled. Control equation on pressure gives residuals having magnitudes that are
typically O(107 ) since all equations are converted to SI units in MRST. Rate equations,
on the other hand, give residuals that typically are O(10−5 ). To be on the safe side, we
overload the solver for linearized problems by a new function that eliminates all equations
that are not posed on grid cells by use of a block-Gaussian algorithm, solve the resulting
problem by the GMRES-ILU(0) method, and then recover the eliminated variables
function [dx, result, report] = solveLinearProblem(solver, problem, model)
keep = problem.indexOfType('cell');
[problem, eliminated] = solver.reduceToVariable(problem, keep);
problem = problem.assembleSystem();

timer = tic();
[result, report] = solver.solveLinearSystem(problem.A, problem.b);
report.SolverTime = toc(timer);

dxCell = solver.storeIncrements(problem, result);

dx = solver.recoverResult(dxCell, eliminated, keep);

We eliminate well equations (and other non-cell equations) by use of member functions
from the LinearSolverAD base class, which in turn relies on block-Gaussian algorithms

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 444 The AD-OO Framework for Reservoir Simulation

implemented in the LinearizedProblem class. This algorithm is not specific to the

GMRES-ILU(0) and in the actual code the content of this function is implemented as a
member function of the LinearSolverAD class.

Constrained Pressure Residual (CPR)

The state-of-the-art approach for solving linear systems arising from black-oil type models
is to use a so-called constrained pressure residual (CPR) method [305, 119]. To explain the
method, we start by writing the equation for the Newton update in block-matrix form




J pp J ps x p Rp
− = .
J sp J ss x s Rs

Here, we have decomposed the unknown Newton increment in primary unknowns (p),
which typically consist of pressure variables, and secondary unknowns (s) that consist of
saturations and mass fractions. The Jacobian matrix blocks read (J m, )i,j = ∂R ,i /∂x m,j ,
where the indices m, run over the variable types {p,s} and i,j run over all cells.
Instead of applying an iterative solver to the full linear system, we first form an (almost)
elliptic equation for the primary unknowns. This equation can be achieved by modifying
the fully implicit system to the form




J ∗pp J ∗ps x p R ∗p
− = ,
J sp J ss x s Rs

where the off-diagonals of the coupling matrix J ∗ps are small, so that J ∗pp x p = R ∗p
resembles an incompressible pressure equation. Because this CPR pressure equation is
almost elliptic, we can use on of the many efficient solvers developed for Poisson-type
equations to compute an approximate increment x p in an inexpensive manner. This
predictor stage is used in combination with a corrector stage, in which we iterate on the
entire system using the approximated pressure update as our initial guess.
The solver implemented in the CPRSolverAD class employs an IMPES-like reduction
(plus dynamic row sums from [119]), in which we determine appropriate weights wc so
that the accumulation terms of the residual equations cancel
   n+1   n
wc Ac = 0, Ac,i = φ bα Sα xc,α − φ bα Sα xc,α .
c α i α i

Here, xc,α gives the volume of component c at standard conditions present in a unit stock-
tank of phase α at reservoir conditions, i.e., xo,o = 1, xo,g = Rv , and so on. The (almost)

decoupled CPR pressure equation is then taken to be R ∗p = c wc Rc . Notice that if compo-
nents are in one phase only, the weight of component c is simply 1/bα , where α is the phase
of component c. An advantage of this decoupling strategy is that the resulting pressure
block J ∗pp is fairly symmetric. Moreover, the decoupling can be done at a nonlinear level,
before the Jacobians are computed, and is thus less prone to errors.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.3 Solving the Discrete Model Equations 445

Aggregation-Based Algebraic Multigrid

The pressure equation in the CPR solver can obviously be solved using a standard direct
solver. However, a more scalable approach is to use an iterative multilevel method. To
present the idea, we consider only two levels: the fine level on which we seek the unknown
solution x and a coarser level on which we seek another unknown vector x̃ having much
fewer elements. To move between the two levels, we have a prolongation operator P and a
restriction operator Q. Then, the iterative method can be written as follows:

x ∗ = x ν + S(b − Ax ν ),
x ν+1 = x ∗ + P A−1 Q(b − Ax ∗ ).

Here, S is a so-called smoother, which typically consists of a few steps of an inexpensive

iterative solver (like Jacobi’s method or ILU(0)). For a multilevel method, this procedure
is performed recursively on a sequence of coarser levels. In classical methods, the coarser
levels are chosen by geometric partitions of the original grid. The state-of-the-art approach
is to select the coarse levels algebraically and build the operator P by analyzing the matrix
coefficients of A. The restriction operator is then chosen as the transpose of the interpola-
tion matrix, Q = P T , and the coarse matrix is Ac = P T AP .
The AGMGSolverAD class is an interface to the aggregation-based algebraic multigrid
(AGMG) solver [238, 15], which builds prolongation operators using recursive aggregates
of cells with constant interpolation so that the operator on each level is as sparse as possible
to avoid that the coarse matrices Ac become more dense the more one coarsens. AGMG
was originally released as free open-source, but recent versions are only freely available
for academic users. The main advantage of AGMG, beyond a simple MATLAB interface,
is that it has low setup cost and low memory requirements. In my experience, it is easy to
integrate and seems to work very well with MRST. For the black-oil equations, this solver
should either be used as a preconditioner for the CPR solver, or as a pressure solver for the
sequentially implicit method implemented in the blackoil-sequential module.
AGMG can also be run in two stages: a setup stage that builds the coarse hierarchies and
the prolongation operators, and a solution stage that employs the prolongation operators to
solve a specific problem. This way, you can hope to reduce computational costs by reusing
the setup stage for multiple subsequent solves. This behavior can be set up as follows:
function [result, report] = solveLinearSystem(solver, A, b)
cleanAfter = false;
if solver.setupDone
solver.setupSolver(A, b); cleanAfter = true;
end
if solver.reuseSetup
fn = @(A, b) agmg(A, b, [], solver.tolerance, ...
solver.maxIterations, [], [], 1);
else
fn = @(A, b) agmg(A, b, [], solver.tolerance, solver.maxIterations);
end

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 446 The AD-OO Framework for Reservoir Simulation

[result, flag, relres, iter, resvec] = fn(A, b);

if cleanAfter, solver.cleanupSolver(A, b); end
end
function solver = setupSolver(solver, A, b, varargin)
if solver.reuseSetup
agmg(A,[],[],[],[],[],[], 1); solver.setupDone = true;
end
end

For brevity, we have dropped a few lines that create the full convergence report. This
example hopefully shows that it is not very difficult to set up an external solver with MRST
if the solver has a standard MATLAB interface.

Interface to External Linear Solvers

Recently, we added a new module called linearsolver to MRST, whose purpose is to
offer a uniform interface to other external solvers on the form:
x = solver(A, b, tolerance, iterations)

So far, the module implements support for AMGCL [81], which is a recent and highly
efficient header-only C++ library for solving large sparse linear systems with algebraic
multigrid methods [296, 285]. AMGCL can either be used as a multigrid solver or as a
smoother, and offers a variety of basic building blocks that can be combined in different
ways:
• smoothers: damped Jacobi, Gauss–Seidel, ILU(0), ILU(k), sparse approximate inverse,
and others;
• coarsening methods: aggregation, smoothed aggregation, smoothed aggregation with
energy minimization, and Ruge–Stüben; and
• Krylov subspace solvers: conjugated gradients, stabilized biconjugate gradients, GMRES,
LGMRES, FGMRES, and induced dimension reduction.
To use AMGCL as a CPR solver, you can, e.g., set it up as follows:
lsolve = AMGCL_CPRSolverAD('block_size', 2, ...
'maxIterations', 150, 'tolerance', 1e-3);
lsolve.setCoarsening('aggregation')
lsolve.setRelaxation('ilu0');
lsolve.setSRelaxation('ilu0');
lsolve.doApplyScalingCPR = true;
lsolve.trueIMPES = false;
:

At the time of writing, AMGCL is not part of the official MRST release and must be
downloaded from a software repository [81], compiled, and integrated via a MEX interface.
Initial tests show that AMGCL is significantly faster than AGMG, and installing it is
definitely worth the effort if you want to use MRST on models with hundreds of thousand
unknowns.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.3 Solving the Discrete Model Equations 447

ECLIPSE Solver: Orthomin and Nested Factorization

The default solver in ECLIPSE is a Krylov subspace method that relies on a set of
search vectors defined so that each new vector q ν+1 is orthogonal to all previous vectors
q 0,q 1, . . . ,q ν . Assuming a preconditioner B and an initial guess x 0 , we start the iteration
process by setting q 0 = B −1 r 0 , where r 0 = b − Ax 0 . The next approximate solution is
then defined as

x ν+1 = x ν + ων B −1 q ν ,

which inserted into the linear system gives the residual

r ν+1 = r ν − ων AB −1 q ν = r ν − ων q̂ ν .

The relaxation parameter ων is optimized by minimizing r ν+1 2 . The resulting orthomin

method is a special version of the conjugate gradient method. For completeness, let us
derive the remaining steps. The optimum value for ων+1 is determined by setting
∂ ∂  
0= r ν+1 2 = r ν · r ν − 2ων r ν · q̂ ν + (ων )2 q̂ ν · q̂ ν
∂ων+1 ∂ων+1
r ν · q̂ ν
=−2r ν · q̂ ν + 2ων q̂ ν · q̂ ν −→ ων = ,
q̂ ν · q̂ ν
The search vectors are set to lie in the space spanned by r ν+1 and all the previous search
vectors, i.e.,

ν
q ν+1 = r ν+1 + αν q ,
=0

subject to the constraint that q̂ ν+1 is orthogonal to all q̂ k for k ≤ ν, i.e.,

 
ν

0 = q̂ ν+1 · q̂  = AB −1 r ν+1 + αν q̂  · q̂ k .
=0

Using the fact that q̂  · q̂ k for ,k ≤ ν and  = k, we can easily show that

(AB −1 r ν+1 ) · q̂ 
αν = − .
q̂  · q̂ 
This completes the definition of the orthomin method.
For grids having a rectangular topology, it is possible to develop a particularly efficient
preconditioner called nested factorization [24]. As we have seen previously, MRST uses
a global numbering of the unknowns, in which the numbering of unknowns first runs
over cells and then over variables; this to utilize highly efficient vectorization from MAT-
LAB. In a simulator written in a compiled language, it is usually more efficient to use a
local numbering that first runs over variables and then over cells. With a standard TPFA

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 448 The AD-OO Framework for Reservoir Simulation

and single-point upwind-mobility discretization, the resulting system will have a block-
heptagonal form,
A = L 3 + L2 + L1 + D + U 1 + U 2 + U 3
where D is a diagonal matrix whose elements are m × m matrices representing the
unknowns per cell. The block matrices Li and U i are lower and upper bands representing
connections to neighboring cells in logical direction i = 1,2,3. Experience shows that the
numbering should be chosen so that direction 1 corresponds to the direction of highest
transmissibilities, which usually is the z-direction; see [271], whose presentation we
henceforth follow. The key idea of nested factorization is now to utilize this special
structure to recursively define a set of preconditioners, whose action can be computed
iteratively by solving block-tridiagonal systems. These can be solved very efficiently
by Thomas’s algorithm that consists of two substitution sweeps, one forward and one
backward. The factorization is built iteratively using the following definition
B = (P + L3 )P −1 (P + U 3 ), (12.3)
−1
P = (T + L2 )T (T + U 2 ), (12.4)
T = (G + L1 )G−1 (G + U 1 ). (12.5)
With a slight abuse of notation, we should understand the terms on the right-hand side as
values from the previous iteration. The matrix G is a block-diagonal matrix, which for a
fully implicit discretization is defined as,
 
G = D − L1 G−1 U 1 − colsum L2 T −1 U 2 − colsum L3 P −1 U 3 . (12.6)
This definition ensures that although the solution may not yet be correct, it will have no
mass-balance errors. Alternatively, in the IMPES and AIM methods, we can construct G
using row sum rather than column sum to ensure accurate pressure.
Let us now look at how the solution proceeds: Starting at the outmost level, we see that
to solve By = r, we need to solve (P + L3 )(I + P −1 U 3 )y = r. This can be decomposed
as follows

ŷ = P −1 r − L3 ŷ , y = ŷ − P −1 U 3 y
If ŷ is known in the first plane of cells, solving these two equations reduces to a forward
elimination and a backward substitution, since the products L3 ŷ and U 3 y only involve
values along planes where the solution is known. In this procedure, we need to invert P
and compute vectors z = P −1 u along planes of cells. Applying the same idea to (12.4),
we obtain

ẑ = T −1 u − L2 ẑ , z = ẑ − T −1 U 2 z,
which we can solve by a forward elimination and a backward substitution provided we
know the solution along a first line of cells in each plane. Here, we must compute w =
T −1 v on each line, which again can be decomposed and solved efficiently as follows:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.4 Simulation Examples 449
−1

ŵ = T v − L1 ŵ , w = ŵ − T −1 U 1 w,
provided we know G along the lines. From the structure of (12.6), it follows that we can
construct G one cell at a time: Starting from a known inverse block G−1 in a cell, we can
compute the contribution L1 G−1 U 1 to the next cell in direction 1, and so on. Once
 G−1 is
known along a line in direction 1, we can compute the contribution of colsum L2 T −1 U 2
using the forward-elimination, backward-substitution algorithm for T −1. When all the
lines
in a plane are known, we can compute the contributions from colsum L3 P −1 U 3 to the
next plane in a similar manner. The method may look laborious, but is highly efficient.
Unfortunately, it is not applicable to fully unstructured grids.

12.4 Simulation Examples

You have now been introduced to all key components that make up the AD-OO framework.
In the following, we demonstrate in more detail how to use the AD-OO framework to
perform simulations with or without the use of structured input decks on the ECLIPSE
format.

12.4.1 Depletion of a Closed/Open Compartment

The purpose of this example1 is to demonstrate how to set up a simulator from scratch with-
out the use of input files. To this end, we consider a 2D rectangular reservoir compartment
with homogeneous properties filled by a single-phase fluid. The reservoir fluid is drained
by a single producer at the midpoint of the top edge. The compartment is either closed
(i.e., sealed by no-flow boundary conditions along all edges), or open with constant pres-
sure support from an underlying, infinite aquifer, which we model as a constant-pressure
boundary condition.
We start by making the geological model and fluid properties
G = computeGeometry(cartGrid([50 50],[1000 100]));
rock = makeRock(G, 100*milli*darcy, 0.3);
pR = 200*barsa;
fluid = initSimpleADIFluid('phases','W', ... % Fluid phase: water
'mu', 1*centi*poise, ... % Viscosity
'rho', 1000, ... % Surface density [kg/m^3]
'c', 1e-4/barsa, ... % Fluid compressibility
'cR', 1e-5/barsa)); ... % Rock compressibility

This is all we need to set up a single-phase reservoir model represented by WaterModel,

which is a specialization of the general ReservoirModel implemented in ad-core. The
only extra thing we need to do is to explicitly set the gravity direction. By default, gravity

1 Complete code: blackoilTutorialOnePhase.m in ad-blackoil.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 450 The AD-OO Framework for Reservoir Simulation
−4
x 10 qWr: Well reservoir rate (water) Open compartment w/aquifer Closed compartment
100 100 200
2

80 80 190
P1 (Aquifer)
m /timestep

P1 (Closed)
60 60 180
1
3

40 40 170

20 20 160

0
5 10 15 20 25 30 35 40 0 0 150
Step # 0 500 1,000 0 500 1,000

Figure 12.6 Drainage from a single reservoir compartment that is either sealed or supported by an
infinite aquifer from below.

in MRST is a three-component vector that points in the positive z-direction. Here, we set it
to a two-component vector pointing in the negative y-direction.
gravity reset on
wModel = WaterModel(G, rock, fluid,'gravity',[0 -norm(gravity)]);

As always, we also need to specify the drive mechanisms by use of the standard data
structures for wells and boundary conditions:
wc = sub2ind(G.cartDims, floor(G.cartDims(1)/2), G.cartDims(2));
W = addWell([], G, rock, wc, ...
'Type', 'bhp', 'Val', pR - 50*barsa, ...
'Radius', 0.1, 'Name', 'P1','Comp_i',1,'sign',1);
bc = pside([],G,'South',200*barsa,'sat',1);

We then create two simulation schedules, one with well and constant pressure on the south
boundary, and the other with well and no-flow conditions on all boundaries:
schedule1 = simpleSchedule(diff(linspace(0,5*day,41)), 'bc', bc, 'W', W);
schedule2 = simpleSchedule(diff(linspace(0,5*day,41)), 'W', W);

The pristine reservoir is in vertical equilibrium, which we compute as a steady-state solu-

tion of the flow equation subject to boundary conditions only
state.pressure = ones(G.cells.num,1)*pR;
state.wellSol = initWellSolAD([], wModel, state);
nonlinear = NonLinearSolver;
state = nonlinear.solveTimestep(state, 10000*day, wModel, 'bc', bc);

We can the compute each flow solution as follows:

[wellSols1, states1] = simulateScheduleAD(state, wModel, schedule1);

Figure 12.6 shows the solution of the two problems. Once the well is turned on, the draw-
down at the wellbore will create an elongated pressure pulse that moves fast downward and

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.4 Simulation Examples 451

expands slowly outward. The open compartment quickly reaches a steady-state in which
the flow rate into the well equals the influx across the bottom boundary. In the closed case,
the pressure inside the compartment is gradually depleted towards the perforation pressure,
which in turn causes the production rate to decline towards zero.

12.4.2 An Undersaturated Sector Model

The next continues in the same vein as the previous: we consider a 1,000 ×
example2
1,000 × 100 m3 rectangular sector model filled with undersaturated oil and compare and
contrast the difference in production when the boundaries are closed or held at constant
pressure. The phase behavior of the oil and dissolved gas is assumed to follow that of the
SPE 1 benchmark, but unlike in the original setup, water is assumed to be mobile, following
a simple quadratic relative permeability.
[, , fluid, deck] = setupSPE1(); fluid.krW = @(s) s.^2;

Permeability is homogeneous and isotropic, and saturation and pressure are for simplicity
assumed to be uniformly distributed inside the reservoir:
[s0, p0] = deal([0.2, 0.8, 0], 300*barsa);
state0 = initResSol(G, p0, s0);
state0.rs = repmat(200, G.cells.num, 1);
model = ThreePhaseBlackOilModel(G, rock, fluid, 'disgas', true);

The reservoir is produced from a single vertical well, placed in the center, operating at a
fixed oil rate, but with a lower bottom-hole pressure (bhp) limit:
ij = ceil(G.cartDims./2);
W = verticalWell([], G, rock, ij(1), ij(2), [], ...
'val', -0.25*sum(model.operators.pv)/T, ...
'type', 'orat', 'comp_i', [1, 1, 1]/3, ...
'sign', -1, 'name', 'Producer');
W.lims.bhp = 100*barsa;

The AD-OO framework assumes no-flow boundary conditions by default, and hence we
do not need to specify boundary conditions to simulate closed boundaries. For the other
case, we assume that the pressure and fluid composition at the boundary are held at their
initial values to model additional pressure support, e.g., from a gas cap or an aquifer with
highly mobile water. Although such boundary conditions have been used often in the past,
there is in reality no physical mechanism that would be able to maintain constant boundary
conditions in a reservoir, so this is a gross simplification that should not be applied when
modeling real cases.

2 Complete code: blackoilSectorModelExample.m in ad-blackoil

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 452 The AD-OO Framework for Reservoir Simulation

bc = []; sides = {'xmin', 'xmax', 'ymin', 'ymax'};

for side = 1:numel(sides)
bc = pside(bc, G, sides{side}, p0, 'sat', s0);
end

We also need to specify the Rs value on all boundaries. For this purpose, MRST employs
a dissolution matrix, which can either be specified per cell interface, or for all interfaces.
Here, the fluid composition is the same on the whole boundary, so we only specify a single
matrix
bc.dissolution = [1, 0, 0;... % Water fractions in phases
0, 1, 0; ...% Oil fractions in phases
0, 200, 1]; % Gas fractions in phases

To define specific dissolution factors on n individual cell faces, we would have to set up an
n × 3 × 3 array.
Turning on the well induces a fast pressure transient that lasts a few days. We do not
try to resolve this, but gradually ramp up the time step, as discussed earlier in the book, to
improve the initial stability of the simulation:
dt = rampupTimesteps(T, 30*day);

Specifically, the routine splits the simulation horizon T into a number of time step of length
30 days (the last step may be shorter if T is not a multiple of the specified time step). It
then uses a geometric sequence 1./2.^[n n:-1:1] to further subdivide the first time step
(default value of n is 8). We can now set up and simulate the two schedules in the same
way as discussed in the previoius subsection
schedule = simpleSchedule(dt, 'W', W, 'bc', bc);
[ws, states] = simulateScheduleAD(state0, model, schedule);

Figure 12.7 shows bottom-hole pressure and surface oil rate for the two simulation,
i.e., the quantities used to control the production well. After the initial transient, the case
with constant pressure boundary settles at a steady state with constant bhp and oil rate
maintained at its target value. With closed boundaries, there is no mechanism to maintain
the reservoir pressure, and hence the bottom-hole pressure starts to gradually decay. After
870 days, the bhp drops below its lower limit of 100 bar, and the well switches control from
oil rate to bhp. As production continues, the reservoir pressure will gradually be depleted
until the point where there is not sufficient pressure difference between the reservoir and
the well to drive any flow into the wellbore; this happens after approximately 4.5 years.
Grossly speaking, qualitative behavior of the pressure is similar to what we observed in the
previous single-phase example.
Figure 12.8 reports cumulative production of oil, gas, and water at surface conditions,
as well as instantaneous production rates of the oleic, gaseous, and aqueous phases at

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.4 Simulation Examples 453

Bottom-hole pressure [bar] Surface oil rate [103 m3/day]

300
2

250
1.5
Pressure Pressure
200
Closed 1 Closed
150
0.5

100
0
0 2 4 6 8 10 0 2 4 6 8 10
Time [years] Time [years]

Figure 12.7 Bottom-hole pressure and surface oil rate for the undersaturated sector model; dashed
lines show field-average pressure. With closed boundaries, the well control switches from oil surface
rate to bottom-hole pressure after 870 days (shown as dotted lines).

reservoir conditions. As the reservoir pressure is gradually lowered, some cells drop below
the bubblepoint and free gas is liberated, starting near the wellbore, where the pressure drop
is largest, and gradually moving outward as shown in Figure 12.9.
Oil is only produced from the oleic phase. When free gas is liberated, the fraction of oil
in the oleic phase increases and the oil reservoir rate decays slightly because the reservoir
volume of the oleic phase required to produce a constant surface rate of oil decreases.
The liberated gas is very mobile and causes an accelerating pressure drop as more gas
starts flowing into the wellbore. Gas production at the surface thus increases dramatically,
because gas is now produced both from the oleic and the gaseous phase. Gas can sometimes
be sold and contribute to generate revenue, but the operator may also have to dispose of it
by reinjecting it into the reservoir to maintain reservoir pressure. Unlike in the original
SPE 1 case, the aqueous phase is mobile. As gas is liberated, the mobility of oil is reduced
and we also see a significant increase in water production. Topside facilities usually have
limited capacity for handling produced water, and the whole production may be reduced
or shut down if the water rate exceeds the available capacity. Here, it might have been
advantageous if the well was instrumented with autonomous inflow devices that would
choke back water and gas and only let oil flow through. We will return to this in the next
example.
As the well switches to bhp control, all rates fall rapidly. For oil in particular, the
additional recovery after the well switches is very limited: at 870 days, the cumulative oil
production has reached 93% of what is predicted to be the total production after 10 years.
The liberated gas (and the water) will continue to flow longer, but also here the production
ceases more or less after 4.5 years.
This idealized example is merely intended to demonstrate use of the AD-OO framework
and illustrate basic flow mechanisms and makes no claim of being representative of real
hydrocarbon reservoirs.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 454 The AD-OO Framework for Reservoir Simulation
6
Cumulative oil production [10 m ]
3
Oil reservoir rate [103 m3/day]

3
6
2.5
2
4 Pressure Pressure
Closed 1.5 Closed

2 1
0.5
0 0
2 4 6 8 10 2 4 6 8 10
Time [years] Time [years]
6 3 3 3
Cumulative gas production [10 m ] Gas reservoir rate [10 m /day]
1,500
30
25
1,000
20
Pressure
15 Closed
500 10
Pressure
5
Closed
0 0
2 4 6 8 10 2 4 6 8 10
Time [years] Time [years]

Cumulative water production [106 m3]

1

0.8

0.6

0.4

0.2 Pressure
Closed
0
2 4 6 8 10
Time [years]

Figure 12.8 Production profiles for the undersaturated sector model. The left column shows
cumulative production of the oil, gas, and water components at surface conditions, whereas the right
column shows instantaneous phase rates at reservoir conditions.

Figure 12.9 Liberation of free gas as the drawdown from the well causes the pressure in the
undersaturated oil to drop below the bubblepoint. (The plot shows Sg >5e-3.)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.4 Simulation Examples 455

Computer exercises
12.4.1 Try to simulate the initial pressure transient, check how long it lasts, and whether
it behaves as in the previous example.
12.4.2 The simulation discussed in this section did not include any consideration of pro-
duction relative to the amount of fluids initially in place. Try to include this to
determine the recovery factor for the closed case. Can you also use this to say
something about the influx across the boundaries in the case with pressure bound-
aries.
12.4.3 With constant pressure boundaries, the water rate starts to increase after approxi-
mately seven years. Can you explain this behavior?
12.4.4 Gravity was not included in this example. Can you set up multilayer simulations
to investigate if gravity has a significant effect on the result?

12.4.3 SPE 1 Instrumented with Inflow Valves

This example3 demonstrates the use of multisegment wells in MRST. The default well
model in MRST assumes that flow in the well takes place on a very short time-scale
compared to flow in the reservoir and thus can be accurately modeled as being instanta-
neous, following the linear inflow performance relationship discussed in Section 11.7.1.
Such models are not adequate for long wellbores that have significant pressure drop due
to friction, or for wells having valves or more sophisticated inflow control devices. The
multisegment well class in MRST supports transient flow inside the well and enables a more
fine-grained representation of the well itself that allows the effects of friction, acceleration,
and nonlinear pressure drops across valves and other flow constrictions to be included; see
Section 11.7.2.
To illustrate, we revisit the SPE 1 case from Section 11.8.1 and replace the vertical
producer by a horizontal well completed in six cells with valves between the production
tubing and each sandface connection. We consider two models, a simple model assuming
instantaneous flow with a standard inflow relationship and hydrostatic pressure along the
wellbore, and a more advanced model that accounts for pressure drop across the valves and
frictional pressure drop along the horizontal wellbore; see Figure 12.10.
First, we load the description of the fluid model and the simulation schedule from an
ECLIPSE input deck:
[G, rock, fluid, deck, state] = setupSPE1();
[nx, ny] = deal(G.cartDims(1),G.cartDims(2));

model = selectModelFromDeck(G, rock, fluid, deck);

model.extraStateOutput = true;
schedule = convertDeckScheduleToMRST(model, deck);

3 Complete source code: multisegmentWellExample.m in ad-blackoil

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 456 The AD-OO Framework for Reservoir Simulation

Figure 12.10 Multisegment well for the modified SPE 1 case. Nodes 1–7 represent the wellbore,
whereas nodes 8–13 represent the void outside of the valves that separate the wellbore from the
sandface.

The model selector examines which phases are present and whether the deck prescribes
dissolved gas or vaporized oil, and then determines the specific black-oil subclass that is
most suitable for simulating the deck. As we already know, the SPE 1 benchmark describes
a three-phase model with dissolved gas but no vaporized oil. The model will thus be an
instance of the general ThreePhaseBlackOilModel class. Likewise, the schedule con-
verter will parse the specification of all wells and create a suitable schedule object, which
here has 120 steps and a single control.
To specify the well models, we first initialize the production well as a standard well
structure completed in the six given grid cells:
c = nx*ny + (2:7)';
prod0 = addWell([], G, rock, c, 'name', 'prod', ...
'refDepth', G.cells.centroids(1,3), ...
'type', 'rate', 'val', -8e5*meter^3/day);

We then define the N1,N2 mappings for the twelve segments, as well as the mapping from
network nodes to reservoir cells represented in the well structure
topo = [1 2345623 4 5 6 7
2 3 4 5 6 7 8 9 10 11 12 13]';
% | tubing | valves |
cell2node = sparse((8:13)', (1:6)', 1, 13, 6);

The cell2node matrix has dimension equal the number of nodes times the number of
reservoir cells perforated by the well. We must also specify the segment lengths and diam-
eters and the node depths and volumes
lengths = [300*ones(6,1); nan(6,1)];
diam = [.1*ones(6,1); nan(6,1)];
depths = G.cells.centroids(c([1 1:end 1:end]), 3);
vols = ones(13,1);

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.4 Simulation Examples 457

With this, we have all the information we need to convert the data structure of the simple
well into a corresponding data structure describing a multisegment network model:
prodMS = convert2MSWell(prodS, 'cell2node', cell2node, 'topo', topo, ...
'G', G, 'vol', vols, 'nodeDepth', depths, ...
'segLength', lengths, 'segDiam', diam);

Finally, we must set up the flow model for each segment. The first six segments are modeled
using the wellbore-friction model (11.29)–(11.31) with roughness e = 10−4 m and dis-
charge coefficient Cv = 0.7. The next six segments represent valves, whose pressure losses
are modeled by use of (11.29) with mixture mass flux. The valves have thirty openings, each
with a nozzle diameter of 25 mm. We then set up the flow model as function of velocity,
density, and viscosity
[wbix, vix] = deal(1:6, 7:12);
[roughness, nozzleD, discharge, nValves] = deal(1e-4, .0025, .7, 30);

prodMS.segments.flowModel = @(v, rho, mu)...

[wellBoreFriction(v(wbix), rho(wbix), mu(wbix), ...
prodMS.segments.diam(wbix), ...
prodMS.segments.length(wbix), roughness, 'massRate'); ...
nozzleValve(v(vix)/nValves, rho(vix), nozzleD, discharge, 'massRate')];

The injector is a simple well modeled by the standard inflow performance relationship, i.e.,
a Peaceman type well model,
inj = addWell([], G, rock, 100, 'name', 'inj', 'type', 'rate', 'Comp_i', ...
[0 0 1], 'val', 2.5e6*meter^3/day,'refDepth', G.cells.centroids(1,3));

To compare the effect of instrumenting the production well with inflow valves, we run two
different schedules: The first is a baseline simulation in which the producer is treated as
a simple well with instantaneous flow and one node per contact between the well and the
reservoir:
schedule.control.W = [inj; prodS];
[wellSolsSimple, statesSimple] = simulateScheduleAD(state, model, schedule);

For the second simulation, we must combine the simple injector and multisegment producer
into a facility model since the two wells are represented with different types of well models
W = combineMSwithRegularWells(inj, prodMS);
schedule.control.W = W;
model.FacilityModel = model.FacilityModel.setupWells(W);

This is normally done automatically by the simulator, but here we do it explicitly on the
outside to view the output classes. These classes are practical if we want to incorporate
per-well adjustments to tolerances or other parameters. The model classes and the well
structures read

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 458 The AD-OO Framework for Reservoir Simulation

SimpleWell with properties: MultisegmentWell with properties:

signChangeChop: 0
W: [1x1 struct] W: [1x1 struct]
allowCrossflow: 1 allowCrossflow: 1
allowSignChange: 0 allowSignChange: 0
allowControlSwitching: 1 allowControlSwitching: 1
dpMaxRel: Inf dpMaxRel: Inf
dpMaxAbs: Inf dpMaxAbs: Inf
dsMaxAbs: 0.2000 dsMaxAbs: 0.2000
VFPTable: [] VFPTable: []
operators: [] operators: [1x1 struct]
nonlinearTolerance: 1.0000e-06 nonlinearTolerance: 1.0000e-06
G: [] G: []
verbose: 0 verbose: 0
stepFunctionIsLinear: 0 stepFunctionIsLinear: 0

W is structure with fields: W is structure with fields:

cells: 100 cells: [6x1 double]
type: ’rate’ type: ’bhp’
val: 28.9352 val: 25000000
r: 0.1000 r: 0.1000
: :
cell2node: [] cell2node: [13x6 double]
connDZ: [] connDZ: [6x1 double]
isMS: 0 isMS: 1
nodes: [] nodes: [1x1 struct]
segments: [] segments: [1x1 struct]

The main difference in the two classes is that the SimpleWell model does not have any
operators defined. Likewise, this model does not have any representation of cell to node
mappings and structures representing the nodes and segments in the W structure. However,
what MRST uses to distinguish simple and multisegment wells is the flag isMS, which is
set to false for the injector and true for the producer. For completeness, let us also look at
the data structures representing nodes and segments for the instrumented producer
segments: nodes:
length: [12x1 double] depth: [13x1 double]
roughness: [12x1 double] vol: [13x1 double]
diam: [12x1 double] dist: [13x1 double]
flowModel: [function_handle]
topo: [12x2 double]
deltaDistance: [12x1 double]

The crucial element here is the function handle flowModel in the segments, which is used
to assemble the well equations.
Figure 12.11 shows bhp in the producer and injector, along with field-average pressure
predicted by the two different well models. Resistance to flow when the flow is constricted
through the nozzles of the valves and by friction along the tubing causes a significantly
higher pressure drop from the sandface to the datum point in the wellbore. Hence, the bhp
must be significantly lower for the multisegment model, compared with the simple well
model to sustain the same surface production rate4 . Looking at the injector bhp, we see that
a relatively high pressure is required initially to force the first gas into the reservoir. As soon
as the first gas has entered the reservoir, the total mobility in the near-well region increases

4 Notice that this comparison is somewhat exaggerated. In reality, one would likely have adjusted either the well index or the
skin factor in the simple well model to account for some of this pressure loss. On the other hand, such representative values
can only be found after one has observed some production history.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.4 Simulation Examples 459
Injector bottom-hole pressure [bar] Producer bottom-hole pressure [bar]

520
400
500
300
480

460 200

440 Multisegment
100
Standard
420
0.5 1 1.5 2 2.5 3 0.5 1 1.5 2 2.5 3
Time [years] Time [years]

Figure 12.11 Comparison of bottom-hole pressures in the injector (left) and the producer (right)
predicted for the modified SPE 1 model with a simple well model and a well model accounting for
pressure drop across valves and caused by friction along the horizontal wellbore. Thin lines in the
right plot show average-field pressure.

and the injection pressure decays. With some gas accumulated, the total flow resistance of
the pore space from injector to producer will decrease, and hence the bhp starts to increase
in the producer after approximately 60 days. On the other hand, the injected gas is not able
to displace a similar amount of reservoir fluids from the reservoir because of lower density
and higher compressional effects, and hence the reservoir pressure will gradually increase.
This effect is observed in both models.
After approximately 20 months, the first gas has reached a neighboring cell of the
producer and will start to engulf the cells above and gradually be drawn down into the
perforated cells; see Figure 12.12. This engulfing process causes a significant decay in the
oil rate and a steeper incline in producer bhp. The rate decay is slightly delayed in the
multisegment model, but the effect on the total oil production is a mere 1.3% increase
(which still could represent significant income).
Figure 12.13 shows difference in pressure between the wellbore and the perforated
cells for both well models. The simple well model assumes instantaneous hydrostatic fluid
distribution, and since the producer is completely horizontal, the pressure is identical in
all the six points where the wellbore is open to the reservoir. The instrumented well has
a significantly larger pressure drawdown near its heel, which also lies furthest away from
the injector and the advancing gas front. Both models use the same inflow performance
relationship and hence the drawdown outside of the valve in the multisegment model will
on average be approximately the same as for the simple well model when both wells are
run on rate controls.

Computer exercises
12.4.5 Our setup had no lower limit on the bhp. What happens if bhp in the producer is
not allowed to drop below 100 bar or 150 bar?
12.4.6 Rerun the two simulations with producer controlled by a bhp of 250 bar. Can you
explain differences in the producer and the bhp of the injector?

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 460 The AD-OO Framework for Reservoir Simulation
Time: 2.0 years Time: 2.5 years Time: 3.3 years

3
Oil reservoir rate [k m /day] Oil surface production [k m3]
10
5
5

4 0

3 10

5
2
0
0.5 1 1.5 2 2.5 3 0.5 1 1.5 2 2.5 3
Time [years] Time [years]

Figure 12.12 Displacement front engulfing the horizontal producer causes decay in oil production.
Because of different pressure in the wellbore, the volumetric reservoir rates is largely different in the
two models, even though surface rates are almost identical.

Figure 12.13 Pressure in perforated cells (top) and wellbore (bottom) as function of time for the
simple well model (left) and the multisegment well model (right). In the latter plot, the middle surface
represents pressure outside of the valves.

12.4.4 The SPE 9 Benchmark Case

In our last example,5 we set up a more realistic simulation case, show how to use CPR
preconditioning together with an algebraic multigrid solver, and compare our simulation
results against ECLIPSE to verify that the AD-OO black-oil simulator is correct. To this

5 Complete source code: blackoilTutorialSPE9.m in ad-blackoil

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.4 Simulation Examples 461

end, we consider the model from the SPE 9 comparative solution project [159], which was
posed in the mid-1990s to compare contemporary black-oil simulators. The benchmark
consists of a water-injection problem in a highly heterogeneous reservoir described by a
24 × 25 × 15 regular grid with a 10-degree dip in the x-direction. By current standards, the
model is quite small, but contains several salient features that must be properly handled
by any black-oil simulator aiming to solve real problems. On one hand, the fluid and
PVT model have certain peculiarities; we have already discussed these models in detail in
Sections 11.3.3 and 11.5, respectively. In particular, the discontinuity in the water–oil cap-
illary pressure curve is expected to challenge the Newton solver when saturations change
significantly. Other challenges include strong heterogeneity, transition from undersaturated
to saturated state, abrupt changes in prescribed production rates, and dynamic switching
between rate and bhp control for the producers. Altogether, this makes the case challenging
to simulate.

Petrophysical Model and Initial Conditions

The basic setup of the grid, petrophysical parameters, fluid model, and initial state is more
or less identical to that of the SPE 1 model discussed in Section 11.8.1. For convenience,
this is implemented in a dedicated function:
[G, rock, fluid, deck, state0] = setupSPE9();
model = selectModelFromDeck(G, rock, fluid, deck);
schedule = convertDeckScheduleToMRST(model, deck);

The reservoir rock is highly heterogeneous, albeit not as bad as the SPE 10 benchmark. The
model consists of 14 unique layers of varying thickness. The permeability is isotropic and
follows a lognormal distribution with values spanning more than six orders of magnitude;
see Figure 12.14. The porosity field has no variation within each layer. One of the high-
permeability layers is represented by two cells in the vertical direction, whereas two of the
other layers have the same porosity, altogether giving 12 unique porosity values.

Figure 12.14 Petrophysical data for the SPE 9 model: the left plot shows absolute permeability in
units md and the right plot shows porosity.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 462 The AD-OO Framework for Reservoir Simulation

Figure 12.15 Initial fluid distribution inside the reservoir. The left figure shows the saturation, which
lies on the two-phase O–W axis in the ternary diagram, whereas the right plot shows the degree of
saturation defined as rs /Rs (p).

Initially, the reservoir does not contain any free gas, but a significant amount of gas is
dissolved into the oleic phase. The dissolved amount rs is constant throughout the whole
model, except for three individual cells. Some of the dissolved gas may be liberated as free
gas if the reservoir pressure drops below the bubblepoint during production. As we know,
the solution gas–oil ratio Rs varies with pressure, which in turn increases with depth. At
the top of the reservoir, the oleic phase is fully saturated (rs = Rs ), whereas the degree of
saturation is only 89% at the bottom of the reservoir (rs = 0.89Rs ). We should therefore
expect that some free gas will form at the top of the reservoir as pressure drops once the
wells start producing.

Wells and Simulation Schedule

The reservoir is produced from 25 vertical wells that initially operate at a maximum rate
of 1,500 STBO/D (standard barrels of oil per day). Pressure is supported by a single water
injector, operating at a maximum rate of 5,000 STBW/D (standard barrels of water per
day) and a maximum bhp of 4,000 psi at reference depth. Between days 300 and 360, the
production rate is lowered to 100 STBO/D, and then raised again to its initial value until the
end of simulation at 900 days. Throughout the simulation, most of the wells switch from
rate control to pressure control.
The simulation schedule consists of three control periods: days 0–300, days 300–360,
and days 360–900, with a ramp-up of time steps at the beginning of the first and third
period. All 26 wells are present during the entire simulation. The injector is injecting a
constant water rate, while the producers are set to produce a target oil rate, letting bhp
and gas/water production vary. In addition to the rate targets, the producers are limited
by a lower bhp value of 1,000 psi, which could, e.g., signify the lowest pressure we can
have in the well and still be able to lift reservoir fluids to the surface. The prescribed
limits are given in the lims field of each well. For producer number two, this would be
schedule.control(1).W(22).lims:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.4 Simulation Examples 463

Controls for PROD2: oil rate [stb/day]

1,600

1,400

1,200

1,000

800

600

400

200

0
0 0.5 1 1.5 2 2.5
Time (years)

Figure 12.16 Well placement in the reservoir (left) and simulation schedule with rate targets for the
producers (right).

Well limits for PROD2:

orat: -0.0028
wrat: -Inf
grat: -Inf
lrat: -Inf
bhp: 6.8948e+06

By convention, the oil rate has negative sign for a producer. The pressure control will be
activated if the wellbore pressure falls below the lower limit.

Customizing Linear Solver and Time-Step Control

The model has 9,000 active cells, which for a three-phase problem corresponds to a linear
system with 27,000 cell unknowns; the total number of unknowns is slightly higher because
of the 26 wells. Models of this size can be solved with the standard direct solvers in
MATLAB on most computers, but runtimes can be reduced if we instead use a CPR
preconditioner. To solve the elliptic pressure subsystem, we use the algebraic multigrid
solver AGMG, with MATLAB’s standard direct solver as potential backup, if AGMG is
not installed on the computer:
try
mrstModule add agmg
pressureSolver = AGMGSolverAD('tolerance', 1e-4);
catch
pressureSolver = BackslashSolverAD();
end
linsolve = CPRSolverAD('ellipticSolver', pressureSolver, ...
'relativeTolerance', 1e-3);

We also adjust the default time-step control slightly to ensure stable simulation by tighten-
ing the restriction on relative changes pressure and absolute changes in saturation

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 464 The AD-OO Framework for Reservoir Simulation

model.dpMaxRel = .1;
model.dsMaxAbs = .1;

Beyond this, we make no modifications to the standard configuration of the simulator,

which employs robust defaults from ECLIPSE.

Running the Simulation

The simulation can now be started by the following statements:
model.verbose = true;
[wellsols, states, reports] =...
simulateScheduleAD(state0, model, schedule,'LinearSolver', linsolve);

We give the schedule with well controls and control time steps. The simulator may use
other time steps internally, but it will always return values at the specified control steps.
Turning on verbose mode ensures that we get an extensive report of the nonlinear iteration
process for each time step:
Solving timestep 01/35: -> 1 Day
Well INJE1: Control mode changed from rate to bhp.
======================================================================================================================
| It # | CNV_W | CNV_O | CNV_G | MB_W | MB_O | MB_G | waterWell| oilWells | gasWells | closure |
======================================================================================================================
| 1 | 2.82e-02 | 1.10e+00 | 5.00e-03 | 1.18e-05 | 2.00e-04 | 1.58e-06 | 6.81e-04 | 5.00e-02 | 1.73e-01 |*0.00e+00 |
| 2 | 1.14e-01 | 7.25e-02 | 5.47e-02 | 5.17e-06 | 1.48e-04 | 8.86e-04 | 4.40e-03 | 2.20e-03 | 1.24e+00 | 1.47e+06 |
Well PROD26: Control mode changed from orat to bhp.
| 3 |*6.37e-04 | 2.38e-02 | 6.68e-02 |*7.78e-09 | 2.24e-04 | 1.13e-03 |*6.04e-07 | 4.58e-05 | 6.02e-02 |*1.73e-18 |
| 4 | 1.89e-03 | 8.12e-03 | 2.55e-02 | 7.23e-06 | 4.74e-05 | 3.56e-04 |*5.09e-06 | 1.85e-04 | 5.16e-02 | 5.67e+05 |
| 5 |*6.04e-04 | 1.69e-03 |*6.50e-04 | 1.32e-06 | 3.47e-06 | 1.90e-06 |*5.17e-07 | 1.92e-05 | 4.85e-03 |*8.67e-19 |
| 6 |*6.31e-04 | 1.46e-03 |*6.72e-04 | 1.16e-06 | 4.47e-06 | 1.74e-06 |*2.28e-10 | 1.56e-05 | 3.87e-03 |*0.00e+00 |
| 7 |*9.16e-05 |*3.79e-04 |*2.20e-04 |*7.16e-08 | 3.82e-07 |*9.41e-09 |*3.58e-11 |*2.41e-07 | 6.84e-05 |*0.00e+00 |
| 8 |*7.96e-05 |*2.45e-04 |*1.08e-04 |*5.50e-08 | 2.40e-07 |*7.63e-08 |*2.87e-14 |*3.82e-11 |*1.10e-08 |*0.00e+00 |
| 9 |*1.37e-05 |*5.70e-05 |*3.43e-05 |*2.15e-09 |*1.20e-09 |*1.22e-09 |*1.79e-14 |*7.86e-11 |*2.40e-08 |*0.00e+00 |
======================================================================================================================
Solving timestep 02/35: 1 Day -> 2 Days

Here, we recognize the mass-balance (MB) errors and the maximum normalized residu-
als (CNV) discussed on page 438 for water, oil, and gas. The next three columns report
residuals for the inflow performance relationship defined per perforations, whereas the last
column reports residual for the control equations defined per well. A star means that the
residual is below the tolerance. In the first time step, mass-balance error for oil was the last
one to drop below the prescribed value of 10−9 . Nine iterations is a quite high number and
indicates that there are significant nonlinearity in the residual equations, primarily caused
by initial transient effects. The next seven control steps all converge within 6 iterations.
Between days 80 and 100, liberated gas reaches the water zone and the next two steps
now require 7 and 12 iterations. Large jumps in the iteration count can typically be traced
back to interaction of different displacement fronts, liberation of gas or other abrupt phase
transitions, and/or steep/discontinuous changes in the well controls or other parameters
affecting the global drive mechanisms.

Switching of Well Controls

Figure 12.17 shows which control is active during each of the 35 control steps. Initially, all
wells are set to be rate controlled. The targeted injection rate of 5,000 STBW/day is much

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.4 Simulation Examples 465
INJE1
PROD10
PROD11
PROD12
PROD13
PROD14
PROD15
PROD16
PROD17
PROD18
PROD19
PROD2
bhp
PROD20
PROD21
PROD22
PROD23 rate
PROD24
PROD25
PROD26
PROD3
PROD4
PROD5
PROD6
PROD7
PROD8
PROD9

100 200 300 400 500 600 700 800 900

Time [days]

Figure 12.17 Well controls active during the simulation of the SPE 9 benchmark.

higher than what can be sustained with reasonable injection pressures. The injector thus
switches immediately from rate to bhp control during the first time step and stays there
throughout the whole simulation. Likewise, producer PROD26 and producers PROD21
and PROD23 switch from rate to bhp control during the first and second control step,
respectively. This switching is reported to screen during simulation, and in the report on
the facing page, we see that the injector switches during the first iteration and producer
PROD26 after the second iteration. The overall picture is that producers are mostly rate
controlled in the beginning and mostly controlled by bhp at the end; the only exception is
PROD14, which runs rate control the whole period. This general switch from rate to bhp
control is a result of the average field pressure dropping during the simulation as fluids are
removed from the reservoir. We also notice the middle period when rates are reduced by a
factor of ten. A smaller drawdown is required to sustain this rate, and hence all producers
are able to stay above the bhp limit.

Analysis of Simulation Results

This reservoir is an example of a combined drive, in which two different mechanisms con-
tribute to push hydrocarbons towards the producers. The first is injection of water below the
initial oil–water contact, which contributes to maintain reservoir pressure and push water
into the oil zone to displace oil. Ideally, this would have been a piston-like displacement that
gradually pushed the oil–water contact upward, but because of heterogeneity and nonlinear
fluid mobilities, water displaces oil as a collection of “fingers” extending from the oil–water
contact and towards the producers; see Figure 12.18. Because the injector operates at fixed
pressure, the injection rate increases monotonically as the reservoir is gradually depleted.
The second drive mechanism is liberation of free gas, which starts near the produc-
ers where the pressure drawdown is largest. As pressure continues to decay, more gas is
liberated at the top of the formation where the pristine oil phase is closest to a saturated
state. The liberated gas forms a gas cap that gradually grows downward and contributes to
push more gas into the wells and reduce the production rates of oil; see the left column in

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 466 The AD-OO Framework for Reservoir Simulation

Water reservoir injection rate [stb/day]

2,000

1,800

1,600

1,400

1,200

1,000

800
200 400 600 800
Time [years]

Injection pressure [psi]

4,000 Injector bhp
Field avg pressure

3,500

3,000

2,500

200 400 600 800

Time [years]

Figure 12.18 Water drive contributing to enhanced production. The left column shows surface
injection rate (top) and bhp compared with the average field pressure (bottom). The right column
shows the bottom water drive, with initial water saturation at the top and water saturation after 900
days at the bottom.

Figure 12.19. Any produced and free gas must come from the gas initially dissolved in oil,
since no gas has been injected. Figure 12.19 reports the amount of dissolved gas initially
and after 900 days.
Figure 12.20 shows production from the 25 wells. First of all, we note that there is
significant spread in the recovery from the individual wells, with some wells that experience
declining oil rates almost immediately and deviate from the upper production envelope that
corresponds to the prescribed rate targets. These are wells at which free gas appears already
during the first step. Since the free gas is more mobile and occupies a larger volume, it will
effectively choke back the production of the oleic phase so much that we see a decay in
the total surface gas rates from these wells. Some of the other wells are able to maintain
the prescribed oil rate for a longer period. These wells experience increasing surface gas
rates until the oil rate is reduced at 300 days. Other wells are of intermediate type and
first experience increasing and then decaying gas rates. Most wells only produce small
amounts of water, and for several of those who produce significant amounts of water, the
breakthrough comes after the period with reduced oil rates has ended. The only exception
in PROD17, which produces almost two orders of magnitude more water than the next

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.4 Simulation Examples 467

Figure 12.19 Liberation of gas. The left column shows initial gas saturation at the top and gas
saturation after 900 days at the bottom. The right column shows the amount of gas dissolved in
the oleic phase.

producer. (PROD17 is not shown in the lower-left plot in Figure 12.20.) We note in passing
that the plot of bhp confirms our previous discussion of switching well controls.

Verification Against ECLIPSE

The default solvers in ad-blackoil have been implemented as to reproduce ECLIPSE
as closely as possible. This does not necessarily mean that ECLIPSE always computes
the correct result, but given the major role this simulator plays in the industry, it is very
advantageous to be able to match results, so that this can be used as a baseline when
investigating the performance of new models or computational methods. We have already
shown that MRST matches ECLIPSE very well for the SPE 1 benchmark. To build further
confidence in the simulator, we also show that MRST matches for the SPE 9 benchmark,
given identical input. To this end, we use the same functionality from the deckreader
module for reading restart files as we used in Section 11.8.2:
addir = mrstPath('ad-blackoil');
compare = fullfile(addir, 'examples', 'spe9', 'compare');
smry = readEclipseSummaryUnFmt(fullfile(compare, 'SPE9'));

We first get the number of report steps and the time at which these are computed

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 468 The AD-OO Framework for Reservoir Simulation
5 6
Cumulative oil production [10 stb] Surface gas rate [10 scf/day]
12 12

10 10

8 8

6 6

4 4

2 2

0 0
200 400 600 800 200 400 600 800
Time [years] Time [years]
Surface water rate [stb/day]
Bhp [psi]
300
3,500
250
3,000
200
2,500
150
2,000
100
1,500
50

0 1,000
200 400 600 800 200 400 600 800
Time [years] Time [years]

Figure 12.20 Fluid production from the 25 producers.

compd = 1:(size(smry.data, 2));

Tcomp = smry.get(':+:+:+:+', 'YEARS', compd);

Let us for instance compare the bhp of PROD13. We extract the corresponding data as
follows:
comp = convertFrom(smry.get('PROD13', 'WBHP', compd), psia)';
mrst = getWellOutput(wellsols, 'bhp', 'PROD13');

and subsequently plot well responses using standard MATLAB commands; see
Figure 12.21. The results computed by the two simulators are not identical. If you run
the example, you will see that the output from ECLIPSE contains 37 steps whereas MRST
reports 35 steps. The reason is that MRST does not include the initial state in its output
and that ECLIPSE has a somewhat different time-step algorithm that adds one extra time
step during the ramp-up after the target oil rates have been increased back to their original
value at 360 days. Nevertheless, the match between the two computed solutions appears to

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.4 Simulation Examples 469
7 PROD13
x 10 PROD13

2.4
2.5
2.2

2 2
Pressure (Pa)

Gas rate (m /s)

1.8

3
1.5
1.6

1.4
1
1.2

1 0.5

0.8
0
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Time (years) Time (years)

7 PROD18
x 10 PROD18
2.6 2.5
MRST
2.4 ECLIPSE
2.2 2

Gas rate (m /s)

Pressure (Pa)

3
1.8 1.5

1.6

1.4 1

1.2
0.5 MRST
1
ECLIPSE
0.8
0
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Time (years) Time (years)

Figure 12.21 Verification of the AD-OO black-oil solver from ad-blackoil against ECLIPSE. The
left column shows bhp for two different producers, whereas the right column shows surface gas rates.

be almost perfect. Similar matches are also seen for other types of well responses and all
the remaining 24 wells.
Bao et al. [30] report similar good match for a set of polymer injection cases including
non-Newtonian fluid rheology, In addition, we have run a number of cases in-house ver-
ifying that the AD-OO framework provides excellent match with commercial simulators.
This includes challenging compositional cases as well as black-oil cases. Altogether, this
shows that the software is a good starting point for case studies. However, when working
with black-oil models, you should expect that you may stumble upon features that are not
yet implemented in MRST, but this is also the point where you can contribute by adding
new features.

Computer exercises

12.4.7 Set up an animation of water and gas to better understand how the aqueous and
gaseous phases displace the oleic phase.
12.4.8 As pointed out on page 466, well PROD17 produces orders of magnitude more
water than the other wells. How would shutting this well in from the start
influence the overall field production.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 470 The AD-OO Framework for Reservoir Simulation

12.4.9 Run the simulation using CPR with AGMG or MATLAB’s direct solver and
compare with just using the direct solver for the whole system. Are there any
noticeable differences in simulation times? Try to install AMGCL and use func-
tionality from the linearsolver module discussed on page 446 to see if you
can reduce runtime even further.
12.4.10 In some cases, produced gas cannot be sold but must be reinjected. Can you set
up such a simulation based on the case discussed in this section? Where would
you place the gas injector? (Hint: Operations involving reinjection generally
require complex topside logic to determine correct rates for the reinjected fluid.
Such logic is not yet available in MRST. As a first approach, you can just assume
that the gas is readily available so that whatever amount of gas produced in step
n can be injected in step n + 1.)

12.5 Improving Convergence and Reducing Runtime

I hope that this and the previous chapters have given you a better understanding of reservoir
simulation, and also provided you with sufficient detail about MRST so that you can
use the software to perform interesting simulations. Although the software is primarily
designed for flexibility and generality, the black-oil simulators implemented by use of the
AD-OO framework offer such a comprehensive set of features that they can be used to
simulate quite complicated models. Out of the box, MRST may not always be as fast
as a simulator written in a compiled language, but with a bit customized setup you can
easily make the simulators much more competitive. A key factor to improve computational
efficiency of MRST is to replace MATLAB’s default direct solver by a modern iterative
solver with appropriate multilevel preconditioning as discussed in Section 12.3.4. With
such capabilities on board, MRST will typically be a few times slower than ECLIPSE or
comparable research simulators like OPM Flow for moderately sized models with O(105 –
106 ) unknowns as long as many-core or multi-node parallelism is not utilized.
As you explore other models than those discussed in the previous section, you will
sooner or later run into convergence problems and/or cases for which the computational
performance is not satisfactory. I end this chapter by briefly discussing how to address
such problems. First, I provide a few general recommendations that apply equally well to
commercial simulators like ECLIPSE, and then end the section with a few issues that are
specific to MRST.

Tweaking Control Parameters for Solvers

Like many other things in the AD-OO framework, default values for the various parameters
that control the behavior of the NonLinearSolver class largely follow what is suggested
for ECLIPSE. These default values have been developed based on many years of expe-
rience on a large variety of simulation models and will work well in most cases. You
can observe improved performance of your simulation by tuning parameters in specific
cases, but it is generally not recommended that you change the default values unless you

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.5 Improving Convergence and Reducing Runtime 471

know what you are doing. In particular, it is rarely necessary to weaken the convergence
tolerances. When convergence problems occur in the nonlinear iterations, the culprit usu-
ally lurks in the time-step selection or the model itself, and it is too late to try to fix the
problem by adjusting the convergence control. Contrary to what you may think, faster
overall convergence can often be observed by tightening convergence controls rather than
weakening them since this may prevent inaccuracies earlier in the simulation, which oth-
erwise may accumulate from one time step to the next and cause severe errors later in the
simulation.
As general advice: the best way to improve convergence and computational performance
is in most cases to choose a more robust linear solver and/or time-stepping scheme. In
some cases, the cause of poor convergence and long runtimes may also lie in the model
parameters, e.g., as discussed in Section 11.8.3.

Selection of Control Steps

How you set up your simulation schedule will largely determine the efficiency of your
simulation. In many cases, the number of control steps and their individual lengths are
determined by external factors. These can be requirements of monthly or quarterly reports
for prediction runs, or the times of data points for history-matching runs. In other cases,
there are no specific requirements, and the schedule can be chosen more freely. The fol-
lowing guidelines may be useful for improving convergence and general computational
performance:

• Make sure that the limit on the maximum time step is compatible with the specified
control steps. A classical mistake is to set the maximum time step to be 30 days, and
design a simulation schedule that reports results at the end of each month. As we all
know, every second month has 31 days . . .
• For cases with complex flow physics like changes between undersaturated and saturated
states, flow reversals, abrupt changes in drive mechanisms, etc., it may be advantageous
to reduce long time steps since reductions in iteration count often will compensate for the
increased number of time steps. As an example, using a gradual ramp-up of time steps
initially and after major changes in well controls and other drive mechanisms will ensure
that transients are better resolved, which in turn means that the simulation runs more
smoothly.
• In some cases, you may be asking for shorter control steps than you actually need. If,
for instance, each control step consists of only one time step and/or the nonlinear solver
converges in very few iterations (say, one to three), it may be advantageous to increase
the length of the control steps to give the time-step selector more freedom to choose
optimal time steps.
For simulations based on ECLIPSE input decks, you should be particularly wary of the
TUNING keyword, which is easy to misuse so that it interferes adversely with the standard
mechanisms for time-step selection. (This keyword is not used by MRST.)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 472 The AD-OO Framework for Reservoir Simulation

Issues Specific to MRST/MATLAB

In implementing MRST, we have tried to utilize MATLAB and its inherent vectoriza-
tion as effectively as possible. However, maintaining both flexibility and efficiency is not
always possible, and there are a few sticky points in the implementation of the AD-OO
framework that may contribute to increased computational time. For vectors with a few
hundred or thousand unknowns, the computational overhead in the AD library and other
parts of the software is significant. The cost of linearization and assembly of linear systems
can therefore completely overwhelm the runtime of the linear solver for small problems.
This computational overhead should in principle, and will also in many cases, gradually
disappear with larger problem sizes. However, this is not always the case.
The first factor that may contribute to maintain a significant overhead also for larger
models, is our handling of wells. Wells are generally be perforated in multiple cells through-
out the grid, and to assemble their models, we need to access and extract small subsets of
large sparse matrices. Such access, which some refer to as slicing, is not as efficient in
MATLAB as one could hope, particularly not in older versions of the software. To be able
to handle different types of well models like the SPE 1 case with inflow valves discussed
in Section 12.4.3, the FacilityModel class requires a somewhat heterogeneous memory
access, which we have not yet figured out how to implement efficiently in MATLAB.
Using this class may therefore introduce significant overhead. For cases where all wells
assume vertical equilibrium (and a Peaceman-type inflow performance relationship), you
can instead use a special facility class called UniformFacilityModel that has been opti-
mized to utilize uniform memory access. Once you have created a reservoir model, you can
convert the well description to this type by the following call:
model.FacilityModel = UniformFacilityModel(model);

For the SPE 9 case discussed in Section 12.4.4, the 26 wells perforated 80 out of the 9,000
cells in the grid. Adding this transformation led to a 25% reduction in runtime on my
computer.
A second factor comes from the way MRST’s vectorized AD library is constructed. By
default, this library always constructs a sparse matrix to represent the derivatives (Jacobian)
of a vector unknown. Constructing and multiplying such matrices has a significant cost that
should be avoided whenever possible. For cases with many independent primary variables,
many of the Jacobian sub-matrices tend to be be diagonal. You can represent a set of
diagonal matrices as a smaller, dense matrix, for which elementary operations are highly
efficient. At the time of writing, work is in progress to develop more efficient ADI backends
that utilize diagonal structures (and other known matrix patterns) to reduce construction,
multiplication, and assembly costs. You can instrument you models by a specific AD back
end through an assignment of the form:
model.AutoDiffBackend = DiagonalAutoDiffBackend;

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 12.5 Improving Convergence and Reducing Runtime 473

I expect that improved support for more efficient backends will continue to develop in
future versions of the software.
Last, but not least, plotting and creating new axes is slow in MATLAB and has in my
experience become even slower after the new graphics engine ("HG2") was introduced in
R2014b. Many of the examples presented in the book have plotting integrated into simula-
tion loops. For brevity and readability I have not optimized this part, and instead explicitly
cleared axes and replotted grids and cell values whenever the plot needs to be updated.
To reduce runtime, you should instead delete and recreate individual graphics handles, or
simply manipulate the color data of existing graphics objects, rather than drawing the whole
scene anew.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Part IV
Reservoir Engineering Workflows

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13
Flow Diagnostics

It is rarely sufficient to study well responses and pressure and saturation distributions to
understand the communication patterns in a complex reservoir model. To gain a better qual-
itative picture, you typically want understand from what region a given producer drains. To
what region does a given injector provide pressure support? Which injection and production
wells are in communication? Which parts of the reservoir affect this communication? How
much does each injector support the recovery from a given producer? Do any of the wells
have backflow? What is the sweep and displacement efficiency within a given drainage,
sweep, or well-pair region? Which regions are likely to remain unswept? Likewise, you
may also want to perform what-if and sensitivity analyses to understand how different
parameters and their inherent uncertainty affect reservoir responses.
Simulating a full reservoir model containing a comprehensive description of geology,
reservoir fluids, flow physics, well controls, and coupling to surface facilities is a compu-
tationally demanding task that may take hours or even days to complete. This limits your
ability to study parameter variations and is at odds with the philosophy of modern reservoir
characterization techniques, which often generate hundreds of equiprobable realizations
to quantify uncertainty in the characterization. In this chapter, we therefore introduce a
set of simple techniques, referred to as flow diagnostics, which you can use to delineate
volumetric communications and improve your understanding of how flow patterns in the
reservoir are affected by geological heterogeneity and respond to engineering controls.
In their basic setup, these techniques only assume that you have a geocellular model of
the reservoir and rely on the solutions of single-phase, incompressible flow problems, as
discussed in Chapters 4 and 5. However, you can equally well employ the techniques to ana-
lyze representative flow field extracted from multiphase simulations of the type discussed
in Chapters 10–12.
My definition of flow diagnostics is a family of simple and controlled numerical flow
experiments that are run to probe a reservoir model, establish connections and basic
volume estimates, and quickly provide a qualitative picture of the flow patterns in the
reservoir and quantitative measures of the heterogeneity in dynamic flow paths. You can
also use these methods to compute quantitative information about the recovery process
in settings somewhat simpler than what would be encountered in an actual field. Using
flow diagnostics, you can rapidly and iteratively perturb simulation input and evaluate the

477

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 478 Flow Diagnostics

resulting changes in volumetric connections and communications to build an understanding

of cause and effects in your model. The following presentation is inspired by [274] and
[218], who developed the concept of flow diagnostics based on standard finite-volume dis-
cretizations for time-of-flight and tracer partitions [221, 98]. Similar ideas have previously
been used within streamline simulation [79] for ranking and upscaling [140, 27, 278],
identifying reservoir compartmentalization [127], rate optimization [287, 244, 144], and
flood surveillance [32].
Flow diagnostics is inexpensive to compute, and I generally recommend that you use
these methods to prescreen your models before you start to conduct more comprehensive
multiphase simulations. MRST already offers a few simple interactive GUIs to this end,
and I expect that similar capabilities will appear in commercial tools in the not too distant
future. Interactive tools are not easy to present in book form, and if you want to see
these work in practice, you should try out some of the examples that follow MRST, or
study the exercises presented throughout this chapter. Much of the same ideas can also be
used to postprocess multiphase flow simulations; such capabilities are currently available
in the OPM ResInsight viewer (http://resinsight.org). Early research [218, 171] indicates
that flow diagnostics offers a computationally inexpensive complement (or alternative)
to full-featured multiphase simulations to provide flow information in various reservoir
management workflows, e.g., as inexpensive objection functions for well placement and
rate optimization, or as a means to perform what-if and sensitivity analyzes in parameter
regions surrounding preexisting simulations.

13.1 Flow Patterns and Volumetric Connections

You have already been introduced to the basic quantities that lie at the core of flow diag-
nostics in Chapters 4 and 5: as you probably recall from Section 4.3.3, we can derive time
lines that show how heterogeneity affects flow patterns for an instantaneous velocity field
v by computing

• the forward time-of-flight, defined by

v · ∇τf = φ, τf |inflow = 0, (13.1)
which measures time it takes a neutral particle to travel to a given point in the reservoir
from the nearest fluid source or inflow boundary; and
• the backward time-of-flight, defined by
− v · ∇τb = φ, τb |outflow = 0, (13.2)
which measures the time it takes a neutral particle to travel from a given point in the
reservoir to the nearest fluid sink or outflow boundary.
The sum of the forward and backward time-of-flight at a given point in the reservoir gives
the total residence time of an imaginary particle as it travels from the nearest fluid source
or inflow boundary to the nearest fluid sink or outflow boundary.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.1 Flow Patterns and Volumetric Connections 479

Studying iso-contours of time-of-flight gives you a time map for how more complex
multiphase displacements may evolve under fixed well and boundary conditions and reveals
more information about the flow field than pressure and velocities alone. We illustrated this
already in Chapter 5, where Figure 5.3 on page 160 showed time lines for a quarter five-
spot flow pattern, and the total residence time was used to distinguish high-flow regions
from stagnant regions. Likewise, in Figure 5.11 on page 173 we used time-of-flight to
identify non-targeted regions for a complex field model; that is, regions with high τf values
that were likely to remain unswept and hence would be obvious targets to investigate for
placement of additional wells. From time-of-flight, you can derive various measures of
dynamic heterogeneity, as we will see in the next section, or compute proxies of economical
measures such as net-present value for multiphase flow models; see [218, 171].
In a similar manner, we can determine all points in the reservoir that are influenced by
a given fluid source or inflow boundary by solving the following equation to compute the
volumetric injector partition:
v · ∇ci = 0, ci |inflow = 1. (13.3)
To understand what this equation represents, let us think of an imaginary painting exper-
iment in which we inject a tracer dye (a massless, non-diffusive ink) of a unique color at
each fluid source or point on the inflow boundary we want to trace the influence from. The
ink will start flowing through the reservoir and paint every point it contacts. Eventually, the
fraction of different inks that flows past each point in the reservoir reaches a steady state,
and by measuring these fractions, we can determine the extent to which each different ink
influences a specific point and find the influence region of an inflow point. Likewise, to
determine how much the flow through each point is influenced by a fluid sink or point on
the outflow boundary, we reverse the flow field and solve similar equations for producer (or
outflow) partitions,
− v · ∇cp = 0, cp |outflow = 1. (13.4)
To summarize, flow diagnostics requires three computations: (i) solution of a pressure
equation, or extraction of fluxes from a previous computation, to determine the bulk fluid
movement; (ii) solution of a set of steady advection equations (13.3)–(13.4) to partition
the model into influence or volumetric flow regions; and (iii) solution of time-of-flight
equations (13.1)–(13.2) to give averaged time lines that describe the flow within each
region. This chapter describes how you can combine and process tracer partitions and time-
of-flight to gain more insight into flow patterns and volumetric connections in the reservoir.
In more advanced methods, we also use residence-time distributions over all flow paths and
analogies with tracer analysis [277, 139] for more accurate analysis of flow paths.

13.1.1 Volumetric Partitions

If each source of inflow (outflow) is assigned a unique “color” value, the resulting color
concentration should in principle produce a partition of unity in all parts of the reservoir

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 480 Flow Diagnostics

that are in communication with the sources of inflow (outflow). In practice, you may not
be able to obtain an exact partition of unity because of numerical errors. Based on the
inflow/outflow partitions, we can further define:

• drainage regions – each such region represents the reservoir volume that eventually will
be drained by a given producer (or outflow boundary), provided that the current flow field
v prevails until infinity;
• sweep regions – each such region represents the reservoir volume that eventually will
be swept by a given injector (or inflow boundary) if the current flow conditions remain
forever;
• well pairs – injectors and producers in communication with each other;
• well-pair regions – regions of the reservoir in which the flow between a given injector
and producer takes place.
When visualizing multiple drainage or sweep regions in 3D, it is common to assign a unique
region to each cell determined by a majority vote over all “colors” that affect the cell.
Well pairs are determined by finding all injectors whose concentration is positive in one
of the well completions of a given producer (or vice versa). Well-pair regions are found
by intersecting injector and producer partitions, possibly followed by a majority vote. Well
allocation factors will be discussed in more detail in Section 13.1.3 and in Chapter 15, in
conjunction with single-phase upscaling.
Our default choice would be to assign a unique “color” to each injector and producer,
but you can also subdivide wells into multiple segments and trace the influence of each
segment separately. You can, for instance, use this to determine if a (horizontal) well has
cross-flow, so that fluid injected in one part of the well is drawn back into the wellbore
in another part of the well, or fluids produced in one completion are pushed out again in
another.
We have already encountered computeTimeOfFlight for computing time-of-flight in
Section 5.3. The diagnostics module additionally offers
D = computeTOFandTracer(state, G, rock, 'wells', W, ...)

that computes forward and backward time-of-flight, injector and producer coloring, as well
as sweep and drainage regions in one go for models with flow driven by wells. These
quantities are represented as fields in the structure D:

• inj and prod give the indices for the injection and production wells in the well
structure W;
• tof is a 2 × n vector with τf in its first and τb in its second column;
• itracer/ptracer contain color concentrations for the injectors and producers;
• ipart and ppart hold the partitions resulting from a majority vote.
Obviously, you can associate similar quantities with boundary conditions and/or source
terms. (MRST does not yet fully implement this.) Well pairs are identified by analyzing the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.1 Flow Patterns and Volumetric Connections 481
Forward time-of-flight Color concentration for I1

Backward time-of-flight Color concentration for P2

Residence time Well-pair regions (I1↔P1) and (I2↔P3)

Sweep regions Drainage regions

Figure 13.1 Time-of-flight, color concentration, and various types of volumetric delineations for a
simple case with two injectors and three producers. Source code necessary to generate the plots is
given in the showDiagnostBasic tutorial in the book module.

itracer and ptracer fields. The following function also computes the pore volume of
the region of the reservoir associated with each pair:
WP = computeWellPairs(G, rock, W, D)

Figure 13.1 shows time-of-flight, color concentrations, volumetric partitions, and a few
combinations thereof for a simple flow problem with two injectors and three production
wells. There are many other ways you can combine and plot these basic quantities to

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 482 Flow Diagnostics

reveal volumetric connection and provide enhanced insight into flow patterns. You can,
for instance, combine sweep regions with time-of-flight to provide a simple forecast of
contacted volumes or to visualize how displacement fronts move through the reservoir. To
this end, you would typically also include estimates of characteristic wave-speeds from
multiphase displacement theory as discussed in Section 8.4.

13.1.2 Time-of-Flight Per Partition Region: Improved Accuracy

It is important to understand that the time-of-flight values computed by a finite-volume
method like the one discussed in Section 4.4.3 are volume-average values, which cannot be
compared directly with the point values you obtain by tracing streamlines, as described in
Section 4.3.3. Because time-of-flight is a quantity defined from a global line integral (see
(4.39)), the point-wise variations can be large inside a single grid cell, in particular near
flow divides and in the near-well regions, where both high-flow and low-flow streamlines
converge. To improve the accuracy of the time-of-flight within each well-pair region, we
can use the color concentrations to recompute the time-of-flight values for influence region
number k,

v · ∇ cik τfk = cik φ. (13.5)
In MRST, this is done by passing the option computeWellTOFs to the time-of-flight solver,
using a call that looks something like
T = computeTimeOfFlight(state, G, rock, 'wells', W, ...
'tracer',{W(inj).cells},'computeWellTOFs', true);

which appends one extra column for each color at the end of return parameter T. You can
also use a higher-order discretization, as discussed in [221, 261].

13.1.3 Well Allocation Factors

Apart from a volumetric partition of the reservoir, one is often interested in knowing the
fraction of the inflow to a given producer that can be attributed to each of the injectors,
or conversely, how the push from a given injector is distributed to the different producers.
We refer to this as well allocation factors. These factors can be further refined so that
they also describe the cumulative flow from the toe to the heel of the well. By computing
the cumulative flux along the wellbore and plotting this flux as a function of the distance
from the toe (with the flux on the x-axis and distance on the y-axis) we get a plot that is
reminiscent of the plot from a production logging tool.
To formally define well allocation factors, we use the notation from Section 4.4.2, so
that x[c] denotes the value of vector x in cell c. Next, we let cin denote the injector color
p
concentration associated with well (or segment) number n, let cm denote the producer
concentration associated with well number m, q the vector of well fluxes, and {wkn }k the
cells in which well number n is completed. Then, the cumulative factors are defined as

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.2 Measures of Dynamic Heterogeneity 483
 p
i
anm [wn ] = q[wkn ] cm [wkn ],
k=1
 (13.6)
p
amn [wm ] = q[wkm ] cin [wkm ].
k=1
The total well allocation factor equals the cumulative factor evaluated at the heel of the
well. Well allocation factors computed by computeWellPair are found as two arrays of
structs, WP.inj and WP.prod, which give the allocation factors for all the injection and
production wells (or segments) accounted for in the flow diagnostics. In each struct, the
array alloc gives the anm factors, whereas influx or outflux that cannot be attributed to
another well or segment is represented in the array ralloc.

13.2 Measures of Dynamic Heterogeneity

Whereas primary recovery can be reasonably approximated using averaged petrophysical
properties, secondary and tertiary recovery are strongly governed by the intrinsic variability
in rock properties and geological characteristics. This variability, which essentially can be
observed at all scales in the porous medium, is commonly referred to as heterogeneity. As
we have seen in previous chapters, both the rock’s ability to store and transmit fluids are het-
erogeneous. However, it is the heterogeneity in permeability that has the most pronounced
effect on flow patterns and volumetric connections in the reservoir. The importance of
heterogeneity has been recognized from the earliest days of petroleum production, and over
the years a number of static measures have been proposed to characterize heterogeneity,
such as flow and storage capacity, Lorenz coefficient, Koval factor, and Dykstra–Parson’s
permeability variation coefficient, to name a few; see, e.g., [176] for a more comprehensive
overview.
In the following, we discuss how some of the static heterogeneity measures from clas-
sical sweep theory can be reinterpreted in a dynamic setting if we calculate them from the
time-of-flight (and color concentrations) associated with an instantaneous flow field [278].
Static measures describe the spatial distribution of permeability and porosity, and large
static heterogeneity means that there are large (local) variations in the rock’s ability to
store and transmit fluids. Dynamic heterogeneity measures, on the other hand, describe
the distribution of flow-path lengths and connection structure, and large heterogeneity
values show that there are large variations in travel and residence times, which again
tends to manifest itself in early breakthrough of injected fluids. Experience has shown
that these measures, and particularly the dynamic Lorenz coefficient, correlate very well
with forecasts of hydrocarbon recovery predicted by more comprehensive flow simulations.
They can thus be used as effective flow proxies in various reservoir management workflows;
see [278, 274, 218].

13.2.1 Flow and Storage Capacity

Dynamic heterogeneity measures are computed from the total residence time, which in turn
is computed as the sum of forward and backward time-of-flight. To define dynamic flow and

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 484 Flow Diagnostics

Figure 13.2 Streamtube analogue used to define dynamic flow and storage capacity. Colors illustrate
different average porosities.

qi F

normalize

Vi 

Figure 13.3 Construction of the F – diagram. The plot to the left shows flow rates qi plotted as
function of streamtube volumes Vi for a homogeneous displacement (green) and for a heterogeneous
displacement (blue). The right plot shows the corresponding F – diagrams, where the flow rates and
the streamtube volumes have been normalized.

storage capacity, we can think of the reservoir as a set of N non-communicating volumetric

flow paths (streamtubes) that each has a volume Vi , a flow rate qi , and a residence time τi =
Vi /qi . The streamtubes are sorted so that their residence times are ascending, τ1 ≤ τ2 ≤
· · · ≤ τN ; see Figure 13.2. Inside each streamtube, we assume a piston type displacement;
think of a blue fluid pushing a red fluid from the left to the right in the figure. We then
define the normalized flow capacity Fi and storage capacity i by,


i 0
N 
i 0
N
i = Vj Vj , Fi = qj qj . (13.7)
j =1 j =1 j =1 j =1

Here, i is the volume fraction of all streamtubes that have broken through at time τi and
Fi represent the corresponding fractional flow, i.e., the fraction of the injected fluid to the
total fluid being produced. These two quantities can be plotted in a diagram as shown in
Figure 13.3. From this diagram, we can also define the fractional recovery curve defined as
the ratio of inplace fluid produced to the total fluid being produced; that is (1 − F ) plotted
versus dimensionless time tD = d/dF measured in units of pore volumes injected.
To understand how the F – diagram can be seen as a measure of dynamic heterogene-
ity, we first consider the case of a completely homogeneous displacement, for which all

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.2 Measures of Dynamic Heterogeneity 485

streamtubes will break through at the same time τ . This means that, (i − i−1 )/(Fi −
Fi−1 ) ∝ Vi /qi is constant, which implies that F = , since both F and  are normalized
quantities. Next, we consider a heterogeneous displacement in which all the streamtubes
have the same flow rate q. Since the residence times τi form a monotonically increasing
sequence, we have that {Vi } will also be monotonically increasing. In general, F () is a
concave function, where the steep initial slope corresponds to high-flow regions giving
early breakthrough, whereas the flat trailing tail corresponds to low-flow and stagnant
regions.
In the continuous case, we can (with a slight abuse of notation) define the storage
capacity as
 τ
(τ ) = φ(
x (s)) ds, (13.8)
0

where x(τ ) represents all streamlines whose total travel time equals τ . By assuming incom-
pressible flow, we have that pore volume equals the flow rate times the residence time,
φ = qτ , and hence we can define the flow capacity as
 τ  τ
φ(x (s))
F (τ ) = q(x (s)) ds = ds. (13.9)
0 0 s
From this, we can define normalized, dynamic flow and storage capacities by

ˆ ) = (τ ) ,
F (τ )
(τ F̂ (τ ) = .
(∞) F (∞)
Henceforth, we only discuss the normalized quantities and for simplicity we drop the hat
symbol. To compute these quantities for a discrete grid model, we would have to first
compute a representative set of streamlines, associate a flow rate, a pore volume, and a
total travel time to each streamline, and then compute the cumulative sums as in (13.7).
Next, we consider how these concepts carry over to our flow diagnostics setting with
time-of-flight computed by a finite-volume method and not by tracing streamlines. Let pv
be an n×1 array containing pore volumes of the n cells in the grid and tof be an n×2 array
containing the forward and backward time-of-flights. We can now compute the cumulative,
normalized storage capacity Phi as follows:
t = sum(tof,2); % total travel time
[ts,ind] = sort(t); % sort cells based on travel time
v = pv(ind); % put pore volumes in correct order
Phi = cumsum(v); % cumulative sum
vt = full(Phi(end)); % total volume of region
Phi = [0; Phi/vt]; % normalize to units of pore volumes

We can compute the flow rate for each cell directly as the ratio between pore volume
and residence time if we assume incompressible flow. With this, the computeFandPhi
calculates the cumulative, normalized flow capacity F as follows

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 486 Flow Diagnostics

q = v./ts; % back out flux based on incompressible flow

ff = cumsum(q); % cumulative sum
ft = full(ff(end)); % total flux computed
F = [0; ff/ft]; % normalize and store flux

The result of the above calculation is that we have two sequences i and Fi that are both
given in terms of the residence time τi . If we sort the points (i ,Fi ) according to ascending
values of τi , we obtain a sequence of discrete points that describe a parametrized curve in
2D space. The first end point of this curve is at the origin: If no fluids have entered the
domain, the cumulative flow capacity is obviously zero. Likewise, full flow capacity is
reached when the domain is completely filled, and since we normalize both  and F by
their value at the maximum value of τ , this corresponds to the point (1,1). Given that both
F and  increase with increasing values of τ , we can use linear interpolation to define a
continuous, monotonic, increasing function F ().

13.2.2 Lorenz Coefficient and Sweep Efficiency

The Lorenz coefficient is a popular measure of heterogeneity and is defined as the difference
in flow capacity from that of an ideal piston-like displacement:
 1
Lc = 2 F () −  d. (13.10)
0

In other words, the Lorenz coefficient is equal twice the area under the F () curve and
above the line F = , and has values between zero for homogeneous displacement and
unity for an infinitely heterogeneous displacement; see Figure 13.4. Assuming that the

Figure 13.4 The definition of Lorenz coefficient from the F – diagram. The green line represents
a completely homogeneous displacement in which all flow paths have equal residence times. Te
blue line is a heterogeneous displacement with significant variation in residence times. The Lorenz
coefficient is defined as two times the gray area.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.2 Measures of Dynamic Heterogeneity 487

flow and storage capacity are given as two vectors F and Phi, the Lorenz coefficient can be
computed by applying a simple trapezoid rule
v = diff(Phi,1);
Lc = 2*(sum((F(1:end-1)+F(2:end))/2.*v) - .5);

This is implemented in the computeLorenz function in the diagnostics module.

We can also use the F – diagram to compute the volumetric sweep efficiency Ev , which
measures how efficient injected fluids are used. Here, Ev denotes the volume fraction of
inplace fluid that has been displaced by injected fluid, or equivalently, the ratio between
the volume contacted by the displacing fluid at time t and the volume contacted at time
t = ∞. In our streamtube analogue, only streamtubes that have not yet broken through will
contribute to sweep the reservoir. Normalizing by total volume, we thus have

q t  
Ev (t) = 1 − F (τ ) dτ
V 0
  F
qt q t qt q 
= − F (τ ) dτ = − F (t)t − τ dF
V V 0 V V 0
 
t 1 d
= 1 − F (t) + τ̄ d =  + (1 − F ) =  + (1 − F )tD .
τ̄ τ̄ 0 dF
The third equality follows from integration by parts, and the fourth equality since τ̄ = V /q
and τ dF = τ̄ d. Here, the quantity d/dF takes the role as dimensionless time. Prior
to breakthrough, Ev = tD . After breakthrough,  is the volume of fully swept flow paths,
whereas (1 − F )tD is the volume of flow paths being swept.
The implementation in MRST is quite simple and can be found in the utility function
computeSweep. Starting from the two arrays F and Phi, we first remove any flat segments
in F to avoid division by zero
inz = true(size(F));
inz(2:end) = F(1:end-1)=F(2:end);
F = F(inz);
Phi = Phi(inz);

Then, dimensionless time and sweep efficiency can be computed as follows:

tD = [0; diff(Phi)./diff(F)];
Ev = Phi + (1-F).*tD;

Figure 13.5 shows the F – and the sweep diagram for the simple example with two
injectors and three producers from Figure 13.1. For this particular setup, the Lorenz coef-
ficient is approximately 0.25, which indicates that we can expect a mildly heterogeneous
displacement with some flow paths that break through early and relatively small stagnant
regions. From the diagram of the sweep efficiency, we see that 70% of the fluids in place
can be produced by injecting one pore volume. By injecting two additional pore volumes
almost all the in-place fluid can be produced.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 488 Flow Diagnostics
1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 1 2 3 4 5 6

Figure 13.5 F – and sweep diagram for the simple case with two injectors and three producers,
which has a Lorenz coefficient of 0.2475.

F 1−F Ev

 td td

Figure 13.6 The three basic flow diagnostics curves: F – diagram, fractional recovery curve, and
sweep efficiency. All quantities , F , Ev , and tD are dimensionless; tD is given in terms of pore
volumes injected (PVI).

Altogether, we have defined three different curves that can be derived from the resi-
dence/travel time. The curves shown in Figure 13.6 are visually intuitive and emphasize
different characteristics of the displacement:

• The F – curve is useful for assessing the overall level of displacement heterogeneity.
The closer this curve is to a straight line, the better is the displacement.
• The fractional recovery curve emphasizes early-time breakthrough behavior and can have
utility as a proxy for fractional recovery of the fluid in place.
• Sweep efficiency highlights the behavior after breakthrough and has utility as a proxy for
recovery factor.
The curves can be defined for the field as a whole, or be associated with sector models,
individual swept volumes, well-pair regions, and so on.
Whereas visually intuitive information is useful in many workflows, others need mea-
sures defined in terms of real numbers. In our work, we have primarily used the Lorenz

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.3 Residence-Time Distributions 489

coefficient, but dynamic analogues of other classical heterogeneity measures can be defined
in a similar fashion. For instance, the dynamic Dykstra–Parsons’ coefficient is defined as
(F  )=0.5 − (F  )=0.841
VDP = , (13.11)
(F  )=0.5
where  = 0.5 corresponds to the mean storage capacity, while  = 0.841 is the mean
value plus one standard deviation. Likewise, we can define the dynamic flow heterogeneity
index,
FH I = F (∗ )/∗,
F  (∗ ) = 1. (13.12)
 
One can show that ( ddF
)i = ττ̄i , where τ̄ = V /q = Vi / qi is the average residence
time for all the streamtubes. These heterogeneity measures have not yet been implemented
in the diagnostics module.

Computer exercises
13.2.1 Implement Dykstra–Parsons’ coefficient (13.11) and the flow heterogeneity index
(13.12) as new functions for the diagnostics module.
13.2.2 Use time-of-flight values defined per influence region as discussed in Section
13.1.2 on page 482 to implement refined versions of the dynamic heterogeneity
measures.
13.2.3 Compute heterogeneity measures for each well pair in the model with two injectors
and three producers. (Original source code: showDiagnostBasics.) Are there
differences between the different regions?
13.2.4 Try to make the displacement shown in Figures 13.1 and 13.5 less heteroge-
neous by moving the wells and/or by changing the relative magnitude of the injec-
tion/production rates.
13.2.5 Define a set of multiphase flow simulations cases, e.g., one five-spot simulation for
each layer of the SPE 10 model (see Section 13.4.2). Which of the heterogeneity
measures above has the strongest correlation with the total hydrocarbon recovery
from multiphase simulations?

13.3 Residence-Time Distributions

You have previously seen how we can compute point-wise values of time-of-flight (and
influence regions) by tracing streamlines or compute cell-averaged values by solving a
steady, linear transport problem using a finite-volume method. This section sheds more light
into the finite-volume approach and discusses how we can overcome some of its potential
limitations. The discussion follows [171] closely and presents functionality that is new in
MRST 2018b.
We assume incompressible flow to simplify the discussion and start by writing the finite-
volume discretization of the time-of-flight equation (13.1) in vector form as V τ = φ. Here,
τ is the vector of unknown time-of-flight values, φ is the vector of pore volumes per cell,

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 490 Flow Diagnostics

and V is the upstream flux matrix corresponding to the operator v ·∇. Consider a cell j with
j
total influx vji , and let cb denote the vector of backward tracer concentrations corresponding
to an imaginary experiment in which a tracer is injected in cell j and allowed to flow in
the reverse direction of v. Let χ j denote the characteristic vector, which equals one in
cell j and zero in all other cells. We can then write the discrete equations for the forward
j
time-of-flight τ and the backward influence region cb as follows
j
V τf = φ and V T cb = χ j vji . (13.13)
From this, we obtain the following expression for the time-of-flight value in cell j
 T
T −1 1 j
τ f [j ] = χ j V φ = c φ. (13.14)
vji b
Hence, τ f [j ] equals the pore volume of the backward influence region of cell j divided by
the influx into the cell. If cell j lies close to a producer, the backward influence region
typically contains flow paths with very different residence times. This implies that the
forward time-of-flight value computed by the finite-volume method represents the average
of a distribution that potentially can have (very) large variance. A similar argument applies
to the backward time-of-flight. The averaging will introduce a systematic bias when we use
averaged time-of-flight values to compute the dynamic heterogeneity measures discussed
in the previous section. This bias may be acceptable in many applications, e.g., if we
use dynamic heterogeneity measures to rank different model realizations or as a simple
reduced-order model to predict recovery of secondary oil recovery [144, 218]. On the other
hand, residence times computed from time-of-flight values will in most cases overestimate
the time to breakthrough in heterogeneous displacements.
For a better description of the dynamic heterogeneity of a reservoir model, we can con-
sider the distribution of time-of-flight for each cell. This is particularly interesting for cells
perforated by production wells, since pointwise time-of-flight values in these cells describe
the residence times, or time to breakthrough, for individual flow paths. To determine this
residence-time distribution (RTD), we start by considering the linear transport equation
∂c
+ v · ∇c = 0, c|i = δ(t),c(
φ x,0) = 0, (13.15)
∂t
which describes the transport of a unit pulse through a domain  from an inflow boundary
i to an outflow boundary o . (The inflow and outflow will obviously be through wells in
most reservoir models, but in the following I refer to these as boundaries for mathematical
convenience.) For each point x, the normalized time-of-flight distribution T (·; x) is now
simply defined by the Dirac function

T (t; x) = c(
x,t) = δ t − τ (
x) . (13.16)
At the outflow, the normalized RTD is given as
 
1
To (t) = c v · n ds, Fo = v · n ds. (13.17)
Fo o o

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.3 Residence-Time Distributions 491
'
It follows from the definition of the Dirac distribution that To (t) dt = 1. From the RTD,
we obtain flow capacity and storage capacity as follows [277]
 t 
Fo t
F (t) = To (s) ds, (t) = s To (s) ds. (13.18)
0 o 0
As before, both quantities are normalized so that F (∞) = (∞) = 1. From this definition,
it also follows that the mean value of To (t) corresponds to the time t¯ = o /Fo it takes to
inject one pore volume (1 PVI).
To compute discrete approximations to the time-of-flight and RTD, we introduce a finite-
volume approximation,
dc qi
+ Mc = 0, c(0) = c0 = . (13.19)
dt φ

Here, M is the discretization of the linear operator φ −1 v · ∇ defined in the same way as for
the time-of-flight equation, and q i is the vector of inflow flow rates. In the diagnostics
module, we use a backward Euler method to solve (13.19) directly. This method is robust
and can take large time steps, but is not very accurate.
Alternatively, you can write the solution formally in terms of matrix exponentials c(t) =
exp(−tM)c0 . Let q o be the vector that collects the flux across all outflow boundaries of
individual cells, 1 be a vector with value one in each cell, and o be the total pore volume
drained by the outflow boundary o . We then represent the discrete counterparts of (13.16)–
(13.17) as follows

T (t)[j ] = χ Tj exp(−tM)c0 and To (t) = q To exp(−tM)c0 /q To 1. (13.20)

The vector T contains time-of-flight distributions in all cells (with value T [j ] in cell j )
and To is the RTD over o . You can then employ a rational Padé approximation to evaluate
the action of the matrix exponential,

T (t + t) = exp(−tM)T (t) ≈ P (−tM) Q(−tM)−1 T (t), (13.21)

for suitable polynomials P and Q. One example would be to use the first-order polynomials
P (x) = 1 + x/2 and Q(x) = 1 − x/2 to reduce fill in, i.e.,
  −1
T (t + t) = I − 12 t M I + 12 t M T (t).

For each successive value you compute, you must solve a linear system and perform a
matrix multiplication. The matrix I + M has the same sparsity as the discretization matrix
for the time-of-flight equation; i.e., is triangular, possibly after a permutation. Each linear
solve is thus highly efficient, but the time step is restricted by a CFL condition because of
the explicit part of the discretization coming from the polynomial P (x). In our experience,
using backward Euler is thus more efficient.
The next two examples illustrate the difference between averaged values and distribu-
tions and how these differences impact measures of dynamic heterogeneity.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 492 Flow Diagnostics

Example 13.3.1 We consider a single horizontal layer from SPE 10 with an injector along
the south boundary and a producer along the north boundary. Assume we have computed
a flow solution with fluxes represented in state, cell-wise pore volumes in pv, time-of-
flight and tracer values represented in the structure D discussed on page 480, and well pair
information in WP. The following routine computes the propagating pulse using (13.19) with
a backward Euler time discretization and extracts the outflux observed in the producer as
function of time:
rtd = computeRTD(state, G, pv, D, WP, W);

The solid lines in Figure 13.7 show an unnormalized version of To (t) as function of time for
two layers; here the integral equals the total allocation. The leading pulse for the Tarbert
layer is spread out and has a small secondary hump. For Upper Ness, the pulse breaks
through earlier and is more focused because of high-permeable channels connecting the

1
1

0.8
0.8

0.6 0.6

0.4 0.4

0.2 0.2
Simulated pulse: L=0.26536 Simulated pulse: L=0.61315
From averaged time-of-flight: L=0.23889 From averaged time-of-flight: L=0.56793
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Figure 13.7 Residence-time distributions and F – diagrams for two permeability field sampled
from the SPE 10 benchmark. Solid lines are computed by tracing a unit pulse through the reservoir
to determine the distribution at the outlet. Dashed lines are obtained by solving the time-of-flight
equation by a finite-volume method and backing out data for representative flow paths through each
of the cells of the model. Thin solid and dotted lines represent the mean of the distribution, i.e., the
time it takes to inject one pore volume, and the time to breakthrough for the fastest flow path.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.3 Residence-Time Distributions 493

south and north boundaries. The mean of each distribution equals 1 PVI by construction.
This may not be apparent from the plots since the distributions have very long tails, partic-
ularly in the channelized case.
We can also estimate this distribution from the average residence times τ r = τ f + τ b ,
obtained by solving the forward and backward time-of-flight equations. We start by using
the relationship fj = φ[j ]/τ r [j ] to back out the flux
D = computeTOFandTracer(state, G, rock, 'wells', W, ..)
f = poreVolume(G,rock)./sum(D.tof,2);

Plotting f against τ r values sorted in ascending order and normalized by 1 PVI gives
a discrete (point) description of how a given total flux through the permeable medium will
distribute itself over a finite number of flow paths. This distribution is highly irregular, and
to get a continuous distribution we sort the τ r values into small bins, sum the total flux
within this bin, and compute an averaged flux density value over the whole length of the
bin. This is implemented in the function
RTD = estimateRTD(pv, D, WP);

Using averaged time-of-flight values introduces a significant delay in the breakthrough

time, in particular for Upper Ness. The two types of distributions also suffer from different
types of numerical errors: Tracing a pulse by a finite-volume method preserves flux allo-
cation and not pore volume and can also contain significant temporal smearing. Backing
out a distribution from averaged time-of-flight values preserves total volume but not flux
allocation. This is taken into account by the following function, which computes F –
diagrams from a RTD,
[f, phi] = computeFandPhi(rtd);

Figure 13.7 also compares the resulting F – diagram with similar diagrams computed
directly from τr and pore volume. As expected, the diagrams computed from the RTD are
more concave and the corresponding Lorenz coefficients are somewhat larger than when
computed from averaged time-of-flight values. However, the differences are not very large,
and the measures computed from averaged time-of-flight values carry a systematic under-
estimation bias and can thus still be robustly used to rank and discriminate different cases.
Notice that we do not report F – diagrams derived from estimated RTDs. The binning
and averaging used in the estimation process introduces a strong smoothing that generally
obfuscates dynamic heterogeneity measures. Hence, we strongly advise against computing
any heterogeneity measure from estimated RTDs; these distributions should only be used
for illustrations like Figure 13.7.
Example 13.3.2 We repeat the same setup as in the previous example, but now with an
injector placed slightly off-center along the south perimeter and two producers placed
symmetrically at the northern perimeter. Figure 13.8 reports residence distributions and
F – diagrams for each of the two injector pairs. In both cases, producer P2 is connected

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 494 Flow Diagnostics

1
1

0.8
0.8

0.6 0.6

0.4 0.4

0.2 Simulated pulse: P1 0.2 Simulated pulse: P1

Simulated pulse: P2 Simulated pulse: P2
Simulated pulse: whole field Simulated pulse: whole field
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Figure 13.8 Residence-time distributions and F – diagrams for two permeability field sampled from
the SPE 10 benchmark with one injector and two producers.

to the injector through more permeable rocks than producer P1, even though the physical
distance from I to P2 is larger than to P1. The fastest flow paths thus break through earlier
in P2 than in P1, and overall, producer P2 allocates a significantly larger portion of the
flux than P1.
Next, we look at the dynamic heterogeneity. For Tarbert, the (I,P2) well pair allocates
approximately twice as much flux and a 30% larger flow region than (I,P1). The majority of
the flow paths in the (I,P2) have small residence times and dynamically, this region is less
heterogeneous than (I,P1), whose RTD contains a heavier tail. Hence, the F – curve of P1
lies above P2 and has a higher Lorenz coefficient. For the reservoir as a whole, these two
effects balance each other: In the beginning, the produced tracer comes predominantly
from fast flow paths connected to P2, whereas the last tracer volumes produced come
predominantly from slow flow paths connected to P1. Hence, fast flow paths from P2 ensure
that the F – curve of the whole reservoir lies below the curve of P1, whereas contributions
from the heavier tail of P1 ensure that the overall curve lies above that of P2.
For Upper Ness, 50% more flux is allocated to P2 even though the associated flow region
is smaller than for P1. In the (I,P1) region, all flow paths traverse regions of both high and

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.4 Case Studies 495
−14
10 P1
P2
Overall

−16
10

0 20 40

−18
10

0 500 1,000 1,500 2,000 2,500 3,000 3,500 4,000

time [years]

Figure 13.9 Residence-time distributions for the Upper Ness case shown on a semi-logarithmic axis.
(One PVI corresponds to 41 years.)

low permeability, which means that this region is more homogeneous than the (I,P2) region,
which contains a large fraction of flow paths that (almost) exclusively traverse cells with
high permeability. Hence, the F – curve of P2 lies above P1. Explaining the F – curve
of the whole reservoir is significantly more difficult. Figure 13.9 shows the RTD on a semi-
logarithmic axis. During the first 20 years, tracer production is predominantly from P2.
Producer P1 has the largest contribution from year 27 to year 1220, and then P2 con-
tributes most for the rest of the simulated period. Obviously, this exercise has nothing to do
with real reservoir engineering but serves to illustrate how tiny details on (unrealistically)
long time scales affect measures like F – diagrams.
Distributions of residence times are used in the study of chemical reactors and tracer
tests [277, 139]. In [218] we demonstrate how you can convolve these distributions with
representative, 1D displacement profiles to develop reduced-order models that, e.g., enable
you to quickly estimate and discriminate improvements in macroscopic and microscopic
displacement efficiency for polymer flooding.

13.4 Case Studies

The use of flow diagnostics is best explained through examples. In this section we therefore
go through several cases and demonstrate various ways flow diagnostics can be used to
enhance our understanding of flow patterns and volumetric connections, tell us how to
change operational parameters such as well placement and well rates to improve recovery,
and so on.

13.4.1 Tarbert Formation: Volumetric Connections

As our first example, we consider a subset of the SPE 10 data set consisting of the top
twenty layers of the Tarbert formation; see Section 2.5.3. We modify the original inverted
five-spot well pattern by replacing the central injector by two injectors that are moved a

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 496 Flow Diagnostics

Figure 13.10 Porosity and well positions for a model consisting of subset of the Tarbert formation in
Model 2 from the 10th SPE Comparative Solution Project

short distance from the model center (see Figure 13.10), assume single-phase incompress-
ible flow, and solve the corresponding flow problem. You can find complete description of
the setup in showWellPairsSPE10.m in the book module.
Given the geological model represented in terms of the structures G and rock, the wells
represented by W, and the reservoir state including fluxes by rS, we first compute the time-
of-flight and volumetric partitions:
D = computeTOFandTracer(rS, G, rock, 'wells', W);

This gives us the information we need to partition the volume into different drainage and
sweep volumes. The simplest way to do this for the purpose of visualization is to use a
majority vote over the injector and producer partitions to determine the well that influences
each cell the most as shown in Figure 13.11. The result of this majority vote is collected in
D.ppart and D.ipart, respectively, and the essential commands to produce the two upper
plots in Figure 13.11 are:
plotCellData(G,D.ipart, ..);
plotCellData(G,D.ppart,D.ppart>1, ..);

Since there are two injectors, we would expect to only see two colors in the to plot of sweep
regions. However, there is also a small blue volume inside the triangular section bounded by
I1, P1, and P2, which corresponds to an almost impermeable part of the reservoir that will
not be swept by any of the injectors. The well pattern is symmetric and for a homogeneous
medium we would therefore expect that the two pressure-controlled injectors would sweep
symmetric volumes of equal size. However, the Tarbert formation is highly heterogeneous
and hence the sweep regions are irregular and clearly not symmetric. The injection rate of
I2 is approximately six times that of I1 because the wells are completed in cells with very
different permeability, hence I2 will sweep a much larger region than I1. In particular, we

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.4 Case Studies 497

Figure 13.11 Sweep (top) and drainage regions (middle) determined by a majority vote over injector
and producer partitions, respectively, for the Tarbert model. The bottom plot shows producer influence
region without majority vote, with gray color signifying cells that affect multiple producers.

see that I2 is the injector that contributes most to flooding the lower parts of the region
near P3, even though I1 is located closer. Looking at the drainage and sweep regions in
conjunction, it does not seem likely that I1 will contribute significantly to support the
production from wells P1 and P2 unless we increase its rate.
When using a majority vote to determine drainage and sweep regions, we disregard that
individual cells may be influenced by more than one well. To better visualize the influence
regions, we can blend in a gray color in all cells in which more than one color has nonzero
concentration, as shown in the bottom plot of Figure 13.11. The plot is generated by the
following call:
plotTracerBlend(G, D.ppart, max(D.ptracer, [], 2), ..);

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 498 Flow Diagnostics
< 1%

20% 22%

< 1% I1, P1
I2, P1
I1, P2 3%
I2, P2
I1, P3
I2, P3
23%
I1, P4
I2, P4 18%

13%

Figure 13.12 Well-pair regions and associated fraction of the total pore volume for the upper 20
layers of the Tarbert formation.

Having established the injection and producer partitions, we can identify well pairs and
compute the pore volumes of the region associated with each pair:
WP = computeWellPairs(rS, G, rock, W, D);
pie(WP.vols, ones(size(WP.vols)));
legend(WP.pairs,'location','Best');

To visualize the volumetric regions, we compute the tensor product of the injector and
producer partitions and then compress the result to get a contiguous partition vector with a
zero value signifying unswept regions:
p = compressPartition(D.ipart + D.ppart*max(D.ipart))-1;
plotCellData(G,p,p>0,'EdgeColor','k','EdgeAlpha',.05);

The result shown in Figure 13.12 confirms our previous observations of the relative impor-
tance of I1 and I2. Altogether, I1 contributes to sweep approximately 16% of the total pore
volume, shown as the light red and the yellow regions in the 3D plot.
It is also interesting to see how these volumetric connections affect the fluxes in and
out of wells. To this end, we should look at the cumulative well allocation factors, which
are defined as the cumulative flux in/out of a well from the bottom to the top perforation
of a vertical well, or from toe to heel for a deviated well. We start by computing the flux
allocation manually for the two injectors (outflow fractions are given from well head and
downward and hence must be flipped):
for i=1:numel(D.inj)
subplot(1,numel(D.inj),i); title(W(D.inj(i)).name);
alloc = cumsum(flipud(WP.inj(i)).alloc,1);
barh(flipud(WP.inj(i).z), alloc,'stacked'); axis tight
lh = legend(W(D.prod).name,4);
set(gca,'YDir','reverse');
end

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.4 Case Studies 499
I1 I2

2 2

4 4

6 6

8 8

P1 P1
10 P2 10 P2
P3 P3
P4 P4
12 12
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
−3 −3
x 10 x 10

Figure 13.13 Well allocation factors for the two injectors of the Tarbert model.

Figure 13.13 shows the resulting bar plots of the cumulative allocation factors. These
plots confirm and extend the understanding we have developed by studying volumetric
connections: I1 will primarily push fluids towards P3. Some fluids are also pushed towards
P4, and we observe that there is almost no outflow in the top three perforations where
the rock has low quality. Injection from I2, on the other hand, contributes to uphold the
flux into all four producers. We also see that the overall flux is not well balanced. Pro-
ducer P1 has significantly lower inflow than the other three. Alternatively, we can use
plotWellAllocationPanel(D, WP) to compute and visualize the well allocation factors
for all the wells in the model, as shown in Figure 13.14. (This function is only useful for a
small number of wells.)
When your simulation case contains only a few well pairs, you may get significantly
more information about the flow patterns in the reservoir if you subdivide the wells into
multiple segments (sets of completions). To exemplify, let us look more closely at the
performance of the different completions along the well paths. To this end, we divide the
completion intervals into bins and assign a corresponding set of pseudo-wells for which we
recompute flow diagnostics. As an example, we split the completions of I1 into three bins
and the completions of I2 into four bins.
[rSp,Wp] = expandWellCompletions(rS,W,[5, 3; 6, 4]);
Dp = computeTOFandTracer(rSp, G, rock, 'wells', Wp);

Figure 13.15 shows the majority-voted sweep regions for the four segments of I2. To better
see the various sweep regions, the 3D plot is rotated 180 degrees compared with the other
3D plots of this model. To obtain the figure, we used the following key statements:
plotCellData(G, Dp.ipart,(Dp.ipart>3) & (Dp.ipart<8),...);
WPp = computeWellPairs(rSp, G, rock, Wp, Dp);
avols = accumarray(WPp.pairIx(:,1),WPp.vols);
pie(avols(4:end));

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 500 Flow Diagnostics

P1 P2

P1 P2 P3

P3 P4

I1 I2

P4 I1 I2

Figure 13.14 Normalized well allocation factors for all wells of the Tarbert model.

27%

48%

18%

7%

Figure 13.15 Majority-voted sweep regions for the Tarbert case with I2 of divided into four segments
that each are completed in five layers of the model. (Notice that the view angle is rotated 180 degrees
compared with Figure 13.11.) The pie chart shows the fraction of the total sweep region that can
attributed to each segment of the well.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.4 Case Studies 501

Notice, in particular, that fluids injected in the lowest segment is the major contributor in
almost half of the well’s total sweep region.

13.4.2 Heterogeneity and Optimized Well Placement

In this example, we first compute the Lorenz coefficient for all layers of the SPE 10 model
subject to an inverted five-spot well pattern. We then pick one of the layers and show how
we can balance the well allocation and improve the Lorenz coefficient and the areal sweep
by moving some of the wells to regions with better sand quality.
To compute Lorenz coefficient for all layers in the SPE 10 model, we first define a
suitable 60 × 220 × 1 grid covering a rectangular area of 1,200 × 2,200 × 2 ft3 . Then, we
loop over all the 85 layers using the following essential lines:
for n=1:85
rock = getSPE10rock(1:cartDims(1),1:cartDims(2),n);
rock.poro = max(rock.poro, 1e-4);
T = computeTrans(G, rock);

for w = 1:numel(wtype),
W = verticalWell(..);
end
rS = incompTPFA(initState(G, W, 0), G, T, fluid, 'wells', W);

D = computeTOFandTracer(rS, G, rock, 'wells', W, 'maxTOF', inf);

[F,Phi] = computeFandPhi(poreVolume(G,rock), D.tof);
Lc(n) = computeLorenz(F,Phi);
end

Transmissibility is recomputed inside the loop because the permeability changes for
each new layer. Likewise, we regenerate the well objects to ensure correct well indices
when updating the petrophysical data. You find the complete source code in the script
computeLorenzSPE10 in the book module.
Figure 13.16 reports Lorenz coefficients for all layers when the injector and the four pro-
ducers are controlled by bottom-hole pressure. The relatively large dynamic heterogeneity
and the variation among individual layers within the same formation is a result of our
choice of well controls. The actual injection and production rates achieved with pressure-
controlled wells are very sensitive to each well being perforated in a region of good sand
quality. This is almost impossible to ensure when using fixed well positions for all layers
of SPE 10, and hence pressure-controlled wells will generally accentuate heterogeneity
effects. To see this, you should modify the script and rerun the case with equal rates in all
four producers.
Our overall setup and somewhat haphazard well placement is not representative of a real
reservoir, but serves well to illustrate our next point. Since the Lorenz coefficient generally
is quite large, most of the cases would have suffered from early breakthrough if we were

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 502 Flow Diagnostics

0.6

0.4

0.2

0
10 20 30 40 50 60 70 80

Figure 13.16 Lorenz coefficient for the 85 horizontal layers of the SPE 10 model subject to an
inverted five-spot pattern with injector and producers controlled by bottom-hole pressure. The dashed
line shows the border between the Tarbert and Upper Ness layers.

Figure 13.17 Well configuration and flux allocation for the four well pairs with initial well
configuration for Layer 61. (Red colors are good sands, while blue colors signify sands of low
permeability and porosity.)

to use this initial well placement for multiphase fluid displacement. Let us therefore pick
one of the layers and see if we can try to improve the Lorenz coefficient and hence also
the sweep efficiency. Figure 13.17 shows the well allocation and the sand quality near the
producers for Layer 61, which is the layer giving the worst Lorenz coefficient. The flux
allocation shows that we have a very unbalanced displacement pattern where producer P4
draws 94.5% of the flux from the injector and producers P1 and P2 together only draw
1.2%. This can be explained by looking at the sand quality in our reservoir. Producer P1 is
completed in a low-quality sand and will therefore achieve low injectivity if all producers
operate at the same bottom-hole pressure. Producers P2 and P3 are perforated in cells with
better sand, but are both completely encapsulated in regions of low sand quality. On the
other hand, producer P4 is connected to the injector through a relatively contiguous region
of high-permeable sand.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.4 Case Studies 503
Lorenz Sweep
1 1

0.8 0.8

Lorenz: 0.78
0.6 0.6
Lorenz: 0.47

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 5 10 15 20 25

Figure 13.18 F – and sweep diagrams before (dashed line) and after (solid line) producers P1 to P4
have been been moved to regions with better sand quality.

Figure 13.19 Well configuration and flux allocation for the four well pairs after all producers have
been moved to regions with better sand quality.

The largely concave F – diagram shown to the left in Figure 13.18 testifies that the
displacement is strongly heterogeneous and is characterized by large differences in the
residence time of different flow paths, or in other words, suffers from early breakthrough of
the displacing fluid. Hence, we must inject large amounts of the displacing fluid to recover
the hydrocarbons from the low-quality sand; this can be seen from the weakly concave
sweep diagram to the right in Figure 13.18. Altogether, we should expect a very unfavorable
volumetric sweep from this well placement.
To improve the displacement, we can try to move each producer to a better location. To
this end, we should look for cells in the vicinity of each well that have better sand quality
(higher porosity and permeability) and are connected to the injector by more contiguous
paths of good quality sands. Figure 13.19 shows the well allocation after we have moved
P1 to P3. Producer P4 was already in a good location and is only shifted one cell. The
result is a much more balanced well allocation, which is confirmed by the rates reported in
Table 13.1. (Notice also that the overall reservoir rate has increased slightly.) Figure 13.18

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 504 Flow Diagnostics

Table 13.1 Volumetric flow rates in units 10−5 m3 /s for each of the
wells in Layer 61 of the SPE 10 model. (These are obviously far
from rates seen in real reservoirs.)

Placement P1 P2 P3 P4 I

Initial −0.0007 −0.0020 −0.0102 −0.2219 0.2348

Improved −0.0719 −0.0660 −0.0604 −0.0398 0.2381

shows how the F – diagram has become significantly less concave and that the sweep
diagram has become more concave. This testifies that the variation in residence times
associated with different flow paths is much smaller. We should thus expect a more efficient
and less heterogeneous volumetric sweep.
Dynamic heterogeneity measures correlate surprisingly well with recovery for water-
flooding processes [278, 144, 259, 260, 218], and are generally easy to use as a guide when
searching for optimal or improved well placement. Because of their low computational
cost, the measures can be used as part of a manual, interactive search or combined with
more rigorous mathematical optimization techniques. In the diagnostics module you can
find examples that use flow diagnostics in combination with adjoint methods to determine
optimal well locations and set optimal injection and production rates. The interested reader
can find more details about optimization based on flow diagnostics in [218].

Computer exercises

13.4.1 Repeat the experiment using fixed rates for the producers (and possibly also for
the injector). Can you explain the differences that you observe?
13.4.2 Compute Lorenz coefficients from RTDs instead and compare.
13.4.3 Use the interactive diagnostic tool introduced in the next section to manually adjust
the bottom-hole pressures (or alternatively the production rates) to see if you can
improve the sweep even further.
13.4.4 Can you devise an automated strategy that uses flow diagnostics to search for
optimal five-spot patterns?
13.4.5 Cell-averaged time-of-flight values computed by a finite-volume scheme can be
quite inaccurate when interpreted pointwise, particularly for highly heterogeneous
media. To investigate this, pick one of the fluvial layers from SPE 10 and use the
pollock routine from the streamline module to compute time-of-flight values
on a 10 × 10 subsample inside a few selected cells. Alternatively, you can use the
finite-volume method on a refined grid.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.5 Interactive Flow Diagnostics Tools 505

13.5 Interactive Flow Diagnostics Tools

The ideas behind most flow diagnostics techniques are relatively simple to describe and
their computation is straightforward to implement. The real strength of these techniques
nevertheless lies in their visual appeal and the ability for rapid user interaction. Together,
MATLAB and MRST provide a wide variety of powerful visualization routines you can use
to visualize input parameters and simulation results. The diagnostics module supplies
additional tools for enhanced visualization, but using a script-based approach to visualiza-
tion means that you each time need to write extra code lines to manually set color map and
view angle or display various additional information such as legends, colorbars, wells, and
figure titles.
We have mostly omitted these extra code lines in our previous discussion, but if you
go in and examine the accompanying scripts, you will see that a large fraction of the code
lines focus on improving the visual appearance of plots. Such code is repetitive and should
ideally not be exposed to the user of flow diagnostics. More importantly, a script-based
approach gives a static view of the data and offers limited capabilities for user interaction,
apart from zooming, rotating, and moving the displayed data sets. Likewise, a new script
must be written and executed each time we want to look a new plot that combines various
types of diagnostic data, e.g., to visualize time-of-flight or petrophysical values within a
given color region, use time-of-flight to threshold the color regions, etc.

Preprocessing: Predicting Fluid Movement

To simplify the user interface to flow diagnostics, we have integrated most of the flow
diagnostics capabilities discussed earlier in the chapter into a graphical tool that enables
you to interact more directly with your data set
interactiveDiagnostics(G, rock, W);

The interesting part of this GUI discussion is not the exact layout of the user interface,
which likely will change in future versions of MRST, but the different quantities it can
produce.
The GUI has primarily been written for preprocessing to predict future fluid movement
and uses the standard two-point incompressible flow solver for a single-phase fluid with
density 1,000 kg/m3 and viscosity 1 cP to compute a representative flow field, which is then
fed to the function computeTOFandTracer to compute the basic flow diagnostics quantities.
Once the computation is complete, the graphical user interface is launched; see Figure
13.20: At the time of writing, this consists of a plotting window showing the reservoir
model and a control window, which contains a set workflow tabs and menus that enable
you to use flow diagnostics to explore volumetric connections, flow paths, and dynamic
heterogeneity measures.
The control window has three different tabs. The “Region selection” tab is devoted
to displaying the various kinds of volumetric regions discussed in Section 13.1.1. The
“Plots” tab lets you compute F – diagram, Lorenz coefficient, and the well allocation

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 506 Flow Diagnostics

Select producers

Select injectors

Different control tabs

How to combine colors

Select quantity to display

Threshold on time-of-flight

Play time-of-flight movie

Figure 13.20 A graphical user interface to flow diagnostics. The plotting window shows forward
time-of-flight, which is the default value displayed upon startup. In the control window, we show
the “Region selection” tab that lets you select which quantity to show in the plotting window, select
which wells to include in the plot, specify how to combine color concentrations to select volumetric
regions, and as well set maximum and minimum time-of-flight values to crop the volumetric regions.

panels shown in Figure 13.14. From this tab, you can also bring up a dialog box to edit
the well settings and recompute a new flow field and the resulting flow diagnostics (unless
the parameter ' computeFlux ' is set to false). In the “Advanced” tab, you can control the
appearance of the 3D plot by selecting whether to display grid lines, well-pair information,
and well paths, set 3D lighting and transparency value, etc. This tab also allows you to
export the current volumetric subset as a vector of boolean cell indicators.
Regions to be displayed in the plotting window are specified by selecting a set of active
wells and by choosing how the corresponding colorings should be combined. That is, for
a given set of active wells, you can either display all cells that have nonzero concentration
for the injector or producer colors, only those cells that have nonzero concentration for
injector and producer colors, only cells with nonzero concentration of producer colors, or
only cells with nonzero concentrations for injector colors. To select a set of active wells,
you can either use the list of injectors/producers in the control window, or mark wells in the
plotting window. In the latter case, the set of active wells consist of the well you selected
plus all other wells this well communicates with. Selecting a well in the 3D view also
brings up a new window that displays a pie chart of the well allocation factors and a graph
that displays cumulative allocation factors per connection. Figure 13.21 shows this type of
visualization for a model where we also have access to a set of states from a full multiphase
simulation. In this case, we invoke the GUI by calling
interactiveDiagnostics(G, rock, W, 'state', state, 'computeFlux', false);

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.5 Interactive Flow Diagnostics Tools 507
Volumetrics for drainage zone of E-1H Simulated production for E-1H
Well allocation factors Allocation by connection Production data for E−1H
reservoir 1 8000 35
C−3H Water BHP
Oil
Injectors and fluid expansion F−1H
2 7000
Gas
F−2H 30
contributing to oil production

Rates [m3/day] (liquid), [1000m3/day] (gas)

6000
C−2H 3
7
x 10 E−1H: Injector distribution for Oil 25
3

Bottom hole pressure [MPa]

reservoir
C−3H
F−1H 4 5000
C−2H
2.5 C−1H C−1H 20
C−4H
F−2H C−4H

Connection #
5 4000
2
15

6 3000
1.5
1
10
0.8 7 2000
1

0.6 5
0.5 8 1000

0.4
0
2 4 6 8 10 12 9 0 0
0 1 2 3 4 5 6 7 8 9 10
0.2 Time [year]

0 10
2 4 6 8 10 12 0 2000 4000 6000 8000
TOF distance in years Accumulated flux [m3/day]

Notice: prediction of
water breakthrough

Figure 13.21 Flow diagnostics used to analyze volumetric connections and predict future inflow of
fluids to well E-1H in a three-phase, black-oil flow simulation of the Norne field. For comparison,
the figure also reports the simulated production from E-1H, where the dashed vertical line represent
present time.

so that the fluxes used to compute time-of-flight and color concentrations are extracted
from the given reservoir state given in state. Left-clicking on producer E-1H brings up a
plot of the well allocation together with a plot of the fluid distribution displayed as function
of the backward time-of-flight from the well. To further investigate the flow mechanism,
we can mark each fluid and get a plot of how various injectors and fluid expansion in the
reservoir contribute to push the given fluid toward the producer. In Figure 13.21 we have
expanded the GUI by another function that plots the simulated well history of each well.
This history confirms what we see in the backward time-of-flight analysis of the drainage
zone: a few years into the future, water arrives at the producer (blue line), and there is a
marked decay in oil and gas production (red and green lines).
The GUI also offers functionality to load additional cell-based data sets that can be
displayed in the 3D plot:
interactiveDiagnostics(G, rock, W, celldata);
interactiveDiagnostics(G, rock, W, celldata, 'state', state);

and has some functionality for postprocessing simulations with multiple time steps, but this
is beyond the scope of the current presentation.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 508 Flow Diagnostics

Figure 13.22 Variation in sweep regions over eight time report steps from an ECLIPSE simulation of
the Brugge benchmark case. Clear color indicate regions swept by a single injector in all time steps,
whereas grayish colors indicate regions being swept by more than one injector over the time interval.

Postprocessing: Enhanced Visualization of Simulation Results

As of release 2018b, MRST also offers another user interface dedicated entirely to postpro-
cessing simulation data
PostProcessDiagnostics()
PostProcessDiagnostics('filename')

Figures 13.22–13.24 illustrate how flow diagnostics is used to enhance the understanding
of the fluid communication for a full ECLIPSE simulation of the Brugge benchmark case
[249].
Figure 13.22 shows the average sweep regions over eight report steps for all injectors,
limited outward by an upper bound on the forward time-of-flight. Regions swept by a single
injector over the whole simulation period are shown as clear colors, whereas diffuse or
grayish colors represent regions swept by multiple injectors. You can use such plots to
analyze how sweep or drainage regions change over time. The matrix plot to the lower left
reports communication strength between each well pair, with rows representing injectors
and columns injectors. The lower right plot shows how the dynamic heterogeneity mea-
sured by the Lorenz coefficient changes over the eight time steps. By selecting a single
injector or producer, you can break up this analysis to see the dynamic heterogeneity of
individual sweep, drainage, or well-pair regions, measured in terms of F – diagrams or
Lorenz coefficients.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.5 Interactive Flow Diagnostics Tools 509

Figure 13.23 Sweep region of injector BR-I-6 for the Brugge benchmark case. The lower plots show
flux allocation for the injector and the RTD for a passive tracer tracer released from the injector.

Figure 13.24 Saturation changes within the sweep region of injector BR-I-6 for the Brugge
benchmark case. The lower-right plot reports flux allocation every half year for the injector from
toe to heel (top to bottom in the plot). The lower-right plot reports water production rates for all
connected producers.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 510 Flow Diagnostics

Figure 13.23 shows oil saturation inside the sweep region of BR-I-6. From the flux
allocation in the lower-left plot, we see that this injector primarily supports BR-P-16, BR-
P-4, and BR-P-1 with the current set of drive mechanisms. Producer BR-P-15 lies closer
than any other wells, but has been shut in (see lower-right plot in Figure 13.24) and is
only shown since significant fluid volumes have been going to this producer earlier in the
simulation. Also BR-P-17 lies closer to BR-I-6 than the other producers with large flux
allocation, but as you can see from Figure 13.22, this well lies in the middle of BR-I-5’s
sweep region and is predominantly supported by that injector. The lower-right plot reports
the first 30 years of an RTD analysis and shows that a passive tracer injected now from
BR-I-6 will first break through in BR-P-16, and then in BR-P-4 and BR-P-1. This gives
time lines for displacement fronts originating from BR-I-6.
Figure 13.24 reports accumulated saturation changes inside the sweep region over the
whole simulation period. We clearly see how the displacement profile has engulfed pro-
ducers BR-P-15 to 17 and caused a decline in production rates. In particular, BR-P-15
hardly managed to stay on plateu and started its decline less than half a year after it started
operating.
The types of flow diagnostics discussed previously are also available in ResInsight,
which is an open-source tool for postprocessing of simulation results on ECLIPSE format.
This software is part of the OPM Initiative and currently functions as an outlet of research
ideas from MRST towards professional use in industry.
In the rest of the chapter, we use the preprocessing GUI to study two reservoir models.
The accompanying scripts do not reproduce all figures directly, but specify the manual
actions you need to reproduce many of the figures.

13.5.1 Synthetic 2D Example: Improving Areal Sweep

The first example considers a slightly modified version of the setup from Figure 13.1. We
have introduced two low-permeable barriers, moved the two injectors slightly to the south,
and switched all wells from rate to pressure control; see Figure 13.25. Table 13.2 reports
specific values for the well controls under the label “base case.”
Our previous setup had a relatively symmetric well pattern in which producers P1 and
P3 were supported by injectors I1 and I2, respectively, whereas producer P2 was supported
by both injectors. This symmetry is broken by the two barriers, and now injector I1 also
provides significant support for producer P3. You can infer this from the plot of producer
partition: the gray areas between the magenta (P2) and red (P3) regions signify parts of the
reservoir that are drained by both producers. The relatively large gray area southeast of I1
indicates that this injector supports both P2 and P3. There is also a gray area southwest
of injector I2, but since this is less pronounced, we should only expect that only a small
portion of P2’s inflow can be attributed to I2. To confirm this, you can load the model in
the interactive viewer (see the script interactiveSimple), and click the names for each
individual producer to bring up a pie chart reporting the corresponding flux allocation.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.5 Interactive Flow Diagnostics Tools 511
Table 13.2 Well controls given in terms of bottom-hole pressure
[bar] for a simple 2D reservoir with five initial wells (I1, I2,
P1, P2, P3) and one infill well (I3).

I1 I2 I3 P1 P2 P3 Lorenz

Base case 200 200 — 100 100 100 0.2730

Case 1 200 200 — 100 130 80 0.2234
Case 2 200 200 — 100 130 80 0.1934
Case 3 200 220 140 100 130 80 0.1887

Figure 13.25 A simple 2D reservoir with two injectors and three producers. The left plot shows
porosity and the right plot the corresponding producer partition.

To figure out to what extent this is a good well pattern or not, we start by looking at
how a displacement front would propagate if the present flow field remains constant. For
a displacement front traveling with a unit speed relatively to the Darcy velocity given by
the flux field, the region swept by the front at time t consists of all cell for which τf ≤ t.
Using the interactive GUI shown in Figure 13.20, you can either show swept regions by
specifying threshold values manually, or use the “Play TOF” button to animate how the
displacement front advances through the reservoir. Figure 13.26 shows four snapshots of
such an advancing front. We notice, in particular, how the sealing fault to the northwest
of the reservoir impedes the northbound propagation of the displacement front from I1
and leads to early breakthrough in P2. There is also a relatively large region that remains
unswept after the two displacement fronts have broken through in all three producers.
As a more direct alternative to studying snapshots of an imaginary displacement front,
we can plot the residence time, i.e., the sum of the forward and backward time-of-flight, as
shown in the upper-left plot of Figure 13.27. Here, we use a nonlinear gray-map to more
clearly distinguish high-flow zones (dark gray) from stagnant regions (white) and other

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 512 Flow Diagnostics

Figure 13.26 Evolution of an imaginary displacement front illustrated by thresholding time-of-flight

iniside the sweep regions for the base case.

regions of low flow (light gray). In the figure, we see that wells I1 and P2 are connected
by a high-flow region, which explains the early breakthrough we observed in Figure 13.26.
The existence of high-flow regions can also be seen from the F – diagram and the Lorenz
coefficient of 0.273.
The interactive diagnostic tool has functionality that lets you modify the well controls
and if needed, add new wells or remove existing ones. We now use this functionality to
try to manually improve the volumetric sweep of the reservoir, much in the same way as
we did for a layer of SPE 10 in Section 13.4.2. We start by reducing the high flow rate in
the region influenced by I1 and P2. That is, we increase the pressure in P2 to, say, 130 bar
to decrease the inter-well pressure drop. The resulting setup, referred to as “Case 1,” gives
more equilibrated flow paths, as you can see from the Lorenz coefficient and the upper-right
plot in Figure 13.27.
Case 2 further decreases the pressure in P3 to 80 bar to increase the flow in the I2–
P3 region. As a result of these two adjustments to the well pressures, we have reduced
the stagnant region north of P2 and also diminished the clear flow-divide that extended
from the south of the reservoir to the region between P2 and P3. To also sweep the large
unswept region east of P3, we can use infill drilling to introduce a new well to the southeast
of this region, just north of the sealing fault. Since the new well is quite close to the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.5 Interactive Flow Diagnostics Tools 513

Base case Case 1

Case 2 Case 3
1

0.8

0.6
Base case
0.4 Case 1
Case 2

0.2 Case 3

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 13.27 In Cases 1 and 2, well controls have been manually adjusted from that of the base case
(see Table 13.2) to equilibrate total travel time throughout the reservoir. Case 3 includes infill drilling
of an additional injector. The bottom plot shows the corresponding F – curves.

existing producer, we should assign a relatively low pressure to avoid introducing too high
flow rates.
In Case 3, we let the well operate at 140 bar and at the same time we have increased
the pressure in I2 to 220 bar. Figure 13.28 shows snapshots of the advancing front at the
same instances in time as was used in Figure 13.26 for the base case. Altogether, the well
configuration of Case 3 gives significantly improved areal sweep and reduces the Lorenz
coefficient to 0.189. We can therefore expect this configuration to give better displacement
if the setups were rerun with a multiphase simulator.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 514 Flow Diagnostics

Figure 13.28 Evolution of an imaginary displacement front illustrated by thresholding time-of-flight

iniside the sweep regions for Case 3.

You can find a more detailed description of how you should use the interactive GUI
to perform the experiments just described in the interactiveSimple.m script of the
book module. I encourage you to use the script to familiarize yourself with interactive
flow diagnostics. Are you able to make further improvements?

13.5.2 SAIGUP: Flow Patterns and Volumetric Connections

The previous example was highly simplified and chosen mainly to illustrate the possibilities
that lie in the interactive use of flow diagnostics. In this example, we take a closer look at
the volumetric connection in the SAIGUP example from Section 5.4.4 on page 169. We
start by rerunning the example script to set up the simulation model and compute a flow
field:
saigupWithWells; close all
clearvars -except G rock W state

Henceforth, we only need the geological model, description of the wells, and the reservoir
state. We clear all other variables and close all plots produced by the script. We can pass
the reservoir state to the interactive GUI and use it in pure postprocessing mode:
interactiveDiagnostics(G, rock, W, 'state', state, 'computeFlux', false);

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.5 Interactive Flow Diagnostics Tools 515

Figure 13.29 Horizontal permeability (log10 Kx ) for the SAIGUP model. The left plot shows the full
permeability field, while the right plot only shows the permeability in cells that have a total residence
time less than 100 years. (The reservoir is plotted so that the north–south axis goes from left to right
in the figure.)

In this mode, we are not able to edit any well definitions and recompute fluxes.
In Figure 5.11 on page 173, we saw that, even though injectors and producers are
completed in all 20 grid layers of the model, there is almost no flow in the bottom half
of the reservoir. A more careful inspection shows that there is almost no flow in the upper
layers in most of the reservoir either because the best sand quality is found in the upper-
middle layers of the reservoir. Figure 13.29 confirms this by comparing the permeability in
the full model with the permeability in cells having a residence time less than 100 years.
Because each injector is controlled by a total fluid rate, large fluid volumes are injected
in completions connected to good quality sand, while almost no fluid is injected into zones
with low permeability and porosity. You can see this in Figure 13.30, which shows overall
and cumulative wellallocation factors for four of the injectors. (These plots are conceptually
similar to what you would obtain by running a production logging tool for all these wells.)
Injectors I3 and I4 are completed in the southern part of the reservoir, and here low-
quality sand in the bottom half of the reservoir leads to almost negligible injection rates
in completions 11–20. In a real depletion plan, the injectors would probably not have been
completed in the lower part of the sand column. Injector I5 located to the west in the
reservoir is completed in a column with low permeability in the top four and the bottom
layer, high permeability in Layers 6–9, and intermediate permeability in the remaining
layers. Hence, almost no fluid is injected through the top four completions, which hence
are redundant. Finally, injector I6 is completed in a column with poor sand quality in the
top three layers, high permeability in Layers 4–9, and intermediate permeability in the
remaining layers.
Figure 13.30 also shows the flux allocation for all well pairs in the reservoir. Each curved
line corresponds to a injector–producer connection, the line color signifies the producer,
and the percentage signifies the fraction of the total flux from each injector that goes to
the different producers. We have truncated the connections to only pairs that correspond

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 516 Flow Diagnostics
Allocation by connection Allocation by connection
Well allocation factors Well allocation factors
P6 2 P4 2
4 4

6 6

Connection #
Connection #
8 8
P5 P3
10 10
P2
12 12

14 P3 14

16 16

18 18

P4 20 20
0 500 1000 1500 0 500 1000
Well: I6 Accumulated flux [m3/day] Well: I4 Accumulated flux [m3/day]

Allocation by connection Allocation by connection

Well allocation factors Well allocation factors
2 2
P4
4 4
P1
6 6
Connection #

Connection #
8 8
10 10
12 12
P6 14 14
P2
16 16
P2
18 18
20 20
0 200 400 600 800 0 500 1000
Well: I5 3
Accumulated flux [m /day] Well: I3 Accumulated flux [m3/day]

Figure 13.30 Flux allocation for all well pairs of the SAIGUP model and cumulative well-allocation
factors from bottom to top layer for injectors I3, I4, I5, and I6.

to at least 1% of the flux. This explains why not all fractions sum up to unity. Let us
take injector I4 as an example. Figure 13.31 shows two plots of the influence region for
this injector. From the well allocation plots, we have already seen that I4 is connected
to producers P2 to P4. Almost all the completions of these wells lie inside the region
colored by I4. Producer P1, on the other hand, is only completed in a single cell inside
the colored region, and this cell is in the top layer of the reservoir where the sand quality
is very poor. It is therefore not clear whether P1 is actually connected to I4 or if this weak
connection is a result of inaccuracies in the computation of color concentrations. Since we
only use a first-order discretization, the color concentrations can suffer from a significant
amount of numerical smearing near flow divides, which here is signified by blue-green
colors.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 13.5 Interactive Flow Diagnostics Tools 517

Figure 13.31 Plot of the influence region for well I4.

Computer exercises
13.5.1 Change all the injectors to operate at a fixed bottom-hole pressure of 300 bar. Does
this significantly change the flow pattern in the reservoir? Which configuration do
you think is best?
13.5.2 Use the flow diagnostic tool to determine well completions in the SAIGUP model
that have insignificant flow rate, eliminate these well completions, and rerun the
model. Are there any apparent changes in the flux allocation and volumetric con-
nections?
13.5.3 Consider the model given in makeAnticlineModel.m. Can you use flow diag-
nostics to suggest a better well configuration?
13.5.4 Go back to the SPE 9 model from Section 12.4.4. Would it make sense to use flow
diagnostics to analyze the flow patterns of this model? If so, how would you do it?

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14
Grid Coarsening

Over the last decades, methods for characterizing subsurface rock formations have
improved tremendously. This has, together with a dramatic increase in computational
power, enabled the industry to build increasingly detailed and complex models to account
for heterogeneous structures on different spatial scales. Using gridding techniques similar
to the ones outlined in Chapter 3, today one can easily build complex geological models
consisting of multiple millions of cells to account for most of the features seen in typical
reservoirs. In most cases, geocellular models used for reservoir characterization contain
more geological layers and model fine-scale heterogeneity with higher resolution than
what is used for flow simulations.
Through parallelization and use of massively parallel computers it is possible to sim-
ulate fluid flow on grid models with up to a billion cells [84, 85, 240], but such simula-
tions require expensive infrastructure and a very high power budget and are rarely seen
in practice. Contemporary high-fidelity models seem to be in the range of a few million
cells, whereas the majority of asset models have ten times fewer cells, since engineers
usually want to spend available computational power on more advanced flow physics or
on running a large number of model realizations instead of a few highly resolved ones.
To obtain computationally tractable simulation models, it is therefore common to develop
reduced models through some kind of upscaling (homogenization) procedure that removes
spatial detail from the geological description. Typically, a coarser model is developed by
identifying regions consisting of several cells and then replacing each region by a single,
coarse cell with homogeneous properties that represent the heterogeneity inside the region
in some averaged sense. We discuss such upscaling methods in more detail in Chapter 15.

14.1 Grid Partitions

You can obviously generate coarse grids in the same way as the original by simply specify a
lower spatial resolution. However, this approach has two obvious disadvantages: first of all,
if the original grid has complex geometry (and topology), it is very challenging to preserve
the exact geometry of the fine grid for an arbitrary coarse resolution. Second, except in
simple cases, there is generally not a one-to-one mapping between cells in the fine and

518

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.1 Grid Partitions 519

coarse grids. Herein, we therefore choose a different approach. In MRST, a coarse grid
always refers to a grid that is defined as a partition of another grid, which is referred to as
the fine grid.
Tools for partitioning and coarsening of grids are found in two different modules of
MRST: The coarsegrid module defines a basic grid structure for representing coarse grids
and supplies simple routines for partitioning grids with an underlying Cartesian topology.
The agglom module offers tools for defining flexible coarse grids that adapt to geological
features and flow patterns, e.g., as discussed in [125, 124, 194, 187]. Coarse partitions also
form a basis for contemporary multiscale methods [195] and are used extensively in the
various multiscale modules of MRST (msrsb, msmfem, msfv, etc). In this chapter, we first
discuss functionality for generating and representing coarse grids found in the coarsegrid
module and then briefly outline some of the more advanced functions found in the agglom
module. In an attempt to distinguish fine and coarse grids, we henceforth refer to fine grids
as consisting of cells, whereas coarse grids are said to consist of blocks.
Coarse grids in MRST are represented by a structure that consists entirely of topological
information stored in the same fields as for the general grid structure introduced in Section
3.4. As a naming convention, we use CG to refer to a coarse-grid structure and G to refer to
the usual (fine) grid. A coarse grid is always related to a fine grid in the sense that
6 6 4 8 8

• each cell in the fine grid G belongs to one, and only one,
1 8
block in the coarse grid CG; 6 4
4

• each block in CG consists of a connected subset of cells 2

1 1 1 3

from G; and
3
• CG is defined by a partition vector p defined such that 2 3
1
1

p(i) =  if cell i in G belongs to block  in CG. 5 1 3 7 7

This concept is quite simple, but has proved to be very powerful in defining coarse grids
that can be applied in a large variety of computational algorithms. We will come back to
the details of the grid structure in Section 14.2. First, let us discuss how to define partition
vectors in some detail, as this is more useful from a user perspective than understanding
the details of how the CG structure is implemented.
To demonstrate the simplicity and power of using partition vectors to define coarse
grids, we go through a set of examples. You can find complete codes for all the following
examples in showPartitions.m from the book module.

14.1.1 Uniform Partitions

For all grids having a logically Cartesian topology, i.e., grids that have a valid field
G.cartDims, we can use the function partitionUI to generate a relatively uniform
partition that consists of the tensor product of a load-balanced linear partition in each index
direction. As an example, let us partition a 7 × 7 fine grid into a 2 × 2 coarse grid:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 520 Grid Coarsening

G = cartGrid([7,7]);
p = partitionUI(G, [2,2]);

plotCellData(G, p, 'EdgeColor', 'y');

outlineCoarseGrid(G, p, 'k');
axis tight off,
caxis([.5 max(p)+.5]);
colormap(lines(max(p)));
set(colorbar,'YTick',1:max(p));

The call to partitionUI returns a vector with one element per cell taking one of the
integer values 1,2,3,4 that represent the four blocks. Since seven is not divisible by two,
the coarse blocks do not have the same size but consist of 4×4, 3×3, 4×3, and 3×4 cells. To
better distinguish different blocks in the plot, we have used outlineCoarseGrid(G, p)
to find and plot all faces in G whose neighboring cells have different values of p.
The same procedure can, of course, also be applied to partition any grid in 3D that has a
Cartesian topology. As an example, we consider a simple box geometry:
G = cartGrid([10,10,4]);
p = partitionUI(G, [3,3,2]);

plotCellData(G, p, 'Edgecolor', 'w');

outlineCoarseGrid(G, p, ...
'EdgeColor','k','lineWidth',4);
colormap(colorcube(max(p)))
view(3); axis off

Here, we have used the colorcube colormap, which is particularly useful for visualizing
partition vectors, since it contains as many regularly spaced colors in RGB color space as
possible. The careful reader will also observe that the arguments to outlineCoarseGrid
changes somewhat for 3D grids.

14.1.2 Connected Partitions

All you need to partition a grid is a partition vector. This vector can be given by the user,
read from a file, generated by evaluating a geometric function, or given as the output of
some user-specified algorithm. As a simple example of the latter, let us partition the box
model [−1,1] × [−1,1] into nine different blocks using the polar coordinates of the cell
centroids. The first block is defined as r ≤ 0.3, whereas the remaining eight are defined by
segmenting 4θ/π :

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.1 Grid Partitions 521

G = cartGrid([11, 11],[2,2]); 5
G.nodes.coords = ...
bsxfun(@minus, G.nodes.coords, 1); 4

G = computeGeometry(G);
3
c = G.cells.centroids;
[th,r] = cart2pol(c(:,1),c(:,2)); 2

p = mod(round(th/pi*4)+4,4);
1
p(r<.3) = max(p)+1;

In the second-to-last line, the purpose of the modulus operation is to avoid wrap-around
effects as θ jumps from −π to π.
The human eye should be able to distinguish nine different coarse blocks in the plot
above, but the partition does unfortunately not satisfy all the criteria we prescribed on
page 519. Indeed, as you can see from the colorbar, the partition vector only has five unique
values and thus corresponds to five blocks, according to our definition of p: cell i belongs
to block  if p(i) = . Hence, what the partition describes is one connected block at the
center surrounded by four disconnected blocks. To determine whether a block is connected
or not, we will use some concepts from graph theory. We first form a local undirected
graph (or a network) in which nodes are cells and edges are cell faces connecting two
cells with the same partition value. A block is then said to be disconnected if the graph
has multiple connected components. A connected component is defined as the subgraph in
which any two nodes are connected to each other through a path. In other words, a coarse
block is disconnected if there exists at least one pair of cells that cannot be connected by a
continuous path in the local grid graph. To get a partition satisfiying our requirements, we
must split the four disconnected blocks. This is done by the following call:
q = processPartition(G, p);

which splits disconnected components that have the same p-value into separate blocks and
updates the partition vector accordingly. In the current case, Blocks 1–4 will be split in two,
whereas Block 5 remains unchanged, giving the nine blocks separated by solid lines in the
figure.
Let us quickly outline how this routine works. We start by identifying all cells having
the same partition value; the left plot in Figure 14.1 shows these cells for the first block.
The grid graph is the same that we used to construct the discrete divergence and gradient
operators, with the edge list given in terms of the two columns C1 (f ) and C2 (f ) from
G.faces.neighbors that specify the cells connected by face f . The only exception is that
we now have to exclude any connections between cells having different p-values. From this
list, we can construct a local adjacency matrix. This symmetric matrix is defined so that all
cells corresponding to nonzero entries in a single row (or column) are directly connected.
In the two adjacency matrices shown in Figure 14.1, we have used different color for cells

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 522 Grid Coarsening
0 0
23 24
5 5
17 18 19 20 21 22
10 10
9 10 11 12 13 14 15 16
15 15

3 4 5 6 7 8
20 20

1 2
25 25
0 5 10 15 20 25 0 5 10 15 20 25

Figure 14.1 Partition of a disconnected block. The left plot shows all cells with the same p-value.
The middle plot shows the adjacency matrix based on the original numbering, whereas the right plot
shows the adjacency matrix after a Dulmage–Mendelsohn permutation has separated the connected
components.

belonging to each of the two components for clarity. The middle plot shows that cell 1 is
connected to cell 3; cell 2 is connected to cell 8; cell 3 is connected to cells 1, 4, and 9; and
so on. To find the connected components, we use a Dulmage–Mendelsohn permutation.
Disconnected components will then appear as diagonal blocks in the permuted adjacency
matrix; see the right plot in Figure 14.1.
Graph operations like this can generally be used to adapt the partition to features in
the geological model. As an example, the processing routine can also take an additional
parameter facelist that specifies a set of faces across which the connections will be
removed before processing:
q = processPartition(G, p, facelist)

Using this functionality one can, for instance, prevent coarse blocks from crossing faults
inside the model.

14.1.3 Composite Partitions

In many cases, it may be advantageous to create partition vectors by combining more than
one partition principle. As an example, consider a heterogeneous medium consisting of two
different facies (rock types), one with high permeability and one with low, that each form
large contiguous regions. If we now let the coarse blocks respect the facies boundaries, we
can assign each block a homogeneous property and avoid upscaling. Within each facies,
we can further use a rectangular partition generated by partitionCartGrid, which is
simpler and less computationally expensive than partitionUI but only works correctly
for a grid having a fully intact logically Cartesian topology with no inactive cells, no cells
that have been removed by removeGrid, and so on. The following code illustrates the
principle:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.1 Grid Partitions 523

G = cartGrid([20, 20], [1 1]);

G = computeGeometry(G);

% Facies partition
f = @(c) sin(4*pi*(c(:,1)-c(:,2)));
pf = 1 + (f(G.cells.centroids) > 0);

% Cartesian partition +
pc = partitionCartGrid(G.cartDims, [4 4]);

% Alternative 1:
[b,i,p] = unique([pf, pc], 'rows' );
% Alternative 2: =
q = compressPartition(pf + max(pf)*pc);

The example also shows two alternative techniques for combining different partitions. The
first alternative collects the partitions as columns in an array A. The call
[b,i,p]=unique(A,'rows') will return b as the unique rows of A so that b=A(i) and
A=b(p). Hence, p will be a partition vector that represents the unique intersection of all the
partitions collected in A. In the second method, we treat the partition vectors as multiple
subscripts, from which we compute a linear index. This index is not necessarily contiguous.
Most routines in MRST require that partition vectors are contiguous to avoid having to
treat special cases arising from noncontiguous partitions. To fulfill this requirement, we
use the function compressPartition that renumbers a partition vector to remove any
indices corresponding to empty grid blocks. The two alternatives have more or less the
same computational complexity, and which alternative you choose in your implementation
is largely a matter of what you think will be easiest to understand for others.
Altogether, the examples presented so far in this chapter explain the basic concepts of
how you can create partition vectors. Before we go on to explain details of the coarse-grid
structure and how to generate this structure from a given partition vector, we show one last
and a bit more fancy example. To this end, we create a cup-formed grid, partition it, and
then visualize the partition using a nice technique. To generate the cup-shaped grid, we
use the fictitious-domain technique we previously used to generate the ellipsoidal grid in
Figure 3.4 on page 61.
x = linspace(-2,2,41);
G = tensorGrid(x,x,x);
G = computeGeometry(G);
c = G.cells.centroids;
r = c(:,1).^2 + c(:,2).^2+c(:,3).^2;
G = removeCells(G, (r>1) | (r<0.25) | (c(:,3)<0));

Assume that we wish to partition this cup model into 100 coarse blocks. We could, for
instance, try to use partitionUI to impose a regular 5 × 5 × 4 partition. Because of the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 524 Grid Coarsening

Figure 14.2 The partition of the cup-formed grid visualized with explosionView.

fictitious method, a large number of the ij k indices from the underlying Cartesian topology
will correspond to cells that are not present in the actual grid. Imposing a regular Cartesian
partition on such a grid typically gives block indices in the range [1, max(p)] that do not
correspond to any cells in the underlying fine grid. In this particular case, only 79 out of the
desired 100 blocks correspond to a volume within the grid model. To see this, we use the
function accumarray to count the number of cells for each block index and plot the result
as a bar chart:
60
subplot(2,1,1);
40
p = partitionUI(G,[5 5 4]);
20
bar(accumarray(p,1)); shading flat
0
0 20 40 60 80 100

q = compressPartition(p); 60

subplot(2,1,2); 40

bar(accumarray(q,1)); shading flat 20

set(gca,'XLim',[0 100]); 0
0 20 40 60 80 100

Figure 14.2 shows the partition obtained after we have compressed the partition vector.
To clearly distinguish the different blocks, we have useed explosionView to create an
explosion view, which is a useful technique for visualizing coarse partitions.

14.2 Coarse Grid Representation in MRST

When working with coarse grids, it is not always sufficient to only know the partition,
i.e., which cells belong to which blocks. It can often be more advantageous to treat the
coarse partition as a general polyhedral grid and have access to the inherent topology and
possibly also the geometry. In MRST, we have chosen a compromise and use a coarse-
grid structure that represents topology explicitly and geometry implicitly. This choice is
motivated by flow solvers, which we previously have seen can be posed on an arbitrary grid
graph provided each node has associated pore volume and depth value and each connection
has an associated transmissibility.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.2 Coarse Grid Representation in MRST 525

Given a grid structure G and a partition vector p, we generate a structure CG representing

the coarse grid by the following call:
CG = generateCoarseGrid(G, p)

The coarse-grid structure consists entirely of topological information stored in the same
way as described in Section 3.4 for G: The fields cells and faces represent the coarse
blocks and their connections. As a result, we can use CG seamlessly with many of the stan-
dard solvers in MRST. Unlike the original grid structure, however, CG does not represent
the geometry of the coarse blocks and faces explicitly and does therefore not have a nodes
field. The geometry information is instead obtained from the parent grid G and the partition
vector p, copies of which are stored in the fields parent and partition, respectively.
The structure, CG.cells, that represents the coarse blocks consists of the following
mandatory fields:

– num: the number Nb of blocks in the coarse grid.

– facePos: an indirection map of size [num+1,1] into the faces array, which is defined
completely analogously as for the fine grid. Specifically, the connectivity infor-
mation of block i is found in the submatrix
faces(facePos(i): facePos(i+1)-1, :)

You can now compute the number of connections of each block using the state-
ment diff(facePos).
– faces: an Nc × 2 array of connections associated with a given block. Specifically, if
faces(i,1)==j, then connection faces(i,2) is associated with block number
j . To conserve memory, only the second column is actually stored in the grid
structure. The first column can be reconstructed by a call to rldecode. Optionally,
one may append a third column that contains a tag inherited from the parent grid.

In addition, the cell structure can contain the following optional fields, which typically
are added by a call to coarsenGeometry, assuming that the corresponding information is
available in the parent grid:

– volumes: an Nb × 1 array of block volumes

– centroids: an Nb × d array of block centroids in IRd

The face structure, CG.faces, consists of the following mandatory fields:

– num: the number Nc of global connections in the grid.

– neighbors: an Nc ×2 array of neighboring information. Connection i is between blocks
neighbors(i,1) and neighbors(i,2). One of the entries in
neighbors(i,:), but not both, can be zero, to indicate that connection i is
between a single block (the nonzero entry) and the exterior of the grid.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 526 Grid Coarsening

– connPos, fconn: packed data-array representation of the coarse → fine mapping.

Specifically, the elements fconn(connPos(i):connPos(i+1)-1) are the con-
nections in the parent grid (i.e., rows in G.faces.neighbors) that constitute
coarse-grid connection i.
In addition to the mandatory fields, CG.faces has optional fields that contain geometry
information and typically are added by a call to coarsenGeometry:
– areas: an Nc × 1 array of face areas.
– normals: an Nc × d array of accumulated area-weighted, directed face normals in IRd .
– centroids: an Nc × d array of face centroids in IRd .
Like in G, the coarse grid structure also contains a field CG.griddim that is used to distin-
guish volumetric and surface grids, as well as a cell array CG.type of strings describing
the history of grid constructor and modifier functions used to define the coarse grid.
As an illustrative example, let us partition a 4 × 4 Cartesian grid into a 2 × 2 coarse grid.
This gives the following structure:
CG =
cells: [1x1 struct]
faces: [1x1 struct]
partition: [16x1 double]
parent: [1x1 struct]
griddim: 2
type: {’generateCoarseGrid’}

with the cells and faces fields given as

CG.cells = CG.faces =
num: 4 num: 12
facePos: [5x1 double] neighbors: [12x2 double]
faces: [16x2 double] connPos: [13x1 double]
fconn: [24x1 double]
Figure 14.3 shows relations between entities in the coarse grid and its parent grid. For
instance, we see that block number one consists of cells one, two, five, and six because
these are the rows in CG.partition that have value equal one. Likewise, we see that
because CG.faces.connPos(1:2)=[1 3], coarse connection number one is made up of
two cell faces that correspond to faces number one and six in the parent grid because
CG.faces.fconn(1:2)=[1 6], and so on.

14.2.1 Subdivision of Coarse Faces

In the discussion so far, we have always assumed that there is only a single connection
between two neighboring coarse blocks and that this connection is built up of a set of
cell faces corresponding to all faces between pairs of cells in the fine grid that belong
to the two different blocks. While this definition is useful for many workflows like in

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.2 Coarse Grid Representation in MRST 527

13 14 15 16 CG.partition =
3 4 ( 1) 1 ( 9) 3
( 2) 1 (10) 3
9 10 11 12 ( 3) 2 (11) 4
( 4) 2 (12) 4
( 5) 1 (13) 3
5 6 7 8 ( 6) 1 (14) 3
1 2 ( 7) 2 (15) 4
( 8) 2 (16) 4
1 2 3 4

37 6 38 39 8 40

16 17 18 19 20 CG.faces.connPos = CG.faces.fconn=
1 1
5 33 34 12 35 36 7 3 6
11 12 13 14 15 5 21
7 22
29 10 30 31 11 32 9 5
6 7 8 9 10 11 10
13 23
1 25 26 9 27 28 3 15 24
1 2 3 4 5 17 11
19 16
21 2 22 23 4 24
: :

Figure 14.3 The relation between blocks in the coarse grid and cells in the parent grid (top) and
between connections in the coarse grid and faces from the parent grid (bottom).

standard upscaling methods, there are also problems for which one may want to introduce
more than one connection between neighboring blocks. To define a subdivision of coarse
faces, we once again use a partition vector with one scalar value per face in the fine
grid, defined completely analogous to vectors used for the volumetric partition. Assum-
ing that we have two such partition vectors, pv describing the volumetric partition and
pf describing the partition of cell faces, the corresponding coarse grid is built through
the call:
CG = generateCoarseGrid(G, pv, pf);

In my experience, the simplest way to build a face partition is to compute it from an

ancillary volumetric partition using the routine:
pf = cellPartitionToFacePartition(G, pv)

which assigns a unique, nonnegative integer for each pair of cell values occurring in the
volumetric partition vector pv, and hence constructs a partitioning of all faces in the grid.
Fine-scale faces that are not on the interface between coarse blocks are assigned zero value.
As an illustration, we continue the example from page 523. We first partition the 8 × 8
fine grid into a 2 × 2 coarse grid and then use facies information to subdivide faces of the
coarse grid so that each coarse connection has a a given combination of facies values on
opposite sides of the interface:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 528 Grid Coarsening

G = computeGeometry(cartGrid([8, 8], [1 1]));

f = @(c) sin(3*pi*(c(:,1)-c(:,2))); 12 13 18 17
pf = 1 + (f(G.cells.centroids) > 0); 32 16
11
plotCellData(G, pf,'EdgeColor','none'); 31
15

10 30
pv = partitionCartGrid(G.cartDims, [2 2]); 14
23 24 25 26 27 28 29
pf = cellPartitionToFacePartition(G,pf); 2 22
7
pf = processFacePartition(G, pv, pf); 21

CG = generateCoarseGrid(G, pv, pf); 1 20

6
CG = coarsenGeometry(CG); 19
3 4 5 9 8
cmap = lines(CG.faces.num);
for i=1:CG.faces.num
plotFaces(CG,i,'LineWidth',6,'EdgeColor', cmap(i,:));
end
text(CG.faces.centroids(:,1), CG.faces.centroids(:,2), ...
num2str((1:CG.faces.num)'),'FontSize',20,'HorizontalAlignment', 'center');

As for the volumetric partition, we require that each interface that defines a connection
in the face partition consists of a connected set of cell faces. That is, it must be possible
to connect any two cell faces belonging to given interface by a path that only crosses
edges between cell faces that are part of the interface. To ensure that all coarse interfaces
are connected collections of fine faces, we have used the routine processFacePartition,
which splits disconnected interfaces into one or more connected interfaces.
The same principles apply also in 3D, here illustrated for a rectangular block with a
rectangular cut-out:

G = computeGeometry(cartGrid([20 20 6]));
c = G.cells.centroids;
G = removeCells(G, ...
(c(:,1)<10) & (c(:,2)<10) & (c(:,3)<3));
plotGrid(G); view(3); axis off

We introduce a volumetric partition and a face partition:

p = partitionUI(G,[2, 2, 2]);
q = partitionUI(G,[4, 4, 2]);
CG = generateCoarseGrid(G, p, ...
cellPartitionToFacePartition(G,q));

plotCellData(CG,(1:max(p))','EdgeColor','none');
plotFaces(CG,1:CG.faces.num,...
'FaceColor' , 'none' , 'LineWidth' ,2);
view(3); axis off

The structure CG also contains lookup tables for mapping blocks and interfaces in the
coarse grid to cells and faces in the fine grid. To illustrate, we visualize one connection of

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.3 Partitioning Stratigraphic Grids 529

a subdivided coarse face that consists of several fine faces, along with the fine cells that
belong to the neighboring blocks:
face = 66;
sub = CG.faces.connPos(face):CG.faces.connPos(face+1)-1;
ff = CG.faces.fconn(sub);
neigh = CG.faces.neighbors(face,:);

show = false(1,CG.faces.num);
show(boundaryFaces(CG)) = true;
show(boundaryFaces(CG,neigh)) = false;
plotFaces(CG, show,'FaceColor',[1 1 .7]);
plotFaces(G, ff, 'FaceColor', 'g')
plotFaces(CG,boundaryFaces(CG,neigh), ...
'FaceColor','none','LineWidth', 2);
plotGrid(G, p == neigh(1), 'FaceColor', 'none', 'EdgeColor', 'r')
plotGrid(G, p == neigh(2), 'FaceColor', 'none', 'EdgeColor', 'b')

14.3 Partitioning Stratigraphic Grids

You can also apply the principles outlined in the previous section to stratigraphic grids.
To demonstrate this, we coarsen two corner-point models of industry-standard complexity:
the sector model of the Johansen aquifer introduced in Section 2.5.4 and the SAIGUP
model from Section 2.5.5. We also apply a few more advanced partition methods to the
sector model with intersecting faults from Section 3.3.1. Full details are given in the scripts
coarsenJohansen, coarsenSAIGUP, and coarsenCaseB4.

14.3.1 The Johansen Aquifer

The Johansen models were originally developed to study a potential site for geological stor-
age of CO2 injected as a supercritical fluid deep in the formation. In Section 2.5.4, we saw
that the heterogeneous NPD5 sector model contains three formations: the Johansen sand-
stone delta bounded above by the Dunlin shale and below by the Amundsen shale. These
three formations have distinctively different permeabilities (see Figure 2.15 on page 46)
and play a very different roles in the sequestration process. The Johansen sandstone has
relatively high porosity (and permeability) and is the container in which the CO2 is to
be kept. The low-permeability Dunlin shale acts as a seal that prevents the CO2 from
escaping back to the sea bottom, and we generally expect that the buoyant CO2 phase will
accumulate as a thin plume that migrates upward under the caprock in the up-dip direction.
To accurately simulate the up-dip migration under the top seal, it is highly important to
preserve the correct interface between the Johansen sandstone and the Dunlin shale in the
coarse model. (In general, it is a good advice to avoid creating coarse blocks containing
large media contrasts, which would otherwise adversely affect upscaling accuracy.) We

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 530 Grid Coarsening
Table 14.1 Permeability values used to distinguish the
different formations in the NPD5 sector model of the
Johansen formation.

Dunlin Johansen Amundsen

K ≤ 0.01mD 0.1 mD < K 0.01 mD < K ≤ 0.1mD

Figure 14.4 A 4 × 4 × 11 coarsening of the NPD5 model of the Johansen aquifer that preserves the
Amundsen, Dunlin, and Johansen formations.

therefore coarsen the three formations separately, using permeability values K as indicator
as shown in Table 14.1. Likewise, to correctly resolve the formation and migration of the
thin plume, it is essential that the grid has as high vertical resolution as possible. Unless
we use a vertically integrated model (as in MRST co2lab [19]), we would normally
only reduce the lateral resolution, say by a factor of four in each lateral direction. Here,
however, we first use only a single block in the vertical direction inside each formation to
more clearly demonstrate how the coarsening can adapt to the individual formations.
Assuming that the grid G and the permeability field K have been initialized properly as
described in Section 2.5.4, the coarsening procedure reads
pK = 2*ones(size(K)); pK(K<=0.1) = 3; pK(K<=0.01)= 1;
pC = partitionUI(G, [G.cartDims(1:2)/4 1]);
[b,i,p] = unique([pK, pC], 'rows');
p = processPartition(G,p);
CG = generateCoarseGrid(G, p);
plotCellData(G,log10(K),'EdgeColor','k','EdgeAlpha',.4); view(3)
outlineCoarseGrid(G,p,'FaceColor','none','EdgeColor','k','LineWidth',1.5);

Figure 14.4 shows the coarse grid obtained by intersecting the partition vector pC, which
has only one block in the vertical direction, with the partition vector pK that represents the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.3 Partitioning Stratigraphic Grids 531

different formations. In regions where all formations are present, we get three blocks in the
vertical direction. In other regions, only the Dunlin and Amundsen shales are present and
we hence have two blocks in the vertical direction.
The aquifer model contains one major and several minor faults. As a result, 1.35% of the
cells in the original grid have more than six neighbors. Coarsening a model with irregular
geometry (irregular perimeter, faults, degenerate cells, etc.) uniformly in index space will in
most cases give many blocks with geometry that deviates quite a lot from being rectangular.
The resulting coarse grid thus contains a larger percentage of unstructured connections than
the original fine model. For this particular model, 20.3% of the blocks have more then six
coarse faces. If we look at a more realistic coarsening that retains the vertical resolution of
the original model, 16.5% of the blocks have more than six neighboring connections. This
model is obtained if we repeat the earlier construction using
pC = partitionUI(G, G.cartDims./[4 4 1]);

Figure 14.5 shows six different coarse blocks sampled from the top grid layer of the
Dunlin shale. Block number one is sampled from a part of the aquifer perimeter that does
not follow the primary grid directions and thus has irregular geometry. The other five blocks
contain (parts of) a fault and will therefore potentially have extra connections to blocks in
grid layers below. Despite the irregular geometry of the blocks, the coarse grid can be
used directly with most of the solvers discussed earlier in the book. In our experience, the
quality of the coarse solution is generally more affected by the quality of the upscaling of
the petrophysical parameters (see Chapter 15) than by the irregular block geometry. In fact,
irregular blocks that preserve the geometry of the fine-scale model respect the layering and
connections in the fine-scale geology and therefore often give more accurate results than a
coarse model with more regular blocks if upscaled correctly.

4
5 3
2

Figure 14.5 Six coarse blocks sampled from the top grid layer of the Dunlin formation in a 4 × 4 × 1
coarsening of the NPD5 sector model of the Johansen formation.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 532 Grid Coarsening

14.3.2 The SAIGUP Model

The SAIGUP model, introduced in Section on page 97, has six user-defined rock types
(also known as saturation regions; see Figure 2.20) that are used to specify different rock-
fluid behavior (relative permeability and/or capillary pressure functions). Depending upon
the purpose of the reduced model, we may want to preserve these rock types using the
same type of technique as described in the previous example. This has the advantage that
if each coarse block is made up of one rock type only, we would not have to upscale
the rock-fluid properties. On the other hand, this typically leads to coarse grids with
(highly) irregular block geometries and large variations in block volumes. To illustrate this
point, we start by partitioning the grid uniformly into 6 × 12 × 3 coarse blocks in index
space:
p = partitionUI(G,[6 12 3]);

This introduces a partition of all cells in the logical 40 × 120 × 20 grid, including
cells that are inactive. To get a contiguous partition vector, we remove blocks that
contain no active cells, and then renumber the vector. This reduces the total num-
ber of blocks from 216 to 201. Some of the blocks may contain disconnected cells
because of faults and other nonconformities, and we must therefore postprocess the
grid in physical space and split each disconnected block into a new set of connected
sub-blocks:
p = compressPartition(p);
p = processPartition(G,p);

The result is a partition with 243 blocks that each consists of a set of connected cells in the
fine grid. Figure 14.6 shows an explosion view of the individual coarse blocks. Whereas
all cells in the original model are almost exactly the same size, the volumes of the coarse
blocks span almost two orders of magnitude. In particular, the irregular boundary near the
crest of the model introduces small blocks consisting of only a single fine cell in the lateral
direction. Large variations in block volumes will adversely affect any flow solver if we later
run a flow simulation on the coarsened model. To get coarse blocks with a more even size
distribution, we therefore pick the smallest blocks and merge them with the neighbor that
has the largest block volume. We repeat this process until the volumes of all blocks are
above a prescribed lower threshold.
The merging algorithm is quite simple: we compute block volumes, select the block
with the smallest volume and merge this block with one of its neighbors. Next, we
update the partition vector by relabeling all cells in the block with the new block number
and compress the partition vector to get rid of empty entries. Finally, we regenerate a
coarse grid, recompute block volumes, pick the block with the smallest volume in the
new grid, and repeat the process. In each iteration, we plot the selected block and its
neighbors:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.3 Partitioning Stratigraphic Grids 533

Figure 14.6 Logically Cartesian partition of the SAIGUP model. The plot to the left shows an
explosion view of the individual blocks colored with colorcube. The bar graph to the right shows
the volumes in units [m3 ] for each of the blocks in the partition.

blockVols = CG.cells.volumes;
meanVol = mean(blockVols);
[minVol, block] = min(blockVols);
while minVol<.1*meanVol

% Find all neighbors of the block

clist = any(CG.faces.neighbors==block,2);
nlist = reshape(CG.faces.neighbors(clist,:),[],1);
nlist = unique(nlist(nlist>0 & nlist=block));
plotBlockAndNeighbors(CG, block, ...
'PlotFaults', [false, true], 'Alpha', [1 .8 .8 .8]);

% Merge with neighbor having largest volume

[,merge] = max(blockVols(nlist));

% Update partition vector

p(p==block) = nlist(merge);
p = compressPartition(p);

% Regenerate coarse grid and pick the block with the smallest volume
CG = generateCoarseGrid(G, p);
CG = coarsenGeometry(CG);
blockVols = CG.cells.volumes;
[minVol, block] = min(blockVols);
end

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 534 Grid Coarsening

To find the neighbors of a given block, we first select all connections that involve block
number block, which we store in the logical mask clist. We then extract the indices
of the blocks involved in these connections by using clist to index the connection list
CG.faces.neighbors. The block cannot be merged with the exterior or itself, so values 0
and block are filtered out. A pair of blocks may generally share more than one connection,
so unique is used to remove multiple occurrences of the same block number.
When selecting which neighbor a block should be merged with, there are several points
to consider from a numerical point of view. We typically want to keep blocks as regular and
uniformly sized as possible, make sure that the cells inside each new block are well con-
nected, and limit the number of new connections we introduce between blocks. Figure 14.7
shows several of the small blocks and their neighbors. In the algorithm shown in the most
recent code box, we have chosen a simple merging criterion: each block is merged with the
neighbor having the largest volume. In iterations one and three, the blue blocks are merged
with the cyan blocks, which is probably fine in both cases. However, if the algorithm later
wants to merge the yellow blocks, using the same approach may not give good results as
shown in iteration ten. Here, it is natural to merge the blue block with the cyan block that
lies in the same geological layer (same K index) rather than merging it with the magenta
block that has the largest volume. The same problem is seen in iteration number four, where
the blue block is merged with the yellow block, which is in another geological layer and
only weakly connected to the blue block. As an alternative, we could choose to merge with
the neighbor having the smallest volume, in which case the blue block would be merged
with the yellow block in iterations one and three, with the magenta block in iteration four,
and with the cyan block in iteration six. This would tend to create blocks consisting of cells
from a single column in the fine grid, which may not be what we want. Another solution
would be to merge across the coarse face having the largest transmissibility (sum of fine-
scale transmissibilities).
Altogether, we see that there are several issues we need to take into consideration
and these could potentially lead to mutually conflicting criteria. This obviously makes it
difficult to devise a good algorithm that merges blocks in the grid in a simple and robust
manner. A more advanced strategy could, for instance, include additional constraints that
penalize merging blocks belonging to different layers in the coarse grid. Likewise, you
may want to avoid creating connections that only involve small surface areas. However,
such connections may already be present in the coarsening we start from, as illustrated by
the yellow and cyan blocks in the lower-right plot in Figure 14.7, which were created when
partitioning the grid uniformly in index space.
The basic processPartition routine only checks that there is a connection between
all cells inside a block. A more sophisticated routine could obviously also check the face
areas associated with each connection, or the magnitude of the associated transmissibility,
and consider different parts of the block to be disconnected if the area or transmissibil-
ity that connects them is too small. As a step in this direction, we could consider the
face area when we postprocess the first uniform partition, e.g., do something like the
following,

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.3 Partitioning Stratigraphic Grids 535

Iteration 4: blue → yellow

Iteration 1:
blue → cyan

Iteration 3: blue → cyan

Iteration 6:
blue → yellow

Iteration 10:
blue → magenta

Figure 14.7 Examples of small blocks (in blue) being merged with one of their neighbors shown in
semitransparent color with fault faces in gray.

p = partitionUI(G,[6 12 3]);
p = compressPartition(p);
p = processPartition(G, p, G.faces.areas<250);

This increases the number of blocks in the final grid obtained after merging small blocks
from 220 to 273, but avoids constructing blocks looking like the yellow and cyan block in
the lower-right plot in Figure 14.7. On the other hand, this approach obviously involves a
threshold parameter that will vary from case to case and has to be set by an expert. The

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 536 Grid Coarsening

result may also be very sensitive to the choice of this parameter. To see this, you can try to
redo the initial partition with threshold values 300 and 301 for the face areas.

14.3.3 Near Well Refinement for CaseB4

In this example1 , we consider the smallest 36×48×12 pillar grid from CaseB4 (see Section
3.3.1) describing a reservoir section with two intersecting dip-slip faults. The reservoir is
known to have four distinct geological layers with k indices 1, 2–7, 8–11, and 12. We start
by making a standard load-balanced partition in index space:
p0 = partitionUI(G, [7 9 1]);
pf = processPartition(G, p0, find(G.faces.tag==1));

In the second line, we identify all fault faces and ask the processing routine to remove any
connections across these faces when examining whether any of the 63 original blocks are
disconnected or not. Blocks that are fully penetrated by a fault will contain at least two
connected components and will hence be split. After splitting, the partition has 79 blocks.
Although not needed for this model, it is quite simple to also split blocks that are only
partially penetrated by faults; see [187] for more details. The disadvantage of imposing
hard constraints like fault faces on the partitioning process is that it tends to increase the
variation in block sizes, as shown in Figure 14.8.
Pressure gradients and flow rates are usually much larger near wells than inside the
reservoir. How accurate we are able to capture the flow solution in the near-well region
strongly influences the accuracy of the overall simulation. It may therefore be desirable to
have higher grid resolution in the near-well zone than inside the reservoir. To refine blocks
near the two wells, we use refineNearWell from the coarsegrid module, which takes
a set of points and partitions these into cylindrical sectors around a single point in the xy-
plane. The first well is placed in the corner of the reservoir section and we partition the
perforated well block into five radial sections. The width of the radial sections is set to
decay as the logarithm of the radial distance:
[wc,p] =
deal( W(i).cells(1), pf); % Pick well cell in top layer
wpt =
G.cells.centroids(wc,:); % Center point of refinement
cells =
(p == p(wc)); % All cells in block
pts =
G.cells.centroids(cells,:); % Points to be repartitioned
out =
refineNearWell(G.cells.centroids(cells,:), wpt, ...
'angleBins', 1, 'radiusBins', 4, ...
'logbins', true, 'maxRadius', inf);
p(cells) = max(p) + out; % Insert new partition in block
p = compressPartition(p); % and compress vector

The second well is perforated in the middle of a coarse block. It is therefore natural to
refine both in the radial and the angular directions. The number of blocks in the angular

1 Complete code: coarsenCaseB4.m in the book module.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.3 Partitioning Stratigraphic Grids 537

Rectangular

7 × 9 × 1 partition + split across faults

+ near-well refinement + adapt to layers

METIS

METIS: 63 blocks + near-well refinement

Figure 14.8 Partitioning of the CaseB4 pillar grid. The rectangular partition starts by a call
to partitionUI, followed by the splitting of blocks across faults, near-well refinement by
refineNearWell, and finally, vertical partition following geological layers. The METIS partition
starts by a graph partitioning with transmissibility as edge weights followed by a near-well
refinement. The large plots show how the partition is gradually refined by including new partition
principles. The small plots show the variation in block volumes relative to the volume of the smallest
block.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 538 Grid Coarsening

direction can be set for each radial section. Here, we therefore let the number of blocks
increase outward, starting with a single block in the innermost layer, and then four, six, and
nine blocks in the next three layers. To be able to place such a refinement, we repartition
not only the well block but also the eight surrounding neighbors. To find these blocks, we
construct a temporary coarse grid structure and use the topology from CG to construct an
adjacency matrix similar to the one we discussed on page 522.
% Adjacency matrix
CG = generateCoarseGrid(G, pw); r= 1

N = getNeighbourship(CG);
A = getConnectivityMatrix(N,true,CG.cells.num);
1
Ar = 1 1 1
% Find all axial/diagonal neighbors 1

rblk = zeros(CG.cells.num,1);
rblk(pwv)=1; 1
2 2 2
rblk((A*A*rblk)>1) = 1; A2 r = 1 2 5 2 1
2 2 2
cells = rblk(pw)>0; 1

We start by defining an indicator vector rblk equal one in the well block and zero in all
other blocks. Multiplying by the adjacency matrix A will set the value in each block equal
the sum over the block and its face-neighbors. After one multiplication, rblk therefore
has ones in all the face-neighbors, and after two multiplications all blocks surrounding the
initial block should have value larger than one. Next, we reset rblk to one in all blocks
with A*A*rblk > 1 and index the resulting vector by the partition vector pw to extract all
cells inside these blocks. These cells can be repartitioned as for the first well:
out = refineNearWell(pts, wpt, 'angleBins', [1,4,6,9],...
'radiusBins', 4, 'logbins', true, 'maxRadius', inf);

Altogether, this increases the number of coarse blocks to 93. Finally, we can impose vertical
refinement that follows the four geological layers. Assuming the indices of these layers are
given as an indirection map, this partition is constructed as follows:
% Here: layers = [1 2 8 12 13]
pK = rldecode((1:numel(layers)-1)',diff(layers));
[,,k]=gridLogicalIndices(G);
pk = compressPartition((pK(k)-1)*max(pw)+pw);
pk = processPartition(G, pk);

The resulting partition shown in Figure 14.8 has 372 blocks and significant variation in
block sizes. This does not necessarily make the coarse grid ill-suited for flow simulations,
but in a more careful implementation it would have been natural to postprocess the grid
to get rid of small blocks away from the near-well regions that can be merged with their
neighbors.
As an alternative to the regular partition in index space, we can use a standard soft-
ware like METIS [154], which offers state-of-the-art partitioning of graphs and meshes
based on multilevel recursive-bisection, multilevel k-way, and multi-constraint partitioning

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.3 Partitioning Stratigraphic Grids 539

algorithms. METIS generally tries to make blocks that are as uniform as possible and
at the same time have as few connections as possible. If we let fine-scale transmissi-
bilities measure connection strengths between cells for the edge-cut minimization algo-
rithm, the software tries to construct blocks without crossing large permeability contrasts.
The coarsegrid module implements a wrapper that greatly simplifies the task of calling
METIS. The following call gives a partition with approximately the same number of blocks
as in the structural partition with four vertical layers, just discussed:
mrstModule add incomp
hT = computeTrans(G, rock);
p0 = partitionMETIS(G, hT, 7*9*4, 'useLog', true);

The partitioning function uses the standard incompTPFA solver from the incomp module
to set up a fine-scale discretization matrix that describes how cells are connected. Here, we
take the logarithm of the transmissibilities before the connection strength is constructed.
Experience shows that this gives better partitions when transmissibilities vary several orders
of magnitude locally. The resulting partition is shown in the lower-left plot in Figure
14.8. Compared with the rectangular partition, the grid blocks are more irregular and
have more connections and larger variation in block sizes. Also for this partition, we can
introduce near-well refinement exactly as described in this subsection. The first well falls
within a smaller grid block and hence gives slightly different local blocks. For the second
well, our detection of neighboring blocks selects a larger region to repartition and hence
refineNearWell is able to generate a refinement that looks more circular compared our
first partition method.
A technical note: To inform the system where to find METIS, we use a global variable
METISPATH that can be set in your startup_user function. If you, like me, use Linux
and have installed METIS in, e.g., /usr/local/bin, you add the following line to your
startup_user.m file:
global METISPATH; METISPATH = fullfile('/usr','local','bin');

Computer exercises

14.3.1 Rerun CaseB4 with the stair-stepped model instead. Why does the code not work?
(Hint: check the number of active cells.) Can you make a partition that splits blocks
properly on opposite sides of faults?
14.3.2 The code example also contains a set up that tries to force METIS to split blocks
across faults, which does not work well. Can you develop an alternative strategy
to ensure that fault faces are also preserved in the METIS partition? (Hint: try to
partition each fault block separately.)
14.3.3 Try to implement the merging technique from the SAIGUP example to reduce the
large difference in cell volumes. (Hint: make sure that you do not merge any blocks
in the near-well regions.)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 540 Grid Coarsening

14.4 More Advanced Coarsening Methods

The previous section outlined two main types of partitioning methods. Systematic graph
partitioning algorithms like those implemented in METIS are widely used and very robust,
but do not necessarily give you the desired control over the coarsening process. In par-
ticular, it can be quite difficult to rigorously formulate the desired coarsening as a set of
cost functionals and constraints that are well posed and computationally tractable. More ad
hoc and straightforward approaches that explicitly impose a structured partition in index
space or use features from the model as partition vectors (Section 14.1.3) leave you in full
control, but can easily lead to undesired artifacts like large variations in block volumes as
we saw for SAIGUP and CaseB4. For large and complex models one cannot rely on visual
quality control and manual repair of individual blocks. A significant body of research has
therefore been devoted to develop automated or semiautomated coarsening procedures that
incorporate certain features of the geology or flow physics.
Most methods reported in the literature try to generate a coarser and more optimal
grid consisting of standard grid blocks with spatially varying resolution. The size of each
grid block is usually determined by equilibrating a density measure or by introducing a
background indicator whose variation is minimized within blocks and maximized between
blocks. In particular, it is common to impose constraints on the geometrical shape of the
new grid blocks (Delaunay/Voronoi) and the degree to which they align with the cells in
the original grid [257, 47, 100]. Density measures and indicators can be defined using
geological quantities like permeability [116, 160]; flow-based quantities like fluid velocity
[101], vorticity [201], or streamlines [301, 72, 59, 310, 126]; local a priori error measures
[162]; or statistical or a posteriori goal-oriented error indicators that measure how a par-
ticular point influences the error in production responses or other predefined quantities.
Flow-based coarsening has been shown to be a powerful approach in combination with
upscaling, and has been developed both for structured and unstructured grids. The basic
goal of flow-based griding [93] is to introduce higher resolution in regions of high flow and
coarser resolution in regions of lower flow.
An alternative approach is to generate coarse blocks by agglomerating cells that are
similar in a sense prescribed by the user. This gives coarse blocks with complex polyhedral
forms that follow the geometries of the original fine grid. An early approach in this direction
[3] was aimed at coarse-scale discretization of transport equations and suggested to group
cells according to the magnitude of the velocity (or flux) field so that each coarse block con-
sists of a collection of cells through which the flow has approximately the same magnitude.
Later research has shown that the original nonuniform coarsening method is just a special
case of a much wider set of heuristic methods for producing coarse partitions that adapt
to various types of geological or flow constraints. These methods have been developed
for three different purposes: (i) to increase the accuracy of multiscale pressure solvers by
adapting to fine-scale geological features [5, 6, 222, 194, 17, 191]; (ii) to define partitions
that preserve rock types or relperm regions and thereby reduces the need for upscaling; and
(iii) to find flow-adapted partitions suitable for coarse-scale transport solvers [3, 125, 124].

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.5 A General Framework for Agglomerating Cells 541

These methods can, in turn, be incorporated into a more general framework that tries to
impose various types of geological/flow constraints in a hierarchical manner [187]. Similar
aggregation-based methods have been applied to fractured media [153, 138]. The agglom
module implements a set of modular components that can be combined in different ways
to create various types of agglomerated partitions as outlined in the next section.

14.5 A General Framework for Agglomerating Cells

This section outlines the general agglomeration framework implemented in the agglom
module. The agglomeration process can be viewed as governed by three rules:

• a neighbor definition that gives the permissible directions one can search for new cells to
agglomerate;
• a set of indicator functions that determine the feasible directions among all the permissi-
ble directions;
• one or more indicators to determine how far the agglomeration should proceed, i.e.,
indicators that can be used to locally determine the size of each grid block.
Permissible directions are given by the grid topology and will in the default setup consist of
all face neighbors. However, the permissible directions can also be extended to include cells
sharing a common edge/node or cells whose centroids lie within a prescribed geometric
distance. Likewise, the permissible directions can be restricted locally so that blocks cannot
be agglomerated across faults or boundaries between different facies types, initialization
regions, saturation regions, etc. The feasible directions are defined through a set of cell-
based indicator functions that may contain geological or petrophysical features, flow-based
quantities, or general user-set expert knowledge. One can also impose additional rules that
prevent the total indicator function from growing too large within blocks, try to regularize
the outline of the blocks, minimize aspect ratios. Altogether, this gives a very flexible
framework that can be used to construct many different types of grids.
To realize this framework, the agglom module implements a set of source functions that
create new partition vectors, and a set of filter functions that take one or more partitions as
input and create a new partition as output. Examples of filter operations include combin-
ing/intersecting multiple partition vectors; performing sanity checks and modifications to
ensure that no blocks are disconnected or contained within other blocks; or modifying the
partition vector by merging small blocks or splitting large blocks.

14.5.1 Creating Initial Partitions

You have already seen several examples of topological partitions, e.g., generated by
partitionUI from the coarsegrid module. Another way to create initial partitions is
to segment indicator functions representing the heterogeneity of the medium. Examples
of such indicator functions include the logarithm of permeability, velocity magnitude,

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 542 Grid Coarsening

Figure 14.9 Segmentation with permeability as indicator for CaseB4 with four lognormally dis-
tributed layers (left). The right plot shows the resulting bins.

time-of-flight or residence time, vorticity, etc. Assuming we have an indicator function I

that takes a single value in each cell of a grid G, the indicator can be segmented into bins
by calling
p = segmentIndicator(G, I, bins)

Constructing indicator functions and making initial partitions will typically involve some
kind of expert knowledge, and hence the MRST module does not offer specific routines.
However, the tutorials contain several examples that should be instructive. To illustrate
this function, let us revisit the CaseB4 example from Section 14.3.3 and see if we can use
this function with permeability as an indicator to automatically detect the four geological
layers in the model. The permeability is lognormal within each layer. The histogram of
permeability values in Figure 14.9 has four distinct peaks, but the different layers are not
clearly separated. We thus pick limits for the bins in the local minima of the distribution (K
is given in md):
p = segmentIndicator(G, K, [0 30 80 205 inf], 'split', false);

Here, we have asked the function to only output the segmented bins. These bins, shown
to the right in Figure 14.9, do not follow the layers exactly since permeability values are
not unique to a single layer. If it is important to preserve the layering exactly, we would be
better off by using the k index for the initial partition. In other cases like the SPE 10 model,
all we have to bin different rock types are the permeability and/or porosity values.

14.5.2 Connectivity Checks and Repair Algorithms

Partitions defined by segmenting indicator functions usually give bins that consist of
multiple connected components. These can be split by the function processPartition as
discussed in Section 14.1.2. This function is called by default from within

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.5 A General Framework for Agglomerating Cells 543

2
3
4

confined blocks

Figure 14.10 The left plot shows coarse blocks from the initial segmentation of CaseB4; the
histogram of cells per block has logarithmic y-scale. Blocks 1,2 and 3,4 are disconnected components
from layers two and three, respectively. The right figure shows isolated blocks that are completely
confined within another block.

segmentIndicator, and if we drop the optional 'split' flag from the call, the initial
bins are split so that they form blocks satisfying MRST’s basic requirements for a coarse
grid. Figure 14.10 shows that the resulting partition consists of six large coarse blocks and
a number of smaller ones. Layers two and three are not fully connected across all faults
and the corresponding bins have therefore each been split in two large blocks (1 and 2, and
3 and 4 in the plot). The remaining small blocks are the result of permeability variations
inside the layers.
Blocks that are entirely confined within other blocks only have a single interface. When
the resulting grid is used for flow simulation, the only flow that can pass the interface
between the block and its surrounding neighbor must be caused by source terms or com-
pressible effects. This tends to create numerical artifacts unless the block contains a well.
For incompressible flow, in particular, the corresponding block interface will act as an
internal no-flow boundary. An additional sanity check should therefore be performed to
detect such blocks:
cb = findConfinedBlocks(G, p);
[cb,p] = findConfinedBlocks(G,p);

The first call just outputs a list of confined blocks, whereas the second also merges these
blocks with their surrounding neighbor. The right plot in Figure 14.10 illustrates confined
blocks for CaseB4.
The detection algorithm is somewhat simplified and the function cannot handle nested
cases properly. In the case of recursively confined blocks, only the outermost and the
innermost blocks are listed in cb. To find a more general solution, we can look to algorithms
for detecting biconnected components from graph theory. Let me explain the idea by a
simple example: Figure 14.11 shows a 7 × 7 grid partitioned into six coarse blocks together

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 544 Grid Coarsening

2 6

5 3 5

3 1 4

1 Coarse grid graph

findConfinedBlocks removeConfinedBlocks

Figure 14.11 Detection and repair of recursively confined blocks by use of functionality from the
MATLAB Boost Graph Library. The simple findConfinedBlocks function fails to detect and repair
cases where a set of blocks are confined within another block.

with the corresponding coarse-grid graph. This graph is not biconnected: node 3 belongs
to two connected components (1–2–3 and 3–4–5–6), and if this node is removed, the graph
becomes disconnected. The function
pn = removeConfinedBlocks(G, p)

merges single confined blocks with their surrounding blocks. In addition, it uses function-
ality from the MATLAB Boost Graph Library2 to determine whether a graph is bicon-
nected or not and to find the corresponding components of the graph. If the graph is not
biconnected, any grid block that completely surrounds a set of other blocks is split in two
(along a plane orthogonal to the x-axis through the block center). The resulting partition
is not necessarily singly connected and may have to be processed by a further call to
processPartition.

14.5.3 Indicator Functions

The agglomeration framework relies on two indicator functions, which we for simplicity
refer to as a volume indicator and a flow indicator. We assume that each indicator function

2 This software is freely available under a BSD License. The matlab_bgl module has a function downloadMBGL that
downloads and installs this library. This is one of the few examples of external dependencies in MRST.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.5 A General Framework for Agglomerating Cells 545

takes value I (ci ) in cell ci and that these values can be interpreted as densities, i.e., that they
are positive, additive, and normalized by the cell volume |ci |. As for the cells, we associate
indicator values I (Bj ) to each block, defined as the volumetric average of the cell indicator
values within the block (or as the arithmetic average if I is not a density).
The volume indicator is only used to determine block volumes and is either set to be
unity for bulk volumes and φ for pore volumes. The flow indicator can in principle be any
user-defined quantity, but typical examples include quantities derived from permeability,
time-of-flight or residence time, velocity or flux, vorticity, and so on (see discussion in
[124]). Error indicators or sensitivities can also be used as flow indicators. Specific exam-
ples are given later.
To understand the role of these two indicators, let us look at four general principles you
can use to control how cells are grouped together:
1. The variation of the flow indicator should be minimized inside each block. That is,
each block should be as homogeneous as possible so that cells with high and low
permeabilities, large and small fluxes, high and low travel times, and so on, are separated
in different blocks. Likewise, cells from different facies, deposition environments, flow
units, initialization regions, relperm/capillary regions, and so on, should preferably not
be agglomerated into the same block. Algorithmically, this means that the flow indicator
is used to pick the most feasible among the permissible neighbors of a block when
adding a new cell to a growing block.
2. The integral of the flow indicator should be equilibrated over blocks. You can use this
principle refine resolution in regions with high flow, low residence time, high error
or sensitivity indicator, and so on. Full equilibration will in general require rigorous
optimization. Algorithms in the agglom framework instead try to ensure that the integral
of the flow indicator is within prescribed upper bounds. These bounds determine when
to stop agglomerating more cells into a block, or alternatively, determine when blocks
are too big and must be split.
3. The size of each block should be with certain upper and lower bounds. Algorithmically,
upper limits on block sizes are imposed by the upper bound on the flow indicator,
whereas lower bounds are imposed through the volume indicator when determining the
blocks that should potentially be merged with their neighbors.
4. Blocks should not have too-large aspect ratios and be as regular as possible. Regularity
is determined in part by the source functions generating initial partitions, and in part
by filter functions that split or merge individual blocks. Controlling regularity may be
difficult in ad hoc algorithms, and if regularity is important, you may be better off using
METIS.

14.5.4 Merge Blocks

We have already seen how manually constructed partitions can give grids with very large
variations in block sizes. Variations in block sizes can be even more pronounced if the
initial partition is created by segmenting an indicator function. One method to remedy the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 546 Grid Coarsening

problem would be to loop through all blocks in the partition and try to merge each small
blocks with one of its neighbors. Our simple implementation in the SAIGUP example from
Section 14.3.2 was not very efficient as it required generation of a new coarse grid structure
in each iteration. Our choice of merging with the neighbor having largest volume was also
somewhat ad hoc.
In the agglomeration framework, we use the volume indicator Iv to measure the size of
cells and blocks with Iv ≡ 1 for bulk volume and Iv = φ for pore volume. (The latter
definition can also be modified to include net-to-gross or other factors that affect the pore
volume.) Let G = {ci }ni=1 denote the whole grid consisting of n cells that each has a bulk
volume |ci |. Then, block Bj = {ci | p(ci ) = j } should be merged with a neighbor if it
violates the following lower bound
1 
n
I¯v =
Iv (Bj ) |Bj | ≥ nl I¯v , Iv (ci ) |ci |, (14.1)
n
i=1

where nl ≥ 1 is a prescribed constant and Iv (Bj )|Bj | = ci ∈Bj Iv (ci ) |ci |. In other words,
block Bj should be merged if its integrated indicator value is less than nl times the average
volume indicator value for all cells. To determine which block to merge with, we use
the flow indicator If and merge block Bj with the block among block Bj ’s permissible
neighbors N (Bj ) that has the closest indicator value, i.e.,
B̃ = argminB⊂N (Bj ) |If (B) − If (Bj )|. (14.2)
This is implemented in the function
q = mergeBlocks(p, G, Iv, Ifl, NL, 'static_partition', ps)

which takes a partition p, a volume indicator Iv, and a positive flow indicator Ifl and
merges all blocks that violate (14.1). In doing so, each merged block is agglomerated into
the neighboring block that has the closest If value. Optionally, the merging operation can
be set to respect a static partition p2 so that no block ends up having different ps values.
The problem with this method is that it does not impose any control on the upper size
of the merged blocks. In particular, merging with the nearest flow indicator tends to work
against the principle that flow indicators should be equilibrated across all the grid blocks.
The idea3 of the flow indicator is that it should enable us to aggressively coarsen the grid
in regions of low flow or regions that has limited impact on the overall simulation result,
while keeping higher resolution in regions of high flow.
To provide upper control on the size of the blocks, we introduce a second requirement,
this time on the flow indicator:
If (B) |B| ≤ nu I¯f , (14.3)

3 In principle, the flow indicator need not have anything to do with flow; I still think it is important that you know the underlying
philosophy so that you can design indicators so that the merging/refinement process works as you desire.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.5 A General Framework for Agglomerating Cells 547
Cell-wise flow indicator mergeBlocks mergeBlocks2

6 1 3 3 3 10 7 9

4 5 1 2 2 20 4 4

10
4
1 5 1 1 1 9
7
3
1 1 5 2 2 9 4

Figure 14.12 Comparison of the two merging algorithms for a simple 5 × 4 grid. The left plot shows
the initial partition and the flow indicator If per cell. The middle and right plot shows partitions after
merging, with numbers giving the integrated flow indicator, which according to (14.3) should not
exceed 7.5.

where nu is a prescribed parameter and I¯f is the flow indicator averaged over all cells,
defined exactly the same way as I¯v . The second merging function
q = mergeBlocks2(p, G, Iv, Ifl, NL, NU)
q = mergeBlocks2(p, G, Iv, Ifl, NL, NU, 'nblock', NB, 'cfac', cfac)

not only provides a more efficient implementation of the merging process, but also imposes
(14.3) as a soft constraint, meaning that the routine tries its best to avoid violating this
condition, but if no other options are available, it merges blocks even if the flow indicator
of the new block becomes too high. It is also possible to specify an optional relative factor
cfac at which (14.3) is turned into a hard constraint. Likewise, the parameter NB gives an
upper bound on the number of cells that can be agglomerated into a single block. (Notice
also that this routine does not support static partitions.)
Figure 14.12 compares the two methods for a small 5×4 grid, called with the parameters
p1 = mergeBlocks(p, G, ones(n,1), I, 2);
p2 = mergeBlocks2(p, G, ones(G.cells.num,1), I, 2, 3);

Here, mergeBlocks identifies four single-cell blocks that are merged with their neighbor
with closest value: 4→6, 5→(2,2), (4,6)→(5,5), and 1→(3,3,3). The result is three new
blocks that all violate (14.3). The alternative routine merges 4→(1,1,1) and 6→1, which
both are acceptable. For cell with value 5, the only option is to merge with (1,1,1,1), even
if this violates (14.3).

14.5.5 Refine Blocks

The next natural step would now be to refine blocks that violate the upper bound (14.3).
You can do this by invoking any of the two calls
p = refineBlocks(p, G, Ifl, NU, @refineUniform, 'cartDims',[Nx Ny Nz])
p = refineBlocks(p, G, Ifl, NU, @refineGreedy)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 548 Grid Coarsening

The first call refines blocks by encapsulating the corresponding cell indices in a rectangular
bounding box and then performing a load-balanced partition inside this box. The second
line uses a greedy algorithm that starts by picking an arbitrary cell c0 on the block boundary
and then agglomerates a new block from the cell on the block boundary that is furthest away
from the first cell. Once this new block is large enough, the routine once again picks the
cell among the remaining ones that is furthest away from c0 and starts to agglomerating a
second block, and so on. The greedy algorithm comes in four different versions:

• refineGreedy – grows a new block inward from the block boundary by adding one entire
ring of permissible neighbors in each iteration [3]. The permissible level-one neighbors
are all cells sharing a face with existing cells in the growing block (optional parameter
'nlevel'=1). The default behavior ('nlevel'=2) is to also include cells that share (at
least) two faces with existing cells in the block or level-one cells in the ring.
• refineGreedy2 – improved algorithm that only adds parts of the ring of level-one or
level-two neighbors to honor the upper bound (14.3) on the flow indicator.
• refineGreedy3 – sorts cells in the neighbor ring according to number of faces shared with
the growing block. Permissible neighbors are defined by the function neighboursByNodes
and includes all cells that share a face or a node with cells in the growing block.
Computing this neighborship is expensive for large grids.
• refineGreedy4 – sorts cells in the ring of neighbors according to difference in Ifl value,
using the same permissible neighbors as refineGreedy3.
To illustrate how these functions operate for permissible level-one neighbors only, we
consider a 4 × 4 grid with uniform flow indicator If ≡ 1. The functions are run with
the following parameters:
p = refineGreedy(I,G,I,4,'nlevel',n);
:
p = refineUniform(I,G,I,4,'CartDims',[2,2]);

where I is a vector of all ones. Figure 14.13 shows that the uniform partition gives
four blocks, as expected. All greedy algorithms pick c0 = 1 and start agglomerating
blocks from cell 16. The simplest greedy algorithm, refineGreedy, only adds face
neighbors and hence tends to create diamond-shaped blocks like the second block,

refineUniform refineGreedy refineGreedy2 refineGreedy3 refineGreedy4

13 14 15 16

9 10 11 12

5 6 7 8

1 2 3 4

1 2 3 4 5

Figure 14.13 Refinement algorithms using level-one permissible neighbors on a 4 × 4 grid with
uniform flow indicator.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.5 A General Framework for Agglomerating Cells 549
refineUniform refineGreedy refineGreedy2 refineGreedy3 refineGreedy4
21 22 23 24 25
16 17 18 19 20
11 12 13 14 15
6 7 8 9 10
1 2 3 4 5

1 2 3 4 5 6 7 8 9 10 11 12 13

Figure 14.14 Refinement algorithms using level-two permissible neighbors (which is the default
setting) on a 5 × 5 grid with uniform flow indicator.

which has six cells (8,11,12,14,15,16) and thus violates (14.3). After agglomerating
three blocks, the algorithm is only left with cell c0 and terminates. For refineGreedy2
and refineGreedy4, the first step agglomerates cells (12,15,16). In the next step, cells
(8,11,14) are possible merge candidates, out of which 8 is chosen since it has the smallest
index. After the algorithm has agglomerated five blocks, there are no remaining cells in
the original block and the algorithm terminates. The empty block is removed by a call to
compressPartition, and blocks two to six are renumbered as blocks one to five. The
third algorithm, refineGreedy3, selects cell 11 in the second step since this cell shares
one more face with cells (12,15,16) than the other two candidates. Once the first block is
formed, the greedy algorithms selects the remaining cell with maximum distance to c0 as
the next seed (cell 4 for the first algorithm, cell 11 for the second and fourth, and cell 8 for
the third.)
Figure 14.14 shows a similar test case with level-two neighbors and a 5 × 5 grid. Here,
the uniform refinement splits the block into four blocks of size 3 × 3, 2 × 3, 3 × 2, and 2 × 2.
The first three violate (14.3) and are thus further split twice in each direction. To understand
the irregular block shapes for the greedy algorithm, we can consider agglomeration of the
fourth block, which starts by cells (4,5). Cell 3 is the only permissible neighbor in step two,
whereas the third step has three candidates (2,7,8). All three are added by the first algorithm,
whereas refineGreedy2 chooses the cell with lowest index, which coincidentally gives
a regular block, albeit with a larger aspect ratio. Why the other two algorithms choose
cell 8 is a random effect caused by how MATLAB picks extremal values for a vector with
identical entries.
The greedy refinement algorithms apply equally well to general polygonal/polyhedral
grids with unstructured topologies. Uniform Cartesian refinement, on the other hand, relies
on an ij k numbering and thus primarily works for grids with rectangular topology. How-
ever, the same idea can be generalized to unstructured grids if we instead of relying on
topology use the cell centroids to sample from a rectangular subdivision of each block’s
geometric bounding box (see function sampleFromBox). This type of refinement is imple-
mented in
p = refineRecursiveCart(p, G, I, NU, cartDims)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 550 Grid Coarsening

The function does not always produce the exact same results as refineBlocks with uni-
form refinement, but usually runs faster.
Refinement and merging are complementary operations that in principle can be applied
in any order. A large number of tests nonetheless show that it is typically better to first
merge small blocks, and then refine large ones, and improved grids can often be obtained
if the merging-refinement process is iterated a few times. Notice also that the greedy
algorithms tend to accentuate irregularity of blocks and should not be used uncritically.

14.5.6 Examples
We end by a few examples that show how the algorithmic components discussed in this
section can be used to generate various types of coarse grids.

Example 14.5.1 (CaseB4) As you may recall from Section 14.3.3, we first partitioned the
CaseB4 grid in the ij coordinates and then split the resulting coarse blocks across the
three faults. Figure 14.15 reports block sizes before and after splitting. To get rid of the
smallest blocks, we can merge all blocks with a pore volume less than 20% of the mean
block volume, which is approximately 50 times the mean cell volume. Applying one of the
merging algorithms directly to the partition risks merging blocks across faults. We avoid
this if we create a new vector that partitions the model into three fault blocks and use this
to constrain the merging algorithm:

areal partition after merging

min [m3 ] mean [m3 ] max [m3 ]

Before split 3.6 · 106 4.3 · 106 5.9 · 106

After split 7.6 · 104 3.4 · 104 5.9 · 106
After merge 0.7 · 106 3.9 · 104 5.9 · 106

Figure 14.15 Areal partition for CaseB4 before and after merging of blocks whose pore volume is
less than 50 times the average cell volume (1.3 · 104 m3 ). The tabular shows pore volumes in various
stages of the algorithm.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.5 A General Framework for Agglomerating Cells 551

Iv = poreVolume(G, rock)./G.cells.volumes;
pm = mergeBlocks(pf, G, Iv, Iv, 50, 'static_partition', ...
processPartition(G,ones(G.cells.num,1),find(G.faces.tag==1)));

Notice, that we here use the same indicator for both volume and flow, which generally
would cause the algorithm to merge with blocks having similar average porosity values.
After merging, the ratio between the smallest and the largest block is decreased by a factor
ten, and we see that the algorithm has merged small blocks with neighbors on the correct
side of the faults. It is important to merge small blocks before we refine around wells to
avoid merging any of the refined blocks in the near-well region.

There are many different ways one can define flow indicators, for instance:

• From Darcy’s law, we know that if a region is subject to a constant pressure gradient,
the local magnitude of flow is proportional to permeability. Pressure gradients are rarely
constant throughout a reservoir, but permeability can still be used as a simple a priori
indicator to separate potential local regions of high and low flow before any flow solution
is computed.
• Given a flow solution, it is natural to use velocity magnitude in each cell as a flow indi-
cator. For a finite-volume method, this quantity must be reconstructed from the intercell
fluxes (see discussion of non-Newtonian flow in Section 7.4), and alternatively one can
use the sum of the absolute values of the fluxes as a measure that is simpler to compute.
In a certain sense, flux/velocity can be considered as local indicators, since they only
measure relative variations between individual cells and do not account for the actual
flow paths.
• To get a global flow indicator that also takes representative flow paths into account, we
can use time-of-flight or residence time.
As all three quantities tend to vary several orders of magnitude, we use their logarithm
instead as suitable indicators:
iK = log10(rock.perm(: ,1) ) ; iK = iK − min(iK) + 1;
v = sqrt(sum(faceFlux2cellVelocity(G, state.flux).^2, 2)) ;
iV = log10(v) ; iV = iV − min(iV) + 1;
iT = −log10(Tf+Tb) ; iT = iT − min(iT) + 1;

assuming that state holds a representative flow solution and that Tf,Tb are the corre-
sponding forward and backward time-of-flight values. Instead of residence time, we could
have used the product of the two time-of-flight values. Let us illustrate these indicators by
an example.

Example 14.5.2 (Adaptive coarsening) Consider a five-spot pattern posed on Layer 25

from the SPE 10 model. For improved readability of plots, we only consider a 120 × 60
excerpt, initially partitioned into a 6 × 3 coarse grid. With a uniform flow indicator, we
would get a coarse grid with 5 × 5 cells per block if we apply a 2 × 2 refinement twice.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 552 Grid Coarsening
Permeability Velocity Residence time

Figure 14.16 Recursive coarsening of a 120 × 60 excerpt of a five-spot well setup on SPE 10, Layer
25. The region is first partitioned uniformly into 6 × 3 coarse grids, and then refined recursively by a
factor 2 × 2 by different flow indicators with nu = 25 in (14.3).

With a heterogeneous flow indicator, we should expect to see local variations in the grid
resolution. Figure 14.16 confirms this. All the initial blocks are refined twice, so that no
grid block has more than 5 × 5 cells. Indicator iK adds one extra refinement to all blocks
having high permeabilities, since these would experience high flow if all cells in the model
were subject to the same pressure drop. However, not all these blocks experience high flow
for the specific five-spot flow pattern, and similarly there will be significant flow through
many blocks with low(er) permeability because the injector and the two producers in the
southeast/southwest corners lie on opposite sides of the low-permeability belt in the middle
of the model.
High-flow connections between the injector and the two producers are detected much
better by the velocity indicator iV. Unlike the indicator iT based on time-of-flight, the
velocity indicator does not take into account the cumulative resistance to flow along flow
paths and may therefore fail to refine blocks that have low local flow velocities but still lie
along a flow path with low residence time. On the other hand, if a partition is built using a
lot of specific flow information, it might be less suitable when applied to simulate radically
different flow patterns.
Example 14.5.3 (NUC algorithm) The original nonuniform coarsening (NUC) algorithm
[3] starts by segmenting a velocity indicator and then uses a sequence of merge-refine-
merge operations as illustrated in Figure 14.17. Here, we have used the upper half of
the model from the previous example, and constructed the NUC partition by the following
sequence of operations:
[volI,flwI] = deal( poreVolume(G, rock)./G.cells.volumes, iV) ;
ps = segmentIndicator(G, flwI, 5);
pm1 = mergeBlocks(ps, G, volI, flwI, 20);
pr = refineGreedy(pm1, G, flwI, 30);
pm2 = mergeBlocks(pr, G, volI, flwI, 20);

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.6 Multilevel Hierarchical Coarsening 553

Figure 14.17 Application of the original nonuniform coarsening method of [3] for a 60 × 60 excerpts
of a five-spot setup on Layer 25 of SPE 10.

The rightmost plot in the figure is obtained by replacing the merge and refinement opera-
tions by mergeBlocks2 and refineGreedy2, respectively. Although this type of partition
originally was developed to coarsen transport equations, it is currently implemented as
part of an industry-standard tool for single-phase upscaling [187] with time-of-flight as
flow indicator.
The agglom module contains several tutorial examples that give a more in-depth dis-
cussion of the NUC family of coarsening methods, and also includes an example of how
the techniques in this section can be used to generate dynamic grid coarsening that adapts
to evolving displacement fronts.

Computer exercises
14.5.1 How would you merge small blocks after the near-well refinement is introduced
in CaseB4 to avoid merging these refined blocks?
14.5.2 Replace the merging in the SAIGUP example (Section 14.3.2) by mergeBlocks.
Which flow indicator would you choose? Can you change the merging algorithm
so that it can use the SATNUM field as flow indicator? (Hint: to this end you should
use a majority vote rather than a volume-weighted average to get representative
flow indicator per block.)
14.5.3 Rerun the NUC example using the other greedy algorithms with level-1 and level-
2 neighbors and permeability and time-of-flight indicators. Try to use the NUC
algorithm on other examples, e.g., CaseB4.
14.5.4 A greedy agglomeration strategy may not always be optimal: Assume that you
have a coarse block with integrated indicator 410, and that the upper bound is 400.
The algorithm then tries to agglomerate two blocks with indicators 400 and 10,
respectively. Can you make a function that instead tries to split the indicator more
evenly among the new blocks?

14.6 Multilevel Hierarchical Coarsening

Geocellular models are realizations of a deeper geological understanding that include struc-
tural, stratigraphic, sedimentologic, and diagenetic aspects. On the way to generate a grid
model and populating it with petrophysical properties, the reservoir may have been divided

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 554 Grid Coarsening

into different units, flow zones, environments of deposition, layering, and lithographic
facies that each represents characteristics of the rock and how it was formed. The geo-
cellular models we have encountered so far in the book have mainly consisted of a grid
populated with a set of petrophysical properties like permeability, porosity, and net-to-
gross. Some geocellular model may also contain geological indicators such as facies or
rock type used to generate petrophysical properties. We have also seen that reservoirs can
be subdivided into various types of regions to model spatial dependence in relative perme-
ability models (PVTNUM), PTV behavior (PVTNUM), rock compressibility (ROCKTAB),
and equilibration regions (EQLNUM).
The purpose of coarsening a grid is usually to develop one or more reduced models
that only contain the most essential properties that affect fluid flow. It is well known
that structural and stratigraphic frameworks have the largest impact on flow patterns, and
preserving key concepts of these frameworks all the way to flow simulation is crucial
to reliably predict flow patterns in the reservoir [48]. Likewise, it is important to pre-
serve initial fluid contacts. Unfortunately, the volumetric partitions imposed by basic geo-
logical characterization are oftentimes lost in geostatistical algorithms and rarely made
available as cell or face properties in the geocellular models. As an example, consider
the distinction between high-permeability sand channels and low-permeability shales and
coal in the Upper Ness formation from SPE 10. Likewise, think of CaseB4, where we
in Figure 14.9 somewhat unsuccessfully tried to delineate four different geological layers
based on permeability. To simplify future creation of coarse models, I therefore recommend
that as much as possible of the basic geological characterization is preserved in terms of
cellular indicators even after the geocellular model has been populated with petrophysical
properties.
The idea of hierarchical multilevel coarsening is to define the various geological char-
acterizations and regions just discussed as partition vectors, give them individual priority,
and apply them recursively. By varying the number of features included, one can create
a hierarchy of coarse partitions of increasing resolution. This approach can be combined
with the agglomeration framework from the previous section to separate models into flow-
dependent compartments representing high-flow and low-flow zones, or zones that are close
to or far away from wells. A detailed discussion of this approach is outside the scope of
this book. Instead, I give two illustrative examples.

Example 14.6.1 The first example is a conceptual 40 × 40 model that has three different
geological features: flow unit, lithofacies assemblage (LFA), and two intersecting fractures.
The two first are represented by cell indicators pu and pl, whereas faults are represented as
an explicit list of faces, faults. When combined, the three geological properties subdivide
the reservoir into 15 different regions; see plots in the upper row of Figure 14.18. Flow
unit and LFA are used to populate the geocellular model with stochastic realizations of
the petrophysical properties. The lower-left plot in Figure 14.18 shows a flow indicator I
derived from permeability, which has mean value approximately equal two in LFA 1 and

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.6 Multilevel Hierarchical Coarsening 555

Lithofacies.tif Flow unit Combined hierarchy

Flow indicator LFA2 refined Merged small blocks

Figure 14.18 Multilevel hierarchical partition for a conceptual geological model with three geologi-
cal properties, flow unit, lithofacies assemblage, and faults, which combined subdivide the reservoir
into fifteen regions. The second row shows flow indicator, hierarchical 2 × 2 refinement, and the same
refinement after blocks with less than four cells in LFA 1 have been merged (shown in a red shade).

ten in LFA 2. We use this flow indicator to introduce a hierarchical Cartesian subdivision
with higher resolution in the second LFA4 :
pc = ones(G.cells.num,4);
for i=1:4
pc(:,i) = partitionCartGrid(G.cartDims,[5 5].*2^(i-1));
end
p = applySuccessivePart(processPartition(G,pu,faults),G, I, 8, [pl pc]);

That is, we first introduce a basic partition that splits the flow units and fault blocks,
and then use the flow indicator to further partition the grid according to lithofacies and
Cartesian blocks when the accumulated block indicator exceeds eight times the average cell
value. This subdivision will create small blocks also inside the first lithofacies. To merge
blocks with less than four cells, we can use mergeBlocksByConnections, which merges
blocks according to connection strength between fine-scale cells. The function does not
merge blocks that have negative connection strength, and hence we set negative connection
strength inside LFA 2 and between cells that belong to different regions in the geological
hierarchy:

4 Complete code: showHierCoarsen.m in book.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 556 Grid Coarsening

Figure 14.19 Conceptual illustration of a reservoir model with aggressive coarsening in the aquifer
zone, modest coarsening above the initial water contact, and original resolution for cells with low
residence time.

qb = [0; processPartition(G,pl+2*(pu-1), faults)];

ql = [0; pl];
T = 2*(qb(G.faces.neighbors(:,1)+1)==qb(G.faces.neighbors(:,2)+1))-1;
T(ql(G.faces.neighbors(:,1)+1)==2) = -1;
pm = mergeBlocksByConnections(G, p, T, 4);

The lower-right plot in Figure 14.18 shows the result. This example is admittedly simple,
but similar principles can be applied to industry-standard models [187].

Example 14.6.2 In the second example, we consider a conceptual model of an anticline

reservoir that overlies an aquifer. We use a hierarchical approach to aggressively coarsen
cells below the initial water contact, which are assumed to be water-filled throughout
the whole simulation. For the cells above the water contact, we apply a more modest
coarsening. To capture the initial displacement between injectors and producers, we pick
all cells with residence time less than twice the median residence time of all cells perforated
by wells. Cells with low residence time are part of the high-flow zones and are thus kept at
their original resolution. In addition, all blocks in the coarsest partition that are perforated
by a well are refined back to their original resolution to ensure that we accurately capture
flow in the near-well zone. Figure 14.19 shows the resulting model. You can find complete
source code in the script showAnticlineModel.m in the book module.

14.7 General Advice and Simple Guidelines

Over the years, I have made the general and perhaps somewhat disappointing observation
that making a good coarse grid is more an art than an exact science. In particular, I do not
believe black-box approaches are able to automatically select the optimal coarsening. As
a result, MRST does not provide well-defined workflows for coarsening, but rather offers
tools that (hopefully) are sufficiently flexible to support you in combining your creativity,
physical intuition, and experience to generate good coarse grids.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 14.7 General Advice and Simple Guidelines 557

Coarse partition are often used as input to a flow simulation. I conclude the chapter by
suggesting a few guidelines that may contribute to decrease numerical errors introduced by
coarsening:
1. Keep the grid as simple as possible: Use regular partitions for mild heterogeneities and
cases where saturation/concentration profiles are expected to be regular.
2. Try to keep the number of connections between coarse-grid blocks as low as possible
to minimize the bandwidth of the discretized systems, and avoid having too many
small blocks, which otherwise would increase the dimension of the discrete system and
adversely affect its stability without necessarily improving the accuracy significantly.
3. Use your understanding of the reservoir to judge what are the important features to
preserve when coarsening the model. In doing so, you should think of what are primary
geological features that control pressure communication and the main flow directions.
4. Make sure you have sufficient vertical resolution if you study systems with large
density differences between the fluid phases (e.g., gas injection or CO2 sequestration)
to be able to capture the vertical segregation and the lighter fluid’s tendency to override
the other fluid(s).
5. Explore alternative partitioning strategies and compare the results.
6. Use relatively robust methods like METIS with transmissibility weights or regular
partitions if you have limited understanding of which features affect the flow patterns.
7. Use basic geological features when available, e.g., facies likely yields more robust
results than segmented permeability values.
8. Adjust your partitioning strategy to the purpose of the coarsening. If you want to run
multiple simulations with more or less the same flow pattern, you can use features of
the specific flow patterns to aggressively coarsen parts of the reservoir with low flow.
Conversely, a fully flow-adapted coarsening may not work well in workflows where
well positions and flow patterns change significantly from one simulation to the next.
9. Use simple indicator functions that have an intuitive interpretation, e.g., permeability,
velocity magnitude, time-of-flight/travel time, etc.
10. Adapt coarse blocks to avoid upscaling large permeability contrasts, constrain blocks
to contiguous saturation regions to avoid upscaling relative permeability and capillary
pressure curves, etc.
11. Check the quality of a coarse grid using single-phase flow, e.g., by computing various
types of flow diagnostics to check that you preserve the volumetric communication
from the fine-scale model.
The next chapter introduces you to methods you can use to compute effective petrophysical
properties on a coarsened grid.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15
Upscaling Petrophysical Properties

In Chapter 2 we argued that porous rock formations are heterogeneous across a wide
span of length scales, from the micrometer scale of pore channels to the kilometer scale
of petroleum reservoirs or even larger scales for aquifer systems and sedimentary basins.
Describing all flow processes pertinent to hydrocarbon recovery or CO2 storage with a sin-
gle model is therefore not possible. Geological characterization therefore usually involves
a hierarchy of models that each covers a limited range of physical scales cover a wide
range of physical scales. Section 2.3 introduced you to the concept of representative ele-
mentary volumes and mentioned various types of models used for flow studies including
flow in core samples (cm scale), bed models (meter scale), sector models (tens to hundreds
of meters), and field models (km scale). Geocellular models at the field/sector scale are
made to represent the heterogeneity of the reservoir and possibly incorporate a measure
of inherent uncertainty in reservoir simulation. Pore, core, and bed models are mainly
designed to give input to the geological characterization and to derive flow parameters
for simulation models. All model types must be calibrated against static and dynamic data
of very different spatial (and temporal) resolution: thin sections, core samples, well logs,
geological outcrops, seismic surveys, well tests, production data, core flooding and other
laboratory experiments, etc. To learn more about the process of building reservoir models,
you should consult the excellent textbook by Ringrose and Bentley [265].
To systematically propagate the effects of small-scale geological variations observed in
core samples up to the reservoir scale, we need mathematical methods that can homogenize
a detailed fine-scale model and replace it with an equivalent model on a coarser scale that
contains much fewer parameters that represent the behavior of the fine-scale model in an
averaged sense; see Figure 15.1.
Such techniques are not only used to incorporate the effect of small-scale models into
macroscale models. High-resolution geocellular models used for reservoir characterization
on field or sector scale tend to contain more volumetric cells than contemporary simulators
can handle without having to resort to massively parallel high-performance computing.
As a result, flow simulations are usually performed with models containing less details
than those used for characterization. Even if you had all the computational power you
needed, there are several arguments why you should still perform your simulations on
coarser models. First of all, one can argue that high-resolution models tend to contain more

558

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.1 Upscaling for Reservoir Simulation 559

Figure 15.1 Upscaling of petrophysical properties, here represented by porosity.

details than what is justified by the information in the available data. Second, in many
modeling workflows the available computational power should be spent running multiple
model realizations to span the range of plausible outcomes rather than on obtaining high
numerical resolution for a few highly uncertain predictions. Third, because coarser models
contain fewer parameters, they are simpler to calibrate to observed reservoir responses like
pressure tests and production data. Finally, a coarser model may be sufficient to predict
flow patterns and reservoir responses with the accuracy needed to make a certain business
decision. It is perhaps tempting to believe that with future increases in computing power,
one will soon be able to close the gap in resolution between models used for reservoir
characterization and models used for flow simulation. The development so far indicates
that such an idea is wrong. The trend is rather that increases in computational power enable
geologists and reservoir engineers to build larger and more complex geocellular models at
a pace that outperforms the improvement in simulation capabilities.
For all these reasons, there is a strong need for mathematical and numerical techniques
that can be used to communicate parameters and properties between models of different
spatial resolution. This need will persist, and maybe even grow stronger, in the foreseeable
future. This chapter discusses techniques for upscaling static rock properties.

15.1 Upscaling for Reservoir Simulation

Upscaling (or homogenization) refers to the process of propagating properties and param-
eters from a model of a given resolution to a model of lower spatial resolution. In this
process, heterogeneous regions in a reservoir model are replaced by homogeneous regions
to make up a coarser model of the same reservoir. Or, in the words of Chapter 14 on grid
coarsening, upscaling is the process in which petrophysical properties in the cells that make
up a coarse block are averaged into effective values for each coarse block; see Figure 15.1.
The effective properties of the new homogeneous regions are defined so that they preserve
the effects of small-scale variations in an averaged sense. How this averaging should be
performed depends on the type of property to be upscaled. We distinguish between addi-
tive properties that can be upscaled using simple volumetric averaging and nonadditive

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 560 Upscaling Petrophysical Properties

properties, for which correct averaging methods only exist in special cases and the best
you can hope for in the general case is to compute accurate approximations.
Downscaling refers to the process of propagating properties from a model with a given
spatial resolution to a model having higher spatial resolution. We can, for instance, be inter-
ested in refining coarse-scale modifications in petrophysical properties obtained during a
history match on a simulation model to update the underlying geomodel. In the downscaling
process, the aim is to preserve both the coarse-scale trends and the fine-scale heterogeneity
structures.
The process of upscaling petrophysical parameters leads to many fundamental ques-
tions. For instance, do the partial differential equations that make up the flow model on
the coarse scale take the same form as the equations modeling flow at the subgrid scale?
And if so, how do we honor the fine-scale heterogeneities at the subgrid level? Even
though upscaling has been a standard procedure in reservoir simulation for more than
four decades, nobody has fully answered these questions rigorously, except for cases with
special heterogeneous formations such as periodic or stratified media.
Homogenization is a rigorous mathematical theory for asymptotic analysis of periodic
structures; see [41, 149, 135]. A relevant result states that for a periodic medium with
permeability K( xε ), there exists a constant symmetric and positive-definite tensor K0 such
that the solutions pε and vε = −K( xε )∇pε to the elliptic problem
x
− ∇ · K( )∇pε = q (15.1)
ε
converge uniformly as ε → 0 to the solution of the homogenized equation
− ∇ · K0 ∇p0 = q. (15.2)
Homogenization theory can be used to derive homogenized tensors for the region, if we
can safely assume that the region we seek to compute an effective property for is part of
an infinite periodic medium. The main advantage of homogenization is that it provides
mathematical methods to prove existence and uniqueness of the coarse-scale solution and
also verifies that the governing equation at the macroscopic level takes the same form as the
elliptic equation that governs porous media flow at the level of the representative elemen-
tary volumes (REVs). However, it is more debatable whether mathematical homogenization
can be used for practical simulations, since natural rocks are rarely periodic.
As you will see shortly, the most basic upscaling techniques rely on local averaging
procedures that calculate effective properties in each grid block solely from properties
within the grid block. Because these averaging procedures do not consider coupling beyond
the local domain, they fail to account for the effect of long-range correlations and large-
scale flow patterns in the reservoir, unless their effect can be represented correctly by the
forces that drive flow inside the local domain. Different flow patterns may call for different
upscaling procedures, and it is generally acknowledged that global effects must also be
taken into consideration to obtain robust coarse-scale simulation models.
Upscaling must also be seen in close connection with griding and coarsening methods,
as discussed in Chapters 3 and 14. To make a fine-scale model, you must decide what kind

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.2 Upscaling Additive Properties 561

of grid you want to represent the porous medium, what (local) resolution you need, and
how you should orient your grid cells. Obviously, you should consider the same questions
for the upscaled model. To make an upscaled simulation model as accurate and robust as
possible, the grid should be designed so that grid blocks capture heterogeneities on the
scale of the block. This often implies that you may need significantly more blocks if you
use a regular grid than if you use an unstructured polyhedral grid. You may also want to use
different coarsening factors in zones of high and low flow, near and far away from wells
and fluid contacts, as discussed in Chapter 14. The importance of designing a good coarse
grid should not be underestimated. Henceforth, however, we simply assume that a suitable
coarse grid is available and focus on upscaling techniques.
The literature on upscaling techniques is extensive, ranging from simple averaging tech-
niques, e.g., [151], via flow-based methods that rely on local flow problems [37, 90] to more
comprehensive global [225, 133] and local–global [63, 64, 117] methods. Some attempts
have been made to analyze the upscaling process, e.g., [31, 318], but so far there is generally
no theory or framework for assessing the quality of an upscaling technique. In fact, upscal-
ing techniques are rarely rigorously quantified with mathematical error estimates. Instead,
the quality of upscaling techniques is usually assessed by comparing upscaled production
characteristics with those obtained from a reference solution computed on an underlying
fine grid. This deficiency is one of the motivations behind the development of modern
multiscale methods [95, 195]. Such multiscale methods are more well developed in MRST
than in any comparable software, but a detailed discussion is outside the scope of this book.
In the following, we only discuss the basic principles and try to show you how you
can implement relatively simple upscaling methods; we also attempt to explain why some
methods may work well for certain flow scenarios and not for others. If you are interested
in a comprehensive overview, you should consult one of the many review papers devoted
to this topic, e.g., [70, 311, 31, 263, 108, 91, 92]. We start by a brief discussion of how
to upscale porosity and other additive properties, before we move on to discuss upscaling
permeability, which is the primary example of a nonadditive property. Because of the way
Darcy’s law has been extended from single-phase to multiphase flow, it is common to distin-
guish the upscaling of absolute permeability K from the upscaling of relative permeability
krα . Upscaling of absolute permeability is often called single-phase upscaling, whereas
upscaling of relative permeability is referred to as multiphase upscaling. The main parts of
this chapter are devoted to permeability upscaling and to upscaling of the corresponding
transmissibilities that account for permeability effects in finite-volume discretizations. As
in the rest of the book, our discussion will to a large extent be driven by examples, for
which you can find complete codes in the upscaling directory of the book module.

15.2 Upscaling Additive Properties

Porosity is the simplest example of an additive property and can be upscaled through a
simple volumetric average. If  denotes the region we want to average over, the averaged
porosity value is given as

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 562 Upscaling Petrophysical Properties

∗ 1
φ = x ) d x.
φ( (15.3)
 

Implementing the computation of this volumetric average can be a bit tricky if the coarse
blocks and the fine cells are not matching. In MRST, however, we always assume that the
coarse grid is given as a partition of the fine grid, as explained in Chapter 14. If q denotes
the vector of integers describing the coarse partition, upscaling porosity amounts to a single
statement
crock.poro = accumarray(q,rock.poro.*G.cells.volumes)./ ...
max(accumarray(q,G.cells.volumes),eps);

We use max(..,eps) to safeguard against division by zero in case your grid contains blocks
with zero volume or your partition vector is not contiguous. Weighting by volume is not
necessary for grids with uniform cell sizes. Similar statements were used to go from the
fine-scale model on the left side of Figure 15.1 to the coarse-scale model shown on the
right. Complete source code is found in the script illustrateUpscaling.m.
Other additive (or volumetric) properties like net-to-gross, (residual) saturations, and
concentrations can be upscaled almost in the same way, except that you should replace the
bulk average in (15.3) by a weighted average. If n denotes net-to-gross, the correct average
would be to weight with porosity, so that

 −1 
n∗ = x ) d x
φ( φ( x ) d x.
x )n( (15.4)
 

and likewise for saturations. In MRST, we can compute this upscaling as follows:
pv = rock.poro.*G.cells.volumes;
N = accumarray(q,pv.*n)./ max(accumarray(q,pv),eps);

To verify that this is the correct average, we simply compute

1   −1 
φ ∗ n∗ = x ) d x
φ( x ) d x
φ( φ( x ) d x
x )n(
 
  
1
= φ( x ) d x = (φ n)∗ .
x ) n(
 

Using the same argument, you can easily propose that concentrations should be weighted
with saturations, and so on. Rock type (or flow unit), on the other hand, is not an additive
property, even though it is sometimes treated almost as if it was by applying a majority vote
to identify the rock type that occupies the largest volume fraction of a block. Such a simple
approach is not robust [324] and should generally be avoided.
To get more acquainted with upscaling of additive quantities, I recommend that you try
to do the following computer exercises.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.3 Upscaling Absolute Permeability 563

Computer exercises

15.2.1 Construct a 8 × 24 × 5 coarse grid CG of the SAIGUP model and upscale the
additive rock properties. Verify your upscaling by computing
pv = accumarray(CG.partition, poreVolume(G,rock));
max(abs(pv - poreVolume(CG, crock)))

Plot histograms of the fine and coarse porosities and compare. How would you
make a coarse grid that preserves the span in porosity values better? To measure the
quality of a given upscaling, you can for instance use the following error measure:
err = sum(abs(rock.poro - crock3.poro(CG.partition)))/...
sum(rock.poro)*100;

It may also be helpful to plot the cell-wise discrepancy between the upscaled and
the original porosity over the fine grid.
15.2.2 The file mortarTestModel from the BedModels1 data set describes a sedimentary
bed consisting of three different facies. Construct a coarse model that has 5 × 5
blocks in the lateral direction and as many blocks in the
vertical direction as required to preserve distinct
facies layers, as shown in the plot to the right. Upscale
the porosity of the model. Do you think the resulting
coarse grid is suitable for flow simulations? (Hint:
Check the guidelines from Section 14.7.)

15.3 Upscaling Absolute Permeability

To study upscaling of absolute permeability, it is sufficient to consider single-phase flow in
the form of a variable-coefficient Poisson equation

∇ · K∇p = 0. (15.5)

Even with such a simple equation, the choice of what is the best method to average absolute
permeability generally depends on a complex interplay between the local permeability
distribution and the characteristic flow directions. In certain special cases, one can develop
simple methods that average permeability correctly, but in the general case, all you can do
is develop computational methods that approximate the true effective permeability of the
upscaled region. How accurate a given approximation is will depend on the coarse grid, the
specific upscaling method, the purpose for which the upscaled values are to be used, and
the complexity of the fine-scale permeability distribution.
Most techniques for upscaling absolute permeability seek an averaged tensor K∗ that
reproduces the same total flow through each homogeneous region you would obtain by
solving the single-pressure equation (15.5) with the full fine-scale heterogeneity. In other

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 564 Upscaling Petrophysical Properties

words, if  is the homogeneous region to which we wish to assign an effective property

K∗ , this property should fulfill
 
∗
x )∇p d x = K
K( ∇p d x. (15.6)
 
This equation states that the net flow rate v through  is related to the average pressure
gradient ∇ p in  through the upscaled Darcy law
v = −K∗ ∇ p. (15.7)
The upscaled permeability tensor K∗ is not uniquely defined by (15.6) for a given pressure
field p, and conversely, there does not exist a unique K∗ so that (15.6) holds for any pressure
field. This reflects that K∗ depends on the flow through , which in turn is determined by
the boundary conditions specified on ∂. The better you know the boundary conditions
that a homogenized region will be subject to in subsequent simulations, the more accurate
estimates you can compute for the upscaled tensor K∗ . In fact, if you know these boundary
conditions exactly, you can compute the true effective permeability. In general, you will not
know these boundary conditions in advance unless you have already solved your problem,
and the best you can do is to make an educated and representative guess that aims to
give reasonably accurate results for a specific or a wide range of flow scenarios. Another
problem is that even though the permeability tensor of a physical system must be symmetric
and positive definite (i.e., z · Kz > 0 for all nonzero z), there is generally no guarantee
that the effective permeability tensor constructed by an upscaling algorithm fulfills the
same properties. The possible absence of symmetry and positive definiteness shows that
the single-phase upscaling problem is fundamentally ill-posed.

15.3.1 Averaging Methods

The simplest way to upscale permeability is to use an explicit averaging formula. power
averages constitute a general class of such formulas,
 1  1/p
∗
K = Ap (K) = K(x )p d x . (15.8)
|| 
Here, p = 1 and p = −1 correspond to the arithmetic and harmonic average, respectively,
whereas the geometric mean is obtained in the limit p → 0 as
 1  
∗
K = A0 (K) = exp x )) d x .
log(K( (15.9)
|| 
The use of power averaging can be motivated by the so-called Wiener-bounds [317], which
state that for a statistically homogeneous medium, the correct upscaled permeability will
be bounded above and below by the arithmetic and harmonic mean, respectively.
To motivate (15.8), we can look at the relatively simple problem of upscaling perme-
ability within a one-dimensional domain [0,L]. From the flow equation (15.5) and Darcy’s
law, we have that

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.3 Upscaling Absolute Permeability 565
 
− K(x)p (x) = 0 "⇒ v(x) = K(x)p (x) ≡ constant.

The upscaled permeability K ∗ must satisfy Darcy’s law on the coarse scale, hence v =
−K ∗ [pL − p0 ]/L. Alternatively, this formula can be derived from (15.6)
 L  L  L
∗  ∗ (15.6) 
K p (x)dx = K (pL − p0 ) = K(x)p (x)dx = − vdx = −Lv.
0 0 0

We can now use (15.6) and Darcy’s law to find the expression for K∗
 L  Lv
p (x)dx = − dx
0 0 K(x)
 L  L  1  L 1 
∗ pL − p0 1 ∗ 
=K dx = K p (x)dx dx ,
L 0 K(x) 0 L 0 K(x)
from which it follows that the correct way to upscale K is to use the harmonic average
1  L 1 −1
K∗ = dx . (15.10)
L 0 K(x)
This result is universally valid only in one dimension, but also applies to the special case of
a perfectly stratified isotropic medium with layers perpendicular to the direction of pressure
drop as illustrated in Figure 15.2. It is a straightforward exercise to extend the analysis just
presented to prove that harmonic averaging is the correct upscaling in this case. (If you are
familiar with Ohm’s law in electricity, you probably realize that this setup is similar to that
of resistors set in parallel.)
Computing the harmonic average is straightforward in MRST:
vol = G.cells.volumes;
for i=1:size(rock.perm,2)
crock.perm(:,i) = accumarray(q,vol) ./ ...
accumarray(q,vol./rock.perm(:,i))
end

Here, we have assumed as in the previous section that q is a partition vector describing the
coarse grid.

p0 K0 K1 K2 K3 K4 K5 K6 K7 K8 K9 pL H

Figure 15.2 Example of a perfectly stratified isotropic medium with layers perpendicular to the
direction of the pressure drop, for which harmonic averaging is the correct way to upscale absolute
permeability.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 566 Upscaling Petrophysical Properties

L
K2
p0 K1 pL H
K0

Figure 15.3 Example of a perfectly stratified isotropic medium with layers parallel to the direction of
the pressure drop, for which arithmetic averaging is the correct way to upscale absolute permeability.

Another special case is when the layers of a perfectly stratified isotropic medium are
parallel to the pressure drop as shown in Figure 15.3. Because the permeability is constant
along the direction of pressure drop, the pressure will be linear in the x-direction with
p(x,y) = p0 + x(pL − p0 )/L. Using (15.6), we can compute
 H L
K∗ ∂x p(x,y) dx dy = K ∗ H (pL − p0 )
0 0
 H L  H L
pL − p0
= K(x,y)∂x p(x,y) dx dy = K(x,y) dx dy,
0 0 L 0 0
from which it follows that the correct upscaling is to use the arithmetic average
 H L
∗ 1
K = K(x,y) dx dy. (15.11)
LH 0 0
These examples show that averaging techniques can give correct upscaling in special cases,
also in three dimensional space. If we now combine these two examples, we see that we
can define the following upscaled permeability tensors for the isotropic media shown in
Figures 15.2 and 15.3,

x 
y 
∗ A−1 (K) 0 ∗ A1 (K) 0
K = and K = y ,
0 Ax1 (K) 0 A−1 (K)
where the superscripts x and y on the averaging operator A from (15.8) signify that the
operator is only applied in the corresponding spatial direction. These averaged permeabili-
ties would produce the correct net flow across the domain when these models are subject to
a pressure differential between the left and right boundaries or between the top and bottom
boundaries. For other boundary conditions, however, the upscaled permeabilities generally
give incorrect flow rates.
To represent flow in more than one direction, also for cases with less idealized heteroge-
neous structures modeled by a diagonal fine-scale tensor, you can compute a permeability
tensor with the following diagonal components:
⎡ yz x ⎤
A1 (A−1 (K)) 0 0
K∗ = ⎣ ⎦.
y
0 Axz
1 (A−1 (K)) 0
xy z
0 0 A1 (A−1 (K))

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.3 Upscaling Absolute Permeability 567

In other words, we start by computing the harmonic average of each of the diagonal perme-
ability components in the corresponding longitudinal direction, i.e., compute the harmonic
average of Kxx along the x direction, and so on. Then, we compute the arithmetic average
in the transverse directions, i.e., in the y and z directions for Kxx , and so on. This average is
sometimes called the harmonic-arithmetic average and may give reasonable upscaling for
layered reservoirs when the primary direction of flow is along the layers. More importantly,
the harmonic-arithmetic average provides a tight lower bound on the effective permeability,
whereas the opposite method, the arithmetic-harmonic average, provides a tight upper
bound.
It is not obvious how to compute the harmonic-arithmetic average for a general unstruc-
tured grid, but it is almost straightforward to do it for rectilinear and corner-point grids in
MRST. For brevity, we only show the details for the case when the grid has been partitioned
uniformly in index space:
q = partitionUI(G, coarse);
vol = G.cells.volumes;
for i=1:size(rock.perm,2)
dims = G.cartDims; dims(i)=coarse(i);
qq = partitionUI(G, dims);
K = accumarray(qq,vol)./accumarray(qq,vol./rock.perm(:,i));
crock.perm(:,i) = accumarray(q,K(qq).*vol)./accumarray(q,vol);
end

The key idea in the implementation above is that to compute the harmonic average in one
axial direction, we introduce a temporary partition qq that coincides with the coarse grid
along the given axial direction and with the original fine grid in the other axial directions.
This way, the call to accumarray has the effect that the harmonic average is computed
for one longitudinal stack of cells at the time inside each coarse block. To compute the
arithmetic average, we simply map the averaged values back onto the fine grid and use
accumarray over the original partition q.

Example 15.3.1 The script averagingExample1 in the book module shows an example of
the averaging methods just discussed. The permeability field is a 40 × 60 sub-sample
of Layer 46 from the fluvial Upper Ness formation in the SPE 10 data set. Figure 15.4
compares the effective permeabilities on a 15 × 15 coarse grid computed by arithmetic,
harmonic, and harmonic-arithmetic averaging. We see that arithmetic averaging has a
tendency to preserve high permeability values, harmonic averaging tends to preserve small
permeabilities, whereas harmonic-arithmetic averaging is somewhere in between.

Although simple averaging techniques can give correct upscaling in special cases, they
tend to perform poorly in practice because the averages do not properly reflect the hetero-
geneity structures. Likewise, it is also generally difficult to determine which averaging to
use, since the best averaging method depends both on the heterogeneity of the reservoir and
the prevailing flow directions. Let us illustrate this by an example.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 568 Upscaling Petrophysical Properties
Fine scale Arithmetic Harmonic Harmonic-arithmetic

.1 10 1000 .1 10 .1 10 1000 .1 10 .1 10 1000 .1 10 .1 10 1000 .1 10

Figure 15.4 Arithmetic, harmonic, and harmonic-arithmetic averaging applied to a 40 × 60 subset of

Layer 46 from the SPE 10 data set.

1000

100

10

0.1

0.01

[mD]

Figure 15.5 Three models used to test the accuracy of averaging techniques. From left to right: a
layered model and two subsets from the Tarbert and the Upper Ness formation, respectively, from the
SPE 10 data set.

Example 15.3.2 Consider an 8 × 8 × 8 reservoir with three different permeability

realizations shown in Figure 15.5, which we upscale to a single coarse block using
arithmetic, harmonic, and harmonic-arithmetic averaging. To assess the quality of the
upscaling, we compare fine-scale and coarse-scale prediction of net flux across the outflow
boundary for three different flow patterns: from west to east, from south to north, and
from bottom to top. Complete source code for this simulation setup is given in the script
averagingExample2.
Table 15.1 reports the ratio between the outflow computed by the coarse models and
the outflow computed on the original fine grid. For the layered model, arithmetic and
harmonic-arithmetic averaging correctly reproduces flow in the lateral directions, whereas

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.3 Upscaling Absolute Permeability 569
Table 15.1 Ratio between flow rate predicted by upscaled/fine-scale models
for three permeability fields, flow scenarios, and upscaling methods.

Model Flow pattern Arithmetic Harmonic Harm-arith

Layered East→West 1.0000 0.2662 1.0000

North→South 1.0000 0.2662 1.0000
Top→Bottom 3.7573 1.0000 1.0000

Tarbert East→West 1.6591 0.0246 0.8520

North→South 1.6337 0.0243 0.8525
Top→Bottom 47428.0684 0.3239 0.8588

Upper Ness East→West 3.4060 0.0009 0.8303

North→South 1.9537 0.0005 0.7128
Top→Bottom 6776.8493 0.0020 0.3400

flow normal to the layers is correctly reproduced by harmonic and harmonic-arithmetic

averaging. For the two anisotropic models from SPE 10, on the other hand, the flow rates
predicted by the arithmetic and harmonic methods are generally far off. The combined
harmonic-arithmetic method is more accurate, with less than 15% discrepancy for the Tar-
bert model and 17–76% discrepancy for the Upper Ness model. Whether this can be con-
sidered an acceptable result will depend on what purpose the simulation is to be used for.

Computer exercises
15.3.1 Set up a set of flow simulations with different boundary conditions and/or well
patterns to test the accuracy of the three effective permeability fields computed in
averagingExample1.
15.3.2 Implement harmonic-arithmetic averaging for the SAIGUP model. (Hint: the
script cpGridHarmonic in the upscaling module computes harmonic averaging
for the SAIGUP model.) Can you also apply it to the coarse grids generated for
CaseB4 in the previous chapter?
15.3.3 Redo the upscaling in Figure 15.4 using a coarse grid that adapts to high-permeable
or high-flow zones. Does this help preserve the distribution of the permeabilities?
15.3.4 Implement the operator Ap in (15.8) as a new utility function.
15.3.5 How would you upscale the well indices in a Peaceman well model?

15.3.2 Flow-Based Upscaling

Thus far in this section, we repeatedly used simple flow problems to argue whether a partic-
ular averaging method was good or not. Taking this idea one step further, we could impose
representative boundary conditions along the perimeter of each coarse block and solve

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 570 Upscaling Petrophysical Properties

Figure 15.6 Illustration of a simple flow-based upscaling method, solving −∇ · (K∇p) = 0, with
p = 1 and p = 0 prescribed along the inflow and outflow boundaries respectively, and no-flow
boundary conditions elsewhere.

the flow problem (15.5) numerically to determine fine-scale pressures and flow rates and
corresponding net flow rates and average pressure gradients. We could then invert the flow
equations discretized on the coarse block to determine effective coarse-scale permeabilities
from Darcy’s law. That is, once p and v have been computed, we use the same type of
argument as we used to determine the formulas for the harmonic and arithmetic averages,
except that we now work with discrete pressures and fluxes. This raises the immediate
question of what kind of boundary conditions we should use.
Different boundary conditions have been suggested over the years. One obvious
approach is to use three different boundary conditions for each block to create a pressure
drop in each axial directions in the same way as for the computations reported in Table 15.1;
see Figure 15.6. By specifying sealing boundary conditions along the other boundaries, we
ensure that the effective flow over the block follows the same axial direction as the pressure
drop. This setup emulates how permeability is measured on core samples in the laboratory
and provides us with three pairs of flow rates in 3D that we can use to compute an upscaled
permeability tensor with diagonal elements

vx Lx vy Ly vz Lz
Kxx = − , Kyy = − , Kzz = − .
px py pz

Here, vx is net flux, Lx the characteristic length of the block, and px the pressure drop
inside the block in the x-direction. With this method, the off-diagonal elements are zero
by construction. Strictly speaking, the assumption of sealing boundaries is only valid in
the idealized case when the permeability field is symmetric with respect to the faces of the
coarse grid, i.e., if the coarse block is surrounded by mirror images of itself. It can be shown
that this method tends to introduce an upscaling bias towards low permeability values by
thickening shale barriers and narrowing sand channels [163, 161].
Unless the grid block is located next to a sealing fault or an impermeable layer, it
therefore more natural to assume that the block has open boundaries so that a pressure

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.3 Upscaling Absolute Permeability 571

differential applied along one of the axial directions will induce a net flow also in trans-
verse directions. This means that the effective permeability is a full tensor. One method
to emulate open boundaries is to prescribe a constant pressure on faces perpendicular to
each flow direction and linear pressure drop along the sides that are parallel to the flow
direction [120, 266] so that the flow can leave or enter these sides. As for the sealing
boundary, a unit pressure drop is applied in each of the axial directions. Use of linear
boundary conditions is only strictly valid if the heterogeneous coarse block is embedded
inside a homogeneous medium and tends to produce permeabilities with a bias towards high
values [90].
Another popular option is to prescribe periodic boundary conditions [90], assuming
that the grid block is sampled from a periodic medium so that the fluxes in and out of
opposite boundaries are equal. In other words, to compute the x-component, we impose
the following conditions:

p(Lx ,y,z) = p(0,y,z) − p, v(Lx ,y,z) = v(0,y,z),

p(x,Ly ,z) = p(x,0,z), v(x,Ly ,z) = v(x,0,z), (15.12)
p(x,y,Lz ) = p(x,y,0), v(x,y,Lz ) = v(x,y,0),

and similarly for the other axial directions. This approach gives a symmetric and positive-
definite tensor and is usually more robust than specifying sealing boundaries. Periodic
boundaries tend to give permeabilities that lie in between the lower and upper bounds
computed using sealing and linear boundaries, respectively.
Let us see how two of these methods can be implemented in MRST for a rectilinear or
corner-point grid. For simplicity, we assume that the d-dimensional domain to be upscaled
is rectangular and represented by grid G and rock structure rock. We start by setting up
structures representing boundary conditions,
bcsides = {'XMin', 'XMax'; 'YMin', 'YMax'; 'ZMin', 'ZMax'};
for j = 1:d;
bcl{j} = pside([], G, bcsides{j, 1}, 0);
bcr{j} = pside([], G, bcsides{j, 2}, 0);
end
Dp = {4*barsa, 0};
L = max(G.faces.centroids)-min(G.faces.centroids);

The first structure, bcsides, just contains name tags to locale the correct pair of boundary
faces for each flow problem. The bcl and bcr structures are used to store a template of
all the boundary conditions we must use. This is not needed to implement the pressure-
drop method, but these structures are handy when setting up periodic boundary conditions.
Finally, Dp and L contain the prescribed pressure drop and the characteristic length in each
axial direction. With the data structures in place, the loop that does the upscaling for the
pressure-drop method is quite simple:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 572 Upscaling Petrophysical Properties

for i=1:d
bc = addBC([], bcl{i}.face, 'pressure', Dp{1});
bc = addBC(bc, bcr{i}.face, 'pressure', Dp{2});
xr = incompTPFA(initResSol(G, 100*barsa, 1),< G, hT, fluid, 'bc', bc);
v(i) = sum(xr.flux(bcr{i}.face)) / sum(G.faces.areas(bcr{i}.face));
dp(i) = Dp{1}/L(i);
end
K = convertTo(v./dp, milli*darcy);

That is, we loop over the axial directions and for each direction we: (i) specify a pressure
from left to right, (ii) compute pressures and fluxes by solving the resulting Poisson prob-
lem, and (iii) compute average velocity v across the outflow boundary and average pressure
drop d. Finally, we can compute the effective permeability by inverting Darcy’s law. The
implementation just described is a key algorithmic component in steady-state upscaling
of relative permeabilities for multiphase flow and is therefore offered as a separate utility
function upscalePermeabilityFixed in the upscaling module.
Periodic boundary conditions are slightly more involved. We start by calling a routine
that modifies the grid structure so that it represents a periodic domain:
[Gp, bcp] = makePeriodicGridMulti3d(G, bcl, bcr, Dp);
for j=1:d, ofaces{j} = bcp.face(bcp.tags==j); end

Technically, the grid is extended with a set of additional faces that connect cells on opposite
boundaries of the domain so that the topology becomes toroidal. The routine also sets up
an appropriate structure for representing periodic boundary conditions, which we use to
extract the faces across which we are to compute outflow. You can find details of how the
periodic grid is constructed in the source code. With the modified grid in place, the loop for
computing local flow solutions reads:
dp = Dp{1}*eye(d); nbcp = bcp;
for i=1:d
for j=1:d, nbcp.value(bcp.tags==j)=dp(j,i); end
xr = incompTPFA(initResSol(Gp, 100*barsa, 1), Gp, hT, fluid, 'bcp', nbcp);
for j=1:d,
v(j,i) = sum(xr.flux(ofaces{j})) / sum(Gp.faces.areas(ofaces{j}));
end
end

Inside the loop over all axial directions, the first for-loop extracts the correct pressure
drop to be included in the periodic boundary conditions from the diagonal matrix dp; that
is, a pressure drop along the current axial direction and zero pressure drop in the other
directions. Then, we solve the flow problem and compute the average velocity across the
outflow boundaries. Because of the periodic conditions, we may have outflow also across
boundaries that have no associated pressure drop, which is why we generally get a full-
tensor permeability. Outside the loop, we compute the average pressure drop in each axial
direction and invert Darcy’s law to compute the permeability tensor:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.3 Upscaling Absolute Permeability 573

dp = bsxfun(@rdivide, dp, L);

K = convertTo(v/dp, milli*darcy)

This implementation is offered as a utility upscalePermeabilityPeriodic in the

upscaling module. Linear boundaries are not supported in MRST, but can be implemented
by combining elements from the two cases just discussed: using the function pside to set a
pressure distribution on the boundaries parallel to the direction of the applied pressure drop
and solving a small linear system to compute the components of the effective permeability.
Setting up the necessary flow problems is relatively simple when the grid corresponds
to just one coarse block. It is a bit more involved to do this upscaling efficiently for
many blocks at a time. In the upscaling module, we therefore offer a utility function
upscalePerm that computes permeability upscaling using the pressure-drop method.

Example 15.3.3 Figure 15.7 shows two different permeability fields upscaled using flow-
based upscaling with sealing and periodic boundary conditions. For the case to the right,
the correlation in the permeability field is along the y-direction, and both methods compute
the same diagonal upscaled tensor. Using pressure drop along the axial directions with
sealing boundaries still gives a diagonal tensor if we rotate the grid so that the correlation
direction is along the diagonal, whereas the periodic boundary conditions give a full tensor.
By computing the eigenvalue decomposition of this tensor, we find that the upscaled tensor
is diagonal with values 15 and 19.5 if we rotate the axial directions 45 degrees clockwise.
Full source code for this example is given in the permeabilityExample1 script.

Example 15.3.4 Both flow-based methods will by design correctly reproduce flow along
or orthogonal to layered media for the experiment in Table 15.1. We therefore replace the
layered permeability field with one having dipping layers. Moreover, we replace the top-to-
bottom pressure drop by a pressure drop between diagonally opposite corners and sample
from a different part of the SPE 10 model. Figure 15.8 shows the new permeability fields

10 mD

30 mD





17.23 0 17.25 −2.25 15.0 0 15.0 0
Ks = ,Kp = , Ks = ,Kp =
0 17.23 −2.25 17.25 0 20.0 0 20.0

Figure 15.7 Upscaling two isotropic permeability fields using flow-based upscaling with sealing
boundaries (s) or periodic boundary conditions (p).

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 574 Upscaling Petrophysical Properties
Table 15.2 Ratio between flow rate predicted by upscaled/fine-scale
models for three different permeability fields, flow scenarios, and
upscaling methods.

Model Flow pattern Harm-arith Sealing Periodic

Layered East→West 0.78407 1.00000 1.02594

North→South 0.49974 1.00000 1.00000
Corner→Corner 1.04273 1.30565 1.34228

Tarbert East→West 0.86756 1.00000 0.56111

North→South 0.89298 1.00000 0.53880
Corner→Corner 0.00003 0.00027 39.11923

Upper Ness East→West 0.83026 1.00000 0.40197

North→South 0.71283 1.00000 0.28081
Corner→Corner 0.09546 0.62377 2183.26643

1000

100

10

0.1

0.01

Figure 15.8 Three models used to test the accuracy of averaging techniques. From left to right: an
isotropic model with dipping layers and two subsets from the Tarbert and the Upper Ness formation,
respectively, from the SPE 10 data set.

and Table 15.2 reports the results of the experiment. The method with sealing boundary
conditions is exact by design for the first two flow fields, but gives incorrect flow rates
for the diagonal flow, in particular for the Tarbert subsample. With periodic boundary
conditions, we get the exact flux when the flow direction follows regions of constant per-
meability from the north to the south side for the isotropic case. For the other two flow
patterns, the flow directions cross dipping layers and we thus get minor deviations in the
lateral direction and somewhat larger deviation for the diagonal flow. For the anisotropic
subsamples, the periodic conditions generally give less accurate results than the other
two methods, except for the diagonal flow on the Tarbert case. Altogether, this example
illustrates quite well how ill-posed the upscaling problem is. Source code is given in the
permeabilityExample2 script.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.4 Upscaling Transmissibility 575

Computer exercises
15.3.1 The data sets BedModels1 and BedModel2 are examples of fine-scale rock models
developed for use in a workflow that propagates petrophysical properties from the
core scale to the reservoir scale. Use the methods presented in this subsection to
upscale the absolute permeability in these models.
15.3.2 Implement a method for upscaling with linear boundary conditions.
15.3.3 Estimates computed by the averaging and flow-based methods have a certain
ordering, from low to high values: harmonic, harmonic-arithmetic, sealing
boundaries, periodic boundaries, linear boundaries, arithmetic-harmonic, and
arithmetic. Make a small test suite of heterogeneity permeability fields and verify
this claim.
15.3.4 For grids that are not K-orthogonal, we know that the TPFA method will compute
incorrect fine-scale flow solutions and that it generally is better to use a consistent
discretization. Go back to Example 15.3.3, perturb the grid (e.g., with twister)
and compare upscaled solutions computed by TPFA and the mimetic or MPFA
method. (At the time of writing, the consistent methods do not yet implement
periodic boundary conditions.)

15.4 Upscaling Transmissibility

In the discrete case, the choice of an appropriate upscaling method depends on the numer-
ical stencil to be used for the spatial discretization of the upscaled model. In previous
chapters we have seen that the two-point finite-volume method is the method of choice in
reservoir simulation. When using this method, we only need grid-block permeabilities to
compute transmissibilities between neighboring coarse blocks. It would therefore be more
convenient if, instead of computing an effective permeability tensor, we could compute the
coarse transmissibilities associated with the interface between pairs of neighboring coarse
blocks directly. These transmissibilities should be defined so that they reproduce fine-scale
flow fields in an averaged sense. That is, instead of upscaled block-homogenized tensors
K∗ , we seek block transmissibilities Tij∗ satisfying
   
1 1
vij = Tij∗ p d x − p d x , (15.13)
|i | i |j | j
'
where vij = − ij (K∇p) · n dν is the total Darcy flux across ij .
We can compute the upscaled transmissibilities Tij∗ much the same way we computed
upscaled permeabilities in Figure 15.9. Here, we use a pressure drop to drive a flow across
the interface ij between two coarse blocks i and j . Thus, by solving (15.5) in the
two-block domain i ∪ j subject to suitable boundary conditions, we can compute the
average pressures Pi and Pj in i and j and then obtain Tij∗ directly from the formula
vij = Tij∗ (Pi − Pj ). (15.14)

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 576 Upscaling Petrophysical Properties

i j

Figure 15.9 Illustration of flow-based upscaling of transmissibility.

Let us see how this upscaling can be implemented in MRST. That is, we discuss how to
upscale the transmissibility associated with a single interface between two grid blocks. To
this end, we assume a fine grid G and a coarse partition vector q that subdivides the grid
into two coarse blocks in the x-direction. We start by constructing a coarse-grid structure
CG, as introduced in Section 14.2. From the data members in this structure we can find all
faces in the fine grid that lie on the interface between the coarse blocks and determine the
correct sign to apply to the coarse-scale flux
i = find(any(CG.faces.neighbors==0,2));
faces = CG.faces.fconn(CG.faces.connPos(i):CG.faces.connPos(i+1)-1);
sgn = 2*(CG.faces.neighbors(i, 1) == 1) - 1;

As we saw in Section 15.3.2, there are different ways we can set up a localized flow problem
that gives a flux across the interface between the two blocks. Assuming that the coarse
interface is more or less orthogonal to the x-axis, we use a pressure drop in this direction
and no-flow boundary conditions on the other sides, as shown in Figure 15.9:
bc = pside([], G, 'XMin', Dp(1));
bc = pside(bc, G, 'XMax', Dp(2));
xr = incompTPFA(initResSol(G, 100*barsa, 1), G, hT, fluid, 'bc', bc);
flux = sgn * sum(xr.flux(faces));
mu = fluid.properties();

All that now remains is to compute pressure values Pi and Pj associated with each coarse
block before we use (15.14) to compute the effective transmissibility. Following (15.13),
Pi and Pj are defined as the average pressure inside each block
P = accumarray(q,xr.pressure)./accumarray(q,1);
T = mu*flux/(P(1) - P(2));

Here, we have assumed that all grid cells have the same size. If not, we need to weight the
pressure average by cell volumes. As an alternative, we could also have used the pressure
value at the block centroids, which would give a (slightly) different transmissibility:
cells = findEnclosingCell(G,CG.cells.centroids);
P = xr.pressure(cells);

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.4 Upscaling Transmissibility 577
1 2 3 4 5

−14
x 10
6
Centroid
2.6 Average
2.4

2.2
7
2

1.8

1.6
1 2 3 4 5 6 7

Figure 15.10 Upscaled transmissibility computed on a series of two-block configurations using either
centroid values or averaged values for the block pressures. Black dots show the location of the block
centroids from which pressure values are sampled.

Example 15.4.1 In Figure 15.10 we have used the code just discussed to compute the
transmissibility between two blocks covering a rectangular domain [0,200] × [0,100] m2 .
We gradually rotate the interface between the two coarse blocks so that it goes from vertical
towards being horizontal. Complete source code is given in the script
transmissibilityExample1 in the book module. For the first four pairs of coarse blocks
shown in Figure 15.10, there is only a slight difference between the transmissibilities
computed using pressure values defined as block averages and pressures defined by
sampling at the block centroids. The transmissibilities start to deviate as the block interface
approaches the diagonal.
A pressure drop along the x-axis is a reasonable drive mechanism for the first block
pairs, but it is highly questionable for the last two block pairs. Let us, for instance, con-
sider block pair number seven. If we instead apply the pressure drop along the y-axis
to create a flow that is more perpendicular to the coarse interface, the transmissibil-
ity computed with block-averaged pressures changes from 2.0e-14 to 8.1e-14, whereas
the value is 8.9e-14 with a pressure drop along both axial directions. (The latter setup
is generally not well-posed, as it will have singularities at the southeast and northwest
corners.) Depending upon what type of flow conditions the block pair is subject to in
subsequent simulations, it might be better to set a pressure drop orthogonal to the coarse-
block interface or use one of the oversampling methods that will be introduced in the next
section.

An important property of the upscaled transmissibilities Tij∗ is that they all should
be positive across all interfaces that can transmit fluids, since this will ensure that the

TPFA scheme defined by j Tij∗ (pi − pj ) = qi reproduces the net grid block pressures
'
p = |1 |  p d x, and hence also the coarse fluxes vij . Unfortunately, there is no
guarantee that the transmissibilities defined by (15.13) are positive, for instance, if the
blocks are such that the chord between the cell centroids does not cross the coarse

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 578 Upscaling Petrophysical Properties

interface. Negative transmissibilities may also occur for regular block shapes if the
permeability is sufficiently heterogeneous (e.g., as in the SPE 10 model). Even worse,
unique transmissibility values may not even exist, since the upscaling problem gen-
erally is not well-posed; see [318] for a more thorough discussion of existence and
uniqueness. To guarantee that the resulting coarse-scale discretization is stable, you
should ensure that the Tij∗ values are positive. A typical trick of the trade is to set
Tij = max(Tij ,0), or in other words, just ignore ill-formed connections between neigh-
boring grid blocks. This is generally not a satisfactory solution, and you should there-
fore either try to change the grid, use different pressure points to define the trans-
missibility, or apply some kind of fallback strategy that changes the upscaling method
locally.
The upscaling module does not contain any simple implementation of transmissibility
upscaling like the one outlined in this section. Instead, the module offers a routine,
upscaleTrans, that is designed for the general case where both the coarse and the fine
grids can be fully unstructured and have faces that do not necessarily align with the
axial direction. To provide robust upscaling for a wide range of geological models,
the routine relies on a more comprehensive approach that will be outlined briefly in
the next section, and also has fallback strategies to reduce the number of negative
permeabilities.

15.5 Global and Local–Global Upscaling

The methods described so far in this chapter have all be local in nature. Averaging meth-
ods derive upscaled quantities solely from the local heterogeneous structures, whereas
flow-based methods try to account for flow responses by solving local flow problems
with prescribed boundary conditions. These boundary conditions are the main factor that
limits the accuracy of flow-based methods. In Table 15.2, we saw how neither of the
local methods were able to accurately capture the correct net flow for the anisotropic
Tarbert and Upper Ness samples with boundary conditions giving diagonal flow. The main
problem for flow-based techniques is, of course, that we do not know a priori the pre-
cise flow that will occur in a given region of the reservoir during a subsequent simu-
lation. Thus, it is generally not possible to specify the appropriate boundary conditions
for the local flow problems in a unique manner unless we already have solved the flow
problem.
To improve the accuracy of the upscaling, you can use a so-called oversampling tech-
nique, which is sometimes referred to as an overlapping method. In oversampling methods,
the domain of the local flow problem is enlarged with a border region that surrounds each
(pair of) grid block(s) and boundary conditions, or other mechanisms for driving flow,
like wells or source terms, are specified in the region outside the domain you wish to
upscale. Local flow problems are then solved in the whole enlarged region, but the effective
permeability tensor is only computed inside the original coarse block. This way, you lessen

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.5 Global and Local–Global Upscaling 579

Figure 15.11 Illustration of oversampling techniques for upscaling. The size of the grid block is
exaggerated in the right plot.

the impact of the localization assumption and the particular mechanism used to drive flow,
and the flow across the boundaries of the coarse block is to a larger extent determined by
the heterogeneity in the local region surrounding the block. The motivation for using an
oversampling technique is to better account for permeability trends that are not aligned
with the grid directions and possible large-scale connectivity in the permeability fields.
To also account for global flow patterns, you can solve the pressure equation once,
or a couple of times, on the full geological model and impose fluxes sampled from the
global flow solution as boundary conditions on the local flow problem used to upscale
permeabilities or transmissibilities. In these so-called global upscaling methods, the pres-
sure solution can either be computed using representative and specific drive mechanisms.
Alternatively, you can compute the pressure solution using a set of generic conditions that,
e.g., specify a flow through the reservoir from east to west, from north to south, and so on.
From a computational point of view, this approach may seem to contradict the purpose of
upscaling, but can be justified for numerical simulation of compressible and/or multiphase
flow. Indeed, for such transient flows, the pressure equation must be solved multiple times
throughout the simulation, and the cost of solving the global pressure equation once (or
even a few times) on a fine grid will in most cases be small compared with the cost of
solving the full multiphase flow problem.
Computing a global pressure solution on the fine grid can be avoided if we use a so-
called local–global method [64, 63, 117]. In these methods, pressure solutions obtained by
solving global flow problems on a coarse grid are used to set boundary conditions for local
flow problems that determine the upscaled transmissibilities. Since the initial coarse-grid
computation may give quite poor boundary conditions for the local flow problems, these
methods use a iterative bootstrapping procedure to gradually ensure consistency between
the local and the global calculations.
Global and local–global upscaling methods may suffer from negative or anomalously
large transmissibility values. To avoid these, Holden and Nielsen [133], who were among
the first to study global upscaling methods, proposed an iterative process involving the
solution of an optimization problem that gradually perturbs the transmissibilities until
they fall within a prescribed interval. A similar technique that avoids the solution of the

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 580 Upscaling Petrophysical Properties

fine-grid problem was presented in [225]. Chen and Durlofsky [63] observed that unphys-
ical transmissibilities occur mainly in low-flow regions. Therefore, instead of using an
optimization procedure that alters all transmissibilities in the reservoir, they proposed to use
a thresholding procedure in which negative and very large transmissibilities are replaced
by transmissibilities computed by a local method. Because the transmissibilities are altered
only in low-flow regions, the perturbation will have limited impact on the total flow through
the reservoir.
The script upscaleTrans offers various implementations of global transmissibility
upscaling using either specific or generic mechanisms to drive flow in the full reservoir
model. This specific approach enables you to tailor the coarse-scale model and achieve
very good upscaling accuracy for a particular simulation setup. The generic approach
gives you a more robust upscaling that is less accurate for a specific simulation setup, but
need not be recomputed if you later wish to simulate setups with significant changes in
the well pattern, aquifer support, and other factors that affect the global flow patterns in
the reservoir. Local–global methods are not yet supported in MRST. Instead, the software
offers various multiscale methods [95], including multiscale mixed finite elements [6, 243],
multiscale finite volumes [211, 214], and the multiscale restriction-smoothed basis method
[216, 215]. These methods offer an alternative approach, in which the impact of fine-
scale heterogeneity is included more directly into coarse-scale flow equations as a set of
basis functions that not only represent the average effect but also incorporate small-scale
variation. The methods also give a natural way to reconstruct fine-scale pressure and
mass-conservative flux approximations.

Computer exercises

15.5.1 Extend the computations reported in Table 15.2 to include oversampling methods.
Does this improve the accuracy of the two flow-based methods?
15.5.2 Implement a function that performs generic global upscaling for uniform partitions
of rectilinear and curvilinear grids. To drive the global flow pattern, you can use a
pressure drop in each of the axial directions. How would you extend your method
to more general grids and partitions?
15.5.3 Extend the upscaling module to include local–global upscaling methods.

15.6 Upscaling Examples

In this section we go through three examples that apply the methods discussed earlier in the
chapter to upscale reservoir models. In doing so, we also introduce you to a simple form of
flow diagnostics [274, 218] you can use to assess the accuracy of single-phase upscaling.
The last example represents a complete workflow, from geological surfaces to grid, via
fine-scale simulation to upscaled model. We end the section with a set of general advice
and simple guidelines.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.6 Upscaling Examples 581

15.6.1 Flow Diagnostics Quality Measure

A challenge with upscaling is to predict upfront whether the upscaling will be accurate or
not. To partially answer this question and give an indication of the quality of the upscaling,
we suggest to compare the cumulative well-allocation factors computed by the fine-scale
and upscaled models. As shown in Section 13.1.3, a well-allocation factor is the percentage
of the flux in or out of a completion that can be attributed to a pair of injection and produc-
tion wells. This information is similar to what is obtained by a production logging tool.
To compute these factors, we need to first compute a global, single-phase pressure
solution for the specific well pattern we want to investigate. From the resulting flow field
we compute influence regions that subdivide the reservoir into subvolumes associated with
pairs of injection and production wells, as discussed in Section 13.1. Global methods tend
to compute at least one fine-scale pressure solution as part of the upscaling procedure, but
if a global solution is not available for the specific well pattern, it is generally not very
expensive to compute compared with the cost of the flow-based upscaling procedure. The
same goes for the computation of a coarse-scale flow field and partition functions on the
fine and the upscaled model.
To improve the predictive power of well-allocation factors, you may have to subdivide
each well into two or more segments that each consists of a connected set of completed
cells. This way, you can measure the communication between different parts of the wells,
which is particularly important if the model to be upscaled includes layers or geological
objects with significantly different permeabilities. You can obviously also investigate to
what extent dynamic heterogeneity measures, sweep/drainage regions, and time-of-flights
are preserved, or compare the fine and coarse fluxes to assess the accuracy of the upscaling.

15.6.2 A Model with Two Facies

We start by considering a small rectangular reservoir containing two facies with contrasting
petrophysical properties. The reservoir is produced by an injector and a producer placed
diagonally opposite each other and completed mainly in the high-permeability parts of the
reservoir. Both wells operate at a fixed rate that amounts to the injection/production of
0.2 pore volumes per year. We neglect gravity forces to avoid potential complications from
cross-flow. The fine grid consists of 40×20×15 cells, which we seek to upscale to a coarse
5 × 5 × 15 model. That is, we upscale by a factor 8 × 4 in the lateral direction and leave the
vertical layers in an attempt to preserve the vertical communication in the model as accu-
rately as possible. Figure 15.12 shows the porosities of the fine and the coarse model. To
also upscale permeability, we use upscalePerm, which implements the flow-based method
from Section 15.3.2 with sealing boundary conditions. The script upscalingExample1 in
the book module contains complete source code.
To get an indication of the accuracy of this upscaling, we compare the volumet-
ric connections between the injector and producer computed for the original and the
upscaled permeability. (The latter is prolongated back onto the fine grid.) With a single

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 582 Upscaling Petrophysical Properties

Figure 15.12 Porosity distribution in a model with two different rock types before and after upscaling.

Figure 15.13 Volumetric subdivision of flooded/drainage regions by majority vote. The left plot
shows the regions flooded by the upper and lower parts of the injector in blue and cyan, respectively.
The right plot shows the drainage regions of the upper and lower half or the producer in yellow
and red, respectively. In both plots, each cell is assigned to the completion that has the highest
concentration value.

injector–producer pair that operates at fixed rates, the only difference we can expect
to see is variation in the distribution of flow rates within each well. To better measure
the communication within the reservoir, we therefore subdivide each well into an upper
and a lower segments, so that we altogether have four well-pairs, whose well-allocation
factors can be used to verify the quality of the upscaling. Figure 15.13 shows the influence
regions of the fine-scale model for each of the four segments, defined as regions with
“color concentration” larger than 0.5. A good upscaling method should preserve these
communication volumes as accurately as possible in the coarse model.
Volumes are not always easy to compare in complex 3D models. Instead, we com-
pare cumulative well-allocation factors. The plots in Figure 15.14 show bar charts of the
cumulative flux in/out of the completions that make up each segment, from bottom to top.
Here, each segment is assigned the same unique color as in Figure 15.13 and each bar is
subdivided into the fraction of the total influx/outflux that belongs to the different well-
pairs the segment is part of; see Section 13.1.3. To compare the fine-scale and coarse-scale

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.6 Upscaling Examples 583

permeability, K and K∗ , we use colors for well allocation computed with the coarse model,
whereas solid lines show the same quantities computed for the fine-scale model. The closer
the color bars and the solid lines are, the better the upscaling.
To explain how this type of flow diagnostics should be interpreted, let us look at the
lower-left bar chart, which reports the allocation factors for the upper segment of the
producer (P:1, yellow color). Here, blue color signifies inflow that can be attributed to
injection from the upper segment I:1 of the injector, whereas cyan color signifies inflow
attributed to the lower segment I:2. Because the colored bars are mostly blue, the inflow
into the upper producer segment is predominantly associated with the I:1–P:1 pair. In
other words, P:1 is mainly supported by the upper segment of the injector and is hardly
connected to the lower segment. You can confirm this by looking at the upper-right bar
chart, which shows that the flow from the lower injector segment mainly contributes to
inflow in the lower segment of the producer. The flooded volumes shown to the left in
Figure 15.13 show that fluid injected from I:2 will mainly sweep the volume surrounding
the lower three cells of P:2. Similarly, the right plot shows that the drainage region of P:1
only engulfs the upper parts of I:1. The flooded regions of I:1, on the other hand, engulfs
both P:1 and the upper parts of P:2, and hence the upper injector segment will contribute
flux to both segments of the producer, as shown in the upper-left chart of Figure 15.14.
Likewise, the drainage region from P:2 engulfs both I:2 and the lower part of I:1 and hence
is supported by flux from both the injector segments, as shown in the lower-right chart of
Figure 15.14.

I:1 I:2

P:1 P:2

Figure 15.14 Normalized cumulative well-allocation factors for the four well completions in the
two-facies model. The color bars show well-allocation factors computed with K∗ (upscaled by
upscalePerm) on the fine grid, while the black lines show the same factors computed with the
original permeability K. The colored bars and the black lines should coincide if the upscaled
permeability reproduces the flux allocation correctly.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 584 Upscaling Petrophysical Properties

So, what does this flow diagnostics tell us about our upscaled permeability? First of
all, K∗ does not fully reproduce the volumetric connections of the original permeability
K. Looking at well-allocation factors for the producers in the bottom row of Figure 15.14,
we see that K∗ predicts too much flux in the upper segment (P:1, yellow color, left chart)
since the colorbars exceed the solid lines at the top. Likewise, in the lower segment of
the producer (P:2, red color, right chart), we get too low influx. Altogether, the allocation
factors seem to suggest that the main problems are that with K∗ we miss some of the vertical
communication and exaggerate the connection between I:1 and P:1. We could have used
a single number to measure the overall discrepancy in flux allocation, but in my opinion
you will benefit more from a graphical presentation like this for problems with only a few
wells, because it may help you identify the cause of the mismatch between the upscaled
and the original permeability. (Lorenz coefficients for K and K∗ are 0.2329 and 0.2039,
respectively.)
To develop a better upscaling strategy, we can upscale transmissibilities and well indices
using a global method. For a given flow scenario, like the one with fixed well positions and
constant well controls, you can determine a set of transmissibilities and well indices that
reproduce a single pressure step exactly. This may seem desirable, but has the disadvantage
that the linear system becomes negative definite, which in turn may lead to unphysical
solutions away from the scenario used for upscaling. To span a reasonable set of global
flow directions, we instead use generic boundary conditions that pressurize the reservoir
from east to west, from north to south, and from top to bottom. We can then combine these
three global flow fields in several different ways to compute a transmissibility value for
each coarse face; see e.g., [152]. Here, we use a simple strategy suggested by [172]: We
pick the one among the three flow field that gives the largest flux across a given coarse
face and insert this into (15.14) to compute the transmissibility associated with the face.
A rationale for this choice is that the accuracy of the coarse model is most sensitive to the
choice of transmissibilities in high-flow regions.
We upscale well indices using a coarse-scale version of the inflow-performance relation-
ship (see (4.27) on page 122)
Ji∗ = Qi /(Pi − pw ), (15.15)
where Qi is the sum of the rates of all the perforations inside block i, Pi is the average
block pressure, and pw is the pressure inside the wellbore. To obtain representative flow
fields for n wells, upscaleTrans solves n−1 pressure equations. Each solution is computed
by setting a positive pressure in one well and zero pressure in all the other. We can then
combine the flow fields using the same strategy as for the transmissibilities by picking the
solution that gives the largest well rate, or as we will do here, add all well rates and divide
by the sum of the corresponding pressure drops. (For models with many wells, we would
rather upscale well indices using local or specific approaches.)
Unfortunately, there is no guarantee that the upscaled transmissibilities are nonnegative.
To avoid creating an ill-conditioned discretization matrix, we therefore set all negative
transmissibilities to zero, thereby blocking the corresponding coarse face for flow. This

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.6 Upscaling Examples 585

I:1 I:2

P:1 P:2

Figure 15.15 Normalized cumulative well-allocation factors for the four well completions in the two-
facies model. The color bars show well-allocation factors computed on a coarse model upscaled using
upscaleTrans, while the black lines show the same factors computed on the original model.

is in our experience an acceptable solution as long as the upscaled transmissibilities are

only intended for a specific scenario, like herein, since negative transmissibilities tend to
appear because flow is tangential to the coarse interface or as a result of numerical noise
in regions with very low flow. If your upscaling, on the other hand, aims to serve a more
general purpose with varying well patterns and/or boundary conditions, you would need to
implement a more sophisticated fallback strategy that recomputes negative transmissibil-
ities by an alternative method like a local flow-based method, an averaging method, or a
combination of these.
Figure 15.15 shows the resulting match in well-allocation factors, which is significantly
better than when using a simple permeability upscaling. In part, this is a result of improved
transmissibilities, and in part because we now have upscaled the well indices.

15.6.3 SPE 10 with Six Wells

You have already encountered Model 2 from the 10th SPE Comparative Solution Project
multiple times throughout the book. Because of its simple grid and strong heterogene-
ity and the fact that it is freely available online, this data set has become a community
benchmark that is used for many different purposes. The original aim of the project was
to “compare upgridding and upscaling approaches and the ability to predict performance
of a waterflood through a million cell geological model,” and in [71] you can read about
the relative merits of the various methods used in the upscaling studies that were submitted
to the project by August 2000. In short, the study showed that the data set is generally
very difficult to upscale accurately with single-phase methods and that the best results are

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 586 Upscaling Petrophysical Properties

P1 P2 P3

P4 I1 I2

Figure 15.16 Model of the Tarbert formation with six wells, two central injectors and four peripheral
producers. The 3D plot shows lateral permeability distribution, whereas the bar charts show
normalized cumulative well-allocation factors.

generally obtained by methods that also account for multiphase effects. Later, however,
several authors have obtained excellent results with local–global and multiscale methods.
In this section, we compare how accurate four different upscaling methods can predict
single-phase flow. That is, we consider two averaging methods (harmonic and harmonic-
arithmetic) and a local flow-based upscaling with sealing boundary conditions for upscaling
permeability, and the global upscaling method discussed at the end of the previous exam-
ple for upscaling transmissibilities and well indices. Unlike in the previous example, we
use maximum flow rate to determine both transmissibilities and well indices. The script
upscalingExample2 contains full details.
We start by considering the Tarbert formation, which is found in the upper 35 layers of
the model. We choose a coarsening factor of 10 × 10 × 3, and to get an even number of cell
layers inside each grid block, we extend the model slightly by repeating the top layer so
that the fine-scale model altogether has 220 × 60 × 36 = 475 200 cells. We also replace the
original five-spot pattern with a pattern consisting of producers in the four corners of the
model and two injectors located near the center of the model; see Figure 15.16. All wells
are controlled by bottom-hole pressure; the producers operate at 200 bar and the producers
at 500 bar.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.6 Upscaling Examples 587

Harmonic Harmonic-arithmetic

20 20

15 15

10 10
P1 P1
P2 P2
5 P3 5 P3
P4 P4

0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02

Flow-based, pressure drop Global, flow-based

20 20

15 15

10 10
P1 P1
P2 P2
5 P3 5 P3
P4 P4

0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02

Figure 15.17 Well-allocation factors for injector I2 in the Tarbert model computed on four different
coarse models.

Figure 15.17 compares well-allocation factors predict by the four upscaling methods
for injector I2. This well has significant communication with all four producers and is
hence more difficult to match than injector I1, which mainly communicates with P3 and
P4. Harmonic averaging not surprisingly gives well rates that are far from the fine-scale
targets. As explained in Section 15.3.1, the harmonic mean has a bias against smaller values
and tends to underestimate the effective permeability. Because all wells are controlled by
bottom-hole pressure, too low permeability means too low well rates. If the wells were
controlled by total rate, as in the previous example, the well rates predicted by the harmonic
average would have been more correct, but the predicted pressure buildup around the
injector and the pressure drawdown in the producers would both be too high.
Comparing the local flow-based method to the harmonic-arithmetic averaging method,
we see that the latter is significantly more accurate. By design, the flow-based method
should give accurate prediction of flow along the axial directions of the grid. Here,
however, we have flow that is strongly affected by heterogeneity and mostly goes in the
lateral, diagonal direction and hence the harmonic-arithmetic method seems to be more
suited, in particular if we consider the computational time of the two methods. In the
current implementation, the flow-based method extracts a subgrid and assembles and
inverts three local matrices for each coarse block. The computational overhead of this
procedure is significant since each new local solve incurs full start-up cost. The result is
that the function upscalePerm has orders of magnitude higher computational cost than the
harmonic-arithmetic averaging, which is implemented using a few highly efficient calls to
accumarray.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 588 Upscaling Petrophysical Properties

The global upscaling method reproduces the fine-scale well-allocation factors almost
exactly. The plots may be a bit deceiving because of the areas where the colorbars extend
beyond the solid lines representing the fine-scale well-allocation. However, bear in mind
that the bars and lines represent cumulative factors, and thus we should only look at the
discrepancy at the top of each colorbar. If you look closely, you will see that here the match
is excellent. There are two reasons for this: First of all, with global boundary conditions,
the boundary conditions used to localize the computation of each transmissibility account
better for correlations that extend beyond the block pair and thus set up a flow that is more
representative of the flow the interface will experience in the subsequent flow simulation.
Secondly, the global method includes upscaling of well indices. This can have a significant
impact on the well rates that determine the flux allocation. In passing, we also note that
the computational cost of the global method is significantly lower than that of the local
method, primarily since our implementation utilizes a highly efficient multigrid solver for
the global flow problems and avoids solving a long sequence of small problems that each
have a large start-up cost.
For completeness, we also include the results of a similar study for the whole SPE 10
model; see Figure 15.18. This includes the fluvial Upper Ness formation, which is noto-
riously difficult to upscale accurately. The harmonic-arithmetic and the global method are
no longer as accurate as for the Tarbert formation. For this 1.1 million grid cell model, an
efficient iterative solver like AGMG is indispensable to be able to compute the fine-scale
solution and perform a global upscaling in MATLAB.

Harmonic Harmonic-arithmetic
50 50

40 40

30 30

20 20
P1 P1
P2 P2
10 P3 10 P3
P4 P4

0 0.01 0.02 0.03 0.04 0.05 0 0.01 0.02 0.03 0.04 0.05

Flow-based, pressure drop Global, flow-based

50 50

40 40

30 30

20 20
P1 P1
P2 P2
10 P3 10 P3
P4 P4

0 0.01 0.02 0.03 0.04 0.05 0 0.01 0.02 0.03 0.04 0.05

Figure 15.18 Well-allocation factors for injector I2 in the full SPE 10 model computed on four
different coarse models.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.6 Upscaling Examples 589

15.6.4 Complete Workflow Example

Our last example shows a complete workflow for a synthetic sector model, from setting
up the geocellular model, via compressible flow simulations, to coarsening/upscaling and
analysis of the results using flow diagnostics. The example can thus be seen as a partial
recap of the content in this book. We therefore include more code details than in the two
previous examples. You can find the complete source code in upscalingExample3 in the
book module.
The reservoir has a layered stratigraphy and contains three faults. The reservoir fluids
form a slightly compressible two-phase oil/water system and hydrocarbons are recovered
by a well pattern consisting of one injector and three producers. The model is mostly
conceptual and is designed to outline and give details of the workflow rather than as
a problem representing a realistic scenario in terms of well locations and fluid physics.
However, unlike many other examples in the book, it is completely self-contained and does
not involve any ECLIPSE files to describe reservoir geometry, wells, fluid properties, and
simulation schedule.
We start by creating the main horizons that deliminate the external and internal geology
of the sector. The top surface is composed of three different trends: a large-scale anticline
structure in the form of an elongated arc-like shape, a set of medium-scale sinusoidal folds,
and a set of small-scale random perturbations. For simplicity, we use the same surface,
shifted by a constant factor, to deliminate the reservoir downward. In between, we introduce
two tilted surfaces with extra small-scale random perturbations. These surfaces are clipped
against the top and bottom surface to create erosional effects. The four horizons are shown
in Figure 15.19. Assuming that these surfaces are given in the arrays zt, zmt, zbt, and zb,
we create a basic corner-point grid as follows (see Section 3.3.1):
horizons = {struct('x', x, 'y', y, 'z', zt), ...
struct('x', x, 'y', y, 'z', zmt), ...
struct('x', x, 'y', y, 'z', zmb), ...
struct('x', x, 'y', y, 'z', zb)};
grdecl = convertHorizonsToGrid(horizons, 'dims', [40 40], 'layers', [3 6 3]);

To introduce the faults, we first extract the coordinates for the corner-points and then shift
the z-coordinates in selected parts of the domain:
[X,Y,Z] = buildCornerPtNodes(grdecl);
i=47:80; Z(i,:,:) = Z(i,:,:) + .022*min(0,Y(i,:,:)-550);
j= 1:30; Z(:,j,:) = Z(:,j,:) + .021*min(0,X(:,j,:)-400);
j=57:80; Z(:,j,:) = Z(:,j,:) + .023*min(0,X(:,j,:)-750);
grdecl.ZCORN = Z(:);

The lower-left plot in Figure 15.19 shows the resulting model. The permeability consists
of four layers of lognormally distributed values with average lateral permeability of 100,
400, 10, and 50 md, from top to bottom. The vertical permeability is set to one-tenth of the
lateral permeability.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 590 Upscaling Petrophysical Properties

Geological horizons. Interpolated 40 × 40 × 12 corner-point grid

Introducing four partial faults. Lateral permeability and well pattern.

Figure 15.19 Creating a synthetic sector model whose geology is described by three volumes (two
of which are eroded), three vertical faults that partially penetrate the sector, and four layers of
distinctively different permeabilities. Oil is recovered by three producers supported by one injector.

The three vertical producers are completed in all layers of the model and operate at a
constant bottom-hole pressure of 250 bar. The injector is also perforated in all layers and
operates at a constant rate that would fill one pore volume of water over a period of ten
years. As you probably recall, such wells are set up as follows:
W = verticalWell([], G, rock, 10, 10, [],...
'Name', 'P1', 'comp_i', [1 0], 'Val', 250*barsa, ..
'Type', 'bhp', 'refDepth', 50);

The fluid system is assumed to consist of two phases, an incompressible water phase
and an oil phase with constant compressibility. If we express the oil compressibility as
an expansion factor on the form, bo (p) = b0 exp[c(p − p0 )], we can model the fluid
system as a special case of the general three-phase black-oil model from the AD-OO
framework discussed in Chapter 12. The simplest fluid object in this framework models an
incompressible fluid with constant densities. To get the correct behavior for the oil phase,
we simply replace the function handle in the bo field by a pointer to an anonymous function
that evaluates the exp function with appropriate arguments. We then call the appropriate
constructor from the AD-OO framework:

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.6 Upscaling Examples 591

fluid = initSimpleADIFluid('mu', [1, 5, 0]*centi*poise, ...

'rho', [1000, 700, 0]*kilogram/meter^3, ...
'n', [2, 2, 0]);
fluid.bO = @(p, varargin) exp((p/barsa - 300)*0.001);

gravity reset on
model = TwoPhaseOilWaterModel(G, rock, fluid);

Initially, the reservoir is assumed to be in an equilibrium state with a horizontal oil–water

(OW) contact separating pure oil from pure water. As explained earlier, MRST uses water,
oil, and gas ordering internally, so in this case we have water in the first column and oil in
the second for the saturations in the initial state
region = getInitializationRegionsBlackOil(model, depthOW, ...
'datum_depth', ddepth, 'datum_pressure', p_ref);
state0 = initStateBlackOilAD(model, region);

The initialization is quite simple and does not perform any sub-cell integration to assign
more accurate saturations in cells intersected by the OW contact; see the upper-left plot in
Figure 15.20.
We define a straightforward simulation schedule consisting of five small initial con-
trol steps followed by 25 larger steps. We keep the well controls fixed throughout
the simulation. If dT contains the time step, the schedule is constructed by the call
simpleSchedule(dT, 'W', W). The simulation model has 16,350 grid cells and the
resulting linear systems are a bit too large for MATLAB’s direct solver to be efficient. We
therefore set somewhat stricter tolerances and use a CPR preconditioner with an algebraic
multigrid solver for the elliptic pressure system as discussed in more detail for the SPE 9
benchmark in Section 12.4.4. Figure 15.20 shows how water displaces oil over time inside
the reservoir.
To reduce computational times, we partition the original 40 × 40 × 12 grid uniformly
in index space into a new 20 × 20 × 4 coarse grid. This will generally give blocks having
cells on opposite sides of faults. To ensure that as many as possible of the coarse blocks
are hexahedral (except for those that are partially eroded along the top or bottom), we split
any coarse block intersected by one of the faults. To do this, we introduce a temporary grid
topology in which faults are set as barriers, and then perform a simple postprocessing to
split any disconnected blocks:
cdims = [20, 20, 4];
Gf = makeInternalBoundary(G, find(G.faces.tag > 0));
p = processPartition(Gf, partitionUI(G, cdims));

Altogether, this produces a coarse grid with 1,437 active blocks (see Figure 15.21), which
corresponds to a little more than a 22 times reduction in the number of unknowns. This
should put us well within the range of models for which we safely can use MATLAB’s
default linear solver without the need for a CPR preconditioner.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 592 Upscaling Petrophysical Properties

Initial state After 0.38 yrs

After 6.15 yrs After 10.0 yrs

Figure 15.20 Time evolution of the water saturation for the sector model.

Figure 15.21 A coarse 20 × 20 × 4 partition of the sector model. New blocks arising when we split
blocks penetrated by one of the faults are marked in green. The right plot shows the initial saturation
as computed by upscaleState.

We can now directly upscale the model, schedule, and initial state. By default, the
upscaling routine uses the simplest possible options, i.e., harmonic averaging of perme-
abilities,
modelC1 = upscaleModelTPFA(model, p);
scheduleC1 = upscaleSchedule(modelC1, schedule);
state0C1 = upscaleState(modelC1, model, state0);

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.6 Upscaling Examples 593

We have already seen that power-averaging of permeability is not very accurate, and thus
we only consider this model as a base case. In addition, we use flow-based transmissibility
upscaling with a specific flow field obtained by solving a single-phase flow problem with
the given well pattern. This approach has proved to be the most accurate in the examples
earlier in the section:
[, TC, WC] = upscaleTrans(GC, model.operators.T_all, ...
'Wells', W, 'bc_method', 'wells', 'fix_trans', true);

modelC2 = upscaleModelTPFA(model, p, 'transCoarse', TC);

scheduleC2 = schedule;
for i = 1:numel(scheduleC2.control)
scheduleC2.control(i).W = WC;
end

To get an idea of how accurate the two upscaling methods will be, we can compare well
allocation with the fine model based on the computation of a single pressure step. The
plots in Figure 15.22 confirm that harmonic averaging also in this case underestimates the
coarse-scale permeability and hence will predict too small flux into the three production
wells. Specific transmissibility upscaling, on the other hand, reproduces the correct flux
allocation very accurately and it therefore seems likely that it will produce quite accu-
rate results. Let us now simulate the two cases. This is done as for a standard fine-scale
model by calling the standard simulator interface with the coarsened model, initial data,
and schedule
[wellSolsC1, statesC1] = simulateScheduleAD(state0C1, modelC1, scheduleC1);

Somewhat surprising, the production curves predicted by the specific flow-based upscaling
are generally not much more accurate than those predicted by the simple harmonic

P1 P2 P3 I1

P1 P2 P3 I1

Figure 15.22 Flow diagnostics used as a quality measure for the two upscaling methods on the 20 ×
20 × 4 upscaled sector model with harmonic upscaling of permeability (top) and specific flow-based
transmissibility upscaling (bottom).

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 594 Upscaling Petrophysical Properties
700
350 Well: P3
Well: P2
600 300
P2 (Fine-scale)
500 250
P2 (Harmonic)

m /day
m /day

400 P2 (Flow-based) 200

3
3

300 150

200 100 P3 (Fine-scale)

P3 (Harmonic)
100 50
P3 (Flow-based)
0 0
500 1,000 1,500 2,000 2,500 3,000 3,500 500 1,000 1,500 2,000 2,500 3,000 3,500
Time [Days] Time [Days]

Figure 15.23 Comparison of oil rate (producer P2) and water rate (producer P3) predicted by the
upscaled 20 × 20 × 4 models and the original 40 × 40 × 12 sector model.

permeability upscaling. Figure 15.23 shows the two production curves for which specific
transmissibility upscaling compares least favorably with harmonic upscaling. This illus-
trates an important point: whereas accurate reproduction of single-phase well allocation
often is a necessary condition for accurate upscaling, it is not a sufficient
condition.
The question now is what has gone wrong here? The primary problem is that the
grid blocks in the coarse grid do not adapt to the layering structure in the permeability
field if we use a uniform subdivision of the K index. You may recall we argued that
adapting to permeability layering is important when we discussed CaseB4 in Section
14.3.3. So let us rerun the upscaling exercise with a new basic partition that respects the
layering
p0 = partitionLayers(G, cdims(1:2), L);

Here, L=[1,3,8,11,13] is a run-length encoding of the layer indices; this vector can
be obtained as the second output of the lognormLayers function we used to generate
the permeability field. The resulting coarse partition has exactly the same number of grid
blocks, but gives significantly more accurate predictions for both upscaling methods, as
you can see in Figure 15.24. We also obtain almost the same accuracy if we instead use a
uniform 10 × 10 × 12 partition, which supports our hypothesis that preserving layering is
very important in this particular example.

Computer exercises

15.6.1 How coarse model are you able to make and still predict the oil and water rates
with reasonable accuracy?
15.6.2 Make a fine-grid model that respects the OW contact exactly. How does this affect
the simulation? How would you coarsen and upscale this model?

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 15.6 Upscaling Examples 595

350
600 Well: P2
300 Well: P3
500
250
P2 (Fine-scale)
400

m /day
m3/day

P2 (Harmonic) 200

3
300 P2 (Flow-based)
150

200 100 P3 (Fine-scale)

P3 (Harmonic)
100 50
P3 (Flow-based)
0 0
500 1,000 1,500 2,000 2,500 3,000 3,500 500 1,000 1,500 2,000 2,500 3,000 3,500
Time [Days] Time [Days]

Figure 15.24 Comparison of oil rate (producer P2) and water rate (producer P3) predicted by the
upscaled 20 × 20 × 4 models where the coarse blocks adapt to the permeability layers of the original
40 × 40 × 12 sector model.

15.6.5 General Advice and Simple Guidelines

Through the various examples in this chapter, I have tried (and hopefully succeeded) to
convince you that accurate and robust upscaling is a challenging problem and that there is
no single upscaling method that is unequivocally better than others. This is one motivation
for the big attention so-called multiscale methods [95, 195] have received in recent years.
These methods offer a promise of a more systematic and consistent way of bringing the
impact of small-scale heterogeneity variations into simulations on a coarser scale.
What is the best upscaling method for a specific project will depend on many factors.
The most important factor is probably what you intend to use the upscaled model for. If
you are upscaling for a specific scenario, I strongly recommended that you use a method
employing specific global information as this generally will enable you to perform more
aggressive coarsening. If you, on the other hand, want to use the upscaled model to simulate
a large variety of flow scenarios, or some yet unspecified scenario, you should pick a
technique that is as robust as possible. Which technique to pick will depend on the type
of heterogeneity you are facing, the flow patterns that exist in a reservoir, and so on.
For relatively homogeneous reservoirs with flow patterns that mainly follow the axial
directions, a simple averaging technique may be sufficient, whereas techniques that utilize
some kind of global information are recommended for highly heterogeneous reservoirs
with strong contrasts and large variations in correlation lengths. Other determining factors
include the time and the computer resources you have available for upscaling, the number
of petrophysical realizations you need to upscale, the required upscaling factor, and so on.
Regardless of your situation, I generally recommended that you try different upscaling
and coarsening methods. Start by simple averaging methods that have a low computational
cost, since these not only give you a first quick estimate of the upscaling uncertainty (or
more generally the uncertainty in the petrophysical properties), but will also provide upper
and lower bounds for more sophisticated upscaling methods. Check to see if there are

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 596 Upscaling Petrophysical Properties

features in your model that need to be resolved by adapting the coarse grid, and ensure that
you do not create a coarse grid that gives problems for subsequent flow simulations; see e.g.,
the discussion in Section 14.7. Then you can gradually move towards more sophisticated
upscaling methods. For local methods, you should try both sealing and linear pressure
boundaries to provide lower and upper bounds. You may also need to check the use of
oversampling methods to lessen the impact of boundary conditions used for localization.
At all stages, you should use representative single-phase simulations to validate your
upscaled model against the original fine-scale model. To this end, you should compare flow
fields computed both with TPFA and a consistent method to get an idea of the amount
of grid-orientation effects. Likewise, I strongly recommend that you use various kinds of
flow diagnostics like the well-allocation factors discussed in the examples in this chapter,
partition functions that give you the volumetric connections between inflow and outflow,
time-of-flight that gives time lines for displacement fronts, and so on. Herein, we have
computed these quantities using finite-volume discretizations that are available in MRST,
but these flow diagnostic measures can also be computed by streamline simulation [79],
which is generally a very efficient tool for comparing geomodels and upscaled simulation
models. Whether you choose one or the other is not that important. What is important is
that you stick to using single-phase flow physics to avoid mixing in the effect of multiphase
flow parameters, such as variations in relative permeabilities and capillary pressures, which
need to be upscaled by other means.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Appendix
The MATLAB Reservoir Simulation Toolbox

Practical computer modeling of porous media constitutes an important part of the book
and is presented through a series of examples that are intermingled with more traditional
textbook material. All examples discussed in the book rely on the MATLAB Reservoir
Simulation Toolbox (MRST), which is a free, open-source software that can be used for
any purpose under the GNU General Public License version 3 (GPLv3).
The toolbox is primarily developed by my research group at SINTEF, which is the
fourth-largest contract research organization in Europe. The software started out as a
research code developed to study consistent discretization and multiscale solvers for
incompressible two-phase flow on stratigraphic and unstructured polyhedral grids. Over
the past 6–7 years the software has been applied to research a wide spectrum of other
problems related to reservoir modeling, and, as a result, the software today has many of
the same capabilities that can be found in commercial reservoir simulators. In addition,
it contains a spectrum of new research ideas; some of these were discussed in Chapters
13 and 14. At least for the time being, our main motivation for continuing to maintain
and develop the software is to have a versatile research platform that enables researchers
at SINTEF to rapidly develop proof-of-concept implementations of new ideas and then
subsequently, and rather effortlessly, turn these into software prototypes that could be
verified and validated for problems of industry-standard complexity with regard to flow
physics and geological description of geology. A second purpose is to support the idea of
replicable/reproducible research, as well as to enable us to effectively leverage results from
one research project in another.
Over the years, we have also seen that MRST is an efficient teaching tool and a good
platform for disseminating new ideas. Maintaining and developing the software as a reliable
community code takes a considerable effort, but by carefully documenting and releasing
our research software as free, open source, we hope to contribute to simplifying the experi-
mental programming of other researchers in the field and give a head start to students about
to embark on a master’s or PhD project.

597

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 598 The MATLAB Reservoir Simulation Toolbox

A.1 Getting Started with the Software

This Appendix provides you with a brief overview of the software and the philosophy
underlying its design. I show you how to obtain and install the software and explain its
terms of use, as well as how we recommend that you use the software as a companion to
the textbook. I also briefly discuss how you can use a scripted, numerical programming
environment like MATLAB (or its open-source clone GNU Octave) to increase the produc-
tivity of your experimental programming and share examples of tricks and ways of working
with MATLAB that we have found particularly useful. I end the chapter by introducing you
to automatic differentiation, which is one of the key aspects that make MRST a powerful
tool for rapid prototyping and enable us to write compact and quite self-explanatory codes
that are well suited for pedagogical purposes. As a complement to the material presented in
this chapter, you should also consult the first of two JOLTS (short online learning modules),
developed in collaboration with Stanford University [188]. The JOLT gives a brief overview
of the software, shows the way it looked a few years ago, tells you why and how it was
created, and instructs you how to download and install it on your computer. If you are not
interested in programming at all, you need not read this Appendix. However, if you choose
to not work with the software alongside the textbook material, be warned, you will miss a
lot of valuable insight.

A.1.1 Core Functionality and Add-on Modules

MRST is a research tool whose aim is to support research on modeling and simulation of
flow in porous media. The software contains a wide variety of mathematical models, com-
putational methods, plotting tools, and utility routines that extend MATLAB in the direction
of reservoir simulation. To make the software as flexible as possible, it is organized quite
similar to MATLAB and consists of a collection of core routines and a set of add-on
modules, as illustrated in Figure A.1. The material presented in Part I of the book relies
almost entirely on general core functionality, which includes routines and data structures for
creating and manipulating grids, petrophysical data, and global drive mechanisms such as
gravity, boundary conditions, source terms, and wells. The core functionality also contains
an implementation of automatic differentiation – you write the formulas and specify the
independent variables, the software computes the corresponding derivatives or Jacobians –
based on operator overloading (see Section A.5), as well as a few routines for plotting cell
and face data defined over a grid. The core functionality is considered to be stable and not
expected to change significantly in future releases.
To minimize maintenance costs and increase flexibility, MRST core does not contain
flow equations, discretizations, and solvers; these are implemented in various add-on
modules. If you have read Chapter 1, you have already encountered the incompTPFA
solver from the incomp module in Section 1.4; this module implements fluid behavior
and standard solvers for incompressible, immiscible, single-phase and two-phase flow.
The mathematical models, discretizations, and solution techniques underlying this module

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A.1 Getting Started with the Software 599
Fully implicit

es
Flow diagnostics Discretizations

ul
od
m
o n
d-
Ad

Grid coarsening Multiscale methods

2
MRST core cells.faces = faces.nodes = faces.neighbors =
1 10 1 1 0 2
6
1 8 1 2 2 3
6 1 7 2 1 3 5
5 6 2 1 2 3 5 0
9 2 2 3 1 6 2
2 5 3 4 0 6
4 12 3 3 4 1 1 3
3 3 7 4 7 4 1
3 2 5 2 6 4
8
2 4 8 5 3 1 8
1 5 4 12 6 2 8 5

Upscaling 2
3
7 1
10
7 14
4
5
5
5 11
6
9
3
4

9
6
7
7
8
8
6
3
4
3
5
4
7
0
7
8
7 CO2 lab
8 13 6 6 9 3
Original permeability Upscaled (x−direction) Upscaled (y−direction) 4 6 5 9 6
7 13 10 4
3 11 7 14 10 5 CO2 saturation
5 : : : : at 500 years 12%
16%
1 4 7
56%
12%

3%
Residual (traps)
Residual
Residual (plume)
Injected volume: Movable (traps)
3
2.185e+07 m Movable (plume)
50 Original Height of CO2−column x 10
7 Leaked
Upscaled (x)
14
40 2
12

30 10 1.5

8
20 1
6

4
10 0.5
2

0 0 0
−1.5 −1 −0.5 0 0.5 1 1.5 2 2.5 0 50 100 150 200 250 300 350 400 450 500

Input decks
Visualization

... ...

Figure A.1 MRST consists of core functionality that provides basic data structures and utility
functions, and a set of add-on modules that offer discretizations and solvers, simulators for
incompressible and compressible flow, and various workflow tools such as flow diagnostics, grid
coarsening, upscaling, visualization of simulation output, and so on.

are extensively discussed in Parts II and III of the book. In particular, Chapters 5 and 10
outline key functionality offered in the module and discuss in detail how the solvers are
implemented. The standard solvers are unfortunately not unconditionally consistent and
may therefore exhibit strong grid-orientation errors and fail to converge. To get a convergent
scheme, one can use one of the methods from the mimetic and mpfa modules discussed in
Chapter 6, which offer consistent discretizations on general polyhedral grids [192].
Solvers for incompressible flow have been part of the software since the beginning and
constitute the first family of add-on modules signified by the “Discretizations” block in
Figure A.1. The incomp, mimetic, and mpfa modules are all implemented using a proce-
dural (imperative) programming model from classical MATLAB, i.e., using mathematical
functions that operate mainly on vectors, (sparse) matrices, structures, and a few cell arrays.
These are generally robust and well documented, have remained stable over many years,
and will not likely change significantly in future releases. The family of incompressible
flow solvers also includes a few other discretization methods from the research front like
virtual element methods and are intimately connected with two of the modules implement-
ing multiscale methods (MsFV and MsMFE). These are not discussed herein; you can
find details in the software itself, on the MRST website, or in the many papers using the
software.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 600 The MATLAB Reservoir Simulation Toolbox

The second family of modules consists of simulators based on automatic differentiation

and is illustrated by the “Fully implicit” block in Figure A.1. The introduction of automatic
differentiation has been a big success and has not only enabled efficient development
of black-oil simulators, but also opened up for a unprecedented capabilities for rapid
prototyping [170, 30]. The process of writing new simulators is hugely simplified by the
fact that you no longer need to compute analytic expressions for derivatives and Jacobians.
This is discussed in Chapters 7 and 11. Initially, we implemented solvers based on
automatic differentiation using an imperative programming model similar to the incomp
family of modules. Soon it became obvious that this was a limiting factor, and a new object-
oriented programming model, implementing a general framework for solvers referred to as
MRST AD-OO [170, 212, 30], was introduced. The individual modules and the underlying
implementations of AD-OO underwent significant changes in the period 2014–2016. From
release 2016a and onward, however, the basic functionality has remained largely unchanged
and has mainly been subject to bug fixes, feature enhancements, and performance
improvements.
The ad-core module is the most basic part in MRST AD-OO and does not contain
any complete simulators, but rather implements the common framework used for many
other modules. The design of this framework deviates significantly from that of the incom-
pressible solvers. The main motivation for introducing an object-oriented framework is
to be able to simulate compressible multiphase models of industry-standard complexity.
Chapters 11 and 12 only discuss how to simulate compressible multiphase models of black-
oil type, but the software also offers simulators for compositional flow. All these multiphase
models are significantly more complex to simulate than the basic incompressible models
implemented in the incomp module. Not only are there more equations and more complex
parameters and constitutive relationships, but making robust simulators also necessitates
more sophisticated solution algorithms involving nonlinear solvers, preconditioners, time-
step control, variable switching, etc. Moreover, industry-standard simulations generally
require lots of bells and whistles to implement specific fluid behavior, well models, and
group controls. A robust implementation also requires a number of subtle tricks of the trade
to ensure that your simulator is able to reproduce results of leading commercial simulators.
Object-orientation enables us to divide the implementation into different numerical con-
texts (mathematical model, nonlinear solver, time-step control, linearization, linear solver,
etc.), hide unnecessary details, and only expose the details that are necessary within each
specific context.
The third family of modules consists of tools that can be used as part of the reservoir
modeling workflow. In Part IV, we go through the three types of tools shown in Figure A.1.
“Diagnostics” signifies a family of computational methods for determining volumetric
connections in the reservoir, computing well-allocation factors, measuring dynamic het-
erogeneity, providing simplified recovery estimates, etc. Tools from the “Grid coarsening”
and “Upscaling” modules can be used to develop reduced simulation models with fewer
degrees of freedom and hence lower computational costs. MRST also offers several other

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A.1 Getting Started with the Software 601

modules of the same type, e.g., history matching and production optimization, but these are
outside the scope of the book.
The fourth family of modules consists of computational methods that have been devel-
oped to study a special problem. In Figure A.1, this is exemplified by the “CO2 lab”
module, which is a comprehensive collection of computational methods and modeling tools
developed especially to study the injection and long-term migration of CO2 in large aquifer
systems. Other modules of the same type include solvers for geomechanics and various
modeling frameworks and simulators for fractured media. These modules are all outside
the scope of this book.
The fifth family consists of a variety of utility modules that offer graphical interfaces and
advanced visualization, more comprehensive routines for reading and processing simula-
tion models and other input data, C-acceleration of selected routines from the core module
to avoid computational bottlenecks, etc. You will encounter functionality from several of
these modules throughout the book, but the modules themselves will not be discussed in any
detail. Last, but not least, there are also modules developed by researchers not employed
by SINTEF.

A.1.2 Downloading and Installing

The main parts of MRST are hosted as a collection of software repositories on Bitbucket.
Official releases are provided as self-contained archive files that can be downloaded from
the website: www.mrst.no/
Assume now that you have downloaded the tarball of one of the recent releases; here,
we use release 2016b as an example. Issuing the following command
untar mrst-2016b.tar.gz

in MATLAB creates a new folder mrst-2016b in your current working director that contains
all parts of the software. Once all code has been extracted to some folder, which we
henceforth refer to as the MRST root folder, you must navigate MATLAB there, either
using the built-in file browser, or by using the cd command. Assuming that the files were
extracted to the home folder, this would amount to the following on Linux/Mac OS:
cd /home/username/mrst-2016b/ % on Linux/Mac OS
cd C:\Users\username\mrst-2016b\ % on Windows

To activate the software, you must make sure that the MRST root folder is on MATLAB’s
search path. This is done by use of a startup script, which also scans your installation and
determines which modules you have installed. When you are in the folder that contains the
software, you the software is activated by the command:
startup;

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 602 The MATLAB Reservoir Simulation Toolbox

Figure A.2 The welcome message displayed when the startup script is run in release 2016a and
later. The careful reader may notice that the user runs a development version of the software and not
one of the official releases.

In MRST 2016a and newer, the startup script will display a welcome message showing
that the software is initialized and ready to use; see Figure A.2. The whole procedure
of downloading and installing the software, step by step, can be seen in the first MRST
Jolt [188] (uses release 2014b).
At this point, a word of caution is probably in order. We generally refer to the software
as a toolbox. By this we mean that it is a collection of data structures and routines that can
be used alongside with MATLAB. It is, however, not a toolbox in the same sense as those
purchased from the official vendors of MATLAB. This means, for instance, that MRST is
not automatically loaded unless you start MATLAB from the MRST root folder, or make
this folder your standing folder and manually issue the startup command. Alternatively, if
you do not want to navigate to the root folder, for instance in an automated script, you can
call startup directly
run /home/username/mrst-2016b/startup % or C:\MyPath\mrst-2016b\startup

In versions prior to MRST 2016a, the startup script only sets up the global search path
so that MATLAB is able to locate MRST’s core functionality and the various modules. To
verify that the software is working, you can run the simple example discussed in Section 1.4
by typing flowSolverTutorial1. This should produce the same plot as in Figure 1.4.

A.1.3 Exploring Functionality and Getting Help

The welcome message shown in Figure A.2 contains links to a number of functions that
are useful if you want to get more acquainted with the software. Upon your first encounter,

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A.1 Getting Started with the Software 603

Figure A.3 Illustration of the MATLAB workbook concept. The editor window shows part of the
source code for the simpleBC tutorial from release 2014b. (In newer versions, this tutorial has been
renamed incompIntro and moved to the new incomp module.) Notice how cells are separated by
horizontal lines and how each cell has a header and a text that describes what the cell does. The
exception is the first cell, which summarizes the content of the whole tutorial. The right window
shows the result of publishing the workbook as a webpage.

the best point to start is by typing (or clicking the corresponding blue text in the welcome
message)
mrstExamples()

This will list all introductory examples found in MRST core. Some of these examples
introduce you to basic functionality, whereas others highlight capabilities found in various
add-on modules that implement specific discretizations, solvers, and workflow tools. These
examples are designed using cell-mode scripts, which can be seen as a type of “MATLAB
workbook” that enables you to break a script into smaller code sections (code cells) that can
be run individually to perform a specific task such as creating parts of a model or making
an illustrative plot; see Figure A.3 for an illustration.
In my opinion, the best way to understand tutorial examples is to go through the corre-
sponding scripts, evaluating one section at the time. Alternatively, you can set a breakpoint
on the first line and step through the script in debug mode, e.g., as shown in the fourth video
of the first MRST Jolt [188]. Some of the example scripts contain quite a lot of text and are
designed to be published as HTML documents, as shown to the right in Figure A.3. If you
are not familiar with cell-mode scripts or debug mode, I strongly urge you to learn these
useful features in MATLAB as soon as possible.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 604 The MATLAB Reservoir Simulation Toolbox

>> help computeTrans

Compute transmissibilities.

SYNOPSIS:
T = computeTrans(G, rock)
T = computeTrans(G, rock, ’pn’, pv, ...)

PARAMETERS:
G - Grid structure as described by grid_structure.

rock - Rock data structure with valid field ’perm’. The permeability
is assumed to be in measured in units of metres squared (m^2).
Use function ’darcy’ to convert from darcies to m^2, e.g.,
perm = convertFrom(perm, milli*darcy)
if the permeability is provided in units of millidarcies.
:
:
RETURNS:
T - half-transmissibilities for each local face of each grid cell
in the grid. The number of half-transmissibilities equals
the number of rows in G.cells.faces.

COMMENTS:
PLEASE NOTE: Face normals are assumed to have length equal to
the corresponding face areas. ..

SEE ALSO:
computeGeometry, computeMimeticIP, darcy, permTensor.

Figure A.4 Most functions in MRST are documented in a standard format that gives a one-line
summary of what the function does, specifies the synopsis (e.g., how the function should be called),
explains input and output parameters, and points to related functions.

To find the syntax and understand what a specific function does, you can type
help computeTrans

which will bring up the documentation for computeTrans shown in Figure A.4. All core
functionality is well documented in a format that follows the MATLAB standard. Func-
tionality you are meant to utilize from any of the add-on modules should also, as a rule,
be well documented. However, the format and quality of documentation may differ more,
depending upon who developed the module and how mature and widely used it is. Starting
with release 2017b, all documentation has been formatted so that it can be auto-extracted
with Sphinx, e.g., as HTML posted on the MRST website (see Figure A.5).
As a general rule, all modules distributed as part of the MRST release are required to
contain worked tutorials highlighting key functionality that most users should understand.
A subset of these tutorials is also available on the MRST website. To list tutorial examples
found in individual modules, you can use the mrstExamples command, which can also list
all the tutorial examples the software offers,
mrstExamples('ad-blackoil') % all examples in the black-oil module
mrstExamples('all') % all examples across all available modules

To learn more about the different modules and get a full overview of all available function-
ality, you can type the command

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A.1 Getting Started with the Software 605

Figure A.5 Auto-generated documentation of all function, as well as all classes from the AD-OO
framework, can also be found on the MRST website.

mrstExploreModules()

This brings up a graphical user interface that lists all accessible modules and outlines their
purpose and functionality, including a short description of all tutorial examples found in
each module, as well as a list of relevant scientific publications, with possibility to view
online versions and export citations to BibTeX. Section A.3 gives a brief description of the
majority of the modules.
MRST also offers simplified access to a number of public data sets. You can use the
following graphical user interface to list all data sets that are known to MRST
mrstDatasetGUI()

The GUI briefly describes each data set, provides functionality for downloading and
installing it in a standard location, lists the files it contains as well as the tutorial examples
in which it is used. More details are given in Section A.2.
Last, but not least, a number of common queries have been listed on MRST’s FAQ
page: www.sintef.no/projectweb/mrst/faq/ For further assistance and discussions you
may visit the user forum (www.sintef.no/projectweb/mrst/forum/) and/or subscribe to
the mailing list ([email protected]). As you get more experience with the
software, I encourage you to help others out by answering questions and contributing to
discussions.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 606 The MATLAB Reservoir Simulation Toolbox

A.1.4 Release Policy and Version Numbers

Over the last few years, key parts of the software have become relatively mature and
well tested. This has enabled a stable biannual release policy with one release in the
spring and one in the fall. The version number of MRST refers to the biannual release
schedule and does not imply a direct compatibility with the same release number for
MATLAB. That is, you do not need to use MRST 2016b if you are using MATLAB
R2016b, or vice versa, you do not need to upgrade your MATLAB to R2019a to use MRST
2019a.
Throughout the releases, basic functionality like grid structures has remained largely
unchanged, except for occasional and inevitable bug fixes, and our primary focus has
been on expanding functionality by maturing and releasing in-house prototype modules.
Fundamental changes will nevertheless occur from time to time, e.g., like when automatic
differentiation was introduced in 2012 and when its successor, the AD-OO framework,
was introduced in 2014b. Likewise, parts of the software may sometimes be reorganized,
like when the basic incompressible solvers were taken out of the core functionality and
put in a separate module in 2015a. In writing this, I (regretfully) acknowledge the fact
that specific code details and examples in books describing evolving software tend to
become somewhat outdated. To countermand this, complete codes for almost all examples
presented in the book are contained in a separate book module that accompanies 2015a
and later releases. These are part of the prerelease test suite and should thus always be up
to date.

A.1.5 Software Requirements and Backward Compatibility

MRST was originally implemented so that the minimum requirement should be MATLAB
version 7.4 (R2007a). However, certain parts of the software use features that were not
present in R2007a:

• The AD libraries use new-style classes (classdef) introduced in R2008a.

• Various scripts use new syntax for random numbers introduced in R2007b.
• Use of the tilde operator to ignore return values (e.g., [,i]=max(X,1)) was introduced
in R2009b.
• Some routines, like the fully implicit simulators for black-oil models, rely on access-
ing sub-blocks of large sparse matrices. Although these routines will run on any
version from R2007a and onward, they may not be efficient on versions older than
R2011b.
When it comes to visualization, things are a bit more complicated, since MATLAB 3D
graphics does not behave exactly the same on all platforms. Moreover, MATLAB intro-
duced new handle graphics in R2014b, which has been criticized by many because it
is slow and because it breaks backward compatibility. We have tried to revise plotting
in newer versions of MRST so that it works well both with the old and the new handle

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A.1 Getting Started with the Software 607

graphics, but every now and then you may stumble across certain tricks (e.g., setting grid
lines semitransparent to make them thinner) that may not work well for your particular
MATLAB version.
MRST can also be used to a certain extent with recent versions of GNU Octave, which is
an open-source scientific programming language that is largely compatible with MATLAB.
The procedural parts should work more or less out of the box, but you may encounter
some problems with some of the (3D) plotting, which works a bit differently in GNU
Octave. The 2017b release adds preliminary support also for the object-oriented AD-OO
framework through the octave module. Unfortunately, the performance seems to be orders
of magnitude behind recent versions of MATLAB, but this will hopefully improve in the
future. The only parts of MRST you cannot expect to work are graphical user interfaces;
their implementation is fundamentally different between MATLAB and GNU Octave.
MRST is designed to only use standard MATLAB. Nevertheless, we have found a few
third-party packages and libraries to be quite useful:

• MATLAB-BGL: MATLAB does not yet have extensive support for graph algorithms. The
MATLAB Boost Graph Library contains binaries for useful graph-traversal algorithms
such as depth-first search, computation of connected components, etc. The library is
freely available under the BSD License from the Mathwork File Exchange1 . MRST has
a particular module (see Section A.3) that downloads and installs this library.
• METIS is a widely used library for partitioning graphs, partitioning finite element meshes,
and producing fill reducing orderings for sparse matrices [154]. The library is released2
under a permissive Apache License 2.0.
• AGMG: For large problems, the linear solvers available in MATLAB are not always suf-
ficient, and it may be necessary to use an iterative algebraic multigrid method. AGMG
[238] has MATLAB bindings included and was originally published as free open-source.
The latest releases have unfortunately only offered free licenses for academic research
and teaching3 .
4
• AMGCL is a header-only library for preconditioning with and without algebraic multigrid .
MATLAB-BGL is required by several of the more advanced solvers that are not part of
the basic functionality in MRST. Installing the other packages is recommended but not
required. When installing extra libraries or third-party toolboxes you want to integrate with
MRST, you must make the software aware of them. To this end, you should add a new
script called startup_user.m and use the built-in command mrstPath to make sure that
the routines you want to use are on the search path used by MRST and MATLAB to find
functions and scripts.

1 http://www.mathworks.com/matlabcentral/fileexchange/10922
2 http://glaros.dtc.umn.edu/gkhome/metis/metis/overview
3 http://homepages.ulb.ac.be/~ynotay/AGMG/
4 http://github.com/ddemidov/amgcl

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 608 The MATLAB Reservoir Simulation Toolbox

A.1.6 Terms of Usage

MRST is distributed as free, open-source software under the GNU Public License
(GPLv3).5 This license is a widely used example of a so-called copyleft license that
offers the right to distribute copies and modified versions of copyrighted creative work,
provided the same rights are preserved in modified and extended versions of the work. For
MRST, this means that you can use the software for any purpose, share it with anybody,
modify it to suit your needs, and share the changes you make. However, if you share any
version of the software, modified or unmodified, you must grant others the same rights to
distribute and modify it as in the original version. By distributing it as free software under
the GPLv3 license, the developers of MRST have made sure that the software will stay
free, no matter who changes or distributes it.
The development of the toolbox has to a large extent been funded by strategic research
grants awarded from the Research Council of Norway. Dissemination of research results
is an important evaluation criterion for these types of research grants. To provide the
developers with an overview of some usage statistics for the software, we ask you to
kindly register your affiliation/country upon download. This information is only used when
reporting impact of the creative work to agencies co-funding its development. If you also
leave an email address, we will notify you when a new release or critical bug fixes are
available. Your email address will under no circumstances be shared with any third party.
If you use MRST in any scholar work, we require that the creative work of the MRST
developers is courteously and properly acknowledged by referring to the MRST website
and by citing this book or one of the overview papers describing the software [192, 170, 30].
Last, but not least, if you have modified parts of the software or added new functionality, I
strongly encourage you to release your code and thus pay back to the community that has
developed it; if not the whole code, then at least generic parts that could be of significant
interest to others.

Computer exercises
A.1.1 Download and install the software
A.1.2 Run flowSolverTutorial1 from the command line to verify that your installation
is working.
A.1.3 Load flowSolverTutorial1 in the editor (editflowSolver
Tutorial1.m) and run it in cell mode to evaluate one cell at the time. Use help or
doc to inspect the documentation for the various functions used in the script.
A.1.4 Run flowSolverTutorial1 line by line: Set a breakpoint on the first executable line
by clicking on the small symbol next to line 27, push the “play” button, and then
use the “step” button to advance a single line at the time. Change the grid size to
10 × 10 × 25 and rerun the example.

5 See www.gnu.org/licenses/gpl.html for more details.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A.2 Public Data Sets and Test Cases 609

A.1.5 Use mrstExploreModules() to locate and load incompIntro from the

incomp module. Run the tutorial line by line or cell by cell, and then publish
the workbook to reproduce the contents of Figure A.3.
A.1.6 Replace the constant permeability in the incompIntro tutorial by a random perme-
ability field
rock.perm = logNormLayers(G.cartDims,[100 10 100])*milli*darcy;

Can you explain the changes in the pressure field?

A.1.7 Run all of the examples listed by mrstExamples() that have the word “tutorial” in
their names. In particular,

• gridTutorialIntro introduces you quickly to the most fundamental parts

of MRST, the grid structure, which is discussed in more detail in Chapter 3;
• tutorialPlotting introduces you to various basic routines and techniques
for plotting grids and data defined over these;
• tutorialBasicObjects will give you a quick overview of a lot of the func-
tionality that can be found in the toolbox.

A.2 Public Data Sets and Test Cases

Good data sets are in our experience essential to enable tests of new computational methods
in a realistic and relevant setting. Such data sets are hard to come by, and when making
MRST we have made an effort to provide simple access to a number of public data sets that
can be freely downloaded. With the exception of a few illustrations in Chapter 3, which are
based on data that cannot be publicly disclosed, all examples discussed in the book either
use MRST to create their input data or rely on public data sets that can be downloaded
freely from the internet.
To simplify dataset management, MRST offers a graphical user interface,
mrstDatasetGUI()

that lists all public data sets known to the software, gives a short description of each, and
can download and unpack most data sets to the correct subfolder of the standard path. A
few data sets require you to register your email address or fill in a license form, and in these
cases we provide a link to the correct webpage. Figure A.6 shows some of the data sets that
are available via the graphical interface. Several of these are used throughout the book.
Herein, I use the convention that data sets are stored in subfolders of a standard path,
which you can retrieve by issuing the query
mrstDataDirectory()

I recommend that you adhere to this convention when using the software as a supplement
to the book. If you insist on placing standard data sets elsewhere, I suggest that you use
mrstDataDirectory(<path>) to reset the default path.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 610 The MATLAB Reservoir Simulation Toolbox

Figure A.6 Examples of the free data sets that are distributed along with MRST.

MRST also contains a number of grid factory routines and simplified geostatistical
algorithms that you can use to make your own test cases. These will be discussed in more
detail in Chapters 3 and 2, respectively. Here, a word of caution about exact reproducibility
is in order. The grid-factory routines are mostly deterministic and should enable you to
create the exact same grids each time your run them (and hence reproduce the test cases
discussed in the book). Routines for generating petrophysical data rely on random numbers
and will not give the same results in repeated runs. Hence, you can only expect to reproduce
plots and numbers that are qualitatively similar whenever these are used, unless you make
sure to store and reset the seed used by MATLAB’s random number generator.

A.3 More About Modules and Advanced Functionality

As you will recall from Section A.1.1, MRST consists of core functionality and a set of
add-on modules that extend, complement, and override basic features, typically in the form
of specialized or more advanced solvers and workflow tools like upscaling, grid coarsening,
etc. This section explains how to load and manage different modules, and tries to explain the
basic characteristics of modules, or in other words the design criteria you can apply if you
intend to develop your own module. For completeness, we also provide a brief overview
of the comprehensive set of modules that currently are in the official release. Most of these
modules are not discussed at all in the book, which for brevity needs to focus on basic
flow models and modules offering standard functionality of a general interest to a wide
audience.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A.3 More About Modules and Advanced Functionality 611

A.3.1 Operating the Module System

The module system is a simple facility for extending and modifying the feature set. Specif-
ically, the module system enables on-demand activation and deactivation of advanced fea-
tures that are not part of the core functionality. It consists of two parts: one that handles
mapping of module names to system paths and one that uses this mapping to manipulate
MATLAB’s search path. The search path is a list of folders used by MATLAB to locate
files. A module in MRST is strictly speaking a collection of functions, object declarations,
and example scripts located in a folder. Each module needs to have a unique name and
reside in a different folder than other modules. When you activate a module, the corre-
sponding folder is added to the search path, and when you deactivate, the folder is removed
from the search path.
The mapping between module names and paths on your computer is maintained by
the function mrstPath. The paths are expected to be full paths to existing folders on your
computer. To determine which modules are part of your current installation, you use the
function as a command
mrstPath

This will list all modules the software is aware of and can activate. To activate a particular
module, you use the function mrstModule. As an example, calling
mrstModule add mimetic mpfa

will load the two modules for consistent discretization methods discussed in Chapter 6. The
mrstModule function has the following additional modes of operation:

mrstModule <list|clear|reset|gui> [module list]

which will list all active modules, deactivate all modules, deactivate all modules except for
those in the list, or bring up a graphical user interface with check-boxes that enable you to
activate and deactivate individual modules. For the latter, you can also use the command
moduleGUI.
All modules that come bundled with the official release will be placed in a predefined
folder structure, so that they can be added automatically to the module mapping by the
startup function. However, module folders can in principle be placed anywhere on your
computer as long as they are readable. To make MRST aware of these additional mod-
ules, you must use mrstPath to register them in the mapping. Assuming, for instance,
that you want to make the AGMG multigrid solver [238] available. This can be done
as follows,
mrstPath register AGMG S:\mrst\modules\3rdparty\agmg

Once the mapping is established, the module can be activated

mrstModule add AGMG

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 612 The MATLAB Reservoir Simulation Toolbox

In my experience, the best way to register modules that are not part of the default mapping
that comes with the official versions of the software is to add appropriate lines to your
startup_user file. The following is an excerpt from mine

mrstPath reregister distmesh ...

/home/kalie/matlab/mrst-bitbucket/mrst-core/utils/3rdparty/distmesh

which adds the mesh generator DistMesh [248] as a module in MRST.

A.3.2 What Characterizes a Module?

MRST does not have a strict policy for what becomes a module and what does not,
and if you look at the modules that are part of the official release, you will see that
they differ quite a lot. Some modules are just small pieces of code that add specialized
capabilities like the mpfa module, whereas others add comprehensive new functionality
such as the incomp, ad-core/ad-blackoil, and co2lab modules. Some of these modules
are robust, well-documented, and contain features that will likely not change in future
releases. As such, they could have been included as part of the core functionality if
we had not decided to keep this as small as possible to simplify maintenance and
reduce the potential for feature conflicts. Other modules are constantly changing to
support ongoing research. Until recently, all modules in the AD-OO family were of
this type.
Semi-independent modules is a simple way to organize software development that pro-
motes software reuse. By organizing a new development as a separate module you will
probably be more careful to distinguish code of generic value from code that is case
specific or of temporary value only. The fact that others, or your future self, may want
to reuse your code can motivate the extra effort needed to document and make examples
and tutorials, whose presence often is decisive when people consider to use or continue
to develop the functionality you have implemented. The module concept is particularly
convenient if you have specific functionality you want to activate or deactivate as you
like. In-house, our team has many modules that are in varying degrees of development
and/or decay. Some are accessible to the whole team, whereas others reside on one person’s
computer only.
You may say that any module that is part of the official release is just some code
that has been organized in a certain way and released to others. However, common
for all such modules is that they attempt to adhere to the following recommended
design rules:

• A module should offer new functionality that is distinctly different from what is already
available in the core functionality and/or in other modules.

• A module should distinguish between functionality exposed to the users of the module
and functionality that is only used internally. The latter can be put in a folder called

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A.3 More About Modules and Advanced Functionality 613

private (so that it is not accessible to functions other than those in the parent folder) or
some other folder that signals that this functionality is for internal use.
• A module should contain a set of tutorials/examples that explain and highlight basic
functionality of the module. The examples should be as self-explanatory as possible and
at most take a few minutes to run through. Examples are easier to comprehend if you
section your scripts using cell mode and accompany each code section by an informative
text that explains the computations to take place or discusses visual output. If special data
sets are required, these should be published along with the module.
• All main routines in a module should be documented, preferably following the format
used elsewhere in the software, describing input and output data, the underlying method,
and assumptions and limitations.
• Modules should, as a general rule, not use functionality from the official toolboxes that
are sold with MATLAB since many users do not have access to these.
• Name conflicts should be minimized to avoid messing up the search path in MATLAB.
When two files with the same name appear on the search path, MATLAB will pick the one
found nearest the top. To avoid potential unintended side effects, it is therefore important
that files have unique names across different modules.
• To the extent possible, the implementation should try to stick to the same naming conven-
tions as used in MRST for readability. This means using camelCaseNames for functions
and CamelCase starting with a capital letter for class objects. Likewise, widely used data
objects can preferably use easily recognizable names like G for grid, rock for petro-
physics, fluid for fluid models, W for wells, etc.
• Preferably, the source code should be kept in a public (or private) repository on a cen-
tralized service like Bitbucket or GitHub so that you can use a version control system to
keep track of its development.
I have recently written a paper [186] that discusses development of open-source software
and good practices for experimental scientific programming in more detail.

A.3.3 List of Modules

This section outlines the various modules that make up the official MRST 2018a release
and explains briefly the purpose and key features of each individual module.

Grid Generation and Partitioning

The core functionality of MRST includes a general data structure for fully unstructured
grids, as well as a number of grid factory and utility routines, including the possibility to
read and construct corner-point grids from ECLIPSE input. In addition, there are a number
of add-on modules for generating and processing grids:

agglom: Offers a number of elementary routines that can be combined in various ways to
create coarse partitions that adapt to geology or flow fields [125, 124, 194, 187]

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 614 The MATLAB Reservoir Simulation Toolbox

based on cell-based indicator values. More details are given in Chapter 14 and in
the additional tutorial examples.
coarsegrid: Extends the unstructured grid format in MRST to also include coarse grids
formed as a partition of an underlying fine grid. Such grids form a key part in
upscaling and multiscale methods, but act almost like any standard MRST grid
and can hence be passed to many solvers in other modules. Coarse grid generated
from a partition are also useful for visualization purposes. More details are given
in Chapter 14.
libgeometry: Geometric quantities like cell volumes and centroids and areas, centroids,
and normals of faces must be computed by the computeGeometry routine. This
module offers a C-accelerated version, mcomputeGeometry, that reduces the com-
putational time substantially for large models. Since 2016a, the need for this
C-acceleration has diminished.
opm_gridprocessing: Corner-point grids can be constructed from ECLIPSE input by
the processGRDECL function, which is part of the core functionality in MRST.
This module offers a C-accelerated version, mprocessGRDECL or processgrid, of
the same processing routines. The implementation is from the Open Porous Media
(OPM) initiative, which generally can be seen as a C/C++ sibling of MRST.
triangle: Provides a mex interface to Triangle, a two-dimensional quality mesh gen-
erator and Delaunay triangulator developed by Jonathan Richard Shewchuk, to
generate grids for MRST.
upr: Contains functionality for generating unstructured polyhedral grids that align to pre-
scribed geometric objects. Control-point alignment of cell centroids is introduced
to accurately represent horizontal and multilateral wells, but can also be used to
create volumetric representations of fracture networks. Boundary alignment of cell
faces is introduced to accurately preserve geological features such as layers, frac-
tures, faults, and/or pinchouts. This third-party module was originally developed
as part of a master thesis by Berge [42]; see also [169].

Incompressible Discretizations and Solvers

This family of modules consist primarily of functionality for solving Poisson-type pressure
equations:

incomp: Implements fluid objects for incompressible, two-phase flow. The module imple-
ments the two-point flux-approximation (TPFA) method and explicit and implicit
transport solvers with single-point upstream mobility weighting, as described in
detail in Parts II and III of the book. Originally part of the core module, but was
moved to a separate module once the software started offering solvers for more
advanced fluid models.
mimetic: The standard TPFA method in incomp is not consistent and may give signif-
icant grid-orientation errors for grids that are not K-orthogonal. The module

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A.3 More About Modules and Advanced Functionality 615

implements a family of consistent, mimetic finite difference methods for incom-

pressible (Poisson-type) pressure equations; see Section 6.4.
mpfa: Implements the MPFA-O scheme (see Section 6.4 on page 188), which is an exam-
ple of a scheme that employs more degrees of freedom when constructing the
discrete fluxes across cell interfaces to ensure a consistent discretization with
reduced grid-orientation effects.
vem: Virtual element methods (VEM) [38, 39] constitute a unified framework for higher-
order methods on general polygonal and polyhedral grids. The module solves
general incompressible flow problems by first- and second-order VEM, with the
possibility to choose different inner products. Originally developed by Klemetsdal
as part of his master thesis [168].
adjoint: Implements strategies for production optimization based on adjoint formula-
tions for incompressible, two-phase flow. Example: optimization of the net present
value constrained by the bottom-hole pressure in wells. For optimization problems
with more complex fluid physics, the newer optimization module from the AD-
OO framework is recommended.

Implicit Solvers Based on Automatic Differentiation

The object-oriented AD-OO framework is based on automatic differentiation and offers a
rich set of functionality for solving a wide class of model equations:

ad-core: does not contain any complete simulators by itself, but rather implements the
common framework used for many other modules. This includes abstract classes
for reservoir models, for time stepping, nonlinear solvers, linear solvers, function-
ality for variable updating/switching, routines for plotting well responses, etc. See
the discussion in Chapters 11 and 12 and [170, 30] for more details.
deckformat: Contains functionality for handling complete simulation decks in the
ECLIPSE format, including input reading, conversion to SI units, and construction
of MRST objects for grids, fluids, rock properties, and wells. Functionality from
this module is essential for the fully implicit simulators in the AD-OO framework;
see Chapters 11 and 12.
ad-props: Functionality related to property calculations for the AD-OO framework.
Specifically, the module implements a variety of test fluids and functions that
are used to create fluids from external data sets. This module is used as part of
almost all simulators in MRST studying compressible equations of black-oil or
compositional type. More details are in Section 11.3.
ad-blackoil: Models and examples that extend the MRST AD-OO framework found
in the ad-core module to black-oil problems. Solvers and functionality from this
module are discussed at length in Chapters 11 and 12.
blackoil-sequential: Implements sequential solvers for the same set of equations as
in ad-blackoil based on a fractional flow formulation wherein pressure and trans-
port are solved as separate steps; see [215] for more details. These solvers can be

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 616 The MATLAB Reservoir Simulation Toolbox

significantly faster than those found in ad-blackoil for many problems, especially
problems where the total velocity changes slowly during the simulation.
ad-eor: Builds on the ad-core and ad-blackoil modules and defines model equations
and provides simulation tools for water-based enhanced oil recovery techniques
(polymer and surfactant injection); see [30] for more details about the polymer
case.
compositional: Introduced in release 2017b and offers solvers for compositional flow
problems with cubic equations of state or tabulated K-values. Both the natural
variables and overall composition formulations are included. More details about
the underlying models and equations can be found in the PhD thesis by Møyner
[212]. The tutorials include validation against commercial and external research
simulators.
solvent: Introduced in release 2017b and offers an extension of the Todd–Longstaff
model for miscible flooding [289] that makes it possible to simulate miscible and
partially miscible displacement without using an expensive fully compositional
formulation.
ad-mechanics: Introduced in release 2017b and includes a mechanical model for linear
elasticity that can be coupled to standard reservoir flow models, such as oil-water
or black-oil models. The global system is either solved simultaneously or by fixed-
stress splitting [22]. The solver for linear elasticity uses the virtual element method
and can handle general grids.
optimization: Routines for solving optimal control problems with (forward and adjoint)
solvers based on the AD-OO framework. The module contains a quasi-Newton
optimization routine using BFGS-updated Hessians, but can easily be set up to
use any (non-MRST) optimization code.
ad-fi: Deprecated module containing our first implementation of AD-based solvers for
black-oil models.

Workflow Tools
The next category of modules contains functionality that can be used either before or after
a reservoir simulation as part of a general modeling workflow.

diagnostics: Flow diagnostics tools are run to establish volumetric connections and
communication patterns in the reservoir and measure the heterogeneity of dynamic
flow paths. Details are discussed in Chapter 13 and in [218, 261, 171].
upscaling: Implements methods and tutorials for averaging and flow-based upscaling of
permeabilities and transmissibilities as discussed in Chapter 15.
steady-state: Functionality for upscaling relative permeabilities based on a steady-state
assumption. This includes general steady state, as well as capillary and viscous-
dominated limits. The functionality is demonstrated through a few tutorial exam-
ples. For more details, see [132, 131].

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A.3 More About Modules and Advanced Functionality 617

Multiscale Methods
Multiscale methods can either be used as a robust upscaling method to produce upscaled
flow velocities and pressures on a coarse grid or as an approximate, iterative fine-scale
solver. Instead of placing these modules into one of the categories discussed so far, I have
given them a separate category, in part because of the prominent place they have played as
a driving force for the development of MRST:
msmfe: Implements the multiscale mixed finite-element (MsMFE) method on stratigraphic
and unstructured grids in 3D [68, 1, 4, 5, 6, 165, 222, 17, 194, 243]. Basis functions
(prolongation operators) associated with the faces of a coarse grid are computed
numerically by solving localized flow problems on pairs of coarse blocks. With
these, you can define a reduced flow problem on a coarse grid and prolongate the
resulting coarse solution back to the fine grid to get a mass-conservative flux field.
Gives a good approximate solver and a robust and accurate alternative to upscal-
ing for incompressible flow problems. MRST was originally developed solely to
support research on MsMFE methods, but the module has nevertheless not been
actively developed since 2012.
msfv: Implements the operator formulation [200] of the multiscale finite-volume (MsFV)
method [146] for incompressible flow on structured and unstructured grids in 3D,
under certain restrictions on the grid geometry and topology [211, 214]. Prolon-
gation operators are defined by solving flow problems localized to dual coarse
blocks and then used to define a reduced flow problem on the coarse primal grid
and map unknowns computed on this grid back to the underlying fine grid. MsFV
can either be used as an approximate solver or as a global preconditioner in an
iterative solver. The module is not actively developed and is offered mostly for
historic reasons. You should use the msrsb module instead.
msrsb: The multiscale restricted-smoothing basis (MsRSB) method [216, 215] is the cur-
rent state-of-the-art within multiscale methods [195]. MsRSB is very robust and
versatile and can either be used as an approximate coarse-scale solver that has
mass-conservative subscale resolution, or as an iterative fine-scale solver that will
provide mass-conservative solutions for any given tolerance. Performs well on
incompressible 2-phase flow [216], compressible 2- and 3-phase black-oil type
models [215, 130], as well as compositional models [217, 212]. MsRSB can uti-
lize combinations of specialized and adapted prolongation operators to accelerate
convergence [191]. In active research, and a more comprehensive version exists
in-house.

Specialized Simulation Tools

This category consists of modules that implement solvers and simulators for other mathe-
matical models than those discussed in this book.
co2lab: Combines results of more than one decade of academic research and development
on CO2 storage modeling into a unified toolchain that is easy and intuitive to

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 618 The MATLAB Reservoir Simulation Toolbox

use. The module is geared towards the study of long-term trapping in large-scale
saline aquifers and offers computational methods and graphical user interfaces
to visualize migration paths and compute upper theoretical bounds on structural,
residual, and solubility trapping. It also offers efficient simulators based on a
vertical-equilibrium formulation to analyze pressure build-up and plume migra-
tion and compute detailed trapping inventories for realistic storage scenarios. Last
but not least, the module provides simplified access to publicly available data
sets, e.g., from the Norwegian CO2 Storage Atlas. For more details, see [19] or
specific references documenting the general toolchain of methods [232, 193, 21],
methods for identifying structural traps [230], and the various vertical-equilibrium
formulations [20, 227, 228].
dfm: Contains two-point and multipoint solvers for discrete fracture-matrix systems
[268, 297], including multiscale methods. This third-party module is devel-
oped by Sandve from the University of Bergen, with minor modifications by
Keilegavlen.
dual-porosity: A module for geologic well-testing in naturally fractured reservoirs has
been developed and is maintained by the Carbonate Reservoir Group at Heriot
Watt University. The module implements tools to generate synthetic transient
pressure responses for idealized and realistic fracture networks.
fvbiot: Developed and maintained by the University of Bergen and implements cell-
centered discretizations of three different equations: (i) scalar elliptic equations
(Darcy flow), using multipoint flux approximations; this is more or less equivalent
to the MPFA implementation in the mpfa module, although the implementation
and data structures are slightly different; (ii) Linear elasticity, using the multipoint
stress-approximation (MPSA) method [237, 233, 156]; (iii) Poromechanics, e.g.,
coupling terms for the combined system of the first two models.
geochemistry: Implements solvers for the following models: aqueous speciation, surface
chemistry, redox chemistry, equilibrium with gas and solid phases. Each of these
models can be readily coupled with a flow solver. The module is developed as a
collaboration between University Texas Austin and SINTEF.
hfm: Implements the embedded discrete fracture method (EDFM) on stratigraphic and
unstructured grids in 2D and 3D. The module also implements a multiscale
restriction-smoothed basis (MsRSB) solver; see [273] for more details. The
module was originally developed as a collaboration between TU Delft and
SINTEF. Recently, researchers from Heriot Watt University have contributed
several new functions and examples. Most notably, this includes greatly improved
support for fracture shapes and orientations in 3D.
vemmech: Offers functionality to set up solvers for linear elasticity problems on irregular
grid, using the virtual element method [39, 114], which is a generalization
of finite-element methods that takes inspiration from modern mimetic finite-
difference methods; see [23, 229]

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A.3 More About Modules and Advanced Functionality 619

Miscellaneous
The last category consists of modules that do not offer any computational or modeling tools
but rather provide general utility functions employed by the other modules in MRST:
book: Contains all the scripts used for the examples, figures, and some of the exercises in
this book.
linearsolvers: Offers bindings to external linear solvers. The initial release includes
tentative support for the AMGCL header-only library for preconditioning with
and without algebraic multigrid methods.
matlab_bgl: Routine for download the MATLAB Boost Graph Library.
mrst-gui: Graphical interfaces for interactive visualization of reservoir states and petro-
physical data. The module includes additional routines for improved visualization
(histograms, well data, etc.) as well as a few utility functions that enable you to
override some of MATLAB’s settings to enable faster 3D visualization (rotation,
etc.).
octave: A number of patches to provide compatibility with GNU Octave.
spe10: Contains tools for downloading, converting, and loading the data into MRST; see
Section 2.5.3. The module also features utility routines for extracting parts of the
model, as well as a script that sets up a (crude) simulation of the full model (using
the AGMG multigrid solver).
streamlines: Implements Pollock’s method [253] for tracing of streamlines on Cartesian
and curvilinear grids based on a set of input fluxes computed by the incompress-
ible flow solvers in MRST.
wellpaths: Functionality for defining wells following curvilinear trajectories.
This list includes all public modules in release 2018a. By the time you read this book, more
modules that are currently in the making will likely have been added to the official list.

Modules Not Part of the Official Release

The following modules are publicly available, but not part of the official release:
enkf: Ensemble Kalman filter (EnKF) module developed by researchers at TNO [182,
181] that contains EnKF and EnRML schemes, localization, inflation, asyn-
chronous data, production and seismic data, updating of conventional and
structural parameters.
remso: An optimization module based on multiple shooting, developed by Codas [87],
which allows for great flexibility in the handling of nonlinear constraints in reser-
voir management optimization problems.
mrst-cap Implements a compositional flow model with capillary pressure; see https://
github.com/sogoogos/mrst-cap.
You can also find other MRST codes online that have not been structured modules or have
not been maintained for many years.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 620 The MATLAB Reservoir Simulation Toolbox

Computer exercises

A.3.1 Try to run the following tutorials and examples from various modules

• simpleBCmimetic from the mimetic module.

• simpleUpscaleExample from the upscaling module
• gravityColumnMS from the msmfem module
• example2 from the diagnostics module
• firstTrappingExample from the co2lab module (notice that this example
does not work unless you have MATLAB-BGL installed).

A.4 Rapid Prototyping Using MATLAB and MRST

How can you reduce the time span from the moment you get a new idea to when you
have demonstrated that it works well for realistic reservoir engineering problems? In our
experience, prototyping and validating new numerical methods is painstakingly slow. There
are many reasons for this. First of all, there is often a strong disconnect between the
mathematical abstractions and equations used to express models and numerical algorithms
and the syntax of the computer language you use to implement your algorithms. This is
particularly true for compiled languages, where you typically end up spending most of your
time writing and tweaking loops that perform operations on individual members of arrays or
matrices. Object-oriented languages like C++ offer powerful functionality that can be used
to make abstractions that are both flexible and computationally efficient and enable you to
design your algorithms using high-level mathematical constructs. However, these advanced
features are usually alien and unintuitive to those who do not have extensive training in
computer sciences. If you are familiar with such concepts and are in the possession of a
flexible framework, you still face the never-ending frustration caused by different versions
of compilers and (third-party) libraries that seems to be an integral part of working with
compiled languages.

Experimental Programming is Efficient in a Scripting Language

Based on working with many different students and researchers over the past twenty years,
I claim that using a numerical computing environment based on a scripting language like
MATLAB/GNU Octave to prototype, test, and verify new models and computational algo-
rithms is significantly more efficient than using a compiled language like Fortran, C, and
C++. Not only is the syntax intuitive and simple, but there are many mechanisms that help
you boost your productivity and you avoid some of the frustrations that come with compiled
languages: there is no complicated build process or handling of external libraries, and your
implementation is inherently cross-platform compatible.
MATLAB, for instance, provides mathematical abstractions for vectors and matrices and
built-in functions for numerical computations, data analysis, and visualization that enable
you to quickly write codes that are not only compact and readable, but also efficient and

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A.4 Rapid Prototyping Using MATLAB and MRST 621

robust. On top of this, MRST provides additional functionality that has been developed
especially for computational modeling of flow in porous media:

• an unstructured grid format that enables you to implement algorithms without knowing
the specifics of the grid;
• discrete operators, mappings, and forms that are not tied to specific flow equations, and
hence can be precomputed independently and used to write discretized flow equations in
a very compact form that is close to the mathematical formulations of the same;
• automatic differentiation enables you to compute the values of gradients, Jacobians, and
sensitivities of any programmed function without having to compute the necessary partial
derivatives analytically; this can, in particular, be used to automate the formulation of
fully implicit discretizations of time-dependent and coupled systems of equations;
• data structures providing unified access enable you to hide specific details of constitutive
laws, fluid and rock properties, drive mechanisms, etc.
This functionality is gradually introduced and described in detail in the main parts of the
book.

Interactive Development of New Programs

An equally important aspect of using a numerical environment like MATLAB is that you
can develop your program differently than what you would do in a compiled language.
Using the interactive environment, you can interactively analyze, change and extend data
objects, try out each operation, include new functionality, and build your program as you
go. This feature is essential in a debugging process, when you try to understand why a
given numerical method fails to produce the results you except it to give. In fact, you can
easily continue to change and extend your program during a test run: the debugger enables
you to run the code line by line and inspect and change variables at any point. You can also
easily step back and rerun parts of the code with changed parameters that may possibly
change the program flow. Since MATLAB uses dynamic type-checking, you can also add
new behavior and data members while executing a program. However, how to do this in
practice is difficult to teach in a textbook. Instead, you should run and modify the various
examples that come with MRST and the book. We also recommend that you try to solve
the computer exercises that are suggested at the end of several of the chapters and sections
in the book.

Ensuring Efficiency of Your MATLAB Code

Unfortunately, all this flexibility and productivity comes at a price: it is very easy to develop
programs that are not very efficient. In the book, I therefore try to implicitly teach program-
ming concepts you can use to ensure flexibility and high efficiency of your programs. These
include, in particular, efficient mechanisms for traversing data structures like vectorization,
indirection maps, and logical indexing, as well as use of advanced MATLAB functions
like accumarray, bsxfun, etc. Although these will be presented in the context of reservoir

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 622 The MATLAB Reservoir Simulation Toolbox

simulation, I believe the techniques are of interest for readers working with lower-order
finite-volume discretizations on general polyhedral grids.
As an illustration of the type of MATLAB programming that will be used, we can
consider the following code, which generates five million random points in 3D and counts
the number of points that lie inside each of the eight octants:
n = 5000000;
pt = randn(n,3);
I = sum(bsxfun(@times, pt>0, [1 2 4]),2)+1;
num = accumarray(I,1);

The third line computes the sign of the x, y, and z coordinates and maps the triplets of
logical values to an integer number between 1 and 8. To count the number of points inside
each octant, we use the function accumarray that groups elements from a data set and
applies a function to each group. The default function is summation, and by setting a unit
value in each element, we count the entries.
Next, let us compute the mean point inside each octant. A simple loop-based solution
could be something like:
avg = zeros(8,3);
for j=1:size(pt,1)
quad = sum((pt(j,:)>0).*[1 2 4])+1;
avg(quad,:) = avg(quad,:)+pt(j,:);
end
avg = bsxfun(@rdivide, avg, num);

Generally, you should try to avoid explicit loops since they tend to be slow in MATLAB.
On my old laptop with MATLAB R2014a, it took 0.47 seconds to count the number of
points within each octant, but 24.9 seconds to compute the mean points. So let us try to
be more clever and utilize vectorization. We cannot use accumarray since it only works
for scalar values. Instead, we can build a sparse matrix that we multiply with the pt array
to sum the coordinates of the points. The matrix has one row per octant and one column
per point. Lastly, we use the indicator I to assign a unit value in the correct row for each
column, and use bsxfun to divide the sum of the coordinates with the number of points
inside each octant:
avg = bsxfun(@rdivide, sparse(I,1:n,1)*pt, accumarray(I,1));

On my computer this operation took 0.53 seconds, which is 50 times faster than the
loop-based solution. An alternative solution is to expand each coordinate to a quadruple
(x,y,z,1), multiply by the same sparse matrix, and use bsxfun to divide the first three
columns by the fourth column to compute the average:
avg = sparse(I,1:n,1)*[pt, ones(n,1)];
avg = bsxfun(@rdivide, avg(:,1:end-1), avg(:,end));

These operations ran in 0.64 seconds. Which solution do you think is most elegant?

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A.5 Automatic Differentiation in MRST 623

Hopefully, this simple example has inspired you to learn a bit more about efficient
programming tricks if you do not already speak MATLAB fluently. MRST is generally
full of tricks like this, and in the book we will occasionally show a few of them. However,
if you really want to learn the tricks of the trade, the best way is to dig deep into the actual
codes.

A.5 Automatic Differentiation in MRST

Automatic differentiation (AD) is a technique that exploits the fact that any function evalu-
ation, regardless of complexity, can be broken down to a limited set of arithmetic operations
(+, −, ∗, /, etc.) and evaluation of elementary functions like exp, sin, cos, log, and so on,
that all have known derivatives. In AD, the key idea is to keep track of quantities and their
derivatives simultaneously; every time an elementary operation is applied to a numerical
quantity, the corresponding differential rule is applied to its derivative.

Implementing Automatic Differentiation

Consider a scalar primary variable x and a function f = f (x). Their AD representations
would then be the pairs $x,1% and $f ,df %, where 1 is the derivative dx/dx and df is
the numerical value of the derivative f  (x). Accordingly, the action of the elementary
operations and functions must be defined for such pairs,

$f ,df % + $g,dg% = $f + g,df + dg% ,

$f ,df % ∗ $g,dg% = $f g,f dg + df g% ,
exp($f ,df %) = $exp(f ), exp(f )df % .

In addition to this, we need to use the chain rule to accumulate derivatives; that is, if
f (x) = g(h(x)), then fx (x) = g  (h(x))h (x). If we have a function f1(x,y) of 2two primary
variables x and y, their AD representations are $x,1,0%, $y,0,1% and f ,fx ,fy . This more
or less summarizes the key idea behind AD; the remaining and difficult part is how to
implement the idea as efficient computer code that has a low user-threshold and minimal
computational overhead.
It is not very difficult to implement elementary rules needed to differentiate the func-
tion evaluations you find in a typical numerical PDE solver. To be useful, however, these
rules should not be implemented as new functions, so that you need to write something
like myPlus(a, myTimes(b,c)) when you want to evaluate a + bc. You can find many
different AD libraries for MATLAB that offer both forward and reverse model for accu-
mulating derivatives, e.g., ADiMat [286, 43], ADMAT [60, 300], MAD [291, 275, 112],
or from MATLAB Central [110, 208]. An elegant solution is to use classes and operator
overloading. When MATLAB encounters an expression a+b, the software will choose one
out of several different addition functions depending on the data types of a and b. All we
now have to do is introduce new addition functions for the various classes of data types that
a and b may belong to. Neidinger [223] gives a nice introduction to how to implement this

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 624 The MATLAB Reservoir Simulation Toolbox

in MATLAB. MRST has its own implementation that is specially targeted at solving flow
equations, i.e., working with long vectors and large sparse matrices.

The ADI Class in MRST

The ADI class in MRST relies on operator overloading as suggested in [223] and uses a rel-
atively simple forward accumulation of derivatives. The ADI class differs from many other
libraries in a subtle, but important way. Instead of working with a single Jacobian of the full
discrete system as one matrix, MRST uses a list of matrices that represent the derivatives
with respect to different variables that will constitute sub-blocks in the Jacobian of the
full system. The reason for this choice is computational performance and user utility. In a
typical simulation, we need to compute the Jacobian of systems of equations that depend on
primary variables that each is a long vector. Oftentimes, we want to manipulate parts of the
full Jacobian that represents specific subequations. This is not practical if the Jacobian of
the system is represented as a single matrix; manipulating subsets of large sparse matrices
is currently not very efficient in MATLAB, and keeping track of the necessary index sets
can also be quite cumbersome from a user’s point-of-view. Accordingly, our current choice
is to let the MRST ADI class represent the derivatives of different primary variable as a list
of matrices.
The following example illustrates how the ADI class works:
Example A.5.1 We want to compute the expression z = 3e−xy and its partial derivatives
∂z/∂x and ∂z/∂y for the values x = 1 and y = 2. Using our previous notation, the AD-
representation of z should be an object of the following form
1 2
z = 3e−xy , − 3ye−xy , − 3xe−xy ≈ $0.4060, − 0.8120, − 0.4060% .
With the ADI class, computing z and its partial derivatives is done as follows:
[x,y] = initVariablesADI(1,2);
z = 3*exp(-x*y)

The first line tells MRST that x and y are independent variables and instantiates two class
objects, initialized with correct values. The second line is what you normally would write
in MATLAB to evaluate the given expression. After the second line is executed, you have
three ADI variables (pairs of values and derivatives):
x = ADI Properties: y = ADI Properties: z = ADI Properties:
val: 1 val: 2 val: 0.4060
jac: {[1] [0]} jac: {[0] [1]} jac: {[-0.8120] [-0.4060]}

∂x ∂x ∂y ∂y ∂z  ∂z 
 
∂x ∂y ∂x ∂y ∂x x=1,y=2 ∂y x=1,y=2

If we continue computing with these variables, each new computation will give a result that
contains the value of the computation as well as the derivatives with respect to x and y.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A.5 Automatic Differentiation in MRST 625

Figure A.7 Complete set of functions invoked to evaluate 3*exp(-x*y) when x and y are ADI
variables. For brevity, we have abbreviated parts of the called functions.

Let us look a bit in detail on what happens behind the curtain. We start by observing
that the function evaluation 3*exp(-x*y) in reality consists of a sequence of elementary
operations: −, ∗, exp, and ∗, executed in that order. In MATLAB, this corresponds to the
following sequence of call to elementary functions
u = uminus(x);
v = mtimes(u,y);
w = exp(u);
z = mtimes(3,w);

To see this, you can enter the command into a file, set a breakpoint in front of the assignment
to z, and use the “Step in” button to step through all details. The ADI class overloads these
three functions (uminus, mtimes, and exp) by new functions that have the same names,
but operate on an ADI class object for uminus and exp, and on two ADI class objects
or a combination of a double and an ADI class object for mtimes. Figure A.7 gives a
breakdown of the sequence of calls to overloaded and internal functions that are invoked
within the ADI class library to evaluate 3*exp(-x*y) when x and y are ADI variables.
The same principles apply for other functions and if x and/or y are vectors, except that the
specific sequence of overloaded and internal functions from ADI will be different.

Computational Efficiency
As you can see from the example, use of AD will give rise to a whole new set of function
calls that are not executed if you only evaluate a mathematical expression and do not
compute its derivatives. Apart from the cost of the extra code lines that are executed,
user-defined classes are fairly new in MATLAB and there is still some overhead in using

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 626 The MATLAB Reservoir Simulation Toolbox

class objects and accessing their properties (e.g., val and jac) compared to the built-in
struct-class. The reason why using the ADI class library still pays off in most examples,
is that the cost of generating derivatives is typically much smaller than the cost of the
solution algorithms they will be used in, in particular when working with equations systems
consisting of large sparse matrices with more than one row per cell in the computational
grid. You should nevertheless still seek to limit the number of calls involving ADI class
functions (including the constructor). We let the following example be a reminder that
vectorization is of particular importance when using ADI classes in MRST:

Example A.5.2 To investigate the efficiency of vectorization versus serial execution of the
ADI objects in MRST, we consider the inner product of two vectors
z = x.*y;

and compare the cost of computing z and its two partial derivatives using four different
approaches:
1. explicit expressions for the derivative, zx = y and zy = x, evaluated using standard
MATLAB vectors of doubles;
2. the overloaded vector multiply (.*) with ADI vectors for x and y;
3. a loop over all vector elements with matrix multiply (*=mtimes) and x and y repre-
sented as scalar ADI variables; and
4. same as 3, but with vector multiply (.*=times).
MATLAB offers a stopwatch timer, which we start by the command tic and whose elapsed
time we read by the command toc:
[n,t1,t2,t3,t4] = deal(zeros(m,1));
for i = 1:m
n(i) = 2^(i-1);
xv = rand(n(i),1); yv=rand(n(i),1);
[x,y] = initVariablesADI(xv,yv);
tic, z = xv.*yv; zx=yv; zy = xv; t1(i)=toc;
tic, z = x.*y; t2(i)=toc;
tic, for k =1:n(i), z(k)=x(k)*y(k); end; t3(i)=toc;
tic, for k =1:n(i), z(k)=x(k).*y(k); end; t4(i)=toc;
end

Figure A.8 reports the corresponding runtimes as function of the number elements in the
vector. For this simple function, using ADI is a factor 20–40 times more expensive than
using direct evaluation of z and the exact expressions for zx and zy . Using a loop will
on average be more than three orders more expensive than vectorization. Since the inner
iterations multiply scalars, many programmers would implement it using matrix multi-
ply (*) without a second thought. Replacing (*) by vector multiply (.*) reduces the cost
significantly for short vectors, but the reduction diminishes as the length of the vectors
increases.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A.5 Automatic Differentiation in MRST 627
2
10
Exact formulat
Vectorized ADI
For loop and mtimes
0
10 For loop and times

−2
10

−4
10

−6
10 0 1 2 3 4
10 10 10 10 10

Figure A.8 Comparison of the time required for computing z=x.*y and derivatives as function of
the number of elements in the vectors x and y.

Vectorization is usually significantly more efficient than loops when built-in or user-
defined functions can be applied to complete vectors. The benefit of vectorization can
be significantly diminished and completely disappear if the vectorized operations must
allocate a lot of temporary memory and does not match the processor cache. Still, I would
recommend that you try to vectorize as much of your code as possible, as this typically
will lead to more readable code. As a general rule, you should always try to avoid loops
that call functions that have non-negligible overhead. The ADI class in MRST has been
designed to exploit vectorization on long vectors and lists of (sparse) Jacobian matrices
and has not been optimized for scalar variables. As a result, there is considerable overhead
when working with small ADI objects.

Using ADI to Assemble Linear(ized) Systems

The main use of ADI objects in MRST is to linearize and assemble (large) systems of
discrete equations. To use ADI to assemble and solve a linear system Ax = b, we must
first write the system on residual form,

r(x) = Ax − b = 0. (A.1)

Since x is unknown, we assume that we have an initial guess called y. Inserting this into
(A.1), we obtain

r(y) = Ay − b = A(y − x),

which we can solve for x to obtain x = y − A−1 r(y). It follows from (A.1) that r  (x) = A,
which means that if we write a code that evaluates equations on residual form, we can use
AD to assemble the system.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 628 The MATLAB Reservoir Simulation Toolbox

Example A.5.3 As a very simple illustration of how AD can be used to assemble a linear
system, let us consider the following linear 3 × 3 system
⎡ ⎤⎡ ⎤ ⎡ ⎤
3 2 −4 x1 −5
⎣ 1 −1 2 ⎦ ⎣ x2 ⎦ = ⎣ −1 ⎦ ,
−2 −2 4 x3 6
whose solution x = [1 2 3]T can be computed by a single line in MATLAB:
x = [3, 2, -4; 1, -4, 2; -2,- 2, 4]\ [-5; -1; 6]

To solve the same system using AD,

x = initVariablesADI(zeros(3,1));
A = [3, 2, -4; 1, -4, 2; -2,- 2. 4];
eq = A*x + [5; 1; -6];
u = -eq.jac{1}\eq.val

or we can specify the residual equations line by line and then concatenate them to assemble
the whole matrix
x = initVariablesADI(zeros(3,1));
eq1 = [ 3, 2, -4]*x + 5;
eq2 = [ 1, -4, 2]*x + 1;
eq3 = [-2, -2, 4]*x - 6;
eq = cat(eq1,eq2,eq3);
u = -eq.jac{1}\eq.val

Here, the first line sets up the AD variable x and initializes it to zero. The next three lines
evaluate the equations that make up our linear system. Evaluating each equation results in
a scalar residual value eq1.val and a 1 × 3 Jacobian matrix eq1.jac. In the fifth line,
the residuals are concatenated to form a vector and the Jacobians are assembled into the
full Jacobian matrix of the overall system.
At this point, you may argue that what I have shown you is a convoluted and expensive
way of setting up and solving a simple system. However, now comes the interesting part.
When solving partial differential equations on complex grids, it is often much simpler to
evaluate the residual equations in each grid cell than assembling the same local equations
into a global system matrix. Using AD, you can focus on the former and avoid the latter.
In Section 4.4.2, we introduce discrete divergence and gradient operators, div and grad.
With these, discretization of the flow equation ∇ · Kμ−1 (∇p − gρ∇z) = 0 introduced in
Section 1.4 can be written in a form that strongly resembles its continuous form
eq = div((T/mu)*.(grad(p) - g*rho*grad(z)));

where T is the transmissibilities (which can be precomputed for a given grid), mu and rho
are constant fluid viscosity and density, g is the gravity constant, and z is the vector of cell
centroids. That single line is all we need to evaluate and assemble the corresponding linear
system of discrete flow equations.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 A.5 Automatic Differentiation in MRST 629

Our primary use of AD is for compressible flow models, which typically give large sys-
tems of nonlinear discrete equations that need to be linearized and solved using a Newton–
Raphson method. As a precursor to the discussion in Chapter 7, we consider a last example
to outline how you can use ADI to solve a system of nonlinear equations.
Example A.5.4 Minimizing the Rosenbrock equation
 2
f (x,y) = (a − x)2 + b y − x 2 (A.2)
is a classical test problem from optimization. This problem is often called Rosenbrock’s
valley or banana function, since the global minimum (a,a 2 ) is located inside a long, narrow
and relatively flat valley of parabolic shape. Finding this valley is straightforward, but it is
more challenging to converge to the global minimum. A necessary condition for a global
minimum is that ∇f (x,y) = 0, which translates to the following two equations for (A.2)


 

∂x f (x,y) −2(a − x) − 4bx y − x 2 0
g(x) = = = . (A.3)
∂y f (x,y) 2b y − x 2 0
In the Newton–Raphson method, we assume an initial guess x and seek a better approxi-
mation x + x by solving for x from the linearized equation
0 = g(x + x) ≈ g(x) + ∇g(x)x. (A.4)
This is quite simple using ADI in MRST:
[a, b, tol] = deal(1, 100, 1e-6);
[x0, incr] = deal([-.5; 4]);
while norm(incr)>tol
x = initVariablesADI(x0);
eq = cat( 2*(a-x(1)) - 4*b.*x(1).*(x(2)-x(1).^2), ...
2*b.*(x(2)-x(1).^2));
incr = - eq.jac{1}\eq.val;
x0 = x0 + incr;
end

This is just a backbone version of a Newton solver that does not contain safeguards of any
kind like checking that the increments are finite, ensuring that the loop terminates after a
finite number of iterations, etc. Figure A.9 illustrates how the Newton solver converges to
the global minimum.
Beyond the examples and the discussion in this Appendix, we will not go more into details
about the technical considerations that lie behind the implementation of AD in MRST. If
you want a deeper understanding of how the ADI class works, the source code is fully open,
so you are free to dissect the details to the level of your own choice.

Computer exercises
A.5.1 As an alternative to using AD, you can use finite differences, f  (x) ≈ [f (x + h) −
f (x)]/ h, or a complex extension to compute f  (x) ≈ Im(f (x + ih))/ h. Use AD
and the function

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 630 The MATLAB Reservoir Simulation Toolbox

Figure A.9 Convergence of the AD-based Newton solver for the Rosenbrock problem. The upper
plot shows the path taken by the nonlinear solver superimposed over the function f (x,y). We have
used a modified colormap to show the valley in which the function f attains its 1% lowest values.
The lower plot depicts f (x,y)1/10 for the initial guess and the six iterations needed to reduce the
norm of the increment below 10−6 .

f = @(x) exp((x-.05).*(x-.4).*(x-.5).*(x-.7).*(x-.95));
to assess how accurate the two methods approximate f  (x) at n equidistant points
in the interval x ∈ [0,1] for different values of h.
A.5.2 The ADI class can compute log and exp, but does not yet support log2, log10,
and logm. Study ADI.m and see if you can implement the missing functions. What
about trigonometric and hyperbolic functions?
A.5.3 Can AD be used to compute higher-order derivatives? How or why not?

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 References

[1] Aarnes, J. E. 2004. On the use of a mixed multiscale finite element method for
greater flexibility and increased speed or improved accuracy in reservoir simulation.
Multiscale Model. Simul., 2(3), 421–439. doi:10.1137/030600655.
[2] Aarnes, J. E., Gimse, T., and Lie, K.-A. 2007b. An introduction to the numerics
of flow in porous media using Matlab. In Hasle, G., Lie, K.-A., and Quak,
E. (eds), Geometrical Modeling, Numerical Simulation and Optimisation: Indus-
trial Mathematics at SINTEF. Berlin, Heidelberg, New York: Springer-Verlag,
pp. 265–306.
[3] Aarnes, J. E., Hauge, V. L., and Efendiev, Y. 2007a. Coarsening of three-dimensional
structured and unstructured grids for subsurface flow. Adv. Water Resour., 30(11),
2177–2193. doi:10.1016/j.advwatres.2007.04.007.
[4] Aarnes, J. E., Kippe, V., and Lie, K.-A. 2005. Mixed multiscale finite elements and
streamline methods for reservoir simulation of large geomodels. Adv. Water Resour.,
28(3), 257–271. doi:10.1016/j.advwatres.2004.10.007.
[5] Aarnes, J. E., Krogstad, S., and Lie, K.-A. 2006. A hierarchical multiscale method
for two-phase flow based upon mixed finite elements and nonuniform coarse grids.
Multiscale Model. Simul., 5(2), 337–363. doi:10.1137/050634566.
[6] Aarnes, J. E., Krogstad, S., and Lie, K.-A. 2008. Multiscale mixed/mimetic methods
on corner-point grids. Comput. Geosci., 12(3), 297–315. doi:10.1007/s10596-007-
9072-8.
[7] Aavatsmark, I. 2002. An introduction to multipoint flux approximations for quadri-
lateral grids. Comput. Geosci., 6, 405–432. doi:10.1023/A:1021291114475.
[8] Aavatsmark, I. 2007. Interpretation of a two-point flux stencil for skew parallelogram
grids. Comput. Geosci., 11(3), 199–206. doi:10.1007/s10596-007-9042-1.
[9] Aavatsmark, I., Barkve, T., Bøe, Ø., and Mannseth, T. 1994. Discretization on non-
orthogonal, curvilinear grids for multi-phase flow. In: ECMOR IV – 4th European
Conference on the Mathematics of Oil Recovery. doi:0.3997/2214-4609.201411179.
[10] Aavatsmark, I., Eigestad, G. T., and Klausen, R. A. 2006. Numerical convergence of
the MPFA O-method for general quadrilateral grids in two and three dimensions. In:
Arnold, D. N., Bochev, P. B., Lehoucq, R. B., Nicolaides, R. A., and Shashkov, M.
(eds), Compatible Spatial Discretizations. New York: Springer.
[11] Aavatsmark, I., Eigestad, G. T., Klausen, R. A., Wheeler, M. F., and Yotov, I. 2007.
Convergence of a symmetric MPFA method on quadrilateral grids. Comput. Geosci.,
11(4), 333–345. doi:10.1007/s10596-007-9056-8.

631

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 632 References

[12] Aavatsmark, I., Eigestad, G. T., Mallison, B. T., and Nordbotten, J. M. 2008. A
compact multipoint flux approximation method with improved robustness. Numer.
Meth. Partial Diff. Eqs., 24(5), 1329–1360. doi:10.1002/num.20320.
[13] Aavatsmark, I., and Klausen, R. 2003. Well index in reservoir simulation for slanted
and slightly curved wells in 3D grids. SPE J., 8(01), 41–48. doi:10.2118/75275-PA.
[14] Abou-Kassem, J. H., Farouq-Ali, S. M., and Islam, M. R. 2006. Petroleum Reservoir
Simulations: A Basic Approach. Houston: Gulf Publishing Company.
[15] AGMG. 2012. Iterative solution with AGgregation-based algebraic MultiGrid.
http://agmg.eu. [Online; accessed July 11, 2018].
[16] Allen III, M. B., Behie, G. A., and Trangenstein, J. A. 1988. Multiphase Flow in
Porous Media: Mechanics, Mathematics, and Numerics. New York: Springer-Verlag.
[17] Alpak, F. O., Pal, M., and Lie, K.-A. 2012. A multiscale method for mod-
eling flow in stratigraphically complex reservoirs. SPE J., 17(4), 1056–1070.
doi:10.2118/140403-PA.
[18] Alvestad, J., Holing, K., Christoffersen, K., and Stava, O. 1994. Interactive mod-
elling of multiphase inflow performance of horizontal and highly deviated wells.
In: European Petroleum Computer Conference. Society of Petroleum Engineers.
doi:10.2118/27577-MS.
[19] Andersen, O. 2017. Simplified models for numerical simulation of geological CO2
storage. PhD. thesis, University of Bergen. url: http://hdl.handle.net/1956/15477.
[20] Andersen, O., Gasda, S. E., and Nilsen, H. M. 2015. Vertically averaged equations
with variable density for CO2 flow in porous media. Transp. Porous Media, 107(1),
95–127. doi:10.1007/s11242-014-0427-z.
[21] Andersen, O., Lie, K.-A., and Nilsen, H. M. 2016. An open-source toolchain
for simulation and optimization of aquifer-wide CO2 storage. Energy Procedia,
86(Jan.), 324–333. doi:10.1016/j.egypro.2016.01.033.
[22] Andersen, O., Nilsen, H. M., and Raynaud, X., 2017a. Coupled geomechanics and
flow simulation on corner-point and polyhedral grids. In: SPE Reservoir Simulation
Conference. doi:10.2118/182690-MS.
[23] Andersen, O., Nilsen, H. M., and Raynaud, X. 2017b. Virtual element method for
geomechanical simulations of reservoir models. Comput. Geosci., 21(5–6), 877–
893. doi:10.1007/s10596-017-9636-1.
[24] Appleyard, J. R., and Cheshire, I. M. 1983. Nested factorization. In: SPE Reservoir
Simulation Symposium. doi:10.2118/12264-MS.
[25] Arbogast, T., Cowsar, L. C., Wheeler, M. F., and Yotov, I. 2000. Mixed finite element
methods on nonmatching multiblock grids. SIAM J. Num. Anal., 37(4), 1295–1315.
doi:10.1137/S0036142996308447.
[26] Arbogast, T., Dawson, C. N., Keenan, P. T., Wheeler, M. F., and Yotov, I. 1998.
Enhanced cell-centered finite differences for elliptic equations on general geometry.
SIAM J. Sci. Comp., 19(2), 404–425. doi:10.1137/S1064827594264545.
[27] Ates, H., Bahar, A., El-Abd, S., et al. 2005. Ranking and upscaling of geostatistical
reservoir models using streamline simulation: A field case study. SPE Res. Eval.
Eng., 8(1), 22–32. doi:10.2118/81497-PA.
[28] Aziz, K., and Settari, A. 1979. Petroleum Reservoir Simulation. London, New York:
Elsevier Applied Science Publishers.
[29] Baker, L. E. 1988. Three-phase relative permeability correlations. In: SPE Enhanced
Oil Recovery Symposium. doi:10.2118/17369-MS.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 References 633

[30] Bao, K., Lie, K.-A., Møyner, O., and Liu, M. 2017. Fully implicit simula-
tion of polymer flooding with MRST. Comput. Geosci., 21(5–6), 1219–1244.
doi:10.1007/s10596-017-9624-5.
[31] Barker, J., and Thibeau, S. 1997. A critical review of the use of pseudorela-
tive permeabilities for upscaling. SPE Reservoir Engineering, 12(2), 138–143.
doi:10.2118/35491-PA.
[32] Batycky, R. P., Thieles, M. R., Baker, R. O., and Chugh, S. H. 2008. Revisiting
reservoir flood-surveillance methods using streamlines. SPE Res. Eval. Eng., 11(2),
387–394. doi:10.2118/95402-PA.
[33] Baxendale, D., Rasmussen, A., Rustad, A. B., et al. 2018. Open porous media: Flow
documentation manual. 2018-10 Rev-2 edn. http://opm-project.org.
[34] Bear, J. 1988. Dynamics of Fluids in Porous Media. Mineola, NY: Dover.
[35] Bear, J. 2007. Hydraulics of Groundwater. Mineola, NY: Dover.
[36] Bear, J., and Bachmat, Y. 1990. Introduction to Modeling of Transport Phenomena
in Porous Media. Dordrecht: Springer.
[37] Begg, S. H., Carter, R. R., and Dranfield, P. 1989. Assigning effective values to
simulator gridblock parameters for heterogeneous reservoirs. SPE Res. Eng., 4(4),
455–463. doi:10.2118/16754-PA.
[38] Beirão da Veiga, L., Brezzi, F., Cangiani, A., et al. 2013. Basic principles
of virtual element methods. Math. Mod. Meth. Appl. Sci., 23(01), 199–214.
doi:10.1142/S0218202512500492.
[39] Beirão da Veiga, L., Brezzi, F., Marini, L. D., and Russo, A. 2014. The hitchhiker’s
guide to the virtual element method. Math. Mod. Meth. Appl. Sci., 24(08), 1541–
1573. doi:10.1142/S021820251440003X.
[40] Beirao da Veiga, L., Lipnikov, K., and Manzini, G. 2014. The Mimetic Finite
Difference Method for Elliptic Problems. New York: Springer.
[41] Benesoussan, A., Lions, J.-L., and Papanicolaou, G. 1978. Asymptotic Analysis for
Periodic Structures. Amsterdam: Elsevier Science Publishers.
[42] Berge, R. L. 2016. Unstructured PEBI grids adapting to geological feautres in sub-
surface reservoirs. M.Sc. thesis, Norwegian University of Science and Technology.
http://hdl.handle.net/11250/2411565.
[43] Bischof, C. H., Bücker, H. M., Lang, B., Rasch, A., and Vehreschild, A. 2002.
Combining source transformation and operator overloading techniques to compute
derivatives for MATLAB programs. Proceedings of the Second IEEE International
Workshop on Source Code Analysis and Manipulation (SCAM 2002). Los Alamitos,
CA, pp. 65–72 doi:10.1109/SCAM.2002.1134106.
[44] Blunt, M. J. 2017. Multiphase Flow in Permeable Media: A Pore-Scale Perspective.
Cambridge: Cambridge University Press.
[45] Braess, D. 1997. Finite Elements: Theory, Fast Solvers, and Applications in Solid
Mechanics. Cambridge: Cambridge University Press.
[46] Branets, L., Ghai, S. S., Lyons, S. L., and Wu, X.-H. 2009a. Efficient and accurate
reservoir modeling using adaptive gridding with global scale up. In: SPE Reservoir
Simulation Symposium. doi:10.2118/118946-MS.
[47] Branets, L. V., Ghai, S. S., Lyons, S. L., and Wu, X.-H. 2009b. Challenges and
technologies in reservoir modeling. Commun. Comput. Phys., 6(1), 1–23.
[48] Branets, L., Kubyak, V., Kartasheva, E., Shmyrov, L., and Kandybor, D. 2015.
Capturing geologic complexity in simulation grid. In: SPE Reservoir Simulation
Symposium. doi:10.2118/173270-MS.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 634 References

[49] Brenier, Y., and Jaffré, J. 1991. Upstream differencing for multiphase flow in
reservoir simulation. SIAM J. Numer. Anal., 28(3), 685–696. doi:10.1137/0728036.
[50] Brenner, S. C., and Scott, L. R. 2007. The Mathematical Theory of Finite Element
Methods. 3rd edn. New York: Springer-Verlag.
[51] Brewer, M., Camilleri, D., Ward, S., and Wong, T. 2015. Generation of hybrid grids
for simulation of complex, unstructured reservoirs by a simulator with MPFA. In:
SPE Reservoir Simulation Symposium. doi:10.2118/173191-MS.
[52] Brezzi, F., Lipnikov, K., and Simoncini, V. 2005. A family of mimetic finite
difference methods on polygonial and polyhedral meshes. Math. Models Methods
Appl. Sci., 15, 1533–1553. doi:10.1142/S0218202505000832.
[53] Brezzi, F., and Fortin, M. 1991. Mixed and Hybrid Finite Element Methods. New
York: Springer-Verlag.
[54] Brooks, R. H., and Corey, A. T. 1966. Properties of porous media affecting fluid
flow. J. Irrigation Drainage Div., 92(2), 61–90.
[55] Buckingham, E. 1907. Studies on the Movement of Soil Moisture. Bul-
letin, no. 38. United States. Bureau of Soils. https://archive.org/details/
studiesonmovemen38buck.
[56] Buckley, S. E., and Leverett, M. C. 1942. Mechanism of fluid displacement in sands.
Trans. AIME, 146(01), 107–116. doi:10.2118/942107-G.
[57] Caers, J. 2005. Petroleum Geostatistics. Richardson, TX: Society of Petroleum
Engineers.
[58] Cao, H. 2002. Development of techniques for general purpose simulators. PhD.
thesis, Stanford University.
[59] Castellini, A., Edwards, M. G., and Durlofsky, L. J. 2000. Flow based modules for
grid generation in two and three dimensions. In: ECMOR VII – 7th European Con-
ference on the Mathematics of Oil Recovery. doi:10.3997/2214-4609.201406120.
[60] Cayuga Research. ADMAT. www.cayugaresearch.com/admat.html. [Online;
accessed Jul 11, 2018].
[61] Chavent, G., and Jaffré, J. 1982. Mathematical Models and Finite Elements for
Reservoir Simulation. Amsterdam: North Holland.
[62] Chavent, G., and Jaffré, J. 1986. Mathematical Models and Finite Elements for
Reservoir Simulation: Single Phase, Multiphase and Multicomponent Flows through
Porous Media. Amsterdam: Elsevier.
[63] Chen, Y., and Durlofsky, L. J. 2006. Adaptive local-global upscaling for general
flow scenarios in heterogeneous formations. Transport Porous Media, 62, 157–182.
doi:10.1007/s11242-005-0619-7.
[64] Chen, Y., Durlofsky, L. J., Gerritsen, M., and Wen, X. H. 2003. A coupled local-
global upscaling approach for simulating flow in highly heterogeneous formations.
Adv. Water Resour., 26(10), 1041–1060. doi:10.1016/S0309-1708(03)00101-5.
[65] Chen, Z. 2000. Formulations and numerical methods of the black oil model in porous
media. SIAM J. Numer. Anal., 38(2), 489–514. doi:10.1137/S0036142999304263.
[66] Chen, Z. 2007. Reservoir Simulation: Mathematical Techniques in Oil Recovery.
Philadelphia: Society for Industrial and Applied Mathematics.
[67] Chen, Z., and Ewing, R. E. 1997. Comparison of various formulations of three-
phase flow in porous media. J. Comput. Phys., 132(2), 362–373. doi:10.1006/jcph.
1996.5641.
[68] Chen, Z., and Hou, T. Y. 2003. A mixed multiscale finite element method
for elliptic problems with oscillating coefficients. Math. Comp., 72, 541–576.
doi:10.1090/S0025-5718-02-01441-2.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 References 635

[69] Chen, Z., Huan, G., and Ma, Y. 2006. Computational Methods for Multiphase Flows
in Porous Media. Philadelphia: Society of Industrial and Applied Mathematics.
doi:10.1137/1.9780898718942.
[70] Christie, M. A. 1996. Upscaling for reservoir simulation. J. Pet. Tech., 48(11), 1004–
1010. doi:10.2118/37324-MS.
[71] Christie, M. A., and Blunt, M. J. 2001. Tenth SPE comparative solution project:
A comparison of upscaling techniques. SPE Reservoir Eval. Eng., 4, 308–317.
doi:10.2118/72469-PA.
[72] Cirpka, O. A., Frind, E. O., and Helmig, R. 1999. Streamline-oriented grid genera-
tion for transport modelling in two-dimensional domains including wells. Adv. Water
Resour., 22(7), 697–710. doi:10.1016/S0309-1708(98)00050-5.
[73] Coats, K. H. 2000. A note on IMPES and some IMPES-based simulation models.
SPE J., 05(03), 245–251. doi:10.2118/65092-PA.
[74] Cordes, C., and Kinzelbach, W. 1992. Continous groundwater velocity fields and
path lines in linear, bilinear, and trilinear finite elements. Water Resour. Res., 28(11),
2903–2911. doi:10.1029/92WR01686.
[75] Courant, R., Friedrichs, K., and Lewy, H. 1928. Über die partiellen Dif-
ferenzengleichungen der mathematischen Physik. Math. Ann., 100(1), 32–74.
doi:10.1007/BF01448839.
[76] Dafermos, C. M. 2010. Hyperbolic Conservation Laws in Continuum Physics.
Berlin, Heidelberg: Springer.
[77] Darcy, H. P. G. 1856. Les Fontaines Publiques de la Ville de Dijon. Paris: Dalmont.
[78] Datta-Gupta, A., and King, M. J. 1995. A semianalytic approach to tracer flow
modeling in heterogeneous permeable media. Adv. Water Resour., 18, 9–24.
doi:10.1016/0309-1708(94)00021-V.
[79] Datta-Gupta, A., and King, M. J. 2007. Streamline Simulation: Theory and Practice.
Richardson, TX: Society of Petroleum Engineers.
[80] DeBaun, D., et al. 2005. An extensible architecture for next generation scal-
able parallel reservoir simulation. In: SPE Reservoir Simulation Symposium.
doi:10.2118/93274-MS.
[81] Demidov, D. 2017 (Oct.). amgcl-sdd-scaling. https://zenodo.org/record/1002948#
.XE_CAFxKiUk.
[82] Deutsch, C. V., and Journel, A. G. 1998. GSLIB: Geostatistical Software Library
and User’s Guide. 2nd edn. New York: Oxford University Press.
[83] Ding, X. Y., and Fung, L. S. K. 2015. An unstructured gridding method for
simulating faulted reservoirs populated with complex wells. In: SPE Reservoir
Simulation Symposium. doi:10.2118/173243-MS.
[84] Dogru, A. H., Fung, L. S.-K., Middya, U., Al-Shaalan, T., and Pita, J. A. 2009.
A next-generation parallel reservoir simulator for giant reservoirs. In: SPE/EAGE
Reservoir Characterization & Simulation Conference. doi:10.2118/119272-MS.
[85] Dogru, A. H., Fung, L. S. K., Middya, U., et al. 2011. New frontiers in
large scale reservoir simulation. In: SPE Reservoir Simulation Symposium.
doi:10.2118/142297-MS.
[86] Douglas Jr., J., Peaceman, D. W., and Rachford Jr., H. H. 1959. A method for
calculating multi-dimensional immiscible displacement. Petrol. Trans. AIME, 216,
297–308.
[87] Duarte, A. C. 2016. Contributions to production optimization of oil reservoirs.
PhD. thesis, Norwegian University of Science and Technology. http://hdl.handle.net/
11250/2383090.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 636 References

[88] Duff, I. S. 2004. MA57–A code for the solution of sparse symmetric def-
inite and indefinite systems. ACM Trans. Math. Softw., 30(2), 118–144.
doi:10.1145/992200.992202.
[89] Duff, I. S., and Reid, J. K. 1983. The multifrontal solution of indefi-
nite sparse symmetric linear. ACM Trans. Math. Software, 9(3), 302–325.
doi:10.1145/356044.356047.
[90] Durlofsky, L. J. 1991. Numerical calculations of equivalent gridblock permeability
tensors for heterogeneous porous media. Water Resour. Res., 27(5), 699–708.
doi:10.1029/91WR00107.
[91] Durlofsky, L. J. 2003. Upscaling of Geocellular Models for Reservoir Flow Simula-
tion: A Review of Recent Progress. Presented at 7th International Forum on Reservoir
Simulation Bühl/Baden-Baden, Germany, June 23–27, 2003.
[92] Durlofsky, L. J. 2005. Upscaling and Gridding of Fine Scale Geological Models for
Flow Simulation. Presented at 8th International Forum on Reservoir Simulation Iles
Borromees, Stresa, Italy, June 20–24, 2005.
[93] Durlofsky, L. J., Jones, R. C., and Milliken, W. J. 1997. A nonuniform coarsening
approach for the scale-up of displacement processes in heterogeneous porous media.
Adv. Water Resour., 20(5-6), 335–347. doi:10.1016/S0309-1708(96)00053-X.
[94] Edwards, M. G., and Rogers, C. F. 1994. A flux continuous scheme for the full tensor
pressure equation. In: ECMOR IV – 4th European Conference on the Mathematics
of Oil Recovery. doi:10.3997/2214-4609.201411178.
[95] Efendiev, Y., and Hou, T. Y. 2009. Multiscale Finite Element Methods: Theory and
Applications. New York: Springer-Verlag.
[96] Eigestad, G. T., and Klausen, R. A. 2005. On the convergence of the multi-point flux
approximation o-method: Numerical experiments for discontinuous permeability.
Num. Meth. Partial Diff. Eqs., 21(6), 1079–1098. doi:10.1002/num.20079.
[97] Eigestad, G., Dahle, H., Hellevang, B., Riis, F., Johansen, W., and Øian, E. 2009.
Geological modeling and simulation of CO2 injection in the Johansen formation.
Comput. Geosci., 13(4), 435–450. doi:10.1007/s10596-009-9153-y.
[98] Eikemo, B., Lie, K.-A., Dahle, H. K., and Eigestad, G. T. 2009. Dis-
continuous Galerkin methods for transport in advective transport in single-
continuum models of fractured media. Adv. Water Resour., 32(4), 493–506.
doi:10.1016/j.advwatres.2008.12.010.
[99] Ertekin, T., Abou-Kassem, J. H., and King, G. R. 2001. Basic Applied Reservoir
Simulation. Richardson, TX: Society of Petroleum Engineers.
[100] Evazi, M., and Mahani, H. 2010a. Generation of Voronoi grid based on vorticity for
coarse-scale modeling of flow in heterogeneous formations. Transp. Porous Media,
83(3). doi:10.1007/s11242-009-9458-2.
[101] Evazi, M., and Mahani, H. 2010b. Unstructured-coarse-grid generation using
background-grid approach. SPE J., 15(2), 326–340. doi:10.2118/120170-PA.
[102] Ewing, R. E., Lazarov, R. D., Lyons, S. L., Papavassiliou, D. V., Pasciak, J., and Qin,
G. 1999. Numerical well model for non-Darcy flow through isotropic porous media.
Comput. Geosci., 3(3-4), 185–204. doi:10.1023/A:1011543412675.
[103] Eymard, R., Gallouët, T., and Herbin, R. 1999. Convergence of finite volume
schemes for semilinear convection diffusion equations. Numer. Math., 82(1), 91–
116. doi:10.1007/s002110050412.
[104] Eymard, R., Gallouët, T., and Herbin, R. 2001. Finite volume approximation of
elliptic problems and convergence of an approximate gradient. App. Numer. Math.,
37(1-2), 31–53. doi:10.1016/S0168-9274(00)00024-6.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 References 637

[105] Eymard, R., Guichard, C., and Herbin, R. 2012a. Small-stencil 3D schemes for
diffusive flows in porous media. ESAIM: Math. Model. Numer. Anal., 46(2), 265–
290. doi:10.1051/m2an/2011040.
[106] Eymard, R., Guichard, C., Herbin, R., and Masson, R. 2012b. Vertex-centred
discretization of multiphase compositional Darcy flows on general meshes. Comput.
Geosci., 16(4), 987–1005. doi:10.1007/s10596-012-9299-x.
[107] Fanchi, J. R. 2005. Principles of Applied Reservoir Simulation. Houston, TX: Gulf
Professional Publishing.
[108] Farmer, C. L. 2002. Upscaling: A review. Int. J. Numer. Meth. Fluids, 40(1–2), 63–
78. doi:10.1002/fld.267.
[109] Fayers, F. J., and Matthews, J. D. 1984. Evaluation of normalized Stone’s meth-
ods for estimating three-phase relative permeabilities. SPE J., 24(2), 224–232.
doi:10.2118/11277-PA.
[110] Fink, M. 2007. Automatic Differentiation for Matlab. MATLAB Central. https://
tinyurl.com/ycvp6n8a. [Online; accessed July 11, 2018].
[111] Floris, F. J. T., Bush, M. D., Cuypers, M., Roggero, F., and Syversveen, A. R. 2001.
Methods for quantifying the uncertainty of production forecasts: a comparative
study. Petroleum Geoscience, 7(S), S87–S96. doi:10.1144/petgeo.7.S.S87.
[112] Forth, S. A. 2006. An efficient overloaded implementation of forward mode
automatic differentiation in MATLAB. ACM Trans. Math. Software, 32(2), 195–222.
doi:10.1145/1141885.1141888.
[113] Fung, L. S. K., Ding, X. Y., and Dogru, A. H. 2014. Unconstrained Voronoi grids
for densely spaced complex wells in full-field reservoir simulation. SPE J., 19(5),
803–815. doi:10.2118/163648-PA.
[114] Gain, A. L., Talischi, C., and Paulino, G. H. 2014. On the virtual element method
for three-dimensional linear elasticity problems on arbitrary polyhedral meshes.
Comput. Meth. App. Mech. Engng., 282, 132–160. doi:10.1016/j.cma.2014.05.005.
[115] Gao, M. 2014. Reservoir and Surface Facilities Coupled through Partially and Fully
Implicit Approaches. M.Sc. thesis, Texas A & M University. http://hdl.handle.net/
1969.1/154076.
[116] Garcia, M. H., Journel, A. G., and Aziz, K. 1992. Automatic grid genera-
tion for modeling reservoir heterogeneities. SPE Reservoir Eng., 7(6), 278–284.
doi:10.2118/21471-PA.
[117] Gerritsen, M., and Lambers, J. V. 2008. Integration of local-global upscaling and grid
adaptivity for simulation of subsurface flow in heterogeneous formations. Comput.
Geosci., 12(2), 193–208. doi:10.1007/s10596-007-9078-2.
[118] Godunov, S. K. 1959. A difference method for numerical calculation of discontinu-
ous solutions of the equations of hydrodynamics. Mat. Sb. (N.S.), 47 (89), 271–306.
[119] Gries, S., Stüben, K., Brown, G. L., Chen, D., and Collins, D. A. 2014. Precondition-
ing for efficiently applying algebraic multigrid in fully implicit reservoir simulations.
SPE J., 19(04), 726–736. doi:10.2118/163608-PA.
[120] Guérillot, D., Rudkiewicz, J. L., Ravenne, C., Renard, D., and Galli, A.
1990. An integrated model for computer aided reservoir description: From out-
crop study to fluid flow simulations. Oil Gas Sci. Technol., 45(1), 71–77.
doi:10.2516/ogst:1990005.
[121] Gunasekera, D., Cox, J., and Lindsey, P. 1997. The generation and applica-
tion of K-orthogonal grid systems. In: SPE Reservoir Simulation Symposium.
doi:10.2118/37998-MS.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 638 References

[122] Hægland, H., Dahle, H. K., Lie, K.-A., and Eigestad, G. T. 2006. Adaptive streamline
tracing for streamline simulation on irregular grids. In: Binning, P. J., Engesgaard,
P. K., Dahle, H. K., Pinder, G. F., and Gray, W. G. (eds), XVI International
Conference on Computational Methods in Water Resources. http://proceedings
.cmwr-xvi.org/.
[123] Hales, H. B. 1996. A method for creating 2-d orthogonal grids which conform to
irregular shapes. SPE J, 1(2), 115–124. doi:10.2118/35273-PA.
[124] Hauge, V. L. 2010. Multiscale methods and flow-based gridding for flow and
transport in porous media. PhD. thesis, Norwegian University of Science and
Technology. http://hdl.handle.net/11250/258800.
[125] Hauge, V. L., Lie, K.-A., and Natvig, J. R. 2012. Flow-based coarsening for
multiscale simulation of transport in porous media. Comput. Geosci., 16(2), 391–
408. doi:10.1007/s10596-011-9230-x.
[126] He, C., and Durlofsky, L. J. 2006. Structured flow-based gridding and upscal-
ing for modeling subsurface flow. Adv. Water Resour., 29(12), 1876–1892.
doi:10.1016/j.advwatres.2005.12.012.
[127] He, Z., Parikh, H., Datta-Gupta, A., Perez, J., and Pham, T. 2004. Identifying
reservoir compartmentalization and flow barriers from primary production using
streamline diffusive time of flight. SPE J., 7(3), 238–247. doi:10.2118/88802-PA.
[128] Heinemann, Z. E., Brand, C. W., and Munka, M. 1991. Modeling reservoir geometry
with irregular grids. SPE Res. Eng., 6(2), 225–232. doi:10.2118/18412-PA.
[129] Helmig, R. 1997. Multiphase Flow and Transport Processes in the Subsurface: A
Contribution to the Modeling of Hydrosystems. Berlin, Heidelberg: Springer.
[130] Hilden, S. T., Møyner, O., Lie, K.-A., and Bao, K. 2016. Multiscale simulation
of polymer flooding with shear effects. Transp. Porous Media, 113(1), 111–135.
doi:10.1007/s11242-016-0682-2.
[131] Hilden, S. T. 2016. Upscaling of water-flooding scenarios and modeling of polymer
flow. PhD. thesis, Norwegian University of Science and Technology. http://hdl
.handle.net/11250/2388331.
[132] Hilden, S. T., Lie, K.-A., and Raynaud, X. 2014. Steady state upscaling of polymer
flooding. In: ECMOR XIV – 14th European Conference on the Mathematics of Oil
Recovery. doi:10.3997/2214-4609.20141802.
[133] Holden, L., and Nielsen, B. F. 2000. Global upscaling of permeability in hetero-
geneous reservoirs; the output least squares (OLS) method. Trans. Porous Media,
40(2), 115–143. doi:10.1023/A:1006657515753.
[134] Holden, H., and Risebro, N. H. 2002. Front Tracking for Hyperbolic Conservation
Laws. New York: Springer-Verlag.
[135] Hornung, U. 1997. Homogenization and Porous Media. New York: Springer-Verlag.
[136] Hoteit, H., and Firoozabadi, A. 2008. Numerical modeling of two-phase flow in
heterogeneous permeable media with different capillarity pressures. Adv. Water
Resour., 31(1), 56–73. doi:10.1016/j.advwatres.2007.06.006.
[137] Hubbert, M. K. 1956. Darcy’s law and the field equations of the flow of underground
fluids. Petrol. Trans., AIME, 207, 22–239.
[138] Hui, M.-H. R., Karimi-Fard, M., Mallison, B., and Durlofsky, L. J. 2018. A
general modeling framework for simulating complex recovery processes in fractured
reservoirs at different resolutions. SPE J., 22(1), 20–29. doi:10.2118/182621-MS.
[139] Huseby, O., Sagen, J., and Dugstad, Ø. 2012. Single well chemical tracer tests –
Fast and accurate simulations. In: SPE EOR Conference at Oil and Gas West Asia.
doi:10.2118/155608-MS.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 References 639

[140] Idrobo, E. A., Choudhary, M. K., and Datta-Gupta, A. 2000. Swept volume calcula-
tions and ranking of geostatistical reservoir models using streamline simulation. In:
SPE/AAPG Western Regional Meeting. doi:10.2118/62557-MS.
[141] Iemcholvilert, S. 2013. A Research on production optimization of coupled surface
and subsurface model. M.Sc. thesis, Texas A & M University. http://hdl.handle.net/
1969.1/151189.
[142] Islam, M. R., Hossain, M. E., Moussavizadegan, S. H., Mustafiz, S., and Abou-
Kassem, J. H. 2016. Advanced Petroleum Reservoir Simulation: Towards Develop-
ing Reservoir Emulators. John Wiley & Sons, Inc. doi:10.1002/9781119038573.
[143] Islam, M. R., Mousavizadegan, S. H., Mustafiz, S., and Abou-Kassem, J. H. 2010.
Advanced Petroleum Reservoir Simulations. Hoboken, NJ: John Wiley & Sons, Inc.
[144] Izgec, O., Sayarpour, M., and Shook, G. M. 2011. Maximizing volumetric sweep
efficiency in waterfloods with hydrocarbon f-φ curves. J. Petrol. Sci. Eng., 78(1),
54–64. doi:10.1016/j.petrol.2011.05.003.
[145] Jansen, J. D. 2017. Nodal Analysis of Oil and Gas Wells-System Modeling and
Numerical Implementation. Richardson, TX: Society of Petroleum Engineers.
[146] Jenny, P., Lee, S. H., and Tchelepi, H. A. 2003. Multi-scale finite-volume method
for elliptic problems in subsurface flow simulation. J. Comput. Phys., 187, 47–67.
doi:10.1016/S0021-9991(03)00075-5.
[147] Jenny, P., Tchelepi, H. A., and Lee, S. H. 2009. Unconditionally convergent
nonlinear solver for hyperbolic conservation laws with S-shaped flux functions. J.
Comput. Phys., 228(20), 7497–7512. doi:10.1016/j.jcp.2009.06.032.
[148] Jenny, P., Wolfsteiner, C., Lee, S. H., and Durlofsky, L. J. 2002. Modeling flow in
geometrically complex reservoirs using hexahedral multiblock grids. SPE J., 7(2).
doi:10.2118/78673-PA.
[149] Jikov, V. V., Kozlov, S. M., and Oleinik, O. A. 1994. Homogenization of Differential
Operators and Integral Functionals. New York: Springer-Verlag.
[150] Jimenez, E., Sabir, K., Datta-Gupta, A., and King, M. J. 2007. Spatial error and
convergence in streamline simulation. SPE J., 10(3), 221–232. doi:10.2118/92873-
MS.
[151] Journel, A. G., Deutsch, C. V., and Desbarats, A. J. 1986. Power averaging for block
effective permeability. In: SPE California Regional Meeting. doi:10.2118/15128-
MS.
[152] Karimi-Fard, M., and Durlofsky, L. J. 2012. Accurate resolution of near-well effects
in upscaled models using flow-based unstructured local grid refinement. SPE J.,
17(4), 1084–1095. doi:10.2118/141675-PA.
[153] Karimi-Fard, M., and Durlofsky, L. J. 2016. A general gridding, discretiza-
tion, and coarsening methodology for modeling flow in porous formations with
discrete geological features. Adv. Water Resour., 96(Supplement C), 354–372.
doi:10.1016/j.advwatres.2016.07.019.
[154] Karypis, G., and Kumar, V. 1998. A fast and high quality multilevel
scheme for partitioning irregular graphs. SIAM J. Sci. Comp., 20(1), 359–392.
doi:10.1137/S1064827595287997.
[155] Keilegavlen, E., Kozdon, J. E., and Mallison, B. T. 2012. Multidimensional upstream
weighting for multiphase transport on general grids. Comput. Geosci., 16, 1021–
1042. doi:10.1007/s10596-012-9301-7.
[156] Keilegavlen, E., and Nordbotten, J. M. 2017. Finite volume methods for
elasticity with weak symmetry. Int. J. Numer. Meth. Eng., 112(8), 939–962.
doi:10.1002/nme.5538.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 640 References

[157] Keilegavlen, E., Nordbotten, J. M., and Aavatsmark, I. 2009. Sufficient criteria are
necessary for monotone control volume methods. Appl. Math. Letters, 22(8), 1178–
1180. doi:10.1016/j.aml.2009.01.048.
[158] Kenyon, D. 1987. Third SPE comparative solution project: Gas cycling of retrograde
condensate reservoirs. J. Petrol. Tech., 39(08), 981–997. doi:10.2118/12278-PA.
[159] Killough, J. E. 1995. Ninth SPE comparative solution project: A reexamination of
black-oil simulation. In: SPE Reservoir Simulation Symposium. doi:10.2118/29110-
MS.
[160] King, M. J. 2007. Recent advances in upgridding. Oil Gas Sci. Technol. – Rev. IFP,
62(2), 195–205. doi:10.2516/ogst:2007017.
[161] King, M. J., and Mansfield, M. 1999. Flow simulation of geologic models. SPE Res.
Eval. Eng., 2(4), 351–367. doi:10.2118/57469-PA.
[162] King, M. J., Burn, K. S., Muralidharan, P. W. V., et al. 2006. Optimal coarsening
of 3D reservoir models for flow simulation. SPE Reserv. Eval. Eng., 9(4), 317–334.
doi:10.2118/95759-PA.
[163] King, M. J., MacDonald, D. G., Todd, S. P., and Leung, H. 1998. Application of
novel upscaling approaches to the Magnus and Andrew reservoirs. In: European
Petroleum Conference. doi:10.2118/50643-MS.
[164] King, M. J., and Datta-Gupta, A. 1998. Streamline simulation: A current perspective.
In Situ, 22(1), 91–140.
[165] Kippe, V., Aarnes, J. E., and Lie, K.-A. 2008. A comparison of multiscale methods
for elliptic problems in porous media flow. Comput. Geosci., 12(3), 377–398.
doi:10.1007/s10596-007-9074-6.
[166] Klausen, R. A., Rasmussen, A. F., and Stephansen, A. 2012. Velocity interpolation
and streamline tracing on irregular geometries. Comput. Geosci., 16, 261–276.
doi:10.1007/s10596-011-9256-0.
[167] Klausen, R. A., and Winther, R. 2006. Robust convergence of multi point flux
approximation on rough grids. Numer. Math., 104(3), 317–337. doi:10.1007/s00211-
006-0023-4.
[168] Klemetsdal, Ø. S. 2016. The virtual element method as a common framework for
finite element and finite difference methods – numerical and theoretical analysis.
M.Sc. thesis, Norwegian University of Science and Technology. http://hdl.handle
.net/11250/2405996.
[169] Klemetsdal, Ø. S., Berge, R. L., Lie, K.-A., Nilsen, H. M., and Møyner, O. 2017.
Unstructured gridding and consistent discretizations for reservoirs with faults and
complex wells. In: SPE Reservoir Simulation Conference. doi:10.2118/182679-MS.
[170] Krogstad, S., Lie, K.-A., Møyner, O., Nilsen, H. M., Raynaud, X., and Skaflestad,
B. 2015. MRST-AD – An open-source framework for rapid prototyping and eval-
uation of reservoir simulation problems. In: SPE Reservoir Simulation Symposium.
doi:10.2118/173317-MS.
[171] Krogstad, S., Lie, K.-A., Nilsen, H. M., Berg, C. F., and Kippe, V. 2017. Efficient
flow diagnostics proxies for polymer flooding. Comput. Geosci., 21(5-6), 1203–
1218. doi:10.1007/s10596-017-9681-9.
[172] Krogstad, S., Raynaud, X., and Nilsen, H. M. 2016. Reservoir management opti-
mization using well-specific upscaling and control switching. Comput. Geosci.,
20(3), 695–706. doi:10.1007/s10596-015-9497-4.
[173] Kružkov, S. N. 1970. First order quasilinear equations in several
independent variables. Mathematics of the USSR-Sbornik, 10(2), 217.
doi:10.1070/SM1970v010n02ABEH002156.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 References 641

[174] Kurganov, A., Noelle, S., and Petrova, G. 2001. Semidiscrete central-upwind
schemes for hyperbolic conservation laws and Hamilton–Jacobi equations. SIAM
J. Sci. Comp., 23(3), 707–740. doi:10.1137/S1064827500373413.
[175] Kwok, F., and Tchelepi, H. 2007. Potential-based reduced Newton algorithm for
nonlinear multiphase flow in porous media. J. Comput. Phys., 227(1), 706–727.
doi:10.1016/j.jcp.2007.08.012.
[176] Lake, L. W. 1989. Enhanced Oil Recovery. Upper Saddle River, NJ: Prentice-Hall.
[177] Lake, L. W. (ed). 2007. Petroleum Engineering Handbook. Richardson, TX: Society
of Petroleum Engineers.
[178] Lax, P., and Wendroff, B. 1960. Systems of conservation laws. Comm. Pure Appl.
Math., 13(2), 217–237. doi:10.1002/cpa.3160130205.
[179] Le Potier, C. 2009. A nonlinear finite volume scheme satisfying maximum and
minimum principles for diffusion operators. Int. J. Finite Vol., 6(2), 1–20.
[180] Lee, S. H., Jenny, P., and Tchelepi, H. A. 2002. A finite-volume method with
hexahedral multiblock grids for modeling flow in porous media. Comput. Geosci.,
6(3-4), 353–379. doi:10.1023/A:1021287013566.
[181] Leeuwenburgh, O., and Arts, R. 2014. Distance parameterization for efficient
seismic history matching with the ensemble Kalman Filter. Comput. Geosci., 18(3–
4), 535–548. doi:10.1007/s10596-014-9434-y.
[182] Leeuwenburgh, O., Peters, E., and Wilschut, F. 2011. Towards an integrated
workflow for structural reservoir model updating and history matching. In: SPE
EUROPEC/EAGE Annual Conference and Exhibition. doi:10.2118/143576-MS.
[183] LeVeque, R. J. 2002. Finite Volume Methods for Hyperbolic Problems. Cambridge:
Cambridge University Press.
[184] Leverett, M. C. 1941. Capillary behavior in porous solids. Trans. AIME, 142, 159–
172. doi:10.2118/941152-G.
[185] Li, X., and Zhang, D. 2014. A backward automatic differentiation framework for
reservoir simulation. Comput. Geosci., 18(6), 1009–1022. doi:10.1007/s10596-014-
9441-z.
[186] Lie, K.-A. 2018. On Holden’s seven guidelines for scientific computing and
development of open-source community software. In: Gesztesy, F., et al. (eds),
Non-Linear Partial Differential Equations, Mathematical Physics, and Stochastic
Analysis: The Helge Holden Anniversary Volume. European Mathematical Society
Publishing House. pp. 389–422.
[187] Lie, K.-A., Kedia, K., Skaflestad, B., et al. 2017a. A general non-uniform coarsening
and upscaling framework for reduced-order modeling. In: SPE Reservoir Simulation
Conference. doi:10.2118/182681-MS.
[188] Lie, K.-A. 2015a. JOLT 1: Introduction to MRST. SINTEF ICT / ICME, Stanford
University. www.sintef.no/mrst-jolts.
[189] Lie, K.-A. 2015b. JOLT 2: Grids and petrophysical data. SINTEF ICT / ICME,
Stanford University. www.sintef.no/mrst-jolts.
[190] Lie, K.-A., Mykkeltvedt, T. S., and Møyner, O. 2018. Fully implicit WENO schemes
on stratigraphic and fully unstructured grids. In: ECMOR XVI – 16th European
Conference on the Mathematics of Oil Recovery.
[191] Lie, K.-A., Møyner, O., and Natvig, J. R. 2017a. Use of multiple multiscale
operators to accelerate simulation of complex geomodels. SPE J., 22(6), 1929–1945.
doi:10.2118/182701-PA.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 642 References

[192] Lie, K.-A., Krogstad, S., Ligaarden, I. S., et al. 2012b. Open-source MATLAB
implementation of consistent discretisations on complex grids. Comput. Geosci., 16,
297–322. doi:10.1007/s10596-011-9244-4.
[193] Lie, K.-A., Nilsen, H. M., Andersen, O., and Møyner, O. 2016. A simulation
workflow for large-scale CO2 storage in the Norwegian North Sea. Comput. Geosci.,
20(3), 607–622. doi:10.1007/s10596-015-9487-6.
[194] Lie, K.-A., Natvig, J. R., Krogstad, S., Yang, Y., and Wu, X.-H. 2014. Grid adap-
tation for the Dirichlet–Neumann representation method and the multiscale mixed
finite-element method. Comput. Geosci., 18(3), 357–372. doi:10.1007/s10596-013-
9397-4.
[195] Lie, K.-A., Møyner, O., Natvig, J. R., et al. 2017b. Successful application of
multiscale methods in a real reservoir simulator environment. Comput. Geosci.,
21(5-6), 981–998. doi:10.1007/s10596-017-9627-2.
[196] Lie, K.-A., Natvig, J. R., and Nilsen, H. M. 2012a. Discussion of dynamics and
operator splitting techniques for two-phase flow with gravity. Int. J. Numer. Anal.
Mod., 9(3), 684–700.
[197] Ligaarden, I. S. 2008. Well models for mimetic finite difference methods and
improved representation of wells in multiscale methods. M.Sc. thesis, University of
Oslo. http://urn.nb.no/URN:NBN:no-19435.
[198] Lipnikov, K., Shashkov, M., Svyatskiy, D., and Vassilevski, Y. 2007. Mono-
tone finite volume schemes for diffusion equations on unstructured triangu-
lar and shape-regular polygonal meshes. J. Comput. Phys., 227(1), 492–512.
doi:10.1016/j.jcp.2007.08.008.
[199] Lipnikov, K., Shashkov, M., and Yotov, I. 2009. Local flux mimetic finite difference
methods. Numer. Math., 112(1), 115–152. doi:10.1007/s00211-008-0203-5.
[200] Lunati, I., and Lee, S. H. 2009. An operator formulation of the multiscale finite-
volume method with correction function. Multiscale Model. Simul., 8(1), 96–109.
doi:10.1137/080742117.
[201] Mahani, H., Muggeridge, A. H., and Ashjari, M. A. 2009. Vorticity as a measure of
heterogeneity for improving coarse grid generation. Petrol. Geosci., 15(1), 91–102.
doi:10.1144/1354-079309-802.
[202] Mallison, B., Sword, C., Viard, T., Milliken, W., and Cheng, A. 2014. Unstruc-
tured cut-cell grids for modeling complex reservoirs. SPE J., 19(2), 340–352.
doi:10.2118/163642-PA.
[203] Manzocchi, T., et al. 2008. Sensitivity of the impact of geological uncertainty on
production from faulted and unfaulted shallow-marine oil reservoirs: Objectives and
methods. Petrol. Geosci., 14(1), 3–15. doi:10.1144/1354-079307-790.
[204] Matringe, S. F., and Gerritsen, M. G. 2004. On accurate tracing of streamlines. In:
SPE Annual Technical Conference and Exhibition. doi:10.2118/89920-MS.
[205] Matringe, S. F., Juanes, R., and Tchelepi, H. A. 2007. Streamline tracing on general
triangular or quadrilateral grids. SPE J., 12(2), 217–233. doi:10.2118/96411-MS.
[206] Mattax, C. C., and Dalton, R. L. (eds). 1990. Reservoir Simulation. Society of
Petroleum Engineers.
[207] McCain, Jr., W. D. 1990. The Properties of Petroleum Fluids. 2nd edn. Tulsa, OK:
PennWell Books.
[208] McIlhagga, W. 2010. Automatic Differentiation with Matlab Objects. MATLAB
Central. https://tinyurl.com/yavlcra4. [Online; accessed July 11, 2018].

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 References 643

[209] Merland, R., Caumon, G., Lévy, B., and Collon-Drouaillet, P. 2014. Voronoi
grids conforming to 3d structural features. Comput. Geosci., 18(3-4), 373–383.
doi:10.1007/s10596-014-9408-0.
[210] Mlacnik, M. J., Durlofsky, L. J., and Heinemann, Z. E. 2006. Sequentially
adapted flow-based PEBI grids for reservoir simulation. SPE J., 11(3), 317–327.
doi:10.2118/90009-PA.
[211] Møyner, O. 2012. Multiscale finite-volume methods on unstructured grids. M.Sc.
thesis, Norwegian University of Science and Technology. http://hdl.handle.net/
11250/259015.
[212] Møyner, O. 2016. Next generation multiscale methods for reservoir simulation.
PhD. thesis, Norwegian University of Science and Technology. http://hdl.handle.net/
11250/2431831.
[213] Møyner, O. 2017. Nonlinear solver for three-phase transport problems
based on approximate trust regions. Comput. Geosci., 21(5-6), 999–1021.
doi:10.1007/s10596-017-9660-1.
[214] Møyner, O., and Lie, K.-A. 2014. The multiscale finite-volume method on strati-
graphic grids. SPE J., 19(5), 816–831. doi:10.2118/163649-PA.
[215] Møyner, O., and Lie, K.-A. 2016a. A multiscale restriction-smoothed basis method
for compressible black-oil models. SPE J., 21(06), 2079–2096. doi:10.2118/173265-
PA.
[216] Møyner, O., and Lie, K.-A. 2016b. A multiscale restriction-smoothed basis method
for high contrast porous media represented on unstructured grids. J. Comput. Phys.,
304, 46–71. doi:10.1016/j.jcp.2015.10.010.
[217] Møyner, O., and Tchelepi, H. A. 2017. A multiscale restriction-smoothed basis
method for compositional models. In: SPE Reservoir Simulation Conference.
doi:10.2118/182679-MS.
[218] Møyner, O., Krogstad, S., and Lie, K.-A. 2014. The application of flow diagnostics
for reservoir management. SPE J., 20(2), 306–323. doi:10.2118/171557-PA.
[219] Muskat, M., and Wyckoff, R. D. 1937. The Flow of Homogeneous Fluids through
Porous Media. Vol. 12. New York: McGraw-Hill.
[220] Natvig, J. R., and Lie, K.-A. 2008. Fast computation of multiphase flow in
porous media by implicit discontinuous Galerkin schemes with optimal ordering of
elements. J. Comput. Phys., 227(24), 10108–10124. doi:10.1016/j.jcp.2008.08.024.
[221] Natvig, J. R., Lie, K.-A., Eikemo, B., and Berre, I. 2007. An efficient discontinuous
Galerkin method for advective transport in porous media. Adv. Water Resour.,
30(12), 2424–2438. doi:10.1016/j.advwatres.2007.05.015.
[222] Natvig, J. R., Skaflestad, B., Bratvedt, F., et al. 2011. Multiscale mimetic solvers
for efficient streamline simulation of fractured reservoirs. SPE J., 16(4), 880–888.
doi:10.2018/119132-PA.
[223] Neidinger, R. 2010. Introduction to automatic differentiation and MATLAB object-
oriented programming. SIAM Review, 52(3), 545–563. doi:10.1137/080743627.
[224] Nessyahu, H., and Tadmor, E. 1990. Nonoscillatory central differencing for hyper-
bolic conservation laws. J. Comput. Phys., 87(2), 408–463. doi:10.1016/0021-
9991(90)90260-8.
[225] Nielsen, B. F., and Tveito, A. 1998. An upscaling method for one-phase flow in
heterogeneous reservoirs; A Weighted Output Least Squares (WOLS) approach.
Comput. Geosci., 2, 92–123. doi:0.1023/A:1011541917701.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 644 References

[226] Nikitin, K., Terekhov, K., and Vassilevski, Y. 2014. A monotone nonlinear finite
volume method for diffusion equations and multiphase flows. Comput. Geos.,
18(3-4), 311–324. doi:10.1007/s10596-013-9387-6.
[227] Nilsen, H. M., Lie, K.-A., and Andersen, O. 2016b. Robust simulation of sharp-
interface models for fast estimation of CO2 trapping capacity. Comput. Geosci.,
20(1), 93–113. doi:10.1007/s10596-015-9549-9.
[228] Nilsen, H. M., Lie, K.-A., and Andersen, O. 2016a. Fully implicit simulation of
vertical-equilibrium models with hysteresis and capillary fringe. Comput. Geosci.,
20(1), 49–67. doi:10.1007/s10596-015-9547-y.
[229] Nilsen, H. M., Nordbotten, J. M., and Raynaud, X. 2018. Comparison between
cell-centered and nodal based discretization schemes for linear elasticity. Comput.
Geosci., 22(1), 233–260. doi:10.1007/s10596-017-9687-3.
[230] Nilsen, H. M., Lie, K.-A., Møyner, O., and Andersen, O. 2015b. Spill-point analysis
and structural trapping capacity in saline aquifers using MRST-co2lab. Comput.
Geosci., 75, 33–43. doi:10.1016/j.cageo.2014.11.002.
[231] Nilsen, H. M., Lie, K.-A., and Natvig, J. R. 2012. Accurate modelling of
faults by multipoint, mimetic, and mixed methods. SPE J., 17(2), 568–579.
doi:10.2118/149690-PA.
[232] Nilsen, H. M., Lie, K.-A., and Andersen, O. 2015a. Analysis of CO2 trapping capac-
ities and long-term migration for geological formations in the Norwegian North
Sea using MRST-co2lab. Comput. Geosci., 79, 15–26. doi:10.1016/j.cageo.2015.
03.001.
[233] Nordbotten, J. M. 2016. Stable cell-centered finite volume discretization for biot
equations. SIAM J. Numer. Anal., 54(2), 942–968. doi:10.1137/15M1014280.
[234] Nordbotten, J. M., and Eigestad, G. T. 2005. Discretization on quadrilateral
grids with improved monotonicity properties. J. Comput. Phys., 203(2), 744–760.
doi:10.1016/j.jcp.2004.10.002.
[235] Nordbotten, J. M., Aavatsmark, I., and Eigestad, G. T. 2007b. Monotonicity of
control volume methods. Numer. Math., 106(2), 255–288. doi:10.1007/s00211-006-
0060-z.
[236] Nordbotten, J. M., and Aavatsmark, I. 2005. Monotonicity conditions for control
volume methods on uniform parallelogram grids in homogeneous media. Comput.
Geosci., 9(1), 61–72. doi:10.1007/s10596-005-5665-2.
[237] Nordbotten, J. M. 2015. Convergence of a cell-centered finite volume dis-
cretization for linear elasticity. SIAM J. Numer. Anal., 53(6), 2605–2625.
doi:10.1137/140972792.
[238] Notay, Y. 2010. An aggregation-based algebraic multigrid method. Electron. Trans.
Numer. Anal., 37, 123–140.
[239] Nutting, P. G. 1930. Physical analysis of oil sands. AAPG Bulletin, 14(10), 1337–
1349.
[240] Obi, E., Eberle, N., Fil, A., and Cao, H. 2014. Giga cell compositional simulation.
In: IPTC 2014: International Petroleum Technology Conference. doi:10.2523/IPTC-
17648-MS.
[241] Odeh, A. S. 1981. Comparison of solutions to a three-dimensional black-oil reservoir
simulation problem. J. Petrol. Techn., 33(1), 13–25. doi:10.2118/9723-PA.
[242] Øren, P.-E., Bakke, S., and Arntzen, O. J. 1998. Extending predictive capabilities to
network models. SPE J., 3(4), 324–336. doi:10.2118/52052-PA.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 References 645

[243] Pal, M., Lamine, S., Lie, K.-A., and Krogstad, S. 2015. Validation of the multi-
scale mixed finite-element method. Int. J. Numer. Meth. Fluids, 77(4), 206–223.
doi:10.1002/fld.3978.
[244] Park, H.-Y., and Datta-Gupta, A. 2011. Reservoir management using streamline-
based flood efficiency maps and application to rate optimization. In: SPE Western
North American Region Meeting. doi:10.2118/144580-MS.
[245] Peaceman, D. W. 1983. Interpretation of well-block pressures in numerical reservoir
simulation with nonsquare grid blocks and anisotropic permeability. Soc. Petrol.
Eng. J., 23(3), 531–543. doi:10.2118/10528-PA.
[246] Peaceman, D. W. 1991. Fundamentals of Numerical Reservoir Simulation. New
York, NY, USA: Elsevier Science Inc.
[247] Peaceman, D. W. 1978. Interpretation of well-block pressures in numerical reservoir
simulation. Soc. Petrol. Eng. J., 18(3), 183—194. doi:10.2118/6893-PA.
[248] Persson, P.-O., and Strang, G. 2004. A simple mesh generator in MATLAB. SIAM
Review, 46(2), 329–345. doi:10.1137/S0036144503429121.
[249] Peters, L., Arts, R., Brouwer, G., et al. 2010. Results of the Brugge benchmark study
for flooding optimization and history matching. SPE Reser. Eval. Eng., 13(03), 391–
405. doi:10.2118/119094-PA.
[250] Pettersen, Ø. 2006. Basics of Reservoir Simulation with the Eclipse Reservoir
Simulator. Lecture Notes. Department of Mathematics, University of Bergen. http://
folk.uib.no/fciop/index_htm_files/ResSimNotes.pdf.
[251] Pettersen, Ø. 2012. Horizontal simulation grids as alternative to structure-based
grids for thin oil-zone problems: A comparison study on a Troll segment. Comput.
Geosci., 16(2), 211–230. doi:10.1007/s10596-011-9240-8.
[252] Pinder, G. F., and Gray, W. G. 2008. Essentials of Multiphase Flow in Porous Media.
Hoboken, NJ: John Wiley & Sons.
[253] Pollock, D. W. 1988. Semi-analytical computation of path lines for finite-difference
models. Ground Water, 26(6), 743–750. doi:10.1111/j.1745-6584.1988.tb00425.x.
[254] Ponting, D. K. 1989. Corner point geometry in reservoir simulation. In: King, P. R.
(ed), ECMOR I – 1st European Conference on the Mathematics of Oil Recovery.
Oxford: Clarendon Press, pp. 45–65. doi:10.3997/2214-4609.201411305.
[255] Potempa, T. C. 1982. Finite element methods for convection dominated transport
problems. PhD. thesis, Rice University. http://hdl.handle.net/1911/15714.
[256] Prevost, M., Edwards, M. G., and Blunt, M. J. 2002. Streamline tracing on curvilin-
ear structured and unstructured grids. SPE J., 7(2), 139–148. doi:10.2118/78663-PA.
[257] Prevost, M., Lepage, F., Durlofsky, L. J., and Mallet, J.-L. 2005. Unstructured 3D
gridding and upscaling for coarse modelling of geometrically complex reservoirs.
Petrol. Geosci., 11(4), 339–345. doi:10.1144/1354-079304-657.
[258] Pyrcz, M. J., and Deutsch, C. V. 2014. Geostatistical Reservoir Modeling. Oxford:
Oxford University Press.
[259] Rashid, B., Muggeridge, A., Bal, A.-L., and Williams, G. J. J. 2012a. Quantifying
the impact of permeability heterogeneity on secondary-recovery performance. SPE
J., 17(2), 455–468. doi:10.2118/135125-PA.
[260] Rashid, B., Bal, A.-L., Williams, G. J. J., and Muggeridge, A. H. 2012b. Using vor-
ticity to quantify the relative importance of heterogeneity, viscosity ratio, gravity and
diffusion on oil recovery. Comput. Geosci., 16(2), 409–422. doi:10.1007/s10596-
012-9280-8.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 646 References

[261] Rasmussen, A. F., and Lie, K.-A. 2014. Discretization of flow diagnostics on
stratigraphic and unstructured grids. In: ECMOR XIV – 14th European Conference
on the Mathematics of Oil Recovery. doi:10.3997/2214-4609.20141844.
[262] Raviart, P. A., and Thomas, J. M. 1977. A mixed finite element method for 2nd order
elliptic equations. In: Galligani, I., and Magenes, E. (eds), Mathematical Aspects of
Finite Element Methods. Berlin, Heidelberg, New York: Springer-Verlag.
[263] Renard, P., and De Marsily, G. 1997. Calculating equivalent permeability: A review.
Adv. Water Resour., 20(5), 253–278. doi:10.1016/S0309-1708(96)00050-4.
[264] Richards, L. A. 1931. Capillary conduction of liquids through porous mediums. J.
App. Phys., 1(5), 318–333. doi:10.1063/1.1745010.
[265] Ringrose, P., and Bentley, M. 2015. Reservoir Model Design: A Practitioner’s
Guide. New York: Springer.
[266] Samier, P. 1990. A finite element method for calculation transmissibilities in n-
point difference equations using a non-diagonal permeability tensor. In: Guérillot, D.
(ed), ECMOR II – 2nd European Conference on the Mathematics of Oil Recovery.
Editions TECHNIP, pp. 121–130. doi:10.3997/2214-4609.201411106.
[267] Samier, P., and Masson, R., 2017. Implementation of a vertex-centered method
inside an industrial reservoir simulator: Practical issues and comprehensive com-
parison with corner-point grids and perpendicular-bisector-grid models on a field
case. SPE J., 22(02), 660–678. doi:10.2118/173309-PA.
[268] Sandve, T. H., Berre, I., and Nordbotten, J. M. 2012. An efficient multi-point flux
approximation method for discrete fracture matrix simulations. J. Comput. Phys.,
231(9), 3784–3800. doi:10.1016/j.jcp.2012.01.023.
[269] Schlumberger. 1999. ECLIPSE 100 User Course. Schlumberger GeoQuest.
[270] Schlumberger. 2014a. ECLIPSE: Reference Manual. 2014.1 edn. Schlumberger.
[271] Schlumberger. 2014b. ECLIPSE Reservoir Simulation Software: Technical Descrip-
tion. 2014.1 edn. Schlumberger.
[272] Schneider, M., Flemisch, B., and Helmig, R. 2017. Monotone nonlinear finite-
volume method for nonisothermal two-phase two-component flow in porous media.
Int. J. Numer. Meth. Fluids, 84(6), 352–381. doi:10.1002/fld.4352.
[273] Shah, S., Møyner, O., Tene, M., Lie, K.-A., and Hajibeygi, H. 2016. The multiscale
restriction smoothed basis method for fractured porous media. J. Comput. Phys.,
318, 36–57. doi:10.1016/j.jcp.2016.05.001.
[274] Shahvali, M., Mallison, B., Wei, K., and Gross, H. 2012. An alternative to stream-
lines for flow diagnostics on structured and unstructured grids. SPE J., 17(3), 768–
778. doi:10.2118/146446-PA.
[275] Shampine, L. F., Ketzscher, R., and Forth, S. A. 2005. Using AD to
solve BVPs in MATLAB. ACM Trans. Math. Software, 31(1), 79–94.
doi:10.1145/1055531.1055535.
[276] Sheldon, J. W., Harris, C. D., and Bavly, D. 1960. A method for general reservoir
behavior simulation on digital computers. In: Fall Meeting of the Society of
Petroleum Engineers of AIME. doi:10.2118/1521-G.
[277] Shook, G. M., and Forsmann, J. H. 2005. Tracer Interpretation Using Temporal
Moments on a Spreadsheet. Tech. rept. INL report 05-00400. Idaho National
Laboratory.
[278] Shook, G., and Mitchell, K. 2009. A robust measure of heterogeneity for ranking
earth models: The F-Phi curve and dynamic Lorenz coefficient. In: SPE Annual
Technical Conference and Exhibition. doi:10.2118/124625-MS.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 References 647

[279] Shubin, G. R., and Bell, J. B. 1984. An analysis of the grid orientation effect in
numerical simulation of miscible displacement. Comput. Methods Appl. Mech. Eng.,
47(1), 47–71. doi:10.1016/0045-7825(84)90047-1.
[280] Spillette, A. G., Hillestad, J. G., and Stone, H. L. 1973. A high-stability sequential
solution approach to reservoir simulation. In: Fall Meeting of the Society of
Petroleum Engineers of AIME. doi:542-MS.
[281] Stephansen, A. F., and Klausen, R. A. 2008. Mimetic MPFA. In: ECMOR XI –
11th European Conference on the Mathematics of Oil Recovery. doi:10.3997/2214-
4609.20146365.
[282] Stone, H. L. 1970. Probability model for estimating three-phase relative permeabil-
ity. J. Petrol. Tech., 22(02), 214–218. doi:10.2118/2116-PA.
[283] Stone, H. L. 1973. Estimation of three-phase relative permeability and residual oil
data. J. Pet. Technol., 12(4). doi:10.2118/73-04-06.
[284] Stone, H. L., and Garder Jr., A. O. 1961. Analysis of gas-cap or dissolved-gas drive
reservoirs. SPE J., 1(02), 92–104. doi:10.2118/1518-G.
[285] Stüben, K. 2001. A review of algebraic multigrid. J. Comput. Appl. Math., 128(1),
281–309. doi:10.1016/S0377-0427(00)00516-1.
[286] Technische Universität Darmstadt. Automatic Differentiation for Matlab (ADiMat).
http://www.adimat.de/. [Online; accessed July 11, 2018].
[287] Thiele, M. R., and Batycky, R. P. 2003. Water injection optimization using a
streamline-based workflow. In: SPE Annual Technical Conference and Exhibition.
doi:10.2118/84080-MS.
[288] Thomas, G. W. 1981. Principles of Hydrocarbon Reservoir Simulation. Boston:
IHRDC.
[289] Todd, M. R. and Longstaff, W. J. 1972a. The development, testing, and application
of a numerical simulator for predicting miscible flood performance. J. Petrol. Tech.,
24(07), 874–882. doi:10.2118/3484-PA.
[290] Todd, M. R., O’Dell, P. M., and Hirasaki, G. J. 1972b. Methods for increased
accuracy in numerical reservoir simulators. Soc. Petrol. Eng. J., 12(06), 515–530.
doi:10.2118/3516-PA.
[291] Tomlab Optimization Inc. Matlab Automatic Differentiation (MAD). http://matlabad
.com/. [Online; accessed July 11, 2018].
[292] Toor, S. M., Edwards, M. G., Dogru, A. H., and Shaalan, T. M. 2015. Boundary
aligned grid generation in three dimensions and CVD-MPFA discretization. In: SPE
Reservoir Simulation Symposium. doi:10.2118/173313-MS.
[293] Toro, E. F. 2009. Riemann Solvers and Numerical Methods for Fluid Dynamics: A
Practical Introduction. 3rd edn. Berlin: Springer-Verlag.
[294] Trangenstein, J. A. 2009. Numerical solution of hyperbolic partial differential
equations. Cambridge: Cambridge University Press.
[295] Trangenstein, J. A., and Bell, J. B. 1989. Mathematical structure of the black-oil
model for petroleum reservoir simulation. SIAM J. Appl. Math., 49(3), 749–783.
doi:10.1137/0149044.
[296] Trottenberg, U., Oosterlee, C. W., and Schüller, A. 2000. Multigrid. Academic press.
[297] Ucar, E., Berre, I., and Keilegavlen, E. 2015. Simulation of slip-induced perme-
ability enhancement accounting for multiscale fractures. In: Fourtieth Workshop on
Geothermal Reservoir Engineering.
[298] van Genuchten, M. T. 1980. Closed-form equation for predicting the hydraulic
conductivity of unsaturated soils. Soil Science Soc. America J., 44(5), 892–898.
doi:10.2136/sssaj1980.03615995004400050002x.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 648 References

[299] Varela, J. 2018. Implementation of an MPFA/MPSA-FV solver for the unsaturated

flow in deformable porous media. M.Sc. thesis, University of Bergen. http://hdl
.handle.net/1956/17905.
[300] Verma, A. 1999. ADMAT: Automatic differentiation in MATLAB using object
oriented methods. In: Henderson, M. E., Anderson, C. R., and Lyons, S. L. (eds),
Object Oriented Methods for Interoperable Scientific and Engineering Computing:
Proceedings of the 1998 SIAM Workshop. Philadelphia: SIAM, pp. 174–183.
[301] Verma, S., and Aziz, K. 1997. A control volume scheme for flexible grids in reservoir
simulation. In: SPE Reservoir Simulation Symposium. doi:10.2118/37999-MS.
[302] Voskov, D. V., and Tchelepi, H. A. 2012. Comparison of nonlinear formulations for
two-phase multi-component eos based simulation. J. Petrol. Sci. Engrg., 82–83(0),
101–111. doi:10.1016/j.petrol.2011.10.012.
[303] Voskov, D. V., Tchelepi, H. A., and Younis, R. 2009. General nonlinear solution
strategies for multiphase multicomponent EoS based simulation. In: SPE Reservoir
Simulation Symposium. doi:10.2118/118996-MS.
[304] Wallis, J. R. 1983. Incomplete gaussian elimination as a preconditioning for gen-
eralized conjugate gradient acceleration. In: SPE Reservoir Simulation Symposium.
doi:10.2118/12265-MS.
[305] Wallis, J. R., Kendall, R. P., and Little, T. E. 1985. Constrained residual accel-
eration of conjugate residual methods. In: SPE Reservoir Simulation Symposium.
doi:10.2118/13536-MS.
[306] Wang, X., and Tchelepi, H. A. 2013. Trust-region based solver for nonlin-
ear transport in heterogeneous porous media. J. Comp. Phys., 253, 114–137.
doi:10.1016/j.jcp.2013.06.041.
[307] Watts, J. W. 1986. A compositional formulation of the pressure and saturation
equations. SPE Res. Eng., 1(3), 243–252. doi:10.2118/12244-PA.
[308] Weiser, A., and Wheeler, M. F. 1988. On convergence of block-centered
finite differences for elliptic problems. SIAM J. Numer. Anal., 25(2), 351–375.
doi:10.1137/0725025.
[309] Welge, H. J. 1952. A simplified method for computing oil recovery by gas or water
drive. J. Petrol. Tech., 4(04), 91–98. doi:10.2118/124-G.
[310] Wen, X. H., Durlofsky, L. J., and Edwards, M. G. 2003. Upscaling of channel
systems in two dimensions using flow-based grids. Transp. Porous Media, 51(3),
343–366. doi:10.1023/A:1022318926559.
[311] Wen, X.-H., and Gómez-Hernández, J. J. 1996. Upscaling hydraulic conductivities
in heterogeneous media: An overview. J. Hydrol., 183, ix–xxxii. doi:10.1016/S0022-
1694(96)80030-8.
[312] Wheeler, J. A., Wheeler, M. F., and Yotov, I. 2002. Enhanced velocity mixed finite
element methods for flow in multiblock domains. Comput. Geosci., 6(3-4), 315–332.
doi:10.1023/A:1021270509932.
[313] Wheeler, M. F., Arbogast, T., Bryant, S., et al. 1999. A parallel multiblock/multido-
main approach for reservoir simulation. In: SPE Reservoir Simulation Symposium,
pp. 51–61.
[314] Wheeler, M. F., and Yotov, I. 2006. A cell-centered finite difference method on
quadrilaterals. In: Arnold, D. N., Bochev, P. B., Lehoucq, R. B., Nicolaides, R. A.,
and Shashkov, M. (eds), Compatible Spatial Discretizations. New York: Springer,
pp. 189–207.
[315] Whitaker, S. 1986. Flow in porous media I: A theoretical derivation of Darcy’s law.
Transp. Porous Media, 1(1), 3–25. doi:10.1007/BF01036523.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 References 649

[316] Whitson, C. H., and Brulé, M. R. 2000. Phase Behavior. Richardson, TX: Society
of Petroleum Engineers.
[317] Wiener, O. 1912. Abhandlungen der Matematisch. PhD. thesis, Physischen Klasse
der Königlichen Sächsischen Gesellscaft der Wissenschaften.
[318] Wu, X.-H., Efendiev, Y., and Hou, T. Y. 2002. Analysis of upscaling
absolute permeability. Discrete Contin. Dyn. Syst. Ser. B, 2(2), 185–204.
doi:10.3934/dcdsb.2002.2.185.
[319] Wu, X.-H., and Parashkevov, R. 2009. Effect of grid deviation on flow solutions.
SPE J., 14(01), 67–77. doi:10.218/92868-PA.
[320] Wyckoff, R. D., Botset, H. G., Muskat, M., and Reed, D. W. 1933. The measurement
of the permeability of porous media for homogeneous fluids. Rev. Sci. Instrum., 4(7),
394–405. doi:10.1063/1.1749155.
[321] Yanosik, J. L., and McCracken, T. A. 1979. A nine-point, finite-difference reservoir
simulator for realistic prediction of adverse mobility ratio displacements. Soc. Petrol.
Eng. J., 19(04), 253–262. doi:10.2118/5734-PA.
[322] Younis, R. 2009. Advances in modern computational methods for nonlinear prob-
lems: A generic efficient automatic differentiation framework, and nonlinear solvers
that converge all the time. Ph.D. thesis, Stanford University.
[323] Younis, R., and Aziz, K. 2007. Parallel automatically differentiable data-types for
next-generation simulator development. In: SPE Reservoir Simulation Symposium.
doi:10.2118/106493-MS.
[324] Zhang, P., Pickup, G. E., and Christie, M. A. 2005. A new upscaling approach
for highly heterogenous reservoirs. In: SPE Reservoir Simulation Symposium.
doi:10.2118/93339-MS.
[325] Zhou, Y., Tchelepi, H. A., and Mallison, B. T. 2011. Automatic differentiation
framework for compositional simulation on unstructured grids with multi-point
discretization schemes. In: SPE Reservoir Simulation Symposium. doi:10.2118/
141592-MS.

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Index

AD-OO framework, 342, 415, 599 solution gas–oil ratio, 371

Add-on modules, 598 vaporized oil–gas ratio, 371
Additive property, 561 water component, 389
Adjacency matrix, 521, 538 Bottom-hole pressure, 122
Adjoint equations, 425 Boundary conditions
Agglomeration, 541 data structure, 146
a priori flow indicator, 551 Dirichlet, 120, 147
feasible directions, 541 hydrostatic, 161
flow indicator, 547 linear pressure, 161
global flow indicator, 551 Neumann, 120, 147
greedy algorithms, 548 no flow, 120, 146
local flow indicator, 551 Bubble point, 362
merging blocks, 546 Bubble-point curve, 362
permissible directions, 541 Bubble-point pressure, 375
refining blocks, 547 Buckley–Leverett, 259, 281, 284, 298, 417
volume indicator, 546 Bundle of tubes, 236
Annular space, 395
Annulus, 396 Capillary fringe, 320, 330
Anonymous function, 205 Capillary pressure, 235
Anticline, 25, 27 Brooks–Corey, 239
API gravity, 372 hysteresis, 237
Aqueous phase, 233 Leverett J, 238
Aquifer, 256 van Genuchten, 239
Areal sweep, 308 Carbonate rock, 24
Assembly, 435 Carman–Kozeny, 37
Automatic differentiation, 204, 623–629 Cartesian grid, 57
Cell-mode script, 603
Basis functions, 178, 180 CFL condition, 275, 278
Bi-stream function, 130 Christmas tree, 365
Biconnected components, 543 Circumcenter, 63
Black-oil equations Circumcircle, 63
gas component, 392 Class object, 401
Black-oil models, 246 Clastic rock, 23
equation of state, 372 CO2 storage, 301, 311
formation-volume factor, 371 Coarse grid, 519
oil component, 390 Cocurrent flow, 251
primary unknowns, 389 Compatible discretization, 188
primary variables, 392 Component conservation, 245
shrinkage factor, 371 Components, 233

650

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Index 651

Compositional flow, 245 coarse grid, 519, 525

Compositional simulation, 338 fluid properties, 144
Compressibility grid, 88
constant, 117 reservoir states, 145
fluid, 117 source terms, 145
gas (isothermal), 373 wells, 147
oil, 376 Dead oil, 247, 342, 380
oil (isothermal), 378 Delaunay, 63
rock, 36 Dew point, 362
slightly, 118 Dew-point curve, 362
total, 117, 252 Disconnected blocks, 521
water, 379 Discrete operators, 135, 203
Compressible flow, 117 average faces to cells, 218
Computational overhead, 418 averaging, 207
Computational steering, 419 divergence, 136, 206
Condensate, 360 divergence (wells), 397
Connate water saturation, 233 gradient, 136, 206
Connected components, 521 gradient (wells), 397
Conservative method, 275 harmonic average, 215
Consistent, 174 upwind, 141, 222
Constitutive equations, 116 Displacement
Constrained pressure residual (CPR), 444 favorable, 262, 308
Contacted volumes, 482 unfavorable, 262, 308
Control step, 403, 416, 438, 471 Dissolution matrix, 452
Convergence DistMesh, 69
nonlinear, 403, 438 Downscaling, 560
rate, 326 Drainage, 237
Convergent, 174 Drainage region, 480
Coordinate system, 91 Dry gas, 382, 383
Corner-point grid, 45, 48, 73 Dulmage–Mendelsohn, 169, 309, 522
Countercurrent flow, 251
Critical point, 361 ECLIPSE input, 46, 97, 339
Critical temperature, 361 ACTNUM, 48, 80
Crossflow, 395 COMPDAT, 402
Curvilinear grid, 58 COORD, 48, 75
Cylinder coordinates, 122, 124 DENSITY, 383
EDIT, 340
Darcy GRAVITY, 383
flow experiment, 113 GRID, 340
Henry, 36, 113 INCLUDE, 341
unit, 36, 40 METRIC, 49
velocity, 115 MULTX, 48
Darcy’s law, 14, 36, 115 MULTZ, 50
Darcy–Buckingham equation, 257 NTG, 48
Data set PERMX, 48
CaseB4, 79, 536, 550 PORO, 48
Johansen, 44, 529 PROPS, 340, 350, 379
Norne, 79, 323 PVCDO, 380
SAIGUP, 46, 97, 171, 514, 532 PVCO, 381
SPE 1, 455 PVDG, 382, 383
SPE 10, 41, 307, 501, 551, 585 PVDO, 380
SPE 10, Tarbert, 495 PVTG, 382
SPE 3, 355, 409 PVTO, 381, 383
SPE 9, 355, 383, 460 PVTW, 382
SPE 1, 350, 355, 383, 399, 409 PVZG, 382
Data structure REGIONS, 340
boundary conditions, 146 ROCK, 383

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 652 Index

ECLIPSE input (cont.) heterogeneity measures, 489

ROCKTAB, 383 heterogeneous q5-spot, 310
RUNSPEC, 340, 380 homogeneous q5-spot, 306
SATNUM, 48, 52 hyperbolic schemes, 282
SCHEDULE, 340, 402 incompressible solvers, 325
SGOF, 351 intro to MRST, 608
SOLUTION, 340, 400 modules in MRST, 619
SUMMARY, 340 multisegment wells, 459
SWOF, 351 non-Newtonian fluids, 220
TABDIMS, 380 numerical errors, 334
TSTEP, 403 quarter five-spot, 159
TUNING, 403 rock properties, 52
WCONINJE, 402 sector model, 453
WCONPROD, 402 single-phase with wells, 173
WELSPECS, 402 single-phase, compressible, 212
ZCORN, 48, 75 single-phase, incompressible, 169
ECLIPSE output, 350 SPE 9 benchmark, 469
Energy equation, 220 structured grids, 61
Energy minimization, 179 three-phase intro, 359
Enhanced oil recovery, 5 unstructured grids, 71
Enthalpy, 221 unstructured stencils, 169
Entropy condition, 261 upscaling sector model, 594
Lax, 261 upstream methods, 286
Oleinik, 261 volume averaging, 563
Entry pressure, 236, 320 wells, incompressible, 173
Equation of state Expansion factor, 371
compressibility, 373 Explicit flux representation, 184
cubic, 119
gas deviation factor, 373 Face mobility, 215
Peng–Robinson, 119 Facies, 238
Redlich–Kwong, 119 Facility model, 433
Redlich–Kwong–Soave, 119 Fanning friction factor, 399
Error Fault trap, 27
capillary dominated flow, 330 Faults, 26, 39, 49
grid-orientation, 176, 187, 330 Fickian diffusion, 245
mass balance, 439, 464 Fictitious domain, 60
normalized residual, 439, 464 Finite-volume method, 132
splitting, 293, 326 Five-spot, 126
temporal, 326 Flow capacity, 483
truncation, 174, 276, 285 Flow diagnostics, 477
Exercises Flow equation, 250
advanced grids, 108 Flow resistance, 458
agglomeration, 553 Fluid object, 16, 290
arithmetic vs. harmonic, 215 black-oil, 355
automatic differentiation, 629 ECLIPSE input, 350
averaging, 569 relative permeability, 144
boundary conditions, 164 saturation, 144
buoyant migration, 312 single phase, 144
consistent schemes, 187 Fluvial, 28, 42
discretization operators, 142 Formation-volume factor, 123, 124, 247
efficiency of prototype code, 228 gas, 371, 373
flow diagnostics optimization, 504 oil, 371, 375
flow diagnostics, SAIGUP, 516 total, 376
flow-based upscaling, 574 water, 379
fluid object, 292 Fractional-flow formulation
global upscaling, 580 global pressure, 254
grid coarsening, 539 phase potential, 255
grid structure, 96 phase pressure, 250–254

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Index 653

Fractional-flow function, 250 tessellation, 63, 68

Fracture, 25 Triangle, 69
Fully implicit, 244 Voronoi, 66, 104
Function spaces Grid-orientation errors, 176, 330
H0div , 178
1 Heat conduction, 221
H 2 , 183 Hele–Shaw cell, 307
L2 , 178 Heterogeneity, 483
Heterogeneity measure
Gas compressibility factor, 373 dynamic, 483
Gas liberation F- diagram, 484
differential, 365 Lorenz coefficient, 486
flash, 365 static, 483
Gas lift, 396 sweep efficiency, 487
Gas reservoir, 365 High-pressure cell, 362
condensate, 366 Homogenization, 560
dew-point, 366 Hybrid, mixed, 182
dry, 366 Hydraulic conductivity, 2, 114, 257
retrograde condensate, 366 Hydraulic head, 113
rich, 366 Hydrocarbon recovery
wet, 366 primary production, 3
Gaseous phase, 233 secondary production, 4
Generating points, 63 tertiary production, 5
Geochemistry, 338 Hysteresis, 237
Geological model, 10, 30
Geostatistics, 40 Ideal gas, 118
Gibbs’ phase rule, 360 Imbibition, 237
Global numbering, 447 Immiscible, 232
Global pressure, 254 Immobile oil, 237
Global-pressure formulation, 254 Implicit discretization, 203
Godunov scheme, 278 Incompressible flow, 117
Gravel pack, 396 Indefinite system, 184
Gravity Indirection map, 89
API, 372 Inflow control device, 396
specific, 372 Inflow-performance relation, 122, 394
Gravity column, 14, 267, 301 Influence region, 479, 497
Gravity override, 315 Injectivity, 310
Grid index, 122
Cartesian, 57 Inner product
coarse, 519 continuous, 178
composite, 101 discrete, 180
conformal, 103 local, 185
corner-point, 73, 97 mixed hybrid, 183
curvilinear, 58 Input deck, 48
data structure, 88 Irreducible saturation, 233
Delaunay, 63
disconnected blocks, 521 Jacobian matrix, 204, 295, 296, 309
DistMesh, 69 Johansen data set, 44
fictitious domain, 60, 523 JOLTS, 598, 602, 603
geometry, 93
multiblock, 104 K-orthogonal, 176, 333
partition, 519 Klinkenberg effect, 239
PEBI (2.5D), 85
PEBI (2D), 67 Laplace operator, 117
PEBI (3D), 104 Lax–Friedrichs, 275
rectilinear, 57 Lax–Wendroff, 276
stair stepped, 79 Leverett J-function, 238
stratigraphic, 72 Limestone, 22

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 654 Index

Line-search method, 296 Ministep, 438

Linear solver Miscible displacement, 246
AGMG, 153, 445, 463 Mixed finite elements, 177
algebraic multigrid, 445 Mixed formulation, 178
condition number, 347 Mixed hybrid, 182
CPR, 444, 463 Mixture velocity, 397
direct, MATLAB default, 152 Model
GMRES, 442 black-oil, 430
ILU preconditioner, 443 facility, 433
iterative methods, 341 physical, 422
Krylov subspace method, 442 reservoir, 426
mldivide, 152, 303, 309 wells (multisegment), 433
multigrid, 153 wells (Peaceman), 433
nested factorization, 447 Model object, 414
orthomin, 447 Modules, 598, 610
preconditioner, 341 ad-blackoil, 430, 615
Thomas algorithm, 303 ad-core, 416, 600
triangular system, 156 ad-eor, 616
tridiagonal system, 303 ad-fi, 616
Linearization, 203 ad-mechanics, 616
Linearized problem, 434 ad-props, 351, 615
Linearly degenerate wave, 261 adjoint, 615
Lithography, 72 agglom, 541, 613
Live oil, 247, 367, 381, 383 blackoil-sequential, 420, 615
Local numbering, 447 book, 10, 157, 619
Local-flux mimetic, 196 co2lab, 617
Logical indexing, 99 coarsegrid, 614
Lorenz coefficient, 486, 501 compositional, 616
deckformat, 48, 350, 405, 615
M-matrix, 197 design rules, 612
Marine rock, 23 dfm, 618
Mass conservation, 116 diagnostics, 153, 480, 616
Mass fraction, 233 dual-porosity, 618
Mass-balance error, 439 enkf, 619
MATLAB fvbiot, 618
accumarray, 621 geochemistry, 618
arrayfun, 346 hfm, 618
bsxfun, 621 incomp, 16, 149, 614
cellfun, 305, 353 libgeometry, 614
classdef, 401, 423 linearsolver, 446
handle classes, 437 linearsolvers, 619
indirection maps, 89 matlab_bgl, 619
logical indexing, 99 mimetic, 186, 599, 614
run-length encoding, 89 mpfa, 196, 599, 615
Mesh generator, 69 mrst-cap, 619
Method of characteristics, 259 mrst-gui, 53, 619
Metis, 538 msmfe, 617
Mimetic finite differences, 188 msrsb, 617
consistency, 189 octave, 607, 619
Mimetic inner product, 189–196 opm_processing, 614
general family, 190 optimization, 616
parametric family, 191 remso, 619
quasi RT0, 195 solvent, 616
quasi two-point, 193 spe10, 42, 619
simple (default), 195 steady-state, 616
two-point, 193 streamlines, 159, 619
Min-max angle criterion, 63 triangle, 69, 614

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Index 655

upr, 104, 614 Johansen, 45

upscaling, 572, 616 modeling in MRST, 39
vem, 615 relative, 240
vemmech, 618 SAIGUP, 51
wellpaths, 619 SPE 10, 44
Monotone, 197 Perpendicular bisector, 66
Mudrock, 22 Phase (dis)appearance, 393
Multicomponent flow, 245 Phase diagram, 360
Multilateral well, 396 binary substances, 362
Multiplier, 39, 50, 323 bubble-point reservoir, 368
Multipoint flux approximation, 196 cricondenbar, 364
Multisegment well, 339, 395, 455 cricondentherm, 364
critical point, 363
Nested factorization, 447 gas reservoirs, 366
Net-to-gross, 38, 50 single-component system, 360
Newton update, 204, 296 Phase mobility, 244
Newton–Raphson, 203, 295 Physical model, 416
Non-neighboring connection, 74 Pillar grid, 73
Non-Newton fluid, 215 Poisson’s equation, 117
Norne data set, 79 Polymer flooding, 216
Numerical context, 415 Pore volumes injected (PVI), 326
Porosity, 35
Octave, 607 Carman–Kozeny relation, 38
Oil reservoir, 365 Johanesen, 45
black-oil, 367 SAIGUP, 51
bubble-point, 367 SPE 10, 43
solution gas, 367 Potential ordering, 309
volatile, 367 Preconditioner, 442
Oil-water contact, 237 Pressure decline, 4
Oleic phase, 233 Pressure equation
Operator splitting, 292 global pressure, 254
phase-potential formulation, 256
Partition two-phase compressible, 253
agglomeration, 541 two-phase incompressible, 250
confined blocks, 543 Primary production, 3
disconnected, 521, 542 combination drive, 4
flow indicator, 544 gas cap drive, 4
graph algorithms, 538 gravity drive, 4
grid faces, 527 solution gas drive, 4
hierarchical, 554 water drive, 4
load balanced, 519 Production-logging tool, 482
METIS, 538 Productivity index, 122, 394
near well refinement, 536 Properties
segmentation, 542 extrapolation, 408
volume indicator, 544 interpolation, 408
Partition of unity, 479 Pseudo components, 233
Partition vector, 519 gas, 246
Pathline, 129 oil, 246
Peaceman, 126, 169
PEBI grid, 67, 85 Quarter five-spot, 126, 157
Periodic medium, 560 heterogeneous, 307, 327
Permeability, 2, 36 homogeneous, 303, 326
absolute, 239 rotated, 330
anisotropic, 37
Carman–Kozeny relation, 38 Radial flow, 122, 167
effective, 239 Rankine–Hugoniot, 260
isotropic, 37 Rarefaction wave, 261

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 656 Index

Raviart–Thomas, 180 Sedimentary environment, 23

Rectilinear grid, 57 Sedimentary rocks, 21
Relative permeability, 240 Self-sharpening wave, 261
Brooks–Corey, 241 Sequential fully implicit, 420
Corey, 241 Sequential solution, 292
default in ECLIPSE 100, 349 Shale, 22, 239
in ad-props, 353 Shallow marine, 24, 28, 42
SPE 1,3,9, 356 Shear factor, 217
Stone I, 242 Shock wave, 261
three-phase, 348 Shrinkage factor, 247, 371
two-phase, 241, 242 Simulation model, 7, 401
van Genuchten, 242 Skin factor, 125, 394
Repeated five-spot, 126 Solution gas-oil ratio, 247, 371, 375, 381
Representative elementary volume, 31, 232, 560 Source terms, 145
Reservoir simulation, 6 Sparsity pattern, 167
Residence time, 478 SPE 10 data set, 41
Residence-time distribution, 479, 490 Specific gas, 373
Residual equations, 294 Specific gravity
Residual saturation, 233 gas, 373
Result handler, 419 liquid, 372
Retrograde condensate gas, 387 Stability analysis, 284
Retrograde gas condensate, 366, 377 Stable, 174
Reynolds number, 399 State object, 145
Rheology flux, 145
Newtonian fluid, 215 pressure, 145
non-Newtonian fluid, 215 saturation, 145
shear thickening, 216 well solution, 145
shear thinning, 216 Stock-tank, 364
Rich gas, 382, 387 Storage capacity, 483
Richards’ equation, 256 Storeage cofficient, 257
Riemann fan, 262 Stratigraphic grid, 72
Riemann problem, 262 Stratigraphic trap, 27
Rock types, 51 Streakline, 129
Root folder, 601 Streamline coordinates, 304
Rosebrock, 629 Streamlines, 128, 158
Rotated gravity, 311 Structural model, 31, 47, 49
Rs -factor, 371, 372, 375, 381, 383, 385 Subset extraction, 99
Run-length encoding, 89 Summary file, 405
Rv -factor, 371, 372, 375 Supercritical state, 361
Surface area, specific, 38
S-shape, 251 Surface tension, 234
SAIGUP data set, 46, 97 Sweep efficiency, 487
Sand screen, 396 Sweep region, 480, 496
Sandstone, 22 Syncline, 25
Saturated, 368, 375
Saturation, 232 Tarbert, 42
connate water, 233 Tessellation, 63
effective, 238, 241 Test function, 178
irreducible, 233 Thermal expansion
residual, 233 adiabatic, 224
Saturation equation free, 223
phase-potential formulation, 256 Joule–Thomson, 224
two-phase compressible, 254 reversible, 224
two-phase incompressible, 250 Thermal: single-phase, 220
Schedule, 340, 402, 418, 457, 462 Thin-tube experiment, 235
Schur-complement, 184 Time loop, 210
Secondary production, 4 Time-of-flight, 129, 153
waterflooding, 4 backward, 478

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Index 657

forward, 478 Vadose zone, 256

per influence region, 482 Vanishing viscosity, 273
Time-step Vaporization curve, 362
chopping, 297, 393 Vaporized oil–gas ratio, 247, 371, 375
control, 297, 415, 463 Variable elimination, 436
selection, 416, 439 Variable switching, 393
Topological sort, 309 Velocity reconstruction, 218
Tortuosity, 38 Vertical equilibrium, 450
Tracer, 129, 156 Viscosibility, 380
Tracer distribution, 130 Viscosity
Transmissibility, 16, 133 gas, 374
Transport equation, 250 oil, 378
Trial function, 180 water, 379
Triple point, 360 Viscous fingering, 307
Truncation error, 174, 285 Volatile oil, 360, 367, 377
Two-point scheme, 16 Volumetric partition, 479
conditionally consistent, 175 Voronoi, 66, 104
derivation, 132–133
Water coning, 313
half-transmissibility, 133
Water content, 257
implementation, 150
Waterflooding, 4
linear system, 133
Well allocation factor, 482, 581
nonlinear TPFA, 200
Well head, 365
transmissibility, 133
Well model
crossflow, 395
Unconditionally stable, 283
flow rate, 208
Unconventional reservoirs, 239
frictional pressure drop, 398, 457
Undersaturated, 368, 375
hydrostatic pressure, 208
Unsaturated flow, 256
inflow-performance relation, 122
Upper Ness, 42 injectivity index, 122
Upscaling, 559 mobility calculation, 395
arithmetic average, 564 multisegment, 397
fallback strategy, 585 Peaceman, 126–128, 169
flow diagnostics, 581 productivity index, 122
flow-based, 570 scaling, 443
generic conditions, 579 simple valve, 398, 457
geometric average, 564 skin, 125
global, 579 surface rate, 209
harmonic average, 564 well control, 209, 428, 451
harmonic-arithmetic, 567 Well pair, 480
ill-posed problem, 564 Well placement, 502
laboratory boundary conditions, Well-pair region, 480
570 Wells, 121, 338
local-global, 579 control, 338, 402, 423, 462
oversampling, 578 control switching, 453, 464
parallel layers, 566 data structure, 147
periodic, 571 flow rate, 147
permeability definition, 564 horizontal, 455
perpendicular layers, 565 multisegment, 339, 455
positive transmissibilities, 577 simple, 457
power average, 564 Wet gas, 382
specific conditions, 579 Wettability, 234
transmissibility, 575 contact angle, 234
volume average, 561 non-wetting, 234
well indices, 584 oil wet, 235
Wiener bound, 564 water wet, 235
Upstream mobility weighting, 279 wetting, 234
Upwind scheme, 278 Workbook, 603

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Usage of MRST Functions

addBC, 146 convertDeckScheduleToMRST, 402

addSource, 145 convertHorizonsToGrid, 77
addWell, 148, 170 convertInputUnits, 49
AGMGSolverAD, 445 convertTo, 17
applySuccessivePart, 555 covert2MSWell, 457
assembleSystem, 435 CPRSolverAD, 444, 463
assignPVTO, 384 cutGrdecl, 81
assignRelPerm, 353
assignROCK, 352 Driving forces, 428
assignSWOF, 352
eliminateVariable, 436
BackslashSolverAD, 403, 442, equationsBlackOil, 390, 431
463 evaluateRelPerm, 428
blackoilSectorModelExample, explicitTransport, 294
451 explosionView, 524
blackoilTutorialOnePhase, 449 extrudedTriangleGrid, 96
blackoilTutorialSPE1, 399
blackoilTutorialSPE9, 460 faceAvg, 389
faceFlux2cellFlux, 171
cartGrid, 15, 57 faceUpstr, 390
checkProperty, 425 FacilityModel, 433, 457
coarsenGeometry, 525 findConfinedBlocks, 543
combineEquations, 436 fluxside, 147, 161
combineMSwithRegularWells,
457 gaussianField, 40
compressPartition, 498, 523 generateCoarseGrid, 527
computeFandPhi, 485 getAdjointEquations, 425
computeGeometry, 15, 93 getDrivingForces, 423
computeLorenz, 487 getEquations, 414, 423, 431
computeMimeticIP, 187, 192 getLinearSystem, 435
computeMultiPointTrans, 196 getPlotAfterStep, 418
computeSweep, 487 getProp, 424
computeTimeOfFlight, 153, 159 getProps, 414
computeTimestep, 440 getUnitSystem, 49
computeTOFandTracer, 480 getVariableField, 424, 427
computeTranMult, 324 getWellOutput, 468
computeTrans, 150 glue2DGrid, 105
computeWellPairs, 481, 498 GMRES_ILUSolverAD, 443
convertDeckSchedule, 456 gravity, 311, 450

658

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 Usage of MRST Functions 659

grdecl2Rock, 49 plotWellAllocationPanel,
griddedInterpolant, 353 499
plotWellSols, 306, 404
implicitTransport, 295 pollock, 159
incompMimetic, 187 processFacePartition, 528
incompMPFA, 196 processGRDECL, 41, 45, 49, 81, 98
incompTPFA, 16, 150, 151 pside, 16, 147, 161
initCoreyFluid, 291 psideh, 162
initDeckADIFluid, 351
initResSol, 145 rampupTimesteps, 452
initSimpleADIFluid, 417, 449 readEclipseDeck, 350
initSimpleFluid, 290 readEclipseSummaryFmt, 406
initSimpleFluidPc, 291 readEclipseSummaryUnFmt,
initSingleFluid, 16, 144, 144 406, 467
initState, 145 readGRDECL, 48, 80
initVariablesADI, 209, 624 reduceToSingleVariableType, 436
initWellSolAD, 417, 429 refineBlocks, 548
insertWellEquations, 432 refineGreedy, 548
interactiveDiagnostics, 505 refineGreedy2, 548
interpReg, 353, 386 refineGreedy3, 548
isPointInsideGrid, 166 refineGreedy4, 548
refineNearWell, 536
LinearizedProblem, 434 refineRecursiveCart, 549
LinearSolverAD, 441 refineUniform, 548
logNormLayers, 41 relPermWO, 354
relPermWOG, 354
makeLayeredGrid, 86 removeCells, 60
makeModel3, 96 removeConfinedBlocks, 544
makeRock, 15, 40 ReservoirModel, 426
moduleGUI, 611 ResultHandler, 419
mrstDataDirectory, 609
mrstDatasetGUI, 355, 605, 609 sampleFromBox, 549
mrstExamples, 603, 604 segmentIndicator, 542
mrstExploreModules, 605 selectModelFromDeck, 401, 456
mrstModule, 611 setProp, 424
mrstPath, 69, 607, 611 setupOperatorsTPFA, 422
MultisegmentWell, 433, 458 simpleGrdecl, 74
multisegmentWellExample, simpleSchedule, 418
455 SimpleTimeStepSelector, 403, 440
SimpleWell, 433, 458
NonLinearSolver, 403, 417, 436 simulateScheduleAD, 404, 418
nozzleValve, 457 solveAdjoint, 425
solveLinearProblem, 441
outlineCoarseGrid, 520 solveTimeStep, 418
solveTimestep, 438
partitionCartGrid, 522 startup, 601, 611
partitionMETIS, 539 startup_user, 69, 607, 612
partitionUI, 520 stepFunction, 424, 438
pebi, 68, 102
PhysicalModel, 422 tensorGrid, 58
pickTimestep, 440 tessellationGrid, 69
plotCellData, 15, 46, 53, 99 tetrahedralGrid, 66
plotFaces, 17 ThreePhaseBlackOilModel, 401,
plotGrid, 15, 46 430, 451
plotToolbar, 53, 404, 419 toleranceCNV, 439
plotTracerBlend, 497 toleranceMB, 439

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894
 660 Usage of MRST Functions

triangleGrid, 64 upscalePerm, 573

twister, 59 upscaleTrans, 578, 580
twophaseJacobian, 295
TwoPhaseOilWaterModel, 417, 431 validateModel, 425, 430
validateState, 425, 430
updateForChangedControls, verticalWell, 149, 170, 172
429
updateState, 429, 431 WaterModel, 449
updateStateFromIncrement, wellBoreFriction, 457
429

Downloaded from https://www.cambridge.org/core. IP address: 189.152.120.32, on 27 May 2021 at 15:38:42, subject to the Cambridge Core terms of use, available at
https://www.cambridge.org/core/terms. https://www.cambridge.org/core/product/F48C3D8C88A3F67E4D97D4E16970F894