Lecture Notes

in Computational Science
and Engineering 32
Editors
Timothy J. Barth, Moffett Field, CA
Michael Griebel, Bonn
David E. Keyes, Norfolk
Risto M. Nieminen, Espoo
Dirk Roose, Leuven
Tamar Schlick, New York
Springer-Verlag Berlin Heidelberg GmbH
Heike Emmerich
Britta Nestler
Michael Schreckenberg
Editors

Interface
and Transport
Dynamics
Computational Modelling

i Springer
Editors
Heike Emmerieh Miehael Schreckenberg
Fachbereieh Chemietechnik Physik von Transport und Verkehr
Universität Dortmund Universität Duisburg-Essen
Emil-Figge-Str.70 Lotharstr. 1
44227 Dortmund, Germany 47048 Duisburg, Germany
e-mail: [email protected] e-mall: [email protected]

Britta Nestler
Fachhochschule Karlsruhe
Moltkestr. 30
76133 Karlsruhe, Germany
e-mail: [email protected]

Cover figure: Experimental photograph of a dendritie solidification structure of an Al-Si alloy

superimposed by a sequence of images showing a three-dimensional simulation result of an
evolving eutectie mierostructure (taken from research results by Britta Nestler, University of
Applied Sciences Karlsruhe, Germany 2003).

Cataloging-in-Publication Data applied for

A catalog record for this book is available from the Library of Congress.
Bibliographie information published by Die Deutsche Bibliothek
Die Deutsche Bibliothek lists this publication in the Deutsche Nationalbibliografie;
detailed bibliographie data is available in the Internet at <http://dnb.ddb.de>.

Mathematics Subject Classification (2000): ooBxx, 3S-XX, 37-XX, 6SCxx, 6SMxx,

68Qxx, 74-XX, 76Rxx, 76T05, 80-XX, 82-XX, 90-06, 90B20
ISSN 1439-73S8
ISBN 978-3-642-07320-5 ISBN 978-3-662-07969-0 (eBook)
DOI 10.1007/978-3-662-07969-0
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifieally the rights of translation, reprinting, reuse of illustrations, recitation, broad-
casting, reproduction on mierofilm or in any other way, and storage in data banks. Duplication of
this publication or parts thereof is permitted only under the provisions of the German Copyright Law
of September 9, 1965, in its current version, and permission for use must always be obtained from
Springer-Verlag Berlin Heidelberg GmbH.
Violations are liable for prosecution under the German Copyright Law.
http://www.springer.de
© Springer-Verlag Berlin Heidelberg 2003
Originally published by Springer-Verlag Berlin Heidelberg New York in 2003
Softcover reprint of the hardcover 1st edition 2003
The use of general descriptive names, registered names, trademarks, etc. in this publication does not
imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.

Cover Design: Friedhelm Steinen-Broo, Estudio Calamar, Spain

Cover production: design & production
Typeset by the authors using a Springer TEX macro package
Printed on acid-free paper 46/31421LK - 543210
Preface

The workshop on "computational physics of transport and interfacial dynam-

ics" was held in Dresden, Germany from February 25 to March 8,2002. The
Max Planck Institute for the Physics of Complex Systems has sponsored the
workshop and the preliminary lecture-based seminar.
The workshop has closely pursued the recent progress of research in com-
putational physics and materials science, particularly in modelling both traf-
fic fiow phenomena and complex multi-scale solidification. These branches
of science have become topics of considerable diversity linking disciplines as
different as physics, mathematical and computational modelling, nonlinear
dynamics, materials sciences, statistical mechanics and foundry technique.
The international workshop brought together experts from different fields in
order to enhance the exchange of knowledge, to assess common interests and
to provide closer cooperation between different communities of researchers.
The workshop intended to create a comprehensive and coherent image of the
current research status and to formulate various possible perspectives for joint
future activities. Special emphases laid on exchanging experiences concerning
numerical tools and on the bridging of the scales as necessary in a variety of
scientific and engineering applications. An interesting possibility along this
line was the coupling of different computational approaches leading to hybrid
simulations. In this sense, we explicitly addressed researchers working with
different numerical schemes as diverse as cellular automata, coupled maps, fi-
nite difference and finite element algorithms for partial differential equations
(e.g. phase-field computations). The following scientific topics were treated
in the contributions:

- Fundamentals and modelling of microstructure formation: Free boundary

problems, phase transitions, solidifiation, interfacial dynamics, Ginzburg-
Landau equations, Cahn-Hilliard equations, phase-field modelling.
- Numerical simulations on different time and length scales: Fluid dy-
namic computations, adaptivity, finite element-discretizations, discrete
modelling techniques such as molecular dynamics and Monte Carlo.
- Multi-scale problems in crystal growth, porous media, solidification and
trafik fiow.

The workshop was attended by 87 participants from 22 different countries.

This volume contains 36 refereed papers having been drawn on from the
oral presentations and the posters at the workshop. The proceedings book
is organized in two main chapters, the first chapter on interface dynamics
and solidification and the second chapter on transport processes. The first
chapter contains contributions on phenomenology and modelling as weIl as
on numerical simulations.
VI Preface

The editors gratefully acknowlegde the granting of the workshop by the

Max-Planck Gesellschaft and thank the Max-Planck Institute in Dresden for
providing space and equipment and for being a very kind host. We wish
to heartily thank all members of the local organizing committee for their
professional and invaJuable help.

Heike Emmerich Britta Nestler Michael Schreckenberg

Table of Contents

I INTERFACE DYNAMICS: Phenomenology and Modelling

Melting Kinetics of Prolate Spheroidal Crystals . . . . . . . . . . . . . . . . . . . . . 3

M.E. Glicksman, A. Lupulescu, M.B. Koss
1 Introduction................................................. 3
1.1 Dendritic Growth ....................................... 3
1.2 Isothermal Dendritic Growth Experiment (IDGE) ........... 3
1.3 Mushy Zones ........................................... 4
2 Video Melting Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3 Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.1 Image Processing ..................... , .. " ..... . .. .. .. .. 5
4 Quasi-static Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
4.1 Background............................................ 6
4.2 Potential Theoretic Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
4.3 Enthalpy Flux .......................................... 9
4.4 Enthalpy Current ....................................... 11
4.5 Kinematic Formulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 11
5 Results..................................................... 13
5.1 Experimental Melting Rates... ... . .. . .. .... .. . . .. .. .. .. .. 13
5.2 Thermal Data .......................................... 15
6 Conclusions................................................. 17
7 Acknowledgements........................................... 17
Deterministic Behaviour in Sidebranching Development . . . . . . . . . . . . .. 20
Ricaro Gonzalez-Cinca, Yves Couder
1 Introduction................................................. 20
2 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 20
3 Model ...................................................... 22
4 Numerical Procedure and Results ...... " .... .. .. .. . . .. .... .... 23
Growth Dynamics during Solidification of Undercooled Melts .. .. .. ... 26
Dieter Herlach
1 Introduction................................................. 26
2 Experimental................................................ 27
3 Description of Dendrite Growth in Undercooled Melts ............ 29
4 Experimental Results on Dendrite Growth in Undercooled Melts ... 32
4.1 Pure Metals ............................................ 32
4.2 Solid Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 34
4.3 Dilute Multiphase Alloys ................. " .. .. .. ... . . ... 35
4.4 Intermetallics with Superlattice Structures. . . . . . . . . . . . . . . . .. 37
4.5 Complex Structures of Polytetrahedral Phases .............. 39
VIII Table of Contents

4.6 Faceted and Non-Faceted Growth of Semiconductors . . . . . . . .. 41

5 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 44
Thermodynaxnics of Diffuse Interfaces ............................. 47
G.J.Schmitz
1 Introduction................................................. 47
2 General Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 49
3 Thermodynamic Approaches to Interfacial Energy. . .. .... .... .. .. 50
4 Crystal Growth Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 50
5 Approaches from Theories of Phase Transitions.... .. .. .. .. .. .. .. 55
6 Statistical Approach.. .. .. .. . . ... . . ... . ... .... .. .. .. .. .. .. .... 57
7 Conclusions................................................. 61
8 Acknowledgments............................................ 63
Computer Investigation of the InHuence of the Internal Structure
Topology on the Percolation Process in Two- and Three-Dimensional
Inhomogeneous Systems ......................................... 65
Aliaksei Konash, Sergey Bagnich
1 Introduction................................................. 65
2 Computational Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 66
3 Results and Discussion ....................................... 67
4 Conclusion.................................................. 73
Electron Transport of Nanoperm Alloys . . . . . . . . . . . . . . . . . . . . . . . . . . .. 75
K. Pekala
1 Introduction................................................. 75
2 Experiments................................................. 75
3 Results and Discussion ....................................... 76
3.1 Electrical Resistivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 76
3.2 Crystalline Fraction ..................................... 77
3.3 Thermoelectric Power. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 79
4 Conclusion.................................................. 79
Self-Organized Formation of Fractal and Regular Pores in Semiconductors 82
Jens Christian Claussen,Jürgen Carstensen, Mare Christophersen,
Sergiu Langa, Helmut Föll (University of Kiel)
1 Electrochemical Etching: Basic Experimental Setup .............. 82
2 The Current-Burst Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 83
3 Consequences on Pore Geometries . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 84
4 Conclusions................................................. 87
Evolution and Shapes of Dunes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 88
Hans J. Herrmann, Gerd Sauermann
1 Introduction................................................. 88
2 Experimental Measurements of a Barchan Dune ................. 89
2.1 Dune Morphology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 89
 Table of Contents IX

2.2 Wind Velocity and Sand Flux . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 92

3 The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 94
3.1 The Wind Shear Stress. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 95
3.2 The Sand Flux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 97
3.3 The Surface Evolution ................................... 98
4 The Shape of the Dune and Outlook ........................... 100
Morphogenesis of Growing Amorphous Films . . . . . . . . . . . . . . . . . . . . . . . 103
Stefan J. Linz, Martin Raible, Peter Hänggi
1 Introduction................................................. 103
2 Basic Concepts .............................................. 104
3 Deposition Equation for Thin Film Growth ..................... 109
4 Conclusions and Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

11 INTERFACE DYNAMICS: Modelling and Simulation

Density Effects and Fluid Flow in Phase-field Models. . . . . . . . . . . . . . . . 121

Massimo Conti
1 Introduction................................................. 121
2 Phase-field Model for a Pure Material .......................... 123
2.1 The Entropy Balance .................................... 123
2.2 The Thermodynamic Potential. . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
2.3 Nondimensional Equations in one Dimension ................ 126
3 Phase-field Model for a Binary Alloy ........................... 127
4 Numerical Solutions ......................................... 130
4.1 Pure Substance ......................................... 130
4.2 Binary Alloy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 133
5 Conclusions................................................. 134
A Lattice Boltzmann Method for the Mesoscopic Calculation of
Anisotropie Crystal Growth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Wolfram Miller and Igor Rasin
1 Introduction................................................. 136
2 Lattice Phase-field Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 136
3 Results..................................................... 138
4 Conclusions................................................. 140
A Phase-field Model for the Solidification Process in Multicomponent
Alloys ......................................................... 142
H. Garcke, B. Nestler, B. Stinner
1 Solidification Effects and Length Scales . . . . . . . . . . . . . . . . . . . . . . . .. 142
2 Description of the Model ...................................... 143
3 The Related Sharp Interface Model. ................. '" ........ 145
4 Examples ................................................... 147
5 Further Generalizations ....................................... 148
X Table of Contents

Planar Solidifieation from Undereooled Melt: An Approximation of a

Dilute Binary Alloy for a Phase-field Model ........................ 150
Denis Danilov
1 Introduction................................................. 150
2 The Model .................................................. 151
3 An Approximation of a Small Impurity Coneentration . . . . . . . . . . .. 152
4 Stationary Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 154
5 Large-Velocity Asymptoties ................................... 154
6 Numerieal Caleulations ....................................... 155
7 Conclusions................................................. 158

Initial Thansients in the Symmetrie Model for Direetional Solidifieation 160

Raul Benitez, Laureano Ramirez-Piscina
1 Introduction................................................. 160
2 The Sharp Interface Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 160
3 The Phase-field Model ........................................ 162
4 Results and Diseussion ....................................... 163
5 Concluding Remarks ......................................... 164

Dynamies of a Faceted Nematic-Smeetie B Front in Thin-Sample

Direetional Solidifieation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 166
Tamils Börzsönyi, Silvere Akamatsu, Gabriel Faivre
1 Introduetion................................................. 166
2 Results..................................................... 167
2.1 Experiments ............................................ 167
2.2 Linear Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 168
2.3 Phase-field Simulations ................................... 169

Last Stage Solidifieation of Alloys: A Theoretieal Study of Dendrite

Arm and Grain Coaleseenee ...................................... 172
M. Rappaz, A. Jacot and W. J. Boettinger
1 Introduetion and Theoretical Aspeets . . . . . . . . . . . . . . . . . . . . . . . . . .. 172
2 Results and Diseussion ....................................... 175
3 Conclusion.................................................. 179
4 Acknowledgements........................................... 180

Phase-field Modeling of Eutectie Solidifieation: From Oscillations to

Invasion ....................................................... 182
Roger Folch, M athis Plapp
1 Introduetion................................................. 182
2 Model ...................................................... 183
3 Simulations................................................. 185
4 Conclusion.................................................. 188
 Table of Contents XI

Phase-field Theory of Nucleation and Growth in Binary Alloys ........ 190

Lasz16 Gronasy, Tamas Börzsönyi*, Tamas Pusztai
1 Introduction: Diffuse Interface and Nucleation ................... 190
2 Phase-field Theory of Nucleation ............................... 191
2.1 Nucleation in 3D .................................. " .... 191
2.2 Multi Particle Solidification in 2D ......................... 192
3 Results and Discussion ....................................... 193

Modelling of Phase Transformations in Titanium Alloys with a

Phase-field Model ................ , .............................. 196
Benoft Appolaire, Elisabeth Gautier
1 Introduction................................................. 196
2 Which Model to Choose? ..................................... 197
3 Can we Rely on the Model to Get Quantitative Predictions? . . . . . .. 198

Spreading of Liquid Monolayers: From Kinetic Monte Carlo Simulations

to Continuum Limit ............................................. 202
M. N. Popescu, S. Dietrich
1 Introduction................................................. 202
2 Model and KMC Simulations .................................. 202
3 Results and Discussion ....................................... 204
4 Patterned Substrates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 206
5 Summary and Conclusions .................................. " 207

A Multi-mesh Finite Element Method for Phase-field Simulations ..... 208

Alfred Schmidt
1 Models for Solidification .................................. . . .. 208
1.1 Sharp Interface Models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
1.2 Diffuse Interface Models ................................ " 209
2 Adaptive Finite Element Methods for Coupled Systems of PDE .... 211
2.1 Aspects of Implementation ............................... 212
2.2 Numerical Analysis and Adaptive Methods ................. 212
3 Adaptive Method for Phase-field Models ........................ 213

Transport of Point Defects in Growing

Si Crystals ..................................................... 218
Axel Voigt, Christian Weichmann
1 Introduction ................................................. 218
2 Point defect modeling ........................................ 219
3 Finite element discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 222
4 Results..................................................... 223
5 Conclusions ................................................. 224
6 Acknowledgment ............................................. 225
XII Table of Contents

Atomistic Simulation of Transport Phenomena in Simple and Complex

Fluids and Fluid Mixtures ....................................... 226
Kurt Binder, Jürgen Horbach, Walter Kob, Fathollah Varnik
1 Introduction ................................................. 226
2 Interdiffusion and Selfdiffusion in Binary Mixtures (A,B) .......... 227
3 Estimation of Selfdiffusion Coefficients in Various Models of Fluids . 233
4 Estimation of the Shear Viscosity . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 240
5 Thermal Conductivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 244
6 Concluding Remarks ......................................... 245

Unusual Viscosity Feature in Spinodal Decomposition Under Shear Flow249

Jian Wang t ,+, Wolfram Gronski+, Christian Friedrick+, Peter
Galenko t and Dieter Herlacht
1 Introduction................................................. 249
2 Theory..................................................... 250
3 Simulation Detail ............................................ 251
4 Conclusion.................................................. 253

Micro-macro Approach to Cluster Formation in Granular Media ...... 255

S. Luding
1 Introduction................................................. 255
2 Models for Multi-particle Simulations ........................... 256
2.1 The Event-driven, Rigid Particle Method ................... 256
2.2 The Time Driven, Soft Particle Technique .................. 257
2.3 The Connection Between Hard- and Soft-sphere Models . . . . . . 258
3 Freely Cooling Granular Media ................................ 258
3.1 Homogeneous and Inhomogeneous Cooling .................. 260
3.2 Cluster Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 261
3.3 Cluster Growth ......................................... 261
3.4 Micro-macro Transition .................................. 263
4 Summary and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 265

III TRANSPORT

Urban Transport Phenomena in the Street Canyon .................. 269

Maciej M. Duras
1 The Field Models of Vehicles and Pollutants . . . . . . . . . . . . . . . . . . . . . 269
2 The Governing Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 270
3 Optimal Control Problems .................................... 272
4 Conc1usions................................................. 273
5 Acknowledgements........................................... 273
 Table of Contents XIII

Walker Behaviour Modelling by Differential Games .................. 275

Serge P. Hoogendoorn
1 Introduction ................................................. 275
2 Pedestrian Behaviour Framework .............................. 276
3 Theory of Walking ........................................... 278
4 Conceptual Pedestrian Walking Task Model . . . . . . . . . . . . . . . . . . . .. 280
4.1 Walking Subtask Hierarchy ............................... 280
4.2 Walking as a Feedback-oriented Control System ............. 281
4.3 Pedestrian Kinematics (Internal Model) .................... 283
4.4 Walking Discomfort (Resistance) .......................... 284
4.5 Cost Components ....................................... 285
4.6 Cost L l due to Deviation from Planned Route .... . . . . . . . . . . 285
4.7 Proximity Cost L 2 • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • 285
4.8 Acceleration and Deceleration Cost La . . . . . . . . . . . . . . . . . . . . . 286
5 Derivation of Walker Model ................................... 286
5.1 Application of the Maximum Principle ..................... 286
5.2 Cooperative Walker Model ............................... 289
5.3 Relation to the Social-forces Model ........................ 290
5.4 Model Refinements ...................................... 290
6 Approach to Model Calibration ................................ 291
7 Application Results .......................................... 292
8 Conclusions and Future Research Directions ..................... 293

Investigations of Vibrations in the Complex Dynamical Systems of

Transmission Pipelines ........................................... 295
Elena Mul, Vladimir Kravchenko
1 Introduction ................................................. 295
2 Model ...................................................... 296
3 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 296
4 Discussion of Results and Conclusions .......................... 300

Information in Intelligent Transportation System .................... 301

J. Wahle, M. Schreckenberg
1 Introduction ................................................. 301
2 Two-route scenario ........................................... 302
2.1 Simulation technique ..................................... 304
3 Floating-Car Data ........................................... 305
3.1 Infiuence of the dynamic drivers ........................... 307
3.2 Infiuence of the static drivers ............................. 309
4 Different criteria ............................................ 310
4.1 Gradient of travel time ................................... 310
4.2 Global density and speed ................................. 311
4.3 Comparison............................................ 313
5 Summary and conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 314
XIV Table of Contents

Experiments on Route Choice Behaviour ........................... 317

Reinhard Selten, Michael Schreckenberg, Thomas Pitz, Thorsten
Chmura, Joachim Wahle
1 Introduction ............................................... 317
2 Experimental Setup ......................................... 318
3 Equilibrium Predictions and Observed Behaviour ................ 318
4 Response Mode .............................................. 319
5 Payoffs and Road Changes .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 320
6 Conclusion .................................................. 320

Transport Out of a Gravitationally Stable Layer with the Help of a

Faster Diffusing Substance: PDE Simulations and Scaling Laws ....... 322
Karsten Koetter, Malte Schmick, Mario Markus
1 Introduction ................................................. 322
2 Analytical Estimates ......................................... 323
3 Experiments................................................. 324
4 PDE Simulations ............................................ 325
5 Conclusion .................................................. 326

Microscopic Parameters and Macroscopic Features of Traffic Flow ..... 329

Peter Berg, Eddie Wilson
1 Microscopic Parameters of Traffic Flow ......................... 329
2 Stochastic and Deterministic Microscopic Description ............. 330
2.1 A Cellular Automata Model .............................. 330
2.2 Car-following Models .................................... 331
3 Optimal-velocity Car-following Models .......................... 331
4 Derivation of Macroscopic Counterparts ........................ 332
5 Similarity of Flow Patterns .................................... 334
6 More Realistic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 335
6.1 Multi-species TraffiCj Temporal and Spatial Variations of
Driver Behaviour ........................................ 337
6.2 Reaction Time .......................................... 339
6.3 Multiple Look-ahead Models .............................. 339
6.4 Multi-lane Flow ......................................... 340
7 Synchronised Flow ........................................... 341
8 Outlook .................................................... 341

An Adaptive Smoothing Method for Traffic State Identification from

Incomplete Information .......................................... 343
Martin Treiber, Dirk Helbing
1 Introduction................................................. 343
2 Description of the Method .................................... 344
3 Application to German Freeways ............................... 350
4 Summary and Outlook ....................................... 357
 Table of Contents XV

Probabilistic Description of Nucleation in Vapours and on Roads ...... 361

Reinhard M ahnke
1 Introduction ................................................. 361
2 Stochastic Master Equation Approach .......................... 363
3 Nucleation in Supersaturated Vapours .......................... 368
4 Car Cluster Formation on Roads ............................... 376
5 An Advanced Model of Car Cluster Formation ................... 382
6 Conclusion .................................................. 385
7 Appendix: Derivation of Thermodynamic Potential and
Corresponding Transition Rate ................................ 385
Cellular Automata Simulation of Collective Phenomena in Pedestrian
Dynamics ...................................................... 390
Andreas Schadschneider, Ansgar Kirchner, Katsuhiro Nishinari
1 Introduction ................................................. 390
1.1 Collective Effects ........................................ 390
1.2 Modelling Approaches ................................... 392
2 Basic Principles of the Model .................................. 393
3 Definition of the Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 395
4 Collective Phenomena ........................................ 397
5 Influence of the Floor Fields ................................... 399
6 Friction Effects .............................................. 401
7 Conclusions ................................................. 403
Modeling, Simulation and Observations for Freeway Traffic and
Pedestrian ..................................................... 406
Yuki Sugiyama, Akihiro Nakayama
1 Introduction ................................................. 406
2 Brief Review of OV Model (Traffic in Circuit) ................... 406
3 Observations in the Bottleneck ................................ 408
4 OV Model with a Bottleneck .................................. 410
4.1 Modelling and Simulation ................................ 410
4.2 The Structure of Flow Upstream of Bottleneck .............. 412
4.3 Summary and Conjecture ................................ 417
5 Modeling Pedestrians in 2-dimensional OV Models ............... 418
Testing Traffic Flow Models ...................................... 422
Elmar Brock/eld, Peter Wagner
1 So Much Models, so Little Time ............................... 422
2 Theory ..................................................... 422
3 Testing Dynamical Models .................................... 423
4 The Trouble with the Data .................................... 426
5 Daganzo's Data .............................................. 426
6 The Contest ................................................. 427
6.1 Building a Reference Model ............................... 427
6.2 Technicalities ........................................... 428
XVI Table of Contents

6.3 The Rest of the Pack .................................... 429

6.4 Finally, the Rest (Preliminary Results!!!) ................... 430
7 Conclusions & Further Plans .................................. 431
 Part I

INTERFACE DYNAMICS:
Phenomenology and
Modelling
 Melting Kinetics of Prolate Spheroidal Crystals

M.E. Glicksman1 , A. Lupulescu1 , and M.B. Koss 2

1 Rensselaer Polytechnie Institute, Materials Science & Engineering Department,
Troy, NY, 12180-3590 USA.
2 Department of Physics, College of the Holy Cross, Worcester, MA 01610-2395
USA.

Abstract. The melting kinetics of a pivalic acid (PVA) dendritic mushy zone was
observed for the first time under convection-free conditions. Video data show that
PVA dendrites melt into fragments that shrink at accelerating rates to extinction.
Individual fragments follow a characteristic time-dependence derived here for the di-
minishing length scales within a melting mushy zone. The melting kinetics against
which the experimental observations are compared is based on the conduction-
limited quasi-static process of melting under shape-preserving conditions. Agree-
ment between analytic theory and experiment was found for the melting of a prolate
spheroidal crystal fragment with an aspect ratio of C/ A = 12.

1 Introduction
1.1 Dendritic Growth
Many theoretical [1], [2], [3], [4] and experimental [5], [6], [7], [8] dendritic
growth studies were reported over the past 40 years. Dendrite growth, as a
pattern-forming process, is much simpler to describe as free crystallization
from a pure molten phase. The thermodynamic driving force for crystalliza-
tion in this case is the melt supercooling, LlT == Tm - T 00, where Tm is the
melting point of the material, and T00 is the temperature of the melt far from
the heat emitting solid-liquid interface. The crystallization process for free
growth requires that the latent heat released be transported away from the
solid-liquid interface by thermal conduction to the cooler (supercooled) melt.

1.2 Isothermal Dendritic Growth Experiment (IDGE)

Theories of dendritic growth in pure materials are best tested under condi-
tions of conduction-controlled heat transfer, a situation possible only where
gravity is reduced to nearly zero, where buoyancy-induced natural convec-
tion is eliminated [10], [11], [12], [13]. The Isothermal Dendritic Growth Ex-
periment, or IDGE, is a basic science experiment, developed under NASA
support, that provides both terrestrial and microgravity data on the kinet-
ics and morphology of dendritic solidification. Before the advent of IDGE in
1994, it was not possible to test quantitatively either Ivantsov's solution [14]
for the steady-state transport field surrounding a dendrite, or evaluate the
H. Emmerich et al. (eds.), Interface and Transport Dynamics
© Springer-Verlag Berlin Heidelberg 2003
4 M.E. Glicksman, A. Lupulescu, M.B. Koss

so-called interfacial ''stability constant," a* ~ 0.02, predicted from several

physics theories of near-equilibrium solid-liquid interfaces [2], [3], [4].
The IDGE flew in near-Earth orbit three times, and provided the first
compelling evidence that lvantsov's heat-conduction solution c10sely descri-
bes steady-state dendritic growth for the test materials succinonitrile (SCN)
and pivalic acid (PVA). These cubic organic crystals offer three major ex-
perimental advantages compared to metals: 1) they are transparent, 2) easily
purified to high levels, and 3) have conveniently low melting points. lOGE
instruments, on board the space shuttle Columbia, provided CCO images
(telemetered to Earth during each flight), NASA-processed 35-mm film neg-
atives (available postflight), and, for the first time, on USMP-4, real-time
streaming of video data telemetered to our laboratory via the K-band (high-
frequency) antenna on the shuttle.

1.3 Mushy Zones

A mushy zone is defined as a finely-mixed two-phase region, consisting of

a dendritic solid phase embedded within its melt phase. Mushy zones are
known to occur in ingots, castings, and welds, and are thought to play a
röle in the evolution of the Earth's semi-solid core, and planetesimals. Mushy
zone, in brief, represent "active" regions in a solidifying body, in which the
microstructure forms and evolves [15], [16]. In order to simplify the behavior
of a melting mushy zone, we studied pure PVA (4-9's purity) under nearly
gravity-free conditions. The melting of a mushy zone in a pure substance uno..
der convection-free conditions should be kinetically controlled by conduction-
limited heat transfer, provided that convection is absent. The kinetics of
melting is, of course, normally studied in the presence of relatively strong
buoyancy-induced melt convection. Convection alters in complicated ways
the thermal fields surrounding dendrites, plus, as an added complexity, sedi-
mentation often takes place because the solid and its melt usually differ (by a
few percent) in their mass densities. Conduction-limited freezing and melting
processes are of importance in both terrestrial and planetary processes. For
example, mushy-zone evolution in castings and ingots, and in fusion welding,
can occur where the length scales available for thermal buoyancy are limited,
and little convection occurs. Similar mushy-zone behavior may occur in eIe-
mentary Earth's core processes [17], [18], [19], [20],[21], [22], and in meteoritic
and planetary genesis [23], [24], [25], all of which occur either where gravity,
pe se, is nearly absent, or reduced below terrestriallevels.

2 Video Melting Data

When melting begins, the dendritic side-arms first shorten, and then larger
secondary branches detach from the primary stem. Figure 1 shows several
video frames representative of a few stages of the total melting sequence. In
 Prolate Spheroidal Crystals 5

microgravity, it is important to note that the individual crystalline elements

of the mushy zone remain motionless. At the late stages of melting, the re-
maining dendritic fragments present a large variety of shapes, from nearly
spherical blobs to elongated needle-like crystals, all of which eventually melt
away.
The mushy-zone melting events, recorded every I/30th second as video
frames, provide to our knowledge the first quantitative data compilation of
convection-free melting over a significant range of length scales: from about
1O-2 m down to about 50 x 1O-4 m. The large dynamic range of these length
scales allows a detailed kinetic analysis of the convection-free melting process.

_ _--' ime

Fig.1. Video sequence (frame height '" 1.5 cm) showing the progressive melting of
PVA dendrites in microgravity. As the dendrites fragment, individual crystallites
remain motionless within the melt. Video data were analyzed frame-by-frame for
the melting of the needle crystal shown in the last panel and reported in Figure 6.

3 Data Analysis
3.1 Image Processing
The shapes of the melting dendritic fragments, as analyzed from the IDGE
video data, were determined using commercial image analysis software. The
fragments, particularly toward the later stages of melting, are best described
as prolate spheroids. Figure 2 shows the shape of a melting dendrite stem out-
lined using this software too!. The equivalent ellipse of the two-dimensional
image of this fragment is calculated on the basis of its major axis, C, and mi. .
nor axis, A. The physical particles are, of course, three-dimensional objects,
approximated over time as prolate spheroids of decreasing size. Image anal-
ysis permitted measurement of the major and minor axes of the equivalent
ellipsoidal fragment as a function of time. Sampling the C/ A ratio over time
towards the end of the melting cycle shows that the values of the C/ A ratio
for a particular fragment do not vary much. For the particular crystal mea-
sured, the initiallength was Co = 0.76cm. Its C/ A ratio varied from a starting
6 M.E. Glicksman, A. Lupulescu, M.B. Koss

value of about 11, rose to a maximum of about 14, and then decreased to
about 10, with almost 95% of the crystal's volume melted. Thus, the C/A
ratio for this melting fragment may be approximated by its volume-weighted
average, which in this instance was (C / A) = 12.0. As shown in Figure 2, this
crystal is needle-like in appearance.

Video Frame: Time =aa

Fig. 2. Sampie digitized video frame for image analysis of a melting PVA dendritic
fragment . The image analyzer calculates the equivalent ellipse for the fragment,
allowing determination of its major axis length, C(t), and its C/A ratio.

4 Quasi-static Theory
4.1 Background
Quasi-static and moving boundary solutions are weH known [26] [27] for the
growth of spherical crystals for which (C / A = 1). For the more general cases
of freezing ellipsoids and hyperboloids, moving boundary solutions were de-
veloped by Harn [28]. However, Ham's similarity solutions for ellipsoids ap-
ply only to the case of crystal growth. Harn clearly states in his paper that
the similarity transformations used in his moving boundary solutions are
inapplicable to cases of melting ellipsoidal particles. Lacking an analytical
moving boundary solution, we developed a quasi-static model for the con-
duction melting of prolate spheroids. Subsequently, we will demonstrate that
the melting kinetics, without convection, of needle-like dendritic fragments
may be accurately described as a quasi-static conduction-limited process.

4.2 Potential Theoretic Formulation

The quasi-static melting or freezing of a crystal is based on determining
the heat current entering or leaving its staUe solid-melt interface, and then
 Prolate Spheroidal Crystals 7

formulating kinematic equations for the crystal's shrinkage or growth. The

rate of change of the particle's volume is assumed to occur slowly enough
so that the surrounding transport field has sufficient time to adjust to the
changing particle. The thermal transport field used only approximates the
time-dependent solution to the conduction equation as a quasi-static solution
based on Laplace's equation. The key condition determining the applicability
of such approximations is that the Stefan number-the intensity measure for
melting or freezing-is small, Le., St « 1.
We first choose cylindrical coordinates, (r, z, cp), to describe a prolate sphe-
roidal crystal with major axis, C, and minor axis A. The crystal's major axis
is aligned with the z-axis. The plane z = 0 is the equatorial plane that divides
the crystal symmetrically, so that its poles (r = 0) occur at the semi-major
axis locations, z = ±C/2. The axis-symmetric temperature field, T(r,z), is
non-dimensionalized as a thermal potential, iJ, defined as

_O( ) = T(r,z) - Tm (1)

'V r,z - T oo -Tm '

where T00 is the temperature of the melt infinitely far away from the particle.
The propensity and rate of melting (or freezing) is determined by the sign and
magnitude of the denominator in Eq.(l). A dimensionless parameter based
on that temperature difference is the Stefan number, defined here as
Too -Tm
(2)
St == LlHf/C:'

where LlHf is the molar enthalpy of fusion, and Cf is the molar specific heat
of the melt at constant pressure. It is convenient to express the potential
deflned in Eq.(l) in an orthogonal, curvilinear coordinate system in which
the solid-melt interface of the prolate spheroid is an isothermal coordinate
e
surface labeled = E, for which the potential is iJ(E) = O. Such a coordinate
system, (e, 11, cp), is discussed by Morse and Feshbach [29], for which the metric
e
coordinates and 11 are sCaled to the focal length of the prolate spheroid,
a/2, where the interfocallength a = C/E = A/VE2 -1. Confocal prolate
spheroidal coordinates with a = 2 are shown as the cp = 0 projection in
Figure 3. In brief, heat ßow to or from a prolate spheroid occurs through
the melt along the hyperboloidal surfaces 1J =const., whereas the isotherms
e
of the melt's temperature field are confocal prolate ellipsoids, ~ E=const.
Specifically, the Laplacian potential surrounding a prolate spheroid with the
fixed interfacial potential iJ(E) = 0, embedded in a melt where the far-field
potential iJ( 00) = 1, is given in these curvilinear-coordinates as

log ill
iJ(e, 11) = 1- log tl·
3-1
(3)
8 M.E. Glicksman, A. Lupulescu, M.B. Koss

Fig.3. Confocal prolate spheroidal coordinates (e,.",<p) on the <p = 0 plane. The
e
solid-melt interface has the profile = 8, where its aspect ratio C/ A = 8/";8 2 - 1.
e
The isotherms surrounding this crystal are surfaces ~ 8 = const., and the heat
flow follows surfaces of the form ." = constant. The foci for these ellipsoids are
e
located at = 1,1'] = ±1, (<p arbitrary) so the interfocallength is a = 2.

The gradient field expressed in prolate spheroidal coordinates associated

with the thermal potential given in Eq.(3) is easily found by taking the partial
derivative of the potential, 'I?(~, 1/), with respect to ~, and dividing the result
by the system's ~-coordinate scale factor, he = (a/2)J(~2 - rp)/(~2 - 1).
Carrying out these mathematical operations yields the desired gradient field

1 a'l?(~,1J) 2 ~ 2 (5+1)-1
h e a~ = 'V'I?(~,1/) = ~y ~ (~+ 1)(~ -1) log 5 -1 ' (4)

which may be simplified to

4 ( =+1)-1
'V'I?(~, 1/) = - J(~2 - 1)(~2 - 1/2) log;'-- (5)
a .::-1
The relationship of the gradient field, Eq.(5), to the actual thermal gradient
in the melt may be found by differentiating Eq.(l) with respect to ~, and
again dividing through by the curvilinear coordinate scale factor, he,
1 8T(r,z) ( 1 ) (6)
'V'I?(r,z) = he a~ Too-Tm ·
 Prolate Spheroidal Crystals 9

Rescaling the potential gradient through the characteristic temperature of

the material, .6.Hf I Cf, and substituting the definition of the coordinate scale
factor, he, yields

.6.Hf
Cf
\1'19(~
,"7
) = .6.Hf . ~ ~. ßT(r,z)
Cf aV~ ß~
( 1)
Too-Tm ·
(7)

If the temperature gradients are evaluated at the solid-melt interface, where

~ = 3, then Eq.(7) may be written in the compact form
(8)

Here the definition of the Stefan number, Eq.(2), has been substituted into
Eq.(7), and the symbol \1Ts denotes the dimensional thermal gradient in the
melt evaluated at the solid-melt interface, ~ = 3.

4.3 Enthalpy Flux

The enthalpy flux, J, for thermal conduction through the melt normal to the
solid-melt interface may be found using Fourier's law, J = -kl\1Ts, where
kl is the thermal conductivity of the melt. The gradient at the solid-melt
interface, Eq.(8), can be substituted into Fourier's law to yield the enthalpy
flux,

J = -kl .6.;f St· \1'19(3, Tl). (9)

l

Inserting the expression for the potential gradient, Eq.(5), evaluated on the
solid-melt interface, ~ = 3, into Eq.(9), gives the thermal flux normal to the
solid-melt interface, as

A nondimensionless flux, .1, may be defined as the last term in Eq.(lO),

namely,

.1 == - [J(3 2 - 1)(32_ Tl2) log ~ ~ ~]-1 (11)

The flux entering the ellipsoidal crystal at its poles, "7 = ±1, is found from
Eq.(ll) to be

[ -2
.1pole=- (.:: -1)log3_1
_]-1
+
.:: 1
' (12)
10 M.E. Glicksman, A. Lupulescu, M.B. Koss

whereas the flux entering the ellipsoid at the equatorial plane, 1/ = 0, is

Jequator =- [V(5 2 _ 1)(52) log ~ ~ ~]-1 (13)

The ratio of the polar-to-equatorial fluxes is

Jpole ~ C
(14)
Jequator = v'52 - 1 = A'

Equation (14) proves that the the flux ratio between the poles and the equator
is equal to the C/ A ratio. This result is eonsistent with the requirement
that the imposed quasi-statie melting rates maintain the particle's shape by
keeping the C/ A ratio fixed in time. Note too that the 1/-dependent terms in
Eqs.(10) and (11) are responsible for what is often termed the ''point-effeet''
of diffusion. As a prolate spheroid inereases its C/ A ratio, it beeomes more
needle-like, and the fluxes around the poles inerease eompared to the fluxes
in the equatorial region. See Figure 4.

--
I

~
~5~
I
I-
CIA:') (sphi'as)

o 0,2 0.4 0.6 1

Equator N. Pole
."

Fig.4. "Point-effect" of diffusion as described by Eq.(l1). The flux (log-scale) at

any latitude, 'T}, is normalized to the equatorial flux. As the spheroid becomes more
needle-like, the fluxes in the polar regions ('T) '" 1) inerease relative to those in the
equatorial region (11 '" 0).
 Prolate Spheroidal Crystals 11

4.4 Enthalpy Current

The total heat current conducted to or from astaUe prolate spheroid may
be expressed as a surface integral of the enthalpy ßux,

Qtot = f J ·ndS, (15)

where the differential element of solid-melt interfacial area on the prolate

e
spheroid, = 3, may be expressed in the curvilinear coordinates as
(16)

where r = a/2V(3 2 - 1)(1- rf) is the radial distance from the major axis of
the spheroid to an arbitrary point, TJ, on its solid-melt interface; du = hf/dTJ ia
the differential of the arc length; and hf/ is the TJ-coordinate scale factor needed
to relate the arc length, 0', along the interface in curvilinear coordinates
defined by Morse & Feshbach as hf/ == a/2V(3 2 -TJ2)/(I-TJ2).
Substituting these transformations into Eq.(16) yields the differential el-
ement of interfacial area on the prolate spheroid, = 3, e
2
dS = ~V(32
4
- 1)(32 -TJ2)dTJd<p. (17)

Inserting Eq.(lO) for the normal ßux vector, and Eq.(17) for the differen-
tial interfacial area, into Eq.(15) yields an integral expression for the total
enthalpy current,

.
Qtot =-
akt.1HfSt (
Cf
3 + 1)-1 [1
log 3 _ 1 LI dTJ
12
0
11"
dcp, (18)

where the integration limits -1 ~ TJ ~ 1 and 0 ~ cp ~ 21l" correspond to the

e
total interfacial area from pole-to-pole on the ellipsoid = 3. Carrying out
the closed surface integration indicated in Eq.(18) introduces the expected to-
tal spherical image, 41l", and yields the total heat current entering the prolate
e
spheroid, = 3, with an interfocallength a,
• akt.1Hf
Qtot = - 1og41l"~
.5'-1
cI.P St. (19)

4.5 Kinematic Formulation

Heat entering (or leaving) the spheroidal crystal results in its shrinkage (or
growth) caused by the S ~ L phase transformation. The time-dependent
volume of a melting or freezing prolate spheroid may be expressed, applying
our notation, in terms of its time-dependent major- and minor-axes through
the formula [30],

V(t) = i A2 (t)C(t) i
= (~) 2 C3 (t), (20)
12 M.E. Glicksman, A. Lupulescu, M.B. Koss

where the time-dependence of the spheroid's fixed aspect ratio, C/A, is ex-
plicitly suppressed here in the notation used for the right-hand equality. Note
that in the limiting case of a sphere of diameter 'D, the aspect ratio C/ A = 1,
so that C = A = 'D, and the familiar result obtains from Eq.(20). H the
crystalline spheroid melts (or freezes) and is subject to a time-rate of change
of its volume, then the time-rate of change of its major-axis length is related
kinematically to the volume derivative as

V(t) = i (:) 2 C2(t) C(t), (21)

where the overdot notation indieates differentiation with respect to time. The
rate of enthalpy change associated with the volume change may be found by
multiplying Eq.(2I) by the volumetrie enthalpy, i1Hf / [J,

•
Htot = 2n
'Ir i1Hf (A)2 2 •
C C (t) C(t), (22)

where [J is the molar volume. This enthalpy expression ignores the impercep-
tible advection occurring in the melt near the interface caused by the small
(~ 3%) density differences of the phases. Corien and McFadden have shown
that this effect is generally negligible in the case of steady-state dendritie
crystallization [31). The kinematic energy balance required under quasi-static
conditions is therefore Htot = Qtot. Equating Eq.(22) with Eq.(I9) gives

?:. i1Hf (A)2 C2(t) C(t) = _ 4~ akti1Hf St. (23)

2 [J C -=.+1
log .:s-1 ctp
Equation (23) is the O.D.E. for the quasi-static evolution of a prolate sphero-
idal crystal of constant aspect ratio contacting a superheated, or supercooled,
melt of Stefan number, St. Equation (23) may be solved for the measurable
quantity C(t) by substituting the transformation for the interfocal length
a = C/ E, and then separating the variables, after which one obtains

lCo
C(t)
CdC = - :;'
8 (C)2 at St
.a:±!. A
~ log 8-1
l
0
t
I
dt . (24)

Carrying out the integrations indicated in Eq.(24) yields, after a few steps of
algebra, the kinetie expression for quasi-static melting or freezing of prolate
spheroids, namely .

C(t) = C2 _ I6atSt (E.-) 2 t (25)

o Elog ~2:~ A .

Equation (25) may be non-dimensionalized as

C(t)
Co = JI - Kprol . St . Fo, (26)
 Prolate Spheroidal Crystals 13

where the kinetic eoefficient, K prol , is defined as

K prol ==::;' 16~ (AC)2, (27)

~ log =-1

and the Fourier number (dimensionless transformation time) is defined as

(Xl
Fa == C2 t. (28)
o
It is eonvenient for the eurrent experimental purposes to relate K prol to the
erystal's CjA ratio. Equation (27) is cross-plotted in Figure 5 against the
aspeet ratio Cj A = 5/ V52 - 1. If one sets aside questions of interfacial
stability, Eq.(25) provides an analytic expression of the quasi-statie kinetic
law for the growth (St 2:: 0) or melting (St ~ 0) of a prolate spheroidal
crystal.

I
c:
,!l
u
~

I
0
100

g
II
I
QJ

J
~

10 I -
0 5 10 15 20

Aspect Ratio . ClA

Fig.5. Quasi-static kinetic coefficient, Kprol, versus the aspect ratio, C/A. Note
that this kinetic coefficient approaches a value of 8 as the C/ A ratio approaches
unity (spheres), but rises rapidly for the case of ellipsoidal needle-like crystals.

5 Results
5.1 Experimental Melting Rates
The PVA crystal that we analyzed in microgravity formed originally as a
part of a dendritic mushy zone created in IDGE growth cycle 4. These den-
14 M.E. Glicksman, A. Lupulescu, M.B. Koss

drites grew from pure molten pivalic acid that was supercooled OAOK. The
mushy zone consisted of approximately 4% dendritic solid when melting was
initiated. The total time needed to melt this mushy zone completely was ap-
proximately 12 minutes. The last large fragment to melt was the one selected
for quantitative analysis. (See again the final panel shown in Figure 1.) This
dendrite fragment (of initiallength Co = 0.76cm) was the residual crystallite
from a melting primary dendrite stern, and exhibited a nearlyconstant aspect
ratio C/ A ~ 12 throughout its final 40s melting sequence. To calculate the
melting kinetics ofthis dendritic fragment, one must know, in addition to the
C/ A ratio, the thermal diffusivity and the Stefan number of the melt. The
thermistors arranged within the IDGE thermostat were optimized for mea-
suring the supercooling prior to dendritic growth quite precisely (±0.002K).
The subsequent temperature history within the growth chamber, especially
during the melting portion of an experimental growth cycle, is, unfortunately,
less certain. Specifically, temperature uncertainties arose because of a lack of
precise synchronization between the video clock and the mission clock. In
view of this situation, one may instead adopt a self-consistent method of
estimating the Stefan number in the melt during the last 40s of the video
melting sequence. The basis of the value of the self-consistent Stefan num-
ber is just the observed extinction time of the fragment. The extinction time
(in this instance, 40s) is the interval needed for the melting crystal to just
disappear. H the left-hand side of Eq.(25) is set equal to zero, a relationship
may be established among the unknown Stefan number, the starting size of
the crystal, and the crystal's observed extinction time, tt = 408, namely
0,2 1
St= 0 .-. (29)
at· Kprol tt
The extinction time can be measured precisely to within 1 video frame
(±0.033s), and the thermal diffusivity of molten PVA at its melting point
is known to be at = 7.0 x 1O-4cm2/s [13]. Inserting these data and the value
for Kprol = 360.6 into Eq.(29), (See again Figure 5 for C/A = 12.0.) yields
the self-consistent Stefan number, St = 0.057, based on the observed ex-
tinction time of 40s . The crystal melting data reported in Figure 6 were fit
in their entirety using the self-consistent Stefan number, which corresponds
to the melt being superheated 0.63K above its equilibrium melting point,
Tm = 35.97K.
The overall prediction of the melting kinetics during the lifetime of this
dendrite fragment, using the quasi-static theory developed in Seetion 4, ap:-
pears to be in good agreement with experiment. Scrutiny of the video data
indicates that only small disparities with the predictions occur over the melt-
ing interval. Particularly as the extinction time is approached, and the speed
of melting accelerates, capillary and interfacial attachment effects may enter
the kinetics and alter the melting rate from that calculated by quasi-static
conduction theory. The departures observed here from quasi-static theory
are, however, minor, and will not be pursued further in this paper.
 Prolate Spheroidal Crystals 15

0,8
I ' I
0.7 -- k /A=12.0 ~- 0 C(l)eKp. -
I ".'--j- I tt=40s

-+--- _,-+---~--'" "

. -
0.6 - --,- - - C(t) 1'Iecty
E
!:.
:0-
0'
!l5
-1- ..

..
ui
~
0.4 --L-. L,--J---,_.
I I
I

I
0 I
ro- 0.3
I I

.
~ J
I I
0.2 - _ _L _
--_._)- "'--- ----
0.1
I
l.
IJ !
I
·1
j
----
iI I

0
0 10 20 30 40 50
Time 151

Fig. 6. Melting kinetics of a dendritic pivalic acid (PVA) crystal fragment (CjA =
12) in microgravity. Experimental melting data (open symbols) obtained through
digital analysis of 230 digitized video frarnes measure C(t) versus time. Solid curve is
the theoretical prediction from Eq.(25) fitted with aStefan number equal to 0.057.
The fit with quasi-static theory is based on the observed extinction time tt = 40s.

5.2 Thermal Data

Thermal data telemetered from the IDGE thermostat indicate independently

that the maximum Stefan number recorded by the high-resolution thermis-
tors interna! to the growth chamber at the end of this experimental cycle was
about 0.16. Melting of the last dendritic fragments clearly occurred at Ste-
fan numbers much less than 0.16, but the precise range of values over which
the observed crystal melted remains imperfectly known. As an alternative
to obtaining direct measurements of the melt temperature within the cham-
ber, one may recognize that the rate of temperature increase in the growth
chamber depended on two countervailing factors: 1) the temperature differ-
ence between the exterior thermostat temperature (fixed at 37.85C during
melting) and the interior melt temperature, and 2) the quantity of crystals
remaining within the chamber that actively absorb latent heat. When melting
of the mushy zone begins near the chamber walls, the temperature difference
is ",1.8K, and the chamber contains "'4% dendritic crystals. The rate oftem-
perature rise is, however, initially slow, because most of the heat added is
absorbed as the latent heat of melting. As the dendrites melt and their total
mass decreases, the rate of latent heat absorption also decreases, and thus the
rate of temperature rise increases. The rate of temperature rise within the
IDGE growth chamber should reach its maximum value at about the time at
16 M.E. Glicksman, A. Lupulescu, M.B. Koss

which the last crystals disappear, i.e., near the time when our detailed melt-
ing data were taken. It is at that point where all latent heat absorption ceases
within the chamber, yet the temperature difference between the interior and
the exterior is still relatively large ("" 1.2K in this case). We emphasize that
using a heating-rate versus time correlation is an approximate procedure.
Nonetheless, when the time derivative of the Stefan number versus time is
calculated for the experiment, a distinct peak occurs about half-way through
the total melting period at t = 720s. The time at which the peak rate oc-
curs corresponds to aStefan number of 0.055. During the 40s period for the
melting sequence of the prolate spheroid we selected, the value of the max-
imum time derivative of the Stefan number is dSt/dt ~ 2 x 1O-4 s- 1 . Thus,
the change of Stefan numbers over the 40s observation period is estimated
to be less than "" 0.008. The average Stefan number during melting based
on the IDGE thermal data is 0.055 ± 0.004-a value in agreement with the
self-consistent Stefan number (St = 0.057 ± 0.001) determined from Eq.(29)
by fitting the kinetic data to the crystal's observed extinction time, tt = 40s.
Possible melting behaviors for prolate spheroidal crystalswith C/ A = 12,
subjected to different Stefan numbers are shown for comparison in Figure 7.

0-8

0.08 0,'

- -olmenslonless
- Me!!'9 Ttme!o:"tIl?'•.

Fig. 7. Major axis length scaled to initial major axis length, C(t)/Co, versus melting
time, predicted by Eq.(25) for C/A = 12. These curves show the sensitivity to the
the Stefan number of the melting kinetics for needle-like prolate spheroidal crystals.

These curves are predicted using Eq.(25) with the indicated Stefan num-
bers, which were selected to cover the full range of values encountered dur-
ing the 12 minute melting portion of the experimental cycle. Note that the
melting kinetic curves for C/ A = 12 change substantially with only modest
variations in the Stefan number. It is interesting to point out that as the as-
 Prolate Spheroidal Crystals 17

pect ratio of a crystal decreases towards unity, its kinetic melting coefficient,
Kprol, rapidly falls toward the limiting value for spheres of 8. Extinction
times, as noted in Figure 7, are rather sensitive to a crystal's aspect ratio.

6 Conclusions
1. Dendritic mushy zones were formed in supercooled high-purity PVA melts
in microgravity. At the end of each experimental cycle the mushy zone was
progressively melted by raising the temperature of the thermostat. Video
data at 30 fps of the melting process for one cycle (OAK supercooling)
are reported here. Upon raising the temperature, the mushy zone melts
into numerous fragments, many of which may be characterized as prolate
spheroids.
2. Individual dendritic fragments remain motionless in the low-gravity envi-
ronment, progressively melting toward extinction at nearly constant C/ A
ratios.
3. Moving boundary solutions, such as those proposed by Ham for ellipsoidal
crystals, are not applicable to the melting process. We developed instead
a more broadly applicable quasi-static analytic theory, using curvilinear
coordinates, in lieu of a numerical solution to the moving boundary prob-
lem. The theory describes the kinetics of conduction melting of prolate
spheroids of arbitrary C/ A ratios. The quasi-static analytic results are,
however, limited to small Stefan numbers (St« 1).
4. The observed melting kinetics in microgravity of a prolate spheroidal crys-
tal with an aspect ratio C/ A = 12 was found to be described accurately
by the quasi-static analysis. Predictions of the crystal size as a function
of time require knowledge of the Stefan number and thermal diffusivity of
the melt, along with the kinetic parameter derived here from the theory
that accounts for the crystal's aspect ratio.
5. As the C/ A increases, and a spheroid becomes more needle-like in shape,
the polar flux increases relative to the equatorial flux, in accord with the
so-called ''point effect" of diffusion.
6. The Stefan number required to compare theory and experiment was esti-
mated using the observed extinction time of the crystal. Measured heating
rates yield an independent Stefan number found to be in good agreement
with the self-consistent value derived from the extinction time. The pre-
dicted convection-free melting kinetics accurately describe the observa-
tions in microgravity that are reported here.

7 Acknowledgements
The authors gratefully ac1mowledge the support provided for this work by the
National Aeronautics and Space Administration, Division of Biological and
Physical Sciences Research, Code U, Washington, D.C., under Grant NAG
8-1696.
18 M.E. Glicksman, A. Lupulescu, M.B. Koss

References

1. Glicksman, M.E., Marsh, S.P., " The Dendrite." Handbook 0/ Crystal Growth.
Ed. D.T.J. Hude, Elsevier Scienee Publishers, Amsterdam (1993) 1,075-1,122.
2. Langer, J.S., Muller-Krumbhaar, H.," Theory of dendritic growth. I. Elements
of astabilityanalysis." Acta Metallurgica, Vol. 26 (1978) 1,681-1,687.
3. Langer,J.S.," Nonequilibrium physies and the origins of eomplexity in nature."
Princeton Series in Physics. Critical Problems in Physics: Chapter 2, Prineeton
University Press 1996.
4. Mullins,W.W., Sekerka S.F., "Morphologieal stability of a particle growing by
diffusion of heat flow." J. Appl. Phys., Vol. 34 (1961) 323-329.
5. Glicksman, M.E., Schaefer, R.J., Ayers, J.D., "Dendritie Growth-A Test ofThe-
ory." Metall. 7rons. A, Vol. 7A (1976) 1,747-1,759.
6. Glicksman, M.E., "Fundamentals of Dendritic Crystal Growth." Crystal Growth
0/ Electronic Materials. Chapter 5. Ed. E. Kaldis, Elsevier Scienee Publishers
(1985) 57-69.
7. Huang, S.C., Glicksman, M.E., "Fundament als of dendritie solidifieation-I.
Steady-state tip growth." Acta Metallurgica, Vol. 29 (1981) 701-715.
8. Trivedi, R, Mason, J.T., "The effects of interface attachment kineties on solid-
ifieation interface morphologies." Metall. 7rons. A, Vol. 22A (1991) 235-249.
9. Lipton, J., Glicksman, M.E., Kurz, W., "Equiaxed Dendrite Growth in Alloys
at Small Supereooling." Metall. 7rons. A, Vol. 18A (1987) 341-345.
10. Glicksman, M.E., Koss, M.B., BushnelI, L.T., Lacombe, J.C., Winsa, E.A., ''The
Isothermal Dendritic Growth Experiment." Materials Science Forum, Vol. 215-
216 (1996) 179-190.
11. LaCombe, J.C., Koss, M.B., Fradkov, V.E., Glicksman, M.E., " Three-
dimensional dendrite-tip morphology." Physical Review, E, Vol. 52, No. 3 (1995)
2,778-2,786.
12. LaCombe, J.C., Koss, M.B., Corrigan, D.C., Lupuleseu, A., Tennenhouse, L.A.,
''lmplieations of the interface shape on steady-state dendritie crystal growth."
Journal 0/ Crystal Growth, Vol. 206 (1999) 331-344.
13. Lupuleseu, A., LaCombe, J.C., Frei, J.E., Glicksman, M.E., Koss, M.B.,
"USMP-4 Microgravity PVA Dendritic Growth Data -Preliminary Analysis-
." Microgmvity Symposium: Proceedings TMS Meeting, Nashville, 2000, CD-
ROM.
14. Ivantsov, G.P., "The Temperature Field Around Spherieal , Cylindrical, and
Needle-Shaped Crystals which Grow in Supereooled Melts." Dokl. Akad. Nauk.
USSR, Vol. 58, No. 4 (1947) 567-569.
15. Marsh, S.P., Glicksman, M.E., "Overview of geometrie effeets on eoarsening
of mushy zones." Metallurgical and Materials 7ronsactions A, Vol. 27A (1996)
557-567.
16. Paradies, C.J., Smith, RN., Glicksman, M.E., "The influenee of eonvection
during solidifieation on fragmentation of the mushy zone of a model alloy."
Metallurgical and Materials 7ronsactions A, Vol. 28A (1997) 875-883.
17. Fearn, D.R, Loper, D.E., Roberts, P.H., "Structure of the Earth's inner eore."
Nature, Vol. 292 (1981) 232-233.
18. Loper, D.E., "Structure of the inner eore boundary." Geophys. Astrophys. Fluid
Dyn. Vol. 25 (1983) 139-155.
 Prolate Spheroidal Crystals 19

19. Wenk, H.R, Takeshita, T., Jeanloz, R., Johnson, G.C., ''Development oftexture
and elastie anisotropy during deformation of hep metals." Geophysical Research
Letters, Vol. 15 (1988) 76-79.
20. Cormier, V.F., ''Inner eore strueture inferred from body waveforms." EOS, Vol.
75 (1994) 67.
21. Souriau, A., Romanowiez, B., "Anisotropy in inner eore attenuation: a new type
of data to eonstrain the nature of the solid eore." Geophysical Research Letters,
Vol. 23 (1996) 1-4.
22. Bergman, M.L, "Measurements of elastie anisotropy due to solidifieation tex-
turing and the implieations for the Earth's inner eore." Nature, Vol. 389 (1997)
60-63.
23. Rushmer, T., Minarik, W.G., Taylor, G.J., "Physical Proeesses of Core For-
mation." Origin of the Earth and Moon, Eds. RM. Canup, K. Righter, The
University of Adzona Press, Tucson, (2000) 227-245.
24. Haack, H., Seott, E.RD., "Asteroid Core Crystallization by Inward Dendritie
Growth." Journal of Geophysical Research, Vol. 97 (1992) 14,727-14,734.
25. Esbensen, K.H., Buchwald, V.F., "Planet (oid) eore erystallization and
fractionation- evidenee from the Agpalilik mass of the Cape York iron me-
teorite shower." Phys. Earth. Planet. Inter., Vol. 29 (1982) 218-232.
26. Sekerka, RF., Jeanfils, C.L., Heckei, RW., ''The Moving Boundary Problem."
Lectures on the Theory of Phase Transformations, Ed. H.1. Aaronson, The
Metallurgical Society, AlME, New York (1982) 117-169.
27. Glicksman, M.E., Diffusion in Solids: Field Theory, Solid-state Principles, and
Applications. John Wiley and Sons, New York, 2000.
28. Harn, F.S., "Shape-Preserving solutions of the time-dependent diffusion equa-
tion." Quart. J. Appl. Math., Vol. 17 (1959) 137-145.
29. Morse, P.M., Feshbach, H., Methods of Theoretical Physics, Vol. 2, MeGraw-
Hill, New York (1953) 1,284-1,292.
30. Eshbach, O.W., Handbook of Engineering Jilundamentals, 2nd edition, John Wi-
ley & Sons, New York (1961) 2-60.
31. MeFadden, G.B., Coriell, S.R., "The Effect of Fluid Flow due to the Crystal-
Melt Density Change on the Growth of a Parabolie Isothermal Dendrite." Jour-
nal of Crystal Growth, Vol. 74 (1986) 507-512.
Deterministic Behaviour in Sidebranching
Development

Ricard Gonzalez-Cinca1 and Yves Couder2

1 Departament de Fisica Aplicada, Universitat Politecnica de Catalunya, Campus

Nord UPC, Ed. B5, J. Girona Salgado s/n, E-08034 Barcelona, Spain
2 Laboratoire de Physique Statistique, Ecole Normale Supeneure, 24 rue
Lhomond, 75231 Paris CEDEX 05, France

Abstract. The development of sidebranching in solidifying dendrites in a regime

of finite diffusion length is studied both experimentally and by means of aphase-
field model. The growth rate of each sidebranch shows a power-Iaw behaviour from
the early stages of its life. From their birth, branches which finally succeed in the
competition process of sidebranching development have a greater growth exponent
than branches which are stopped.

1 Introduction
Dendritic patterns in diffusion-controlled systems have received much atten-
tion in the past decades [1-3]. In crystal growth, the study of sidebranches
that appear at the flanks of a growing dendrite has been developed more
recently. Most of the work devoted to sidebranching has been focussed in the
scenario of selective amplification of fluctuations in the region elose to the tip
in which linear theories are valid [4-7]. However, the shape of the dendrite
further down from the tip and, in particular, the development of sidebranches
there have been scarcely studied [8,9].
In this paper we present an experimental and numerical study on the
growth process of sidebranches at initial stages, which turns out to be of
crucial importance in the selection of the final shape of a two-dimensional
dendrite. We consider solidification systems where growth is controlled by a
finite-range diffusive field. Numerical simulations were carried out by means
of a phase-field model.
In the next Section we present some experimental results on the behaviour
of competing sidebranches. In Sec. 3, the phase-field model employed in this
work is presented. The numerical procedure as weIl as the simulation results
and comparison with experiments are presented in Sec. 4.

2 Experimental Results
We grew ammonium bromide crystals in conditions similar to those described
in Refs. [10,11], but from a solution with different concentration. The system
is driven by mass diffusion, and we limit to the case when the diffusion length

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 Deterministic Behaviour in Sidebranching Development 21

lD = DjV, D and V being the diffusion coefficient and the dendrite velo city
respectively, is small when compared to the region occupied by the dendrite.
In Fig. 1 a portion of the considered dendrite is shownin which fully
developed sidebranching can be observed. We focus in the smaller region
shown in Fig. 2 and we choose three sidebranches (labeled a, b and c in the
figure ) growing from the main dendrite.

Fig. 1. Ammonium bromide dendrite grown in a regime of finite diffusion length.

Fig. 2. Region of the dendrite containing selected sidebranches a, b and c.

We have recorded the shortest distance of the tip of each sidebranch to

the axis of the main dendrite, y, as a nmction of time from the birth of
22 R. GonzaIez-Cinca, Y. Couder

the sidebranch. The plot of the three curves in Fig. 3a shows the saturation
in the growth of branch b caused by the competition process taken place
with a and c. Fig. 3b corresponds to a log-log plot of the same curves. By
eliminating the saturation part and fitting each of them to y rv t n , it is
found n = 1.24,0.82 and 1.01 for branches a, b and c, respectively. Thus, the
growth rate of sidebranches follows a power-Iaw behaviour and the exponent
is smaller in those branches which are screened off in the competition process .

..

r
a , c
/.
I a
.c
.
!OO

I
:
,
-'" I
f
b
,: 1'00
1*I I
.... ..
.. ' .
b . ," !'.l'

.*
!
..
Sill :IC

",,'
' J 1. "

..-
~~'----~----~~----~'OO-----J
,=·

Fig. 3. a) Distance between the tip of each sidebranch and the axis of the main
dendrite vs. time; b) log-log plot.

3 Model
We have performed simulations of dendritic growth by means of aphase-
field model for solidification [12,13]. The corresponding equations for the time
evolution of the phase field ifJ and the dimensionless diffusive field u(r, t) can
be written in the following form [12]:

(.2 r (O) ~~ = ifJ(l- ifJ) (ifJ - ~ + 30wLluifJ(1- ifJ))

_(.2 :x [11(0)11' (0) ~:] + (.2 :y [11(0)11' (0) ~~] + (.2V [11 2 (0)VifJ] (1)

8u
8t + Ll1 (30ifJ
2- 60ifJ
3+ 30ifJ4) 8ifJ 2
8t = V u + 'l/J(x,y, t), (2)
where Ll is the dimensionless undercooling. Lengths are scaled in some arbi-
trary reference length w, while times are scaled by w2 / D. In these equations
 Deterministic Behaviour in Sidebranching Development 23

o is the angle between the x-axis and the gradient of the phase field and
7J(O) = a(O)/a(O) , a(O) being the surface energy. The kinetie coefficient has
been taken as isotropie, ß(O) = ßo, whieh leads to r(O) = m7J(O) with con-
stant m = cpDßo/ Ldo, Cp being the specific heat per unit volume, L the
latent heat per unit volume and do the capillary length. a is equal to ~~.
The parameter E controls the interface thiekness and by E -+ 0 the classieal
sharp-interface model is recovered. The term 'IjJ introduces an external noise
that induces sidebranching [7] .

Fig. 4. Dendrite at different times obtained with a phase-field model.

4 Numerical Procedure and Results

The phase-field model equations have been solved on rectangular grids with
mesh spacing Llx using first-order finite differences. A four-fold surface ten-
sion anisotropy a = a(O)(l + 'Ycos(40)) has been considered. The noise term
is evaluated at each uncorrelated cell of lateral size Llx simply as I . r, where
I denotes the amplitude of the noise, and r is a uniform random number in
the interval [-0.5,0.5]. We have used a set of parameters that gives rise to
a needle without sidebranching when I = O. The employed parameters are
a = 400, m = 20, 'Y = 0.06, E = 0.003, I = 19, Llt = 10-4 , Llx = 0.01 and
24 R. GonzaIez-Cinca, Y. Couder

Ll = 0.42. In Fig 4 a dendrite is shown which is obtained with this set of

parameters at different times. Labels A and B indicate the sidebranches that
we will consider. It can be observed that growth of branch B saturates at
later times.
A more detailed picture of the region containing the considered branches
is shown in Fig. 5. We have computed the position y of the tips of both
branches as a function of time, similarly to what we did with the selected
sidebranches among those grown in the experiment.

Fig. 5. Region containing branches A and B at different times.

In Fig. 6 the points (t,y) for branches A and B as weIl as fitting curves
y ,...., t n for both data sets are represented. The value of the growth exponent is
1.06 and 0.93 for A and B respectively, which indicates that the winner branch
grows from the initial stages of its life faster than the looser branch. This
conc1usion is obtained systematically when comparing different branches.
The results presented in this paper suggest that the noise mechanism plays
a less important role in the determination of the final sidebranching structure
than expected. We have found that the competition process between branches
is fully determined from the beginning and the branch which starts growing
faster succeeds.
We acknowledge A. Hernandez-Machado for fruitful discussions.

References
1. Langer, J .S., in Chance and Matter, Souletie, J., Vannimenus, J., Storra, R.,
Eds. North Holland, (1987) 629
 Deterministie Behaviour in Sidebranching Development 25

1.4

12

0.8

B
0.6

0.4

O~------~-------L------~------~
o o.s t 1.5
time (dmensiorIass)

Fig.6. Position of the tip of branches A and B as a function of time (points) and
the corresponding fitting curves y '" t n (lines).

2. Dynamics 0/ Curved Pronts, Pelce, P., Ed. Academic Press (1988)

3. Growth and Form. Nonlinear Aspects, Ben Amar, M., Pelce, P., Tabeling, P.,
Eds. Plenum Press (1988)
4. Brener, E., Temkin, D.: Noise-induced sidebranching in the three-dimensional
nonaxisymmetrie dendritic growth. Phys. Rev. E 51 (1995) 351-359
5. Karma, A., Rappel, W.-J.: Phase-field model of dendritic sidebranching with
thermal noise. Phys. Rev. E 60 (1999) 3614-3625
6. Pavlik, S.G., Sekerka, R.F.: Fluctuations in the phase-field model of solidifica-
tion. Physiea A 277 (2000) 415-431
7. Gonzalez-Cinca, R., Ramirez-Piscina, L., Casademunt, J., Hernandez-Machado,
A.: Sidebranching induced by external noise in solutal dendritie growth. Phys.
Rev. E 63 (2001) 051601-1-9
8. Bisang, U., Bilgram, J.H.: Shape of the tip and the formation of sidebranches of
xenon dendrites. Phys. Rev. E 54 (1996) 5309-5326
9. Li, Q., Beckermann, C.: Evolution of the sidebranch structure in free dendritie
growth. Acta Mater. 47 (1999) 2345-2356
10. Maurer, J., Bouissou, P., Perrin, B., Tabeling, P.: Europhys. Lett. 6 (1989) 67
11. Couder, Y., Argoul, F., Arneodo, A., Maurer, J., Rabaud, M.: Statistieal prop-
erties of fractal dendrites and anisotropie diffusion-limited aggregates. Phys.
Rev. A 42 (1990) 3499-3503
12. Wheeler, A.A., Murray, B.T., Schaefer, R.J.: Computation of dendrites using
a phase-field model. Physica D 66 (1993) 243-262
13. McFadden, G.B., Wheeler, A.A., Braun, R.J., Coriell, S.R., Sekerka, R.F.:
Phase-field models for anisotropie interfaces. Phys. Rev. E 48 (1993) 2016-2024
Growth Dynamics during Solidification of
U ndercooled Melts

Dieter Berlach1
Institut fuer Raumsimulation, DLR, D-51170 Institut fuer Raumsimulation, DLR,
D-511 Koeln, Germany

Abstract. Electromagnetic levitation technique is applied for containerless under-

cooling of liquid metals, metallic alloys and semiconductors as weIl. The velo city of
the solidification front is directly measured as a function of undercooling. Measure-
ments on pure metals reveal growth velocities ranging up to 100 m/s at the largest
undercooling. The dependence of the growth velo city on undercooling is analysed
within current theories of dendrite growth. The experimental findings are weIl de-
scribed without any free parameters assuming the validity of the coIlision limited
growth model and deviations from local equilibrium at the solid-liquid interface.
Solid solutions behave similar as pure metals with the difference that also devia:-
tions from the chemical equilibrium at the interface have to be taken into account.
As tested on binary alloys of strongly partitioning elements the deviations from
chemical equilibrium are weIl understood by solute trapping described by a veloc-
ity dependent partitioning coefficient. A phenomenon analogous to solute trapping
is disorder trapping during solidification of superlattice structures in intermetallics.
Examples of measurements on intermetallic alloys are presented. Results of exper-
iments on quasicrystalline and polytetrahedral alloys are presented. The sluggish
growth behaviour in such systems with complex crystalline structure is understood
by a growth dynamics, which is not collision but rather diffusion limited. The tran-
sition from faceted to non-faceted growth is discussed on the basis of experimental
investigations of the growth dynamics of semiconducting materials.

1 Introduction

Most of all materials are prepared from the liquid state as their parent phase.
Therefore, solidification is one of the most important phase transformations
in materials science. H a melt is cooled one may distinguish between two dif-
ferent cases. Under the conditions of directional solidification a positive tem-
perature gradient is applied externally, which stabilizes a planar solid/liquid
interface. Crystal growth dynamics is externally controlled by the change
of the temperature gradient. The dynamics of the crystal growth behaves
very sluggish preserving local thermodynamic equilibrium at the solidifica-
tion front. This means the velocity with which the interface is proceeding is
very small compared to the atomic diffusion speed in the liquid. On the other
hand if a melt is undercooled below its equilibrium melting temperature a
negative temperature gradient is arising. Such a negative temperature gradi-
ent is destabilizing a planar interface. Any instantaneous perturbation tends

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 Growth Dynamies during Solidification 27

to grow since a curved interface is more favourable for the heat redistribution
at the solid/liquid interface. In case of alloys a concentration gradient is built
up because of the different solvability of the solute in liquid and solid sol-
vent. This concentration gradient enhances the instability of the solid/liquid
interface caused by the negative temperature gradient in undercooled melts.
The gradients in front of the interface during solidification of undercooled
melts give rise to dendritic crystal growth. The morpholögy of a dendritic
solidification front ensures a high interface area to volume ratio that is more
favourable for heat and mass redistribution compared to a planar interface. As
a consequence the velo city of the dendritic interface is higher compared with
the kinetics of a planar interface. Provided large undercoolings are achieved,
deviations from local equilibrium are occurring, which lead to a kinetic un-
dercooling of the interface and a reduction of the concentration profiles in
aIloys.
In the present work electromagnetic levitation technique is applied to process
melts in diameter of several mm in containerless state. Owing to the complete
avoidance of heterogeneous nucleation on container walls large undercoolings
are accessible ranging up to about 25 % of the melting temperatures of the
respective metals [1]. A freely suspended drop offers the additional benefit
that it is directly observable for rapid solidification by using proper diagnostic
means. In particular, different techniques are applied to measure quantita-
tively the dendritic growth velocity of levitation processed melts of metals
and semiconductors as a function of undercooling. Experimental results will
be presented for pure metals, solid solutions, dilute non-miscible alloys, in-
termetallies and quasicrystals as weIl. In addition, also the growth dynamics
of levitation undercooled semiconductors is investigated. The experimental
findings are discussed within dendrite growth theory and stability analysis of
the solidification front.

2 Experimental

Electromagnetic levitation is a very efficient tool to process bulk melts in

containerless state. A schematic view of an electromagnetic levitation cham-
ber for containerless undercooling and solidification experiments is shown in
Fig. 1. This technique gives the extra benefit that the freely suspended drops
are not only accessible for direct diagnostics but also for external manip-
ulation of the crystallization process. The levitation coil together with the
sampie in diameter of about 6 nun are placed within an ultra-high-vacuum
chamber which can be evacuated to pressures of p ~ 10-8 mbar before back-
filling with gases such as He or He-H2 mixture. The gases are purified by
an oxysorb system and, additionally, by passing them through a liquid ni-
trogen cold trap. The sampie is processed within the levitation coil which is
connected to a high frequency generator. The maximum power output of the
28 D. Herlach

rf-generator is 24 kW and the frequency can be changed in the range between

300 kHz and 1.2 MHz. Temperature control in a limited range is possible by
varying the sam pIe position along the symmetry axis of the levitation coil
and by using forced convection due to cooling gases. The temperature of the
sampie is measured by means of a two-colour pyrometer with an absolute
accuracy of ± 3 K and a sampling rate of 100 Hz. However, such a frequency
is not sufficient to monitor the rapid temperature rise during recalescence.
Therefore, a high speed photosensing device has been developed to record
relative temperature changes during rapid solidification with a frequency of
1 Mhz . Solidification of the undercooled melt can be externally initiated by
touching the lower bottom of the sampie with a crystallization trigger needle.

Owing to the avoidance of heterogeneous nucleation on container walls and ist

far reduction on free surfaces by foreign phases by e.g. oxides thanks to pro-
cessing in an ultra-high-vacuum chamber under high purity environment large
undercoolings are achieved. Thus, the growth dynamics can be studied as a
function of temperature over extended ranges of undercooling. To measure
the growth velocity a rapid photosensing device was developed. It measures
the time L1t = t2 - tl that is needed by the solidification front to propagate
through the observation window of the photosensing device. Because of the
fact that solidification is externally stimulated by touching the lower bottom
of the sample with the sampie holder the geometry of the solidification path-
way is weIl defined. The difference of the solidification path L1d = d2 - d1
is therefore easily determined and the growth velo city V is obtained from
the ratio V = L1d/ L1t [2]. As checked by a double sensor the growth veloc-
ity is independent on the loeation of the solidifieation front. Solidification in

. .t r e _
T-.e..u reM enu
b yoptto _ _
( ' WIl l)
ßl-l:uJ-b
OystBJ,h l8QOn
'---,r---:---.-..J triggc r needlc

Fig. 1. Left: Eleetromagnetic levitation chamber for containerless undercooling and

solidifieation of metals.
Right: Experimental techniques to measure the growth dynamics of the solidifi-
eation front in levitation undercooled sampIes: The Photodiode method (left hand
side) and the eapaeity sensor technique, CPS (right hand side). More details are
given in the text.
 Growth Dynamies during Solidification 29

undercooled melts is by dendritic growth (see chapter 3). The rapid photo-
sensing technique is based upon the assumption that the solidification front
is approximated by the envelope of the tips of all dendrites, which start to
propagate at the external nucleation point. This assumption is fulfilled in
the medium and large undercooling range at which many small dendrites are
nucleated at the same location at sampie surface. But the assumption may
be violated at small undercoolings where only a few fat dendrites are formed
which slowly migrate through the volume of the undercooled melt. In order
to measure the growth velo city even in this undercooling range the capac-
ity senSor technique (CPS) was developed. The nucleation sensing trigger is
made of a nickel needle that is part of an electrical resistance Ü capacity (RC)
circuit. The frequency of the RC circuit depends On the capacity C that is
changed if the trigger needle is touching the metallic liquid drop and initiates
externally crystallizatioo. Thesignal R-rec&d€d as-a- f"uootiGn-of time--oll--a
transient recorder marks the time at which the needle touches the sampie by
a sharp cusp at time ti. At the nucleation point a central dendrite is formed
which propagates from the lower end of the sampie to its upper end. As SOOn
as the dendrite is arriving at the top of the sampie the pyrometer measures a
rapid increase of pyrometer signal S at time t J. The growth velo city of this
central dendrite is then easily determined by the ratio of the time difference
and the sampie diameter D o, V = (tJ - ti)/Do [3].

3 Description of Dendrite Growth in Undercooled

Melts
Solidification of undercooled melts takes place by free dendritig growth. The
release of the heat of crystallization leads to an increase of the temperature
at the solid-liquid interface and the undercooled melt acts as a heat sink
itself. As a consequence a negative temperature gradient arises that destabi-
lizes a planar solidification front. In pure metals the negative temperature is
the physical origin of the dendritic growth morphology. The strongly curved
interface of a dendrite is most favourable for the heat redistribution at the
interface during solidification and causes rapid crystal growth, which is much
faster than any crystallization kinetics of a planar interface during e.g. direc-
tional solidification. Classical theory for free dendrite growth is basing upon
the solution of the heat and mass transport equations for liquid and solid
phase [4]. Accordingly, the total undercooling of a melt, Ll.T, which canbe
measured by experiment is consisting of several contributions:
(1)
The thermal undercooling Ll.Tt is calculated by the solution of the heat trans-
port equation resulting in
V·R
Pt=-- (2)
2al
30 D. Herlach

with ilHf the heat of fusion, C~ the specific heat of liquid phase, Iv(Pt ) the
Ivantsov function in terms ofthe thermal Peclet-number Pt, V the velo city of
the solidification front, R the curvature radius at the tip of a dendrite, and a,
the thermal diffusivity. Due to strong curvature at the tip of a dendrite the
equilibrium melting temperature is reduced because of the Gibbs-Thomson
effect that leads to a local interface undercooling ilTr of the solid-liquid
interface. The curvature undercooling ilTr is given by

ilTr = 2 . r = 2. q / ilS f (3)

R R
with r the Gibbs-Thomson coefficient, q the interfacial energy and ilSf the
entropy of fusion.

At growth velocities comparable or larger than the a~omic diffusive speed Vd,
local deviations at the interface are occurring due to a kinetic undercooling
ilTk • The kinetic undercooling is estimated within rate theory by calculating
the transition rates of atoms going from liquid to solid state and vice versa
[5]. It results

J.L = f· ilHf . Va (4)

kBTl
J.L is the kinetic growth coefficient, TL the liquidus temperature and kB Boltz-
manns constant and f the atomic side factor. The velocity Va is determined
by Va = 8 {P exp( -ilGa/kBTi ). 8 represents an interatomic spacing, {P a typ-
ical atomic frequency, ilG a the activation energy for atomic diffusion and Ti
the interface temperature. In the case of collision limited growth (ilG a = 0)
Va is approximated by the speed of sound in liquid state Va. In pure metals
and solid solutions the condition of collision limited growth is evidenced [6].
However, the situation may change if complex systems with large unit cells
are considered. Here, short-range diffusion processes are required to built up
the solid from the liquid. In such a case Va should approximate the atomic
diffusive speed, Vd, rather than Va. Since the speed of sound, Va, is about
three orders of magnitude larger than Vd, fast growth dynamics is expected
in simple metallic systems while in alloys with complex structure a rather
sluggish growth kinetics should be observed. The atomic side factor f will
depend on the similarity of short-range order in undercooled liquid and solid.
Provided there is a similar short-range order in both phases f is given by
unity, f = 1, i. e. each atomic jump from the liquid to the solid is successful
in the sense that the atom is directly integrated into the solid. But if the
short-range order differs from liquid to solid f should be smaller than unity,
f < 1, since each atom needs several jumps to find the equilibrium lattice
side. In case of alloys a constitutional undercooling, ilTe, has to be taken into
account. The solubility of the alloy solute is often smaller in solid than in liq-
uid. This causes a pile up of solute in front of the interface and establishes a
concentration gradient, which leads to the constitutional undercooling ilTe
 Growth Dynamics during Solidifieation 31

owing to the loeal change of liquidus und solidus temperature in front of

the solid-liquid interface. It also destabilizes the interface in addition to the
negative temperature gradient. LlTe reads

(5)

rnl denotes the slope of the liquidus line, Co the nominal concentration of the
alloy, kE the equilibrium partition coefficient, Iv(Pc) the Ivantsov function
in terms of the chemical Peclet number Pe = (V R) /2D with D the chemical
diffusion coefficient. Similar as in case of the kinetic undercooling, deviations
occur from the local chemical equilibrium if the solidification velo city becomes
comparable or even exceeds the atomic diffusive speed. Solute is trapped
into the solid beyond the chemical equilibrium. This non-equilibrium effect
is taken into account in dendrite growth theory by a velocity dependent par-
tition coefficient and a velo city dependent liquidus slope [7],[8]. The velo city
dependent partition coefficient k(V) is determined within the solute-trapping
model of Aziz [9] as:

k(V) = kE + V/Vd (6)

1 + V/Vd
Equation 1 describes the undercooling LlT as a function of growth velo city V
and curvature radius R in terms ofthe Peclet numbers Pt and Pe, respectively.
In order to find a unique correlation between these quantities an additional
relation has to be found. An independent equation correlating the growth
velocity V with the tip radius R is inferred from morphological stability
analysis. According to the hypothetical criterion of marginal stability [10] an
equation is found for the dendrite tip radius, R, as

R2 = rju* (7)
rn,Ge-Gt
Ge and Gt denote the chemical and thermal gradient in front ofthe interface,
and u* is the stability constant. The marginal stability analysis leads to an
expression for the (1*:
* 2doa, 1
(1 = R2V = 411"2 (8)

with do = (1smC~/LlHJ the capillary length (u s': solid-liquid interfacial en-

ergy). The criterion of marginal stability is a hypothetical one. The solvability
theory leads to a self-consistent solution of the problem by introducing the
anisotropy of the interface as a parameter that dissolves the degeneration of
the solutions of the transport equation [11],[12],[13]. According to solvabil-
ity theory, the stability constant differs from the expression of equ. (8) by
an anisotropy parameter €, which is however elose to unity for systems with
metallic bonding.
32 D. Herlach

4 Experimental Results on Dendrite Growth in

U ndercooled Melts
4.1 Pure Metals
Figure 2 shows the dendrite growth velo city V as a function of undercooling
i1T for pure Ni, in a double logarithmic scale. The closed squares represent
early measurements by using the photo-sensing device while the open squares
give results of measurements particularly in the low undercooling range. The
dots correspond to measurements using the capacity sensor. The solid line
gives the prediction of dendrite growth theory taking into account a kinetic
undercooling of the interface [3]. In the medium undercooling range 100 K
< i1T < 200 K there is a good agreement between experiment and theory. At
large undercoolings i1T > 200 K deviations between theory and experiment
are observed, which lead to a break in the temperature dependence of the
growth velocity-undercooling relation at a critical undercooling i1T* ~ 175
K [14]. At small undercoolings i1T < 100 K the measured growth velo city
is systematically larger than the theoretically predicted values. This finding
becomes obvious by the measurements using the photo-sensing device even
though the data are showing a considerable scatter. But using the capacity
sensor the accuracy of the measurements is essentially improved. It is clearly
discriminated that at velocities V < 1 mjs the experimentally determined
values of the growth velocities are enhanced to the predictions of theory.
It is interesting to note that these deviations are observed at growth veloc-
ities that are comparable with the fluid flow velocities induced externally
in electro-magnetically processed melts [15]. Fluid flow will certainly cause
changes in the heat and mass transport in front of the solid-liquid interface
and therefore will lead to a corresponding change in the growth dynamics.
This will be subject of future research also including experiments on board
of the International Space Station. These experiments make use of reduced
natural convection and electromagneticaHy induced stirring in melts, which
are processed containerlessly in the Electro-Magnetic Levitator of the Materi-
als Science Laboratory in the ESA module (Columbus) [16]. The positioning
forces needed for containerless processing in space are by about three orders
of magnitude smaller than the forces required to levitate a sampie on Earth
against the gravitational forces. The reduced positioning fields, in turn, guar-
antee a very quiescent state of the melt whose solidification kinetics is not
very much disturbed by externally stimulated heat and mass transport in the
liquid.

The experimental data in the medium undercooling range are weH described
within dendrite growth theory provided the condition of local equilibrium is
relaxed by taking into account a kinetic undercooling. Such kinetically caused
deviations from local equilibrium at the solid-liquid interface are not surpris-
ing at growth speeds as high as V ~ 70 mjs at the largest undercooling i1T
= 330 K. Such a growth velo city is by about one order of magnitude larger
 Growth Dynamics during Solidification 33

than the atomic diffusive speed Vd in metallic systems [17]. In calculating

the kinetic undercooling iJ.Tk according to equ. (4) a value for the speed of
sound of liquid Ni, Vs = 2000 m/ s, has been taken. It leads to a satisfactory
agreement with the experiment, i. e. the atomic side factor is equal to unity,
f = 1. This situation differs from that of pure Fe melts.
Figure 2, right hand side, illustrates the dependence of the dendrite growth
velocity of undercooling for pure Fe. The open circ1es represent measured
data applying the photosensor technique, and the c10sed circ1es correspond
to measurements using the capacity sensor. The solid lines give the predic-
tions of dendrite growth theory if the prefactor Va of the kinetic undercooling
is decreased from the Va = Vs = 4000 m/ s, the speed of sound for liquid Fe
to a value of Va = 250 m/ s. Only such a value being by more than one order
of magnitude smaller than Vs is capable to describe the experimental data.
But it is hard to understand that Va should be essentially smaller than the
speed of sound Vs for pure Fe [18]. According to equ. (4) a larger kinetic

Photodjode Ni
• Schleip et al. 1988
• Eelder et aI. 1993

o
~
is " Vo"4IDmIJ
nolv." tmmh

1lÖ~--~Q~~~~~--~--~
uJ Hf SO 100 ISO JOD
oOvo"ZlJIIIIf

~SO 300 350

----_0 Undercooling AT (K) UndercoalJn, t.T [K]

Fig. 2. Left hand side: Dendrite growth velocity V as a function of undercooling

LlT for Ni. Closed squares: measured by the photo-sensing device [2], open squares:
measured by the photo-sensing device with special emphasis placed on the small
undercooling range, solid circ1es: measured by the capacity sensor technique (3],
solid line: prediction of growth theory.
Right hand side: Dendrite growth velocity V as a function of undercooling Ll.T for
pure Fe. Open circles: measured by the photosensing device, closed circles: measured
by the capacity sensor. The solid lines give the predictions of dendrite growth theory
for different values of the prefactor Vo .

undercooling may also be caused by a change of the atomic side factor f. If

the difference between the prediction of dendrite growth theory assuming the
validity of the collision limited growth model for pure Fe and the experimen-
tal data is entirely attributed to f, it corresponds to a change from f = 1 to
f ~ 0.1. This means that each atom in the liquid boundary adjacing to the
interface needs several jumps (in average 10) to find the correct lattice place
34 D. Herlach

in the solid. A possible source of such a situation is a different short-range

order in the liquid and the solid. Recent neutron scattering experiments on
pure Fe melts both in the stable and undercooled melt regime indicate a
coordination number of Z ~ 12 [19], which is close to Z = 12 for fcc struc-
ture but different from Z = 8 for bcc structure of the solid. Undercooling
and nucleation experiments on pure Fe reveal that primary solidification of
bcc phase is energetically favoured compared to fcc because of an interfa-
cial energy O"sz(bcc), which is smaller than the interfacial energy of O"sz(fcc)
[20]. Hence, there must be some rearrangements of atoms to occur, which
attempt to go from the liquid to the solid. This effect becomes even more
pronounced in systems as Ge or Si which behave metallic like in liquid and
show semiconducting behaviour in the solid state (see chapter 4.5).

4.2 Solid Solutions

In alloys constitutional effects have to be taken into account. The constitu-

tional undercooling leads to a reduction of the growth dynamics because it
arises due to redistribution of solute at the interface, which requires energy.
On the other hand the concentration gradient in front of the interface leads
to a narrowing of the dendrite. Since thin dendrites propagate faster than fat
dendrites this effect of stability counteracts the effect of constitutional under-
coolings. The stability effect however, only dominates in the dilute regime of
alloys containing a solute with a large diffusion coefficient as e.g. C in Ni [21].
Usually, the addition of solute to solvent leads to a reduction of the growth
dynamics. Figure 4 shows the growth velo city V as a nmction of undercooling
l1T for C7oNi30, (left) and C050 Pd50 (right), alloys with complete miscibil-
ity of the components in solid state. For both alloys the growth velo city
increases with undercooling according to apower law up to a critical under-
cooling .1T*. At this critical undercooling there is a discontinuous change
of the slope of velocity-undercooling relationship. The critical undercooling
.1T* amounts to 200 K for CU7oNi30 and to 150 K for C050 Pd50 respectively.
At undercoolings .1T > .1T* the velo city linearly rises with undercooling.
The solid lines are calculated within dendrite growth theory [7],[8]. In the
case of CU7oNi30 alloy an equilibrium partition coefficient kE = 0.81 leads to
an underestimation of the growth dynamics, while the assumption of com:"
plete partitionless solidification (k = 1) results in an overestimation of growth
dynamics. Assuming the validity of the solute trapping model by Aziz [9] en-
ables a satisfactory description of the experimental results up to the critical
undercooling. The break in the temperature dependence of V at .1T* is not
weH understood so far. A special result has been achieved for C050 Pd50 . This
alloy possesses a small heat of fusion, thus, a relatively small hypercooling
limit .1Thyp = .1Hf /C~ = 268 K, which only amounts to 65 % of the av-
eraged value of the ideal solution [22]. However, no anomalous behaviour is
observed if the velo city V exceeds the hypercooling limit.
 Growth Dynamics during Solidification 35

. ..•
Co.;d'<\;u
.:e ~~ I
.E.
:>- 30
CI'J1NI",
.
> 15
.. .: r
~
,p I_
I

!e :
• e.

..
~ :.,~
-;;; .. .. • EX'perunent ~ 0 I
0 20 :!O I

~
;>-
j 15 AT""

l 10 ~

I
10

0 5

100 100 50 100 150 :!JJl ~50 30J 15O

Undercoolingt.T (K] U_<OCIIhle AT I)C]

Fig. 3. Dendrite growth velocity V as a function of undercooling LlT for solid

solutiollS; the dots give results of measurements while the solid lines represent the
predictions of dendrite growth theory.
Left side: CU7oNi30 alloy, the various lines correspond to growth velocities under
the assumption of local chemical equilibrium (equilibrium partition coefficient kE
= 0.81), a velocity dependent partition coefficient k(V) (according to the solute
trapping model by Aziz) and for a partition coefficient k = 1 (complete partitionless
solidification) [14).
The right hand side gives equivalent results for CosoPdso alloy [23). Similar to
CU7oNi30 the experimental results are weIl described by theory, but a break in
the temperature dependence is observable at a critical undercooling LlT* = 200 K
(Cu70NiaO) and LlT* = 150 K (CosoPdso), respectively. In case of CosoPdso alloy
undercoolings have been achieved exceeding the hypercooling limit LlTThyp . No
change in the Velocity-undercooling relation is found.

4.3 Dilute Multiphase Alloys

Ni-B and Ni-Zr alloys show eutectic points in their phase diagrams and be-
long to the glass forming alloys at their eutectic compositions. Another char-
acteristic feature is the steep fall of liquidus temperature with B and Zr
composition and the very small equilibrium partition coefficient kE on the
Ni-rich side. Therefore, strong partitioning is expected during solidification of
dilute Ni-B and Ni-Zr alloys, which should lead to a correspondingly growth
dynamics. Figure 4 illustrates the dependence of the growth velo city V on
undercooling for dilute Ni-B (left) [24] and one dilute Ni-Zr (right) alloy [25].
The experimental results are represented by the closed symbols. In case of
Ni-B two dilute alloys were investigated, containing 0.7 at% B (squares) and
1.0 at% B (circles) . For comparison the results for pure Ni [14] are given in
the right hand diagram. The solid lines correspond to predictions of dendrite
growth theory. For both dilute alloy systems the growth dynamics is very
much reduced compared to pure Ni at undercoolings below a critical under-
cooling t1T; . The sluggish growth in this undercooling range is interpreted to
be chemical diffusion limited growth. According to the equilibrium partition
coefficient kE solute is redistributed in front of the interface in the regime
 36 D. Herlach

30 'M'
]. 30
~
g, Ni,.oo-1' :z >
t-
>
..
20
:z- o '0
0 10
~ p(J:J 'iI
'EI
0 ,..,UI >
öl 05
I> 10
~ 10
~co ts
(; .!l
0 ~
0 100 200 SOO
~
a 0
0 lOG 200 300
Unde:rcooling.1 T (1<1 Undt!l"c:oolinc tll (K]

Fig. 4 . Dendrite growth velocity V as a function of undercooling LlT for dilute

Ni-B [24J and Ni-Zr [25J a1loys. Closed symbols: experimental results, solid lines:
predictions of dendrite growth theory. For comparison, the growth velocity of pure
Ni is also given (x=O). In both a1loys, the growth velo city of the a1loys is very
much reduced compared to pure Ni at undercoolings less than a critical under-
cooling LlT;, which depends crucially on the a1loy system and its concentration.
If the undercooling reaches the critical undercooling the growth dynamics is weil
described by theory taking into account a velocity dependent partition coefficient
k(V) according to the solute trapping model by Aziz [9J.

of very small undercoolings. As soon as the undercooling reaches the critical

undercooling the growth velo city steeply rises with increasing undercooling.
In this regime the growth dynamics is weH described by theory provided a
velo city dependent partition coefficient k(V) according to the solute trap-
ping model by Aziz [9] is taking into account. In particular in this range of
solute trapping the description of the growth velocity-undercooling relation-
ship crucially depends on the parameters of the chemical diffusion coefficient
D and the atomic diffusive speed Vd • Therefore, both of these parameters
were independently determined by laser surface re-solidification experiments
in combination with Rutherford Backscattering experiments on NiggZrl al-
loys [25]. As can be seen in the right hand diagram of Fig. 4 the experimental
results obtained from Ni-Zr are in good agreement with the dendrite growth
theory described without any free adjustable parameter. At largest under-
coolings a break in the temperature dependence of the growth velocity is
observed similar as for pure Ni, whose physical origin is subject of on going
research [26]. The critical undercooling depends essentiallyon the concentra-
tion of the solute. It changes from a value of ilT; = 223 K for Ni99.3BO.7 to
ilT; = 256 K for Nig9Bl. This means, only a slight variation of the concen-
tration by 0.3 at% B changes the critical undercooling ilT; by 33 K! Also,
the critical undercooling depends on the solute element. At equal concentra-
tion of 1 at% B and Zr, t:1T; is altered from ilT; = 256 K for NiggB l to
ilT; = 185 K for Ni99Zrl alloy.
 Growth Dynamics during Solidification 37

4.4 Intermetallics with Superlattice Structures

A phenomenon quite analogous to solute trapping is disorder trapping. It
occurs during rapid solidification in melts of intermetallic compounds, which
form superlattice structures under equilibrium solidification conditions. Dis-
order trapping means that the growth velo city is too high to build up the
superlattice of the intermetallic phase. It is theoretically described by a model
of Boettinger and Aziz [27]. Figure 5 shows the results of measurements on
the equiatomic intermetallic compounds FeSi and CoSi [28]. The squares
represent the experimental results. Dendrite growth theory is used to anal-
yse the experiments. First, it is assumed that the interface undercooling
i1Tk = VI J.L is determined from the model of collision-limited growth [6].
This leads to the dashed line in Fig. 5. Apparently, the assumption of the
validity of collision-limited growth overestimates the velocity. Alternatively,
diffusion-limited growth is supposed. Under these conditions, the atomic dif-
fusive speed is the limiting factor, and J.L is small, which results in a large
interface undercooling i1Tk. The experimental data can be fitted under the
assumption of diffusion-limited growth. The best fit is obtained for J.L = 0.021
m/sK, which is much less than J.L = 4.2 m/sK for collision-limited growth.
The resulting conclusion is that at undercoolings below a critical undercooling
i1T* ~ 310 K growth appears to be diffusion limited. This may be under-
stood by the fact that the build-up of a superlattice structure requires atomic
diffusion over at least two interatomic spacings. However, if the undercooling

:g 2.B FeSi Vi'

]
,
5
oE. 2A
:> 2-D :>
.~ 1.11 .~
0 0 3
öl 12 'öl
:> :> 2
!..
CI
o.s
DA
==tt ry
~
0
5
1

D 100 2DD 3DD D 100 3DD

UndercoolingA T \I<) UndercoollußA T [K)

Fig.5. The dendrite growth velocity V as a function of undercooling ilT for the
intermetallic compound FeSi (Ieft) and CoSi (right), respectively. The squares give
the measured velocities. The solid lines represent the theoretical predictions assum-
ing either collision-limited growth or diffusion-limited growth (28].

exceeds a critical value i1T* ~ 310 K for CoSi alloy, the velocity V starts
to rise rapidly. The slope of the V(i1T) curve beyond the critical undercool-
ing is comparable to that of the curve calculated under the assumption of
collision-limited growth. This transition is understood by the onset of disor-
der trapping at i1T* ~ 310 K when the growth velocity becomes comparable
38 D. Herlach

to, or even larger than the atomic diffusive speed. The atoms can no longer
sort themselves out onto the various sublattices of the intermetallic phase.
This leads to a disordered structure. In the present levitation experiments,
the cooling rate is too small to avoid reordering of the as-solidified sampies.
Therefore, the disordered structure, primarily solidified at the largest under-
cooling, cannot be revealed by, e.g., diffraction studies on the as-solidified
material cooled down to room temperature. However, it is evidenced in Ni-Al
based intermetallic compounds by investigations both of the solidification ve-
locity versus undercooling relation [29] and transmission electron microscopy
(TEM) studies as weH [30]. The TEM micrographs clearly show the existence
of antiphase domains [31].

Even in the regime of sluggish growth of intermetallic aHoys dendritic growth

can be collision limited [32]. Based upon a study of the changes of interface
undercooling due to non-equilibrium effects as solute trapping and solute
drag, the interface undercooling .1Ti is expressed as sum of various terms:

(9)
with .1Tk the kinetic undercooling (cf. equ. (4)), .1Tst a contribution by so-
lute trapping, .1Tsd a term, which is originating from solute drag and .1Tdt
the part by disorder trapping. The various contributions of equ. (9) have
been calculated on the basis of the concept of maximum driving force [32].
Accordingly, .1Tsd only prevails at small growth velocities V /Vd < 10- 2 • If
the velo city is approaching the atomic diffusive speed Vd the terms of solute
and disorder trapping, .1Tst and .1Tdt beeome dominant, but they vanish
if V exceeds Vd. At V > Vd, the true kinetie undereooling .1Tk becomes
most prominent. Based upon this analysis the V - .1T relations measured
for Ni-Al intermetallic aHoys are interpreted. Figure 6 shows V as a funetion
of .1T for Ni 73 . 7Ab6.3 and Ni 76 .5Ab3.5 alloy, respectively. According to the
metastable phase diagram of Ni-Al as calculated by Ansara and Dupin [33] a
phase competition of various phases exists depending on concentration and
undercooling: ß the ordered bcc B2 compound (NiAl), "{' the ordered ccp L12
compound (NiaAI) and "{ the disordered ccp solid solution (Ni). NiAI melts
congruently at about 46 at% Al, Ni3 Al incongruently in a peritectic reaction
"{' -+ "{ + L at ab out 24 at% Al.
Due to the significant undereoolings and consequently high measured growth
velocities, the phase selection behaviour is assumed to be mainly nucleation-
controIled. Onee a eritieal nuc1eus has formed, the first stage of solidification,
inv01ving recalescence, is expected to be eompleted without further nuc1eation
events occurring. The c10sed circles represent the experimental1y determined
growth velocities while the lines give the predietions of theory for the growth
dynamics of the various phases. For the composition Ni-26.3 Al (1eft hand
side), the growth velocity versus undercooling relation measured in levita-
tion experiments are understood if primary erystallization of bcc phase is
supposed. It is weIl known from undercooling experiments on Fe-Ni [34) and
 Growth Dynamics during Solidification 39

~r---------~.-----~---.
lSr------------.--------Tö
! Ni73.?AJ263 &&1
I
I

;-
:0 10 """ I
I

!
I
1
I
I
oS

" ..
I
~ 5 ..!
I,) :,..
4
/'.

..... ~~ ,
D~~~~~~ ________ ~

150 200 250 300 3S0 o 300 350

UndorcoollngbT IKJ Undonoaling6T [KJ

Fig.6. Growth velocity V as a function of undercooling LlT for Ni73.7Ab6.3 (left)

and Nh6.5Abs.5 (right) intermetallic alloy. The thicker dashed curve has been calcu-
lated for the ordered bcc-phase, the thinner dashed line for the disordered bcc-phase,
the dotted line for the unstable regime of the disordered bcc-phase, the thinner solid
line for the disordered ccp-phase and the thicker solid line for the ordered ccp-phase
[31].

Ni-V [35] alloys that, due to the lower energy of the solid/liquid interface,
bcc crystal structures are often favoured over ccp with respect to nudeation.
The theoretical transition point from partially ordered to completely disordf
the ccp phase are somewhat larger than for an alloy Ni-25 Al. The velocities
measured for the ordered bcc structure are located just near the transition
point to disorder trapping [31]. Apparently, the studies in Ni-Al reveal that
even in case of ordered superlattice structures in intermetallic compounds
the atomic attachement kinetics at the liquid-solid interface is described by
collision limited growth provided several contributions to the interface under-
cooling LlTi are taken into consideration. But, we assurne that a transition
from collision limited growth (with the speed of sound as the upper veloc-
ity limit) to diffusion limited growth (with the atomic diffusion speed as the
upper limit) should occur if the complexity of the unit cell of crystalline
structure is enhanced.

4.5 Complex Structures of Polytetrahedral Phases

Alloy systems as Al-Fe showing large unit cells with polytetrahedral ordering
are belonging to the dass of so called approximant phases. The size of the
unit cell with polytetrahedral order differs essentially depending on the alloy
system. In case of jL-phase in intermetallic compound Al5Fe2 the unit cell
consists of 15 atoms, while in case of 'x-phase in intermetallic compound
Ah3Fe4 it is increased to a size of 102 atoms. An infinitely large unit cell of
polytetrahedral order is present in quasicrystalline alloys. To built up large
unit cells of polytetrahedral order atomic migration is needed not only over
40 D. Herlach

one interatomic spacing as for e.g. simple cubic structure but atoms have to
jump over more than one interatomic distance to sort them out selves for
the construction of a large unit cell with polytetrahedral order. Therefore,
growth dynamics in undercooled melts of such alloys with a complex unit cell
structure is expected to be diffusion controlled rather than collision limited.
Figure 7 illustrates such diffusion controlled growth on a binary Al5Fe2 alloy
with a unit cell of polytetrahedral order containing 15 atoms. In addition,
also the growth dynamics of an AI72Pd21Mn7 alloy was measured, which
forms a stable icosahedral phase with quasiperiodicity in three dimensions
[36]. The alloys were undercooled and the growth velocity was measured by
the photo-sensing device. In addition, high speed video camera technique was
also applied to record the propagation of the solidification front through the
undercooled melt [37]. The growth velocities of the polytetrahedrally ordered

"'" IAI#.., I
~
,. Vao
111) - 1KTG1G>
um;w>

I~
..
:>

Ion
i '"
" ::0 ;:
.-. . .. .
.....
~ '.

111) 1:11 :xl) °o~~::o~~,"~~~~~m--~I~II)--~'::O~~I,"

U:tlra:ra1J AT [K] u.-q6T[K]

Fig.1. Growth velo city V as a function of undercooling f1T for the polytetra-
hedral ",-phase formed from an AI sFe2 melt (left) and for an icosahedral I phase
formed from an AI 72 Pd21Mn7 melt. The symbols represent experimental results as
obtained by measurements with the photo-sensing device (closed circles) and with
the high speed video camera (closed squares), respectively. The lines give the pre-
dictions of dendrite growth theory assuming collision-limited growth (dashed line)
and diffusion-limited growth (solid line) for the binary A15Fe2 alloy. The dendrite
growth theory allows for calculations of the velocity - undercooling relation only
for pure metals and binary alloys. For the ternary alloy, a quantitative analysis is
therefore lacking [37]

phases are sm aller by one to two orders of magnitude if compared with growth
velocities measured in pure metals and solid solutions (cf. Fig. 2 and Fig. 3).
The growth dynamics is even more sluggish than that of the intermetallic
compounds (cf. Fig. 5 and Fig. 6). In case of Al5Fe2 alloy there is a fair
description of the results by assuming diffusion limited growth, at least in
the undercooling range iJ.T < 150 K. At larger undercoolings a steep rise of
the velocity is observed somehow similar as for CoSi alloy (cf. Fig. 5). The
temperature dependence of the growth velocity at undercoolings iJ.T > 150
K becomes comparable to the dashed curve calculated for collision limited
growth. This behaviour may indicate the onset of disorder trapping during
 Growth Dynamics during Solidification 41

rapid solidification of this alloy. For the ternary icosahedral alloy a monotonie
increase of V is recorded without any change in ist temperature derivative.
The maximum growth velo city is smaller than 0.1 m/s even at the largest
undercooling of ..1T ~ 130 K. The sluggish growth of the polytetrahedral
phases is understood by the chemical ordering in the construction of the
large unit cells with polytetrahedral order. The atoms in the liquid cannot
occupy each crystal site but have to find their correct sites in solid state. This
requires short-range diffusion of the atoms at the solid-liquid interface and
consequently the growth kinetics is controlled by a diffusion limited growth
process. The dynamics of growth is the more sluggish the larger the unit cell,
and consequently slowest for quasicrystalline alloys with icosahedral order.

4.6 Faceted and Non-Faceted Growth of Semiconductors

From a more general perspective, there are two idealized mechanisms for
crystal growth from the melt. (i) Systems with rough solid-liquid interface
are thought to grow by the continuous growth mechanism, in which atom
transfer from liquid can occur at any site on the interface [38]. (ii) Strongly
faceting materials of high entropy of fusion exhibit edgewise growth, in which
the interface is atomically smooth except for the presence of ledges or steps,
and in which atom transfer can occur only at a few special sites on the in-
terface. Considering solidification into an undercooled melt, Cahn et al. [39]
have suggested a model featuring a transition from edgewise growth (EG) to
continuous growth (CG) when the driving force for crystallization is raised.
Figure 8 gives an overview of the various microstructures formed in pure Ge
upon solidification of differently undercooled melts. At a small undercooling
of ..1T = 40 K (a), a faceted structure becomes apparent which is indicative
for edgewise growth; at a medium undercooling of..1T = 316 K (b), a coarse
grained micrastructure is substituted far lamellar twins that are ubiquitous
in Ge crystals grown under normal conditions (dendritic substructure is vis-
ible at moderate undercoolings if introducing minute amount of tin, e.g., in
the resultant grain structure of Ge99.61SnO.39 at ..1T = 277 K, not shown). At
the largest undercooling, ..1T = 426 K (c), an equiaxed, grain refined struc-
ture is observed [40]. By investigations on a whole series of sampies of pure
Ge, a critical undercooling ..1T; was determined that separates two growth
regimes: ..1T < ..1T; ~ 300 K, edgewise growth and a faceted structure;
..1T > ..1T;, dendritic growth resulting in coarse grains. According to Evans
et al. [41], the first microstructural conversion from twins to coarse grains
is explained as follows: the critical radius of two-dimensional nucleation be-
comes comparable with the step size or with the so-called correlation length
at the solid-liquid interface when ..1T exceeds the critical value, suggesting
that the step at the interface willloose its ''identity''. Regarding the second
microstructural transition to refined grains it has been recognized that the
origin of grain refinement in undercooled materials is due to dendrite break
up. The physical mechanism of this microstructural transformation in under-
42 D. Herlaeh

cooled melts has been investigated by theoretical analysis [42] and experi-
ments as weH [43]. The models (LKTjBCT) of dendrite growth [7],[8] have

Fig.8. Morphologieal transformation of pure Ge with inereasing undereooling LlT:

(a) /lT = 40 K, faceted structure, (b) /lT = 316 K, eoarse grains with some
dendritie elements, (e) /lT = 426 K, grain refined strueture [40].

been developed to describe continuous crystal growth. Two points merit con-
sideration before using these models for faceting materials. The first concerns
the transition from edgewise to continuous growth. The LKT jBCT models
have been proven valid for dendritic solidification (see previous chapters). For
the Ge-based systems they should be applicable in the regime lJ.T > lJ.T:
(dendritic growth), provided an offset of the total (bath) undercooling lJ.T
by a critical interface undercooling lJ.Tt for the EG f- CG transition is taken
into account. Evans et al. [41] gave a plot of the interface temperature at the
onset radius for instability as a function of the bulk undercooling lJ.T for Ge,
according to their computation on stability of spherical crystal growth. From
the computed curve, one derives lJ.Tt ~ 153 K at lJ.T: = 300 K for pure
Ge. The numerical value of lJ.Tt ~ 153 K should be independent on alloy
concentration in dilute Ge-Sn aHoys. The total undercooling eq. (1) reads:
(10)
To test this equastion the growth velocity was measured as a function of
undercooling for pure Ge and dilute Ge-Sn alloys [40]. The results of the
measurements are shown in Fig. 9. Two salient features are visible for pure
Ge: first, in the range lJ.T < lJ.T: ~ 300 K, the growth velocity is very small
and rises slowly. In this region, the crystal grows in the edgewise mode. It
results in the faceted structure and twins (cf. Fig. Ba). Second, at lJ.T: =
300 K, a change takes place in the temperature dependence of the growth
behaviour. In the range lJ.T > lJ.T:, the velocity rapidly increases with lJ.T.
Such a behaviour is typical for the continuous growth mode of dendritic so-
lidification as described by Peteves et al. [44]. Adding a small amount of 0.39
 Growth Dynamics during Solidification 43

.......
1'" 5 G~OlJ..xSo"
> 4 --0--X - 2.110 f- OJl9 ••
:
i'
0 3
__ x - 0.39 f- O.o45
____ x-o f-O.ol
a ,

i
t
2

1
~
0
200 Dl 400
Underc:oolin.: t.T 1.KJ

Fig.9. Growth velocity, V, as a function of undercooling, LlT, for pure Ge (trian-

gles), a very dilute Ge99.61Sno.39 alloy (closed circles), and a dilute Ge98Sn2 alloy
(squares). The symbols and curves correspond to measured values and predictions,
based on the modified dendrite growth model, respectively. f denotes the fraction
of interfacial sites at which growth can occur. The critical undercoolings LlT; for
the microstructural transition from faceted to non-faceted growth are LlT; :::::: 300
K (pure Ge), LlT; :::::: 255 K (Ge99.61SnO.39), LlT; :::::: 230 K (Ge9sSn2).

at% Sn to pure Ge leads to a drastic enhancement of the crystal growth ve-

locity by a factor of five at the largest undercooling. Further increasing the
Sn concentration up to 2 at% Sn slightly reduces this effect.
The experimental results were analysed within the frame of dendrite growth
theory. Concerning the Ge-based systems the interface undercooling merits
special attention since it is known that semiconductors grow with large in-
terfacial undercooling. According to equ. (4), the interfacial undercooling is
given by iJ.Tk = V / J.L with J.L the kinetic growth coefficient. In case of colli-
sion limited growth Vo is given by the speed of sound, Vs, and for metals,
the atomic side factor f = 1. Since liquid Ge is metallic, it is realistic to
assurne the model of collision limited growth model. However, solid Ge is
a semiconductor of strongly directional bonds. Because of the differences of
bonding in metallic like liquid Ge and semiconducting solid Ge, the factor
f may be less than unity. The factor f is used as the only free parameter
to describe the growth velocities measured at iJ.T > iJ.T": within dendrite
growth theory. The results of the calculations are represented by the solid
lines in Fig. 9. The numerical values of the factor f as inferred from the fits
are changing from f : : : 0.01 for pure Ge to f : : : 0.045 for Ge99.61SnO.39 and
f : : : 0.09 for Ge9SSn2. This means that the fraction of interface sites at which
atomic attachment can occur increases by the addition of the metallic compo-
nent with the consequence of acceleration of the growth dynamics. This effect
prevails the counteracting effect of the contribution of the constitutional un-
dercooling iJ.Tc , which causes a decrease of the growth velo city. This finding
is supported by the results of the alloy containing 2 at% Sn. Even though
f is further increased in comparision with the very dilute alloy, the growth
44 D. Herlach

velo city is slightly redueed. This indieates that with inereasing coneentration
of Sn, the eonstitutional undercooling beeomes more important.

5 Summaryand Conclusions
Eleetromagnetic levitation technique in eombination with proper diagnostie
means to obserbe the propagation of the solidifieation front into undereooled
melts was applied to measure the growth dynamies of eontainerlessly pro-
eessed sampies as a function of undereooling. Deviations from loeal equilib-
rium at the solid-liquid interface were evideneed owing to the large driving
forees for rapid erystallization at the undereoolings achieved by eontainerless
undercooling. The deviations are diseussed within a kinetieally undereooled
interface on pure metaJs and solid solutions, solute and disorder trapping in
alloys of solid solutions, dilute Ni-BjNi-Zr alloys, intermetallics and diffusion
eontrolled growth in alloys with unit eells showing polytetrahedral order. Fi-
nally, the transition from edgewise to eontinuous growth in Ge and dilute
Ge-Sn alloys was diseussed.

Acknowledgements
This work benefited from serious collaboration with my eoworkers H. Assadi,
M. Barth, K. Eekler, D. Holland-Moritz, D. Li, J. Scbroers, M. Schwarz, T.
Volkmann, B. Wie and R. Willnecker. The author expresses his sincere thanks
to all of them for the always pleasant and fruitful eooperation. The author
is very mueh indebted to the Deutsche Forschungsgemeinschaft (DFG) for
financial support of the present work within several projects of the priority
program SPP 729 "Undereooled Metallic Melts".

References
1. Herlach, D. M.: Annual Review of Materials Scienee 21 (1991) 23.
2. Schleip, E., Willnecker, R., Herlach, D. M. and Görler, G. P.: Mat. Sei. Eng. A
98 (1988) 39.
3. Eckler, K. and Herlach, D. M.: Mat. Sei. Eng. A 178 (1994) 159 - 162.
4. Kurz, W. and Fisher, D. J.: Fundamentals of Solidification, (Trans Tech Pub-
lieations Switzerland-Germany-UK-USA 1989).
5. Christian, J. W.: The Theory of Transformations in Metals and Alloys, Perga-
mon, Oxford, 1975.
6. Turnbull, D.: Acta Metall. 30 (1982) 2135.
7. Trivedi, R., Lipton, J. and Kurz, W.: Acta Metall. 35 (1987) 965.
8. Boettinger, W. J., Coriell, S. R. and Trivedi, R. in Mehrabian, R. and Parrish,
P. A.: (eds.), Rapid Solidification Processing: Principles and Technologies IV,
Claitor's, Baton Rouge, LA, 1988, p. 13.
9. Aziz, M. J.: J. Appl. Phys. 53 (1982) 1158.
 Growth Dynarnics during Solidification 45

10. Langer, J. S., and Müller-Krumbhaar, H.: Acta Metall. 79 (1978) 1689, 1697.
11. Peke, P. and Pomeau, Y.: Stud. Appl. Math. 74 (1986) 245.
12. BenAmar, M. and Pomeau, Y.: Europhys. Lett. 2 (1986) 307.
13. Barbieri, A., Hong, D. C. and Langer, J. S.: Phys. Rev. A, 35 (1986) 1802.
14. Willnecker, R., Herlach, D. M. and Feuerbacher, B.: Phys. Rev. Lett. 62 (1989)
2707.
15. EI-Kaddah, N. and Szekely, J. in: Mat. Res. Soc. Proc. Symp. "Materials Pro-
cessing in the Reduced Gravity Environment of Space", ed.: G.E. Rindone, 9
(1982), p. 191.
16. Piller, J., Seidl, A., Stauber, M. and Dreier, W. in: Solidification 1999, eds.
Hofmeister, W. H., Rogers, J. R., Singh, N. B., March, S. P., Vorhees, P. W.:
Proc. TMS Fall Meeting (Rosemont 1998) p. 3.
17. Smith, P. M. and Aziz, M. J.: Acta Metall. Mater. 42 (1994) 3515.
18. Eckler, K., Gärtner, F., Assadi, H., Norman, A. F., Greer, A. L. and Herlach,
D. M.: Mat. Sci. Eng. A 226-228 (1997) 410.
19. Schenk, T., Holland-Moritz, D., Simonet, V., Bellissent, R. and Herlach, D. M.:
to be published.
20. Barth, M., Eckler, K., Herlach, D. M. and Alexander, H.: Mat. Sei. Eng. A 133
(1991) 790.
21. Eckler, K., Harnerton, R. G., Herlach, D. M. and Greer, A. L.: Mat. Sei. Eng.
A 133 (1991) 730.
22. Wilde, G., Görler, G. P. and Willnecker, R.: Appl. Phys. Lett. 69 (1996) 2995.
23. Volkmann, T., Wilde, G., Willnecker, R. and Herlach, D. M.: J. Appl. Phys.
83 (1998) 3028.
24. Eckler, K., Cochrane, R. F., Herlach, D. M., Feuerbacher, B. and Jurisch, M.:
Phys. Rev. B 45 (1992) 5019.
25. Arnold, C. B., Aziz, M. J., Schwarz, M. and Herlach, D. M.: Phys. Rev B. 59
(1999) 334.
26. Karma, A.: private communication, and this issue.
27. Boettinger, W. J. and Aziz, M. J.: Acta Metall. 37 (1989) 3379.
28. Barth, M., Wei, B. and Herlach, D. M.: Phys. Rev. B 51 (1995) 3422.
29. Goetzinger, R., Barth, M., Herlach, D. M., Hunziker, O. and Kurz, W.: Mat.
Sci. Eng. A 226-228 (1997) 415.
30. Assadi, H., Barth, M., Greer, A. L. and Herlach, D. M.: Materials Seience
Forum 215-216, 3 (1996) 7.
31. Assadi, H., Barth, M., Greer, A. L. and Herlach, D. M.: Acta Mater. 46 (1998)
491.
32. Assadi, H. and Greer, A. L.: J. Crystal Growth 172 (1997) 249.
33. Dupin, N.: Ph.D. thesis, Institut Nastional Polytechnique de Grenoble, 1995.
34. Schleip, E., Herlach, D. M. and Feuerbacher, B.: Europhys. Lett. 11 (1990)
751.
35. Notthoff, C., Feuerbacher, B., Frans, H., Herlach, D. M. and Holland-Moritz,
D.: Phys. Rev. Lett. 86 (2001) 1038.
36. Tsai, A. -P., Inoue, A. and Masumoto, T.: Philos. Mag. Lett. 62 (1990) 95.
37. Schroers, J., Holland-Moritz, D., Herlach, D. M. and Urban, K.: Phys. Rev. B
61 (2000) 14 500.
38. Jackson, K. A.: Mat. Sei. Eng. 65 (1984) 7.
39. Cahn, J. W., Hillig, W. B. and Sears, G. W.: Acta Metall. 35 (1964) 1421.
40. Li, D. and Herlach, D. M.: Phys. Rev. Lett. 77 (1996) 180l.
46 D. Herlach

41. Evans, P. V., Vitta, S., Hamerton, R. G., Greer, A. L. and Turnbull, D.: Acta
Metall. Mater. 38 (1990) 233.
42. Karma, A.: Int. J. Non-Equilibrium Processing 11 (1998) 201.
43. Schwarz, M., Kanna, A., Ed:ler, K. and Herlach, D. M.: Phys. Rev. Lett. 73
(1994) 1380.
44. Peteves, S. D. and Abbaschian, R.: Metall. Trans. 22A (1991) 1259.
 Thermodynamics of Diffuse Interfaces

G.J.Schmitz l

ACCESS e.V. , D-52056 Aachen, Germany

1 Introduction
Classical thermodynamics describe phase-equilibria between bulk phases
rather than kinetics of phase transitions and interface dynamics. Especially,
c1assical thermodynamies do not comprise the definition of a specific length
scale, as all energy formulations in thermodynamics hold on any length scale
if it is only large enough to allow for reasonable statistics.
As no specific length is defined in the free energy formulations, there is
no hope to deduce information about the evolution of structures-which do
have length scales-by minimization of c1assical thermodynamic free energy
functionals. Accordingly, c1assical thermodynamics yields only information
about the fractions of the bulk phases being in equilibrium, but not about
their spatial distribution the microstructure.
Structures may be defined as interfaces/transitions between different, ho-
mogeneous bulk areas Le. between regions not revealing any structure them-
selves. The inc1usion of contributions of interfaces to the total energy of the
thermodynamic system therefor is mandatory if the evolution of structures
shall be accessible at al1.
As interfacial energies are several orders of magnitude smaller than the
energies of the respective bulk phases they often however are not considered.
If interfacial energies are considered, they are inc1uded into the systems free
energy description as mathematically sharp surfaces contributing a related
surface energy. It is important to note, that the thickness of the interface
represents a characteristic length scale in the system, which however is de-
generate in the case of mathematically sharp interfaces.
Interfaces in physical systems often have a finite extension of several
monolayers and therefor the extension of thermodynamics inc1Uding sharp
interfaces towards a more general thermodynamic description involving dif-
fuse interfaces seems meaningfu1.
Scope of the present paper is to derive a free energy functional for diffuse
interfaces starting from a thermodynamic approach. The principle idea of
the proposed approach is most easily demonstrated for a pure substance
undergoing a phase transition e.g. from solid "s" to liquid "1". The total Gibbs
free energy Gtot for such system is commonly defined as the sum of the Gibbs
free energies of the individual phases Gl and GB:

Gtot = <p!otG + <pfotGI =

B
III
v~
<Ps Ci, t)GsdV + III
v~
<Plei, t)GldV (1)

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
48 G.J.Schmitz

where indices l, s denote liquid and solid respectively and 4>/,4>s denote the
volume fractions of these phases as function of space :t and the time t. For a
fixed ratio 4>1/ 4>s several possible arrangements of the two phases are depicted
in Fig. la which, according to equation (1), all have the same total energy
Gto t . However, these arrangements obviously vary in the total interfacial area

- ..
(a)

(b) (c)

Fig.1. Several arrangements of two phases coexisting in a finite volume. Note

that the volume fraction is in all cases the same. The arrangements however differ
significantly with respect to the amount of interface between the two phases which
may be a sharp interface (b) or a diffuse interface (c).

between the two phases involved. This maybe accounted for by adding a
surface energy term 'YA to the bulk free energy of the system, Fig. 1b:

G tot = III
val
4>sG sdV + III
val
4>IG l dV + II
sur
'YdA (2)

A diffuse interface, however, is not characterized by a two dimensional

surface element dA but by a finite volume in which the two phases coexist.
As a consequence, the surface integral in equation (2) turns into a volume
 Thermodynamies of Diffuse Interfaces 49

integral, its integrand ,A then being some nmction of 4Js and 4J" Fig. lc:

(3)

III
'Vol 'Vol

+ f (4J:Ct, t), 4J;ct, t)) dV

'Vol

The 4JB) 4J, have been primed here to distinguish them from from their sharp
interface counterparts. The primes " , " will however be neglected in the
following. Scope of the present work is to identify the contribution of the
interface i.e. the ftmction f (4Js (?t, t), 4J, ct, t)) and the amount ofrenormal-
ization necessary for the fractions of the bulk phases, i.e. to give relation
between 4J and 4J'. Within this paper the identification of this function and
an overall free energy functional shall proceed by:
- general considerations
- theories of crystal growth (Kossel-model, Jackson-model, Temkin-model)
- theories of phase transitions (Ginzburg-Landau theories)
- statistical considerations

2 General Considerations
As the desired ftmction f is a measure of interface resp. interface density, it
has to vanish in either of the bulk phases, which means:

f (4Js, 4J,) == 0 for 4Js = 4J, = 0)

1 (and (4)
and for 4Jz = 1 (and 4Js = 0) (5)
for any point ?t being located in the bulk regions and at any time t.
A simple ansatz fullfiling this requirement for are products:

f '" 4J? * 4J': (6)

The exponents n and m have to be positive but not necessarily integers- as
otherwise f would not be defined in the bulk regions where either 4J1 or 4Js
are O. More general would be a linear combination:

(7)
n,m

(n, m "I 0 for all interface terms, n or m being 0 will correspond to the other
term to represent the bulk phase)
The inclusion of other terms vanishing in the bulk regions (4Js =
1 and 4J, = 0 and vice versa) without being products like 4Js ln(4Js) or
4J,ln(4Jz) or linear combinations of such logarithmic terms like e.g. occurring
for concentrations in ideally mixed thermodynamic systems is also possible.
50 G.J.Schmitz

As within the bulk phases the values of the 4>-fields do not vary and thus
gradients of these fields only do occur at interfaces, the nlllction f may also
contain terms proportional to \1<ps and/or \1<PI. Most general would be a
combination of any of above terms-products, logarithms, gradients.

3 Thermodynamic Approaches to Interfacial Energy

Classical thermodynamics are based on Gibbs formulation of the following
variational problem [1]:

(8)

with A, J.L being Lagrange multipliers accounting for the constraints of energy
and probability conservation. The solution of this variational problem yields
the weIl known relation for the probability of specific state <Pi:

(9)

with S being the thermodynamic sum being necessary to normalize the prob-
abilities.
It is interesting to note instead of the variation 8 also the fuH functional
derivative 8/81> = [\1 (8/8(\11») - 8/81>1 = 0 being used to derive the equa-
tions of motion for the phase field leads to this result as there are no explicit
dependencies on gradients in the-equivalent-formulation:

Replacing the lo~arithmic terms E 1>i In(1)i) by gradient terms/ scalar prod-
ucts Cit\11» In(ä\11» as physically motivated in the following chapters dras-
tically changes that situation. No more exponential terms are expected in
t
the solution for the and under some simplifying assumptions a gradient en-
ergy expression lat I IV' ,W being the gradient energy known from phase--field
models occurs instead:

To motivate this replacement of the logarithmic terms by gradient terms it

is helpful to consider models of crystal growth.

4 Crystal Growth Models

Thermodynamics of interfaces are especially important in the case of growing
crystals, where the transition of atoms from the liquid to the solid state at
the interface largely determines the structure and the properties of the grown
 Thermodynamics of Diffuse Interfaces 51

crystal. More information ab out the function f describing the interfacial en-
ergy density may therefor be expected from models of crystal growth. Two
models describing the crystal growth of aso called Kossel-crystal, see Fig. 2,
will be analyzed in more detail:
- the J ackson model of faceted growth and [7]
- the Temkin model for the growth of a diffuse interface [6]

+(x) ~ +n
18 -~- __
14 -
12 -
10 -

6
-
4-
2-
o L -_ _-,-.,.-,.......,.-....,---?'-_ _ __ . . n ~ x
·2 -1 0 1 2 3

Fig.2. Kossel's model of the growing crystal. Atoms attach to the interface in
layers. The model assumes that the atoms may only adhere on the top of already
solid atoms ('solid on solid'). A smooth transition with a finite interface thickness
is identified when averaging the fraction of solid atoms parallel to the interface.

Both models allow for a region of interface between the pure solid and the
pure liquid, which in case of the Jackson model is restricted to one monolayer
and extends across severallayers in the Temkin model.
The spatial distribution of the two phases in this interface region is a
priori unknown, Fig. 3(top). At first order ideal mixing between the solid
and liquid phases may be assumed generating a spatially homogeneous dis-
tribution of solid and liquid in the entire interface volume, Fig.3(middle). In
classical thermodynamics an ideal mixture is described by logarithmic terms,
which in case of deviations from ideal mixture are supplemented by addi-
tional terms (e.g. Redlich-Kister polynomials in the case of mixed solute).
52 G.J.Schmitz

~~
Fig. 3. Spatial distribution of solid and liquid phases coexisting in a region between
the pure bulk phases. Nothing is apriori known about this distribution (top). Intro-
ducing ideal mixing in this region corresponds to the model of Jacksön introduced to
describe faceted growth by assuming ideal mixing in a monolayer between bulk solid
and bulk liquid (middle). Gradual mixing across severallayers was first introduced
by Temkin in a multilayer extension of Jacksons approach.

For a phase transition in a unary system this situation basically corresponds

to the conditions given by Jackson in his theory of faceted growth, were he
considered the mixt ure between a pure solid and a pure liquid to be restricted
tö one monolayer. For this monolayer, the Gibbs free energy Gint is (adapted
from Woodruff [5]) :
G int = kT [4>ln(4)) + (1- 4»ln(l - 4» + a4>(l - 4»] (12)
where 4> denotes the fraction of solid atoms in the interface layer and (1- 4»
represents the fraction of atoms in the liquid state respectively. a = 118f / kT
is the so called Jackson-factor enabling differentiation between diffuse (a ~ 2)
and faceted (a > 2) growth. Note, that in the case of a sharp interface where
4>s and 4>1 are either identical 0 or 1 (obeying 4>s + 4>1 = 1) all of the terms in
f vanish. They only contribute to the the total energy Gtot in region of the
diffuse interface where 4>, 1-4> == {4> ER: 0 < 4> < I} and thus are related to
the interfacial energy.
It seems reasonable to assume that the transition from solid to liq-
uid across a diffuse interface is not a "two step" process, (100% -t
 Thermodynamics of Diffuse Interfaces 53

50% and 50% -+ 0%) but involves at least several steps, Fig.3(bottom), and
may eventually be described by a monotonous, continuous function. A re-
spective ansatz has been given by Temkin in his multilayer model assuming
a finite number n of transition layers between solid and liquid to describe
the diffuse interface of a Kossel- crystal. According to Temkin the Gibbs free
energy related to this "diffuse" interface reads:

(13)

L
00

+ kT (</>~-1 - </>~) In (</>~-1 - </>~)

n=-oo

where the term in brackets with L11-" (difference in chemical potential) is

related to the bulk phases </>s and </>1 (i.e. : 1- </>s) and represents the driving
force for a transition. In equilibrium (L11-" = 0) we get:

L
00

L
00

G int = 4w </>~ (1 - </>~) +kT (</>~-1 - </>~) In (</>~-1 - </>~)

n=-oo n=-oo

This formulation recovers Jackson's model as a limiting case for a single

interface layer as can be easily verified by setting 0 < </>n < 1 for only a single
value of n.
Besides obviously representing a more general formulation, Temkin's
model reveals following interesting features:
- summation (resp. integration) across alllayers (Le the entire volume)
- description of a coupling to the bulk phases for non-vanishing L11-"
- introduction of a gradient and accordingly of a length scale
Although the gradient is not immediately obvious in above formulation it
may identified by L1</> = </>~-1 - </>~ = d</>/dz * L1z in one dimension. General-
ization to three dimensions then yields d;s\l</>s with d;s representing a vector
characterizing the metric imposed by the crystal structure (its components
a x , ay, a z representing the lattice spacings in x, y, z directions.)
Reformulating Temkin's expansion in a continuous way accordingly yields:

(14)
vol vol

giving the first hints how the desired function f might look like.
Remarkable is the loss of symmetry as no equivalent terms d;l\l</>l do
appear, which may be a peculiarity of the second phase in Temkin's model
being a liquid with no characteristic length scale d;l. In case of two interact-
a
ing solids 81 and 82 with different crystal structures Sl and S2 the sit- a
uation corresponds to stressed interfaces and a more general -because more
54 G.J.Schmitz

symmetric-formulation reads:

G =kT JJJ [(Vqll et81 ) In (VrjJ8 ets1 ) + (VrjJ82 ets2 ) In (VrjJS2et S2 )]

1 (15)
val

This formulation can be rewritten under specific assumptions (see appendix)

as:

val

w is a constant related to bond energies and the difference lets21-letsll in

the latter part of the sum corresponds to a lattice mismatch between the two
solid phases and can be neglected in case of solid/liquid phase transition as
the spacing of fictitious liquid lattice would immediately adapt to the solid
lattice leading to a vanishing difference let'l- letsI. A detailed treatment
of this term wiH however be interesting in solid-solid phase transitions like
e.g. in Solid-Phase-Epitaxy (SPE) or in precipitation phenomena in solids.
Neglecting these stress-related terms by approximating letli'" letsI'" letl
yields a free energy formulation similar to the weH known double-obstacle
formulation of the phase-field free energy density functional:

G = kT JJJ 21etl 2 1VrjJ 12 + JJJ 4wrjJ81rjJS2

81 (17)
val val

To summarize the implications of above interpretation of Temkin's formu-

lation it should be noted that logarithmic terms known from classical ther-
modynamics have evolved into and thus have been related to gradient terms
known /rom Time- dependent Ginzburg-Landau theory. This fact introduces
a length scale into thermodynamics and has strong impact on the variational
minimization of the respective functional.
To finish the interpretation of Temkin's model it is moreover interesting
to have a closer look at the formulation of the driving force term related to
the bulk energies:

with LlfJ. representing the driving force for the transition, which has not been
treated further by Temkin.
 Thermodynamics of Diffuse Interfaces 55

For a single component system undergoing a transition from liquid to

solid L1JL corresponds to L1G = GI - GS. The extraction of (L1G = GI - GS)
however obviously pro duces some artifacts relating the Gibbs energy of one
phase with the fraction of the respective counter phase:
0 0
L rjJs + GI L rjJl - GI L rjJs - L rjJl
00 00

Gbulk = G S GS
1 -00 1 -00

Neglecting these artifacts (Le the terms GI L: rjJs and GS L: rjJl) extending the
sums to infinity yields

L rjJs + GI L rjJl
00 00

G bulk = G S
-00 -00

which is in compliance with equation (1) when written in continuous form in

three dimensions:

vol vol

5 Approaches from Theories of Phase Transitions

Time and space resolved modeling of phase-transitions has gained more and
more importance in modeling of solidification phenomena and evolving mi-
crostructures in the recent years. The fundamental difference between sharp
interface models and diffuse interface theories is the finite interface thickness
yielding an additional degree of freedom for the system. The evolution of this
continuous parameter in time may be used to describe as weH the first for-
mation of interfaces-i.e. nucleation-as also their stability during propagation
in a continuum model. Actual diffuse interface theories of phase-transitions
make only use of a stationary value of the interface thickness. They do not
address fluctuations or evolution of this parameter in the time.
Shape and thickness of the transition layer are largely determined by
the choice of the potential energy formulation in the phase-field free energy
functional. EspeciaHy the interface thickness corresponds to the height of
the energy barrier and various potentials lead to the same sharp interface
limit. As by now the thickness of the interface, its internal dynamics and its
mesoscopic shape in the direction normal to the interface have not in depth
been considered, the choice of the potential actually entering the phase-field
free energy formulation has been of minor interest. Actual phase-field mod-
els accordingly treat the interface thickness as a fixed, stationary parameter
not exhibiting its own dynamics and not having a physical meaning at aH.
Interfaces in physical systems-e.g. in thin film deposition or in solidification
56 G.J.Schmitz

of many metals - however often have a finite thickness of several monolay-

ers, the number of which varies as a function of extemal conditions like e.g.
temperature.
The general structure of the phase-field free energy formulation, from
which the equations of motion are derived by functional differentiation, con-
sist of three contributions:
- a gradient energy being related to gradients of the order parameter
- a potential energy ( in general a double-well-potential with minima either
at +/ - 1 or -in more recent work at 0 or 1. Sometimes a double obstacle
potential is used.)
- a driving force term describing the deviation from equilibrium and in
general being proportional to an undercooling LlT
The common aspect of all formulations [2-4,8] is identified in the driving
force terms, which all involve deviations from an equilibrium state described
by an undercooling LlT. This corresponds to a Taylor expansion around an
equilibrium state:
Gtot = Gequi + (8G/8T)LlT
Thus by now all the free energy functionals are parameterized by this value. A
free energy formulation not referencing to the equilibrium state, but only be-
ing based on the energies of the phases involved, their volume fractions, their
interfacial area (being proportional to gradients of the fraction solid/liquid
respectively) and interfacial energy u would represent a more comprehensive
description:
Gtot = Gtot(Gs,G',U,</>s,</>/, '1</>s, '1</>/)
To approach a respective formulation, the potential and driving force parts
of the free energy functional used by Steinbach et al [4]. may serve as an
instructive example when being compared to Temkin's model. As in Temkin's
model, terms relating the Gibbs energy of one phase to the fraction of the
counter phase appear when expressing LlG as Gi - Gj :
(</>i</>j)2 +Gi(1/3</>: +</>~</>j -</>i</>; -1/3</>~) -Gj(1/3</>: +</>~</>j -</>i</>; -1/3</>~)
which, when considering the pure bulk contributions separately Le neglecting
terms"", </>i</>j, yields:
Gbu/k = 1/3Gj</>~ + 1/3Gi</>: - 1/3Gi</>~ - 1/3Gj</>:
This expression looks similar to Temkin's expression for the bulk energy con-
tributions besides the fact of different powers of </> occurrs in either expres-
sion. This fact - meaning that superposition of the bulk energies might not be
linear in phase-fraction in mixed phase systems- and the fact that different
forms of potentials (double-well, double obstacle) may equivalently be used
to describe the phase transition leads to more detailed considerations about
the origin of the potential in the following chapter.
 Thermodynamics of Diffuse Interfaces 57

6 Statistical Approach
To get a statistical approach to the problem, the fractions 4>s are now con-
sidered as the probabilities of finding an atom in the solid state (i.e. being in
an environment typical for asolid state) when being drawn out of a control-
volume centered around -t at time t. FoHowing very general statement can
be made for any finite size of this control volume at any length scale:

4>s(-t,t) + 4>l(-t,t) = 1 (18)

and accordingly (since In = 1)

4>; + 24>s4>l + 4>? = 1 (19)
4>~ + 34>;4>l + 34>s4>? + 4>r = 1 (20)
4>! + 44>~4>1 + 64>;4>? + 44>s4>r + 4>t = 1
and so forth
Here the extension has been continued up to fourth order as to allow for
comparison with the classical Ginzburg-Landau theory that uses the double-
weH fourth order polynom.
This procedure can be considered as defining interface states (i.e. product
terms 4>s4>l) between the two states 4>s and 4>l and simultaneously renor-
malizing the probabilities for occurrence of the bulk phases in the interface
regions. In contrast to renormalization group procedures, however, where
starting from a large number of degrees of freedom irrelevant degrees of free-
dom are averaged out of the system, here the starting point is a low number
of macroscopic states and necessary additional degrees of freedom are intro-
duced:

4>~ + 24>s4>l + 4>~ = 1 may be reinterpreted as Ps + Pi! + Pl = 1 (21)

with Ps being the probability of an atom being in a solid environment, Pl being

the prob ability of an atom being in a liquid environment and Pi! being the
probability of an atom to belong to the interface. The definition of products
terms 4>s4>l as a prob ability of both bulk phases to be present in the interface
region has already been stated in the chapter on general considerations. It
is interesting to note that such products also appear in both Jackson and
Temkin models as also in classical and Time-dependent Ginzburg- Landau
theories.
The physical meaning of these interface terms may be best explored using
the picture of a three dimensional cubic crystal with a sixiold coordination
(Z = 6) revealing 5 (resp. 7) different types of configurations, Fig. 4, each
occurring with different prob ability. These probabilities can be calculated by
the binomial coefficients of (4) s + 4>1) z :

1= 4>~ + 64>~4>l + 154>!4>? + 204>~4>r + 154>;4>[ + 64>s4>7 + 4>r (22)

58 G.J.Schmitz

Fig. 4. Three dimensional view of a Kossel erystal. The oeeurrenee of each of the
cubes has a different probability and is related to a speeifie energy level .. (see text) .

It is easily seen, that e.g. the term <ps<pr may be identified with a eube facing
the liquid with 5 sides and the solid with 1 side (e.g. eube 1 in Fig. 4) . This
state is sixfold degenerate as no face of the attaehed eube is preferred. The
degree of degeneracy- indieated by the binomial eoefficients will be less in
the case of anisotropie systems. The degree of degeneraey of the other cubes
may be verified accordingly.
Given the probabilities for the possible states from above considerations,
the determination of the systems energy becomes possible by weighting each
state with its specific energy:

<Ps + <PI = 1 <PsG S + <PIG1 = Gtot

(23)
<p~G8 + 2<Ps<PIG 81 + <p~GI = Gtot
(24)
<p:Gs + 3{<Psss <PI S1 )<Ps<PI + <prGI = Gtot
(25)
<p!Gs + 4{<p~ss<p~SI)<Ps<PI + 6<p~<p~wsl + <PtGI = Gtot
(26)

This proeedure does not affeet the pure solid bulk regions where {<Ps)4 =
{<Ps)3 = {<Ps)2 = <Ps = 1 (and thus Gtot = GS) and <P = 0 (or viee versafor the
liquid bulk regions) and thus reeovers the classieal energy formulation given
 Thermodynamies of Diffuse Interfaces 59

in equation (1) for vanishing interfacial area/volume 4Js4J, and/or negligible

interfacial energy contributions GSI , S8, Si, W 81 •
This total energy obviously describes either of the bulk phases, since G =
G8 for 4Js = 1 and accordingly a 4J1 == 0, and G = G' for 4J1 = 1 and 4Js == 0,
which fixes the selection of the energy coefficients related to the pure bulk
phases to GS and G' , respectively. All other coefficients are constants related
to interfacial energy or regions of liquid/solid coexistence respectively.
Equation (19) can be written in the more general form:

Z ZI
1
- t:o
~,/..n,/..Z-n
'l's 'f'/
•
(Z - n)!n!
(27)

where Z corresponds to the coordination-number. For a cubic 2d symmetry

Z would equal 4, for the above example of cubic 3d symmetry Z equals
6. Higher values of Z can be interpreted as correlations with next-nearest
neighbor atoms, which however will significantly differ in bond energy. In
order to obtain the total energy of the system each of the degenerate states
has to be weighted with its specific energy G(n, Z - n) in more general form:

Z ZI
Gtot = ~ ,/..n,/..Z-n
L..J'I's'l'l
•
(Z-)I
n .n.1 * G(n, Z -n ) (28)
n=O

To obtain more information about the coefficients G(n, Z - n) related to the

interface we take again the picture of the sixfold coordinated cube described
above. We may then identify a pure solid state as astate where an atom is
bond to six other atoms and their mutual energy per bond corresponds to
the strength of a solid-solid bond VSS ' A pure liquid state accordingly is char-
acterized by six liquid-liquid bonds each being of strength Vll. These specific
energies are obtained by summing up the bond energies of the individual
cubes counting one liquid/liquid bond-energy Vll for each side of the cube
facing the liquid and one solid/solid bond energy for each side facing the
solid (see Tab. 6):

G(n, Z - n) = (Z - n) * Vll + n * V ss (29)

corresponding to a linear transition from Z * VZI to Z * V ss with increasing n.

In this formulation solid/liquid bonds VIs have been treated as mean values
of the bulk bond energies V ss and Vu defined by Fig. 6:

1
Vis = '2 (v ss + Vll) (30)

More generally, the energy of an l / s bond may be less then this mean value
promoting demixing or exceed the mean value leading to superstructures.
60 G.J.Schmitz

This fact is accounted for by introducing a variable <p:

1
Vl s = 2 (v ss + vu) + <p (31)

~<p = Vl s - ~ (v ss + vu) (32)

with <p characterizing the miscibility gapjmixing potential (<p = 0 ideal mix-
ing, <p > 0 superstructures, <p < 0 pronounced demixing). In the formulation
of (G, Z - n) this leads to additional terms proportional to <p that correspond
to deviations from ideal mixture:

G(n, Z - n, <p) = (Z - n)vll + nvss + (Z -IZ - 2nD <p

n ideal G(n, Z - n)(Z = 6) real G(n, Z - n, cp)(Z = 6)

0 6Vll 6Vll
1 5Vll + VBS 4Vll + 2VIs (= 5vIl + Vs s + 2cp)
2 4Vll + 2vss 2Vll + 4VIs (= 4Vll + 2vss + 4cp)
3 3VlZ + 3vss 6VIs (= 3vII + 32vS8 + 6cp)
4 2vII + 4vss 2vss + 4VIs (= 2VlZ + 4Vs8 + 4cp)
5 VII + 5vss 4vII + 2VIs (= VII + 5vss + 2cp)
6 6vss 6vss

D o D o
(a) (b)

Fig. 5. Definition of the bond energies Vu,Vss and Vis used in the text. For a " solid
"square being bound by four other solid squares (left) all the bonds are virtually cut
allowing to treat the square independant of ist surrounding. The dangling bonds
than are rejoined yielding saturated bonds again. In the case of a "solid" square
being surrounded by three solid squares and one liquid square (right) this procedure
leads to liquid-solid bonds with a specific energy Vis which may differ from the mean
value of V ss and VII. The table left summarizes respective data for three dimensional
case of cubes as depicted in Fig 4.
 Thermodynamies of Diffuse Interfaces 61

The total free energy of a single component system coexisting in two states
(8 and I) accordingly in terms of atomic bonds reads:

Z Z'
Gtot = ~ <p:<pf- n (Z _ ~)!n! [(Z - n)vlI + nV BB + (Z -IZ - 2nD * cp] (33)

Setting <Ps = 1(=<Pl == 0), Le. looking at a volume in the solid single phase
region leads to the identification of the bulk energies:
GB
Gtot = <P: * Z * Vss = Z * Vss = G S
=? VBS =Z (34)

or for <Ps = 0 and <PI = 1:

Gtot = Z * Vll = GI =? Vll = -G'
Z
(35)

and replacing Vll and V BB accordingly in 33,yields:

Gtot = ~ <p:<pf-n (Z :~)!n! [(Z - n)~ + n a; + (Z -IZ - 2nD * cp]

(36)
GB and G' themselves in solidification problems depend on various
other parameters like temperature, solute concentration, pressure, elec-
tric/magnetic field e.g.:

(37)
These dependencies are considered when the partial derivatives:
8Gs 8Ni 8GB 8T
(38)
8Ni 8<PB' 8T 8<Ps'···
are included which have been neglected through out the paper. Note, that in
the sharp interface limit (Ll<ps = 1) these terms read classically:

(39)

A detailed treatment of these derivatives as e.g. occurring in multicomponent

systems will be subject of a forthcoming paper.

7 Conclusions
The different approaches towards a comprehensive free energy formulation
presented in this paper are summarized, combined and compared in Tab. 1.
62 G.J.Schmitz

The proposed form for the free energy density functional is:

(?ot = !!!
1101
7tV<p ln 7tV<p + >..(<p~Gs + <p~G' + 2G's<ps<Pl)

The sca1ar produsts in the logarithmic terms represent a more general formu-
lation of classic logarithmic terms as the Jackson model has heen shown to he
a limiting case of the Temkin model. The gradient energy of phase-field free
energy functionals can be derlved from this form by a Taylor expansion uder
some simplifying assumptions making also this gradient energy a special case
of the more general formulation being proposed. The potential energy part
of the functional has been modified by deducing it from a binomial ansatz
for reasons being detailed in the paper.
In summary, a formulation for the total free energy of a mixed phase
system including interfaces could be derived. The formulation being suggested
is fully compatible with classical thermodynamies in the sense of recovering
its sharp interface limit. .
The energy density functional has been derived purely based on following
considerations:

- coexistence of two phases in a distinct volume including interfacial energy

terms
- identifications of interfacial energy to be a volume related quantity in
diffuse interfaces depending on volume fraction of bulk phases
- assumption of ideal mixing of phase fractions within the diffuse interface
(Jackson) replaced by a gradual mixing of phases (Temkin)
- extension of Temkin's one dimensional discrete model of an interface to
a continuum description of the entire system in three dimensions

Entropie terms of the probabilities of aphase state have been replaced by log-
arithmic terms including gradients of the respective probabilities. This leads
to the interpretation of maximum entropy (e.g. ideal mixing) to be identi-
cal to absence of gradients (especially absence of interfaces) which seems to
be a reasonable way of thought not contradicting classical thermodynam-
ics. Further studies of the proposed formulations are necessary in the future,
but implications for other fields of physics related to phase transitions, like
e.g. spontaneous symmetry breaking in particle physics, superconductivity,
solitons, non-linear optics, critical point phenomena, chemical reactions, self-
organized criticality, non-linear dynamics, fuzzy logic and many others are
anticipated.
 Thermodynamics of Diffuse Interfaces 63

8 Acknowledgments
This work is based on joint discussions with the ACCESS microstructure
simulation group and their work mainly being funded by: BMBF, DFG and
the Corus group.
Thanks are due to Toni lvas for commenting on the manuscript and get-
ting it into a readable form.

References
1. Jaynes, E.T: Probability Theory: The logic of Science, chapter 11, (internet:
http:// omega.albany.edu:8008/JaynesBook.html)
2. Kobayashi, R.: Modeling and numerical simulations of dendritic crystal growth.
Physica D 63 (1993) 410-423
3. Caginalp, G., Fife, P.C.: Phys. Rev. B 33 11 (1986) 7792
4. Steinbach, I., Pezzola, F., Nestler, B., Seesselberg, M., Prieler, R., Schmitz, G.J.,
Rezende, J.L.L.: APhase field concept for multiphase systems. Physica D 94
(1996) 135-147
5. Woodruff, D.: The solid liquid interface. Cambridge University Press, 1973
6. Temkin, D.E.: Crystallization Processes (ed. N.N. Sirota, F.K. Gorskü and
V.M. Varikash). English translation published by consultants Bureau, New York
(1966)
7. Jackson, K.A.: Liquid Metals and solidification ASM. Cleveland Ohio, (1958)
174
8. Karma A., Rappel, W.J.: Numerical Simulation of Three dimensional Dendritic
Growth. Phys. Rev. Lett. vol. 77 No. 19 (1996) 4050-4053
9. Fix, G.: Phase field models for free boundary problems. in Free boundary prob-
lems Vol. II, (ed. A.Fasano, M.Primicerio) Piman Boston (1983)
10. Wang, S.L., Sekerka, R.F., Wheeler, A.A., Murray, B.T., Coriell, S.R., Braun,
R.J., McFadden, G.B.: Thermodynamically-consistent phase-field models for so-
lidification. Physica D 69 (1993) 189-200
 Table 1. The different free energy formulations addressed in this paper. Note the arrows indicating the introduction of a length scale
into entropy
Terms related to Interface terms Bulk terms; Constraints
Model
field gradients (vanish in bulks) driving forces
IVcf>I" Ci/V cf» In(aV cf» cf>dncf>. cf>.cf>j cf>., cf>t, etc
GibbsjShanon Ecf>dncf>i EEicf>. Ecf>.-l=O
Jackson ErI>dncf>i cf>icf>j not found
Temkin I C7tVcf» In(ttVcf» I...J cf>icf>j LlG(cf>. - cf>j)
GinzburgjLandau I Acf>icf>j + B(cf>.cf>j)"
Time-dependent t
Ginzburg-Landau IVcf>1 2 Acf>icf>j + B(cf>.cf>j)2
Phase-field (cf>icf>j)~ + LlG(lj3cf>'t
(double-well) IVcf>1 2 +cf>~cf>j - cf>icf>j- 1/3cf>J)
Phase-field
(double-obstacle) IVcf>1 2 cf>.cf>j Ecf>.-l=O
J
Proposed form (71V cf» In(aV cf» cf>icf>j Gicf>t +Gjf1 (Ecf>SoI-l =0
00
..;
ci
"'"
<0
 Computer Investigation of the Inftuence of the
Internal Structure Topology on the Percolation
Process in Two- and Three-Dimensional
Inhomogeneous Systems

Aliaksei Konash and Sergey Bagnich

Institute of Molecular and Atomic Physics, Minsk 220072, Belarus
Email: [email protected]

Abstract. In our work the percolation process in two- and three-dimensional inho-
mogeneous lattices is studied. The inhomogeneous lattice is simulated by a random
distribution of obstacles differing in size and number. The influence of obstacles on
the parameters (critical concentration, average number of sites in finite clusters,
percolation probability, critical exponents, and fractal and spectral dimensions of a
finite cluster) characterizing the percolation in the system is analyzed. It is demon-
strated that all these parameters essentially depend on the linear size and relative
area of the obstacles.

1 Introduction
Nowadays developing of the computer technique for imitating such complex
processes as the process of percolation in disordered systems and related phe-
nomena is an actual task. Rock fracture and fragmentation [1], conducting in
a random resistance grating [2], strongly inhomogeneous media [3], propaga-
tion of forest fires [4] and doped semiconductors [5] were successfully analyzed
in terms of the percolation theory. And in [6] Kopelman et. al. developed a
cluster formalism for describing the electronic excitation energy transfer in
inhomogeneous systems. In such a presentation to characterize a condition
of investigating a system, the function of the percolation probability P 00 and
the average number IAv of sites in the cluster were chosen. The dependence
of them on the concentration C is determined by next scaling relationships
[7]
I~v oe IG/Gc _11-7 (1)

Poo oe IG/Gc -l1 ß (2)

where Co is the critica1 concentration of the sites and ß and 'Y are the critical
exponents(depend only on space dimension).
An application of this theory for transport of electronic excitation energy
in the mixed molecular crystals [8] and solutions of organie compounds in
low-molecular vitrifying solvents [9] has demonstrated good correspondence

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
66 A. Konash, S. Bagnich

Table 1. The values of the critical exponents

Critical Percolation Isotopically Chemically Solid solu- Ethanol

exponents theory mixed molec- mixed tion of solution
ular crystals molecular benzalde- of benzal-
crystals hyde in dehyde in
ethanol porous glass
2D 3D
ß 0.14 0.41 0.13 0.13 0.41 0.25
I 2.1 1.6 2.1 2.09 1.7 1.95

between experimental results and the theoretical predictions of the percola-

tion theory for the critical exponents (tabl.1).
However, recent experimental studies [10] of similar processes in porous
matrices revealed an influence of the internal inhomogeneous structure of the
system on the topology of energy transport.
In our work, we performed the computer simulation of the percolation
process on a square and cubic lattices with obstacles differing in size and rel-
ative area (volume) in order to elucidate the possible effect of these obstacles
on the critical concentration Ce, the average number lAVof sites in a cluster,
the percolation probability Poo , the fractal dimension of an infinite cluster
and the critical exponents.

2 Computational Technique
We will solve the site percolation problem, because it is the most important
one from the viewpoint of energy migration in heterogeneous systems. Let us
consider a lattice formed by a set of sites and bonds. It is assumed that C
is the part of sites painted black in a random manner. An aggregate of black
sites connected to one another either directly or through chains of connected
black sites is referred to as a cluster. At C=1, black clusters are absent in the
system. At C«1, black clusters consist of a small number of sites: single sites,
pairs, triads, etc. However, as the percolation threshold is approached, par-
ticular clusters merge together and their average size increases. The average
number of sites in finite clusters is defined by the expression

Eim m 2
lAV= 'E. ,
'tmm
(3)
m

where im is the number of clusters containing m sites. The analytical depen-

dence of JAV on the fraction C is unknown. Numerical calculations showed
that, at G-+ Ge - 0, the quantity lAv goes to infinity [see relationship (1)].
According to [11], the percolation probability is defined as the ratio between
 Computer Investigation of the Influence of Internal Structure 67

the number of sites forming an infinite cluster and the total number of sites
in the lattice. In the numerical simulation, the number of sites contained in
the maximum cluster (mmaz) is calculated and the percolation probability
is estimated from the formula

(4)

where I is the linear size of the lattice. Extensive simulations and theoreti-
cal considerations show that, near C-+ Ce + 0, the percolation probability
decreases according to the power law (2).
All the results presented in this work were obtained from simulations of
the percolation process on 200*200 and 100*100*100 lattices for two- and
three- dimension cases respectively. Despite the currently available methods
of computer reconstruction of Vycor porous glasses (see, for example, [12]),
the inhomogeneous system was initially simulated by introducing square ob-
stacles of a specified size into the lattice, as was done by Bujan-Nunez et al.
[13]. For this purpose, a new lattice with a cell size equals to the linear size of
obstacles was constructed on the primary lattice. In the simulation, some of
the cells were randomly chosen as obstacles and all the primary lattice sites
lying within the chosen obstacles were eliminated (removed) from consider-
ation. Fig. 1 displays variants of the model matrices with obstacles differing
in linear size and relative area.

Sc\'. =100/0 S ob. = 20% S ob. = 33%

10 = 1 10 = 5 10 = 20

Fig.1. Variants of matrices with obstacles differing in linear size and relative area.

3 Results and Discussion

First and foremost, we analyzed haw the inhomogeneity of the matrix affects
the critical concentration. In each case, the value of Ce was determined by
68 A. Konash, S. Bagnich

two methods. According to Hoshen et al. [14], the critical concentration can
be determined from the position of the maximum in the dependence of the
reduced average number I'av of sites in clusters on the site concentration C:

(5)

The dependence obtained by averaging over 200 realizations (2D space)

exhibit a rather smeared maximum, even though the concentration in the
course of the experiment was changed with the step LlC= 0.001. For this
reason, the critical concentration for each realization was taken as the con-
centration corresponding to the onset of the percolation between opposite
sides of the lattice. The critical concentration determined from these data
agrees closely with the value obtained using other methods for a square lattice
[4,5,7,11]. The introduction of obstacles into the lattice considerably affects
the critical concentration: its value increases substantially (Fig.2).
The dependence of the average number of sites in a cluster on the reduced
concentration CjCAV of occupied sites is weH described by the power law
predicted by formula (1). The deviation from the power dependence near
the critical concentration is caused by the finite sizes of the lattice. In fact,
as foHows from formula (1), this quantity should increase to infinity at the
critical point, which, in principle, is impossible in systems of finite size. The
critical exponent "( determined from this dependence coincides with the value
obtained by the same method in [14] and is slightly less than the exponents
derived from other techniques [7]. Figure 3 displays the dependencies of the
critical exponent "( (determined in a similar manner) on the linear size of
obstacles at different values of their relative areas.
It can be seen, that the critical exponent for obstacles with 10 =1 coincides
with the exponent for the homogeneous lattice. The critical exponent "( char-
acterizes the cluster growth with an increase in the concentration. The larger
the size and the larger the relative area of obstacles, the higher the concentra-
tion at which sites begin coalesce into clusters and small-sized clusters merge
into large-sized clusters. To state this differently, an increase in the average
cluster size with an increase in the concentration is more pronounced than
that in the system with a homogeneous matrix; in fact this corresponds to
an increase in the critical exponent "(. The observed decrease in the exponent
"( for matrices with obstacles of size lo=20 can be explained by the finite
sizes of the lattices. As can be seen from Fig.1, an increase in the size of the
obstacles brings ab out the transformation of the inhomogeneous matrix into
the matrix with a specially conflned structure.
This is essentially pronounced for lattices with the maximum relative
area of the obstacles used in the computer experiment. Numerical simulation
revealed that, for 2D matrices at lo=20 and Sobs = 38%, an increase in the
lattice size to 400*400 is accompanied by increase in the critical exponent "(
 Computer Investigation of the Influence of Internal Structure 69

0,3

o 10 20 30 S %40 50 60
obs'

Fig.2. Dependence of the critical concentration Ce on the homogeneity fraction

Bobs in the matrix at different linear sizes lo of obstacles.

to 2.94(for 200*200 lattice, ')'=2.50). Note that an increase in the size of the
obstacles to 40 results in a further decrease in the exponent.
Figure 4 depicts the dependence of the percolation prob ability on the
reduced concentration for the lattices with different fractions of the obstacles.
It is clearly seen that the introduction of obstacles into the lattice is attended
by a more rapid increase in the percolation probability with an increase in
the concentration.
The fractal dimension df is a principal characteristic of the infinite cluster
at the critical point. Mandelbrot [15] was the first to introduce the notion
of a fractal. Subsequently, he specified the tentative concept and defined the
fractal as a structure consisting of parts that, in some sense, are similar to
a unity [16]. An infinite cluster at the critical point exhibits a statistical
self-similarity [11]. The fractal geometry of the infinite cluster and its statis-
tical self-similarity are interrelated. This interrelation leads to the following
70 A. Konash, S. Bagnich

2,8

2,6
3D
---- 5
- - 5 ...=3D%
- o - S,.JI!!I.45%.
5_~6D%
-
_15%

2,4

2,2

2,Q

\ .8
- J - ....... .L- ..... - - - ' - -11 _ .. L_ .. _
o 5 \0 \6 20
10

3,0 2D
- ....
- S - 10%
5 ....- 20%
5 ....-30%
2,6 - S ....-36%

2,11
~

2,04-

2.2

2,0
0 10 1!5
10

Fig. 3. Dependence of the critical exponent 'Y on the linear obstacle size 10 in lattices
for different values of their relative areas.

relationship between the mass and the linear size of the cluster:
M =ld, (6)
Sokolov in [17] showed that the fractal dimension in virtually all physical
problems is defined as the exponent in the relationship (4). At present, the
fractal dimension is determined using different methods. One of them is the
embedded square method proposed in [11] . Forrest and Witten [18] proposed
to bring the central point into coincidence with the center of gyration of
the studied object in order to improve the reproducibility of the results. In
our work, we also determined the fractal dimension of the infinite cluster
by using the embedded square method. To accomplish this, among all the
 Computer Investigation of the Influence of Interna! Structure 71

possible realizations, we chose clusters whose centers of gyration were close

to L/2. The results presented below were obtained by averaging over 300
different clusters. For lattices with large -sized obstacles, additional averaging
was performed over 20 configuration of distribution of the obstacles in these
lattices.

5 4
1.0 3D 3
1 S..,-o
0.8
2 S..,=15 2
38..,=30
0.8 4 S..,>=45
:;
E
CL
5 S..,--60
0.4

0.2

0,0
0 z
CIC.

1, 0
5 4 3 2 1
2D
o.a IS... =O
2S...- 10
0.8 38... =10
i
"-
4S...- 30
5S.....38
0.4

0.2

0.0
0 2
CIC.

Fig.4. Dependence of Pmaz on the reduced concentration C/Cc for lattices with
obstacles of the linear size lo=l at different fractions of the obstacles in the matrix.

The dependence of the cluster mass (the number of sites) exhibits a linear
behavior beginning with square sizes of the order of 20*20. The fractal dimen-
72 A. Konash, S. Bagnich

sion determined from the slope of this dependence is equal to 1.8. This value
is slightly less than the exact dimensional df = 91/48, which was calculated
in [4,7] in terms of the scaling theory. This difference can arise for two reasons.
First, the fractal dimension determined by the embedded square method is
underestimated compared to that obtained by other methods. Second, the
underestimated value of df can be dictated by finite size of the lattice. In
any case, our prime concern is with the influence of the inhomogeneous prop-
erties of the lattice on the fractal dimension rather than in its absolute value.
This influence is illustrated by the data shown in Fig.5.

~=S

2.2

IGo 0
_2P
-C

1,,8 0

1,,6

1,,4
0 10 40 50 60

Fig. 5. Dependence of the fractal dimension of a percolation cluster on the fraction

of the obstacles Sobs in the matrix at different linear sizes 10 of the obstacles.

Another dimension associated with random walks on a cluster is a spectral

dimension ds. In [19], it has been shown that at the percolation threshold
the number SN of distinct sites visited during an N-step random walk on an
infinite cluster varies asymptotically as:

(7)

in all Euclidean dimensions (Alexander-Orbach hypothesis) [20].

 Computer Investigation of the Infiuence of Internal Structure 73

In our investigation we obtained the values of spectral dimension closed

to 4/3 for lattices differing in relative area and linear size of the obstacles.
This result is in a good agreement with the Alexander-Orbach hypothesis.

4 Conc1usion
The results obtained in this work confirm the assumption made earlier (on
the basis of the available data on the energy transfer in disordered system,
specifically, in matrices with different structures on the microscopic level)
that the inhomogeneous properties of matrices substantially affect the perco-
lation process. The introduction of the obstacles causes strong influence on all
percolation parameters in both two- and three- dimensional spaces with next
laws. Increase of the relative area of the obstacles leads to increase of the crit-
ical concentration value, increase of the critical exponent 'Y value, increase of
the growth rate of percolation probability with increase of the concentration
and decrease of the fractal dimension of percolation cluster. In turn, increase
of the linear size of the obstacles (with the same the relative area of them)
results in a fall of the critical concentration value, increase of the critical
exponent 'Y value, a fall of the growth rate of percolation probability with
increase of the concentration and decrease of the fractal dimension of perco-
lation cluster. The spectral dimension of percolation cluster does not depend
on the presence of the obstacles in lattice what accords with the conception
of super universality for this parameter.

References
1. Fragmentation Form and Flow in Fractured Media, Eds. R. Engelman and Z.
Jaeger (IPS, Bristol, 1986).
2. Percolation Structures Processes, Ann. Isr. Phys. Soc., Vol.. Eds. G. Deutsher,
R. Zallen and J. Adler, (Hilger, Bristol, 1983), 5.
3. B. I. Shklovskiy and A. L. Efros. Percolation theory and conductivity of highly
inhomogeneous media, Sov. Phys. Usp. 18 (1975) 845-878.
4. D. Stauffer. Introduction to Percolation Theory, (Taylor & Francis, London,
1985) 420.
5. B. I. Shklovskiy, A. L. Efros. Electronic Properties of Doped Semiconductors,
(Springer -Verlag, New York, 1984).
6. R. Kopelman. in Spectroscopy and Exitation Cynamics in Condenced Molecular
Systems, Ed. by V.M. Agranovich and R.M. Hochstrasser (Amsterdam, North-
Holland, 1983).
7. D. Stauffer. Scaling theory of percolation cluster, Phys. Rep., 54 (1979) 1-74.
8. Borczyskowski, von C., Kirski, T., Bunsenges, Ber.; Dispersive energy transport
in disordered molecular crystals, Phys. Chem. 93 (1989) 1373-1377.
9. S. A. Bagnich. Triplet excitation migration for compound moleeules in solid
solutions, Chem. Phys. 185 (1994) 229-236.
10. S. A. Bagnich. Migration of benzaldehyde triplet excitation in porous matrices,
SPIE Proc. 3176 (1997) 212-218.
74 A. Konash, S. Bagnich

11. J. Feder. Fractals (Plenum, New York, 1988) 322.

12. M. E. Kainourgiakis, E. S. Kikkinides, A. K. Stubos, N. K. Kanellopoulos.
Simulation of self-diffusion of point-like and finite-size tracers in stochastically
reconstructed Vycor porous glasses, J. Chem. Phys. 111 (1999) 2735-2743.
13. M. C. Bujan-Nunez, A. Miguel-Fernandez, M. A. Lopez-Quintela. Diffusion-
infiuenced controlled reaction in an inhomogeneous medium: Small concentra-
tion of reagents, J. Chem. Phys. 112 (2000) 8495-8501.
14. J. Hoshen, R. Kopelman, E. M. Monberg. Percolation and cluster distribution.
11. Layers, variable-range interaction, and exciton cluster model, J. Stat. Phys.
19 (1978) 219-242.
15. B. B. Mandelbrot. The Fractal Geometry of Nature, (W.H. Freeman, New-York,
1983) 486.
16. B. B. Mandelbrot. Fractals Encyclopedia of Physics and Technology, 5 (1987)
579.
17. 1. M. Sokolov. Dimension and other geometrical critical exponents in the per-
colation theory, Sov. Phys. Usp. 29 (1986) 924-948.
18. S. R. Forest, T. A. Witten. Long-range correlation in smoke-particle aggregates,
J. Phys. A 12 (1979) LI09 - L117.
19. R. Rammal, G. Toulose. Random walks on fractal structures and percolation
clusters, J. Phys. Lett. 44 (1983) 13-22.
20. S. Alexander, R. Orbach. Density of states on fractals: "tractons", J. Phys.
Lett. (Paris) 43 (1982) 625-631.
 Electron Transport of Nanoperm Alloys

K. Pekala

Department of Physics, Warsaw University of Technology Chodkiewicza 8,

PL-02-524 Warsaw, Poland.

Abstract. Electrical resistivity and thermoelectric studies of nanocrystallizing

Fe - Zr - Cu - B alloys allow to separate the surface and bulk crystallization
processes. Nucleation mechanisms specific for both crystallization channels are cor-
related to different effects of boron and copper atoms. A temperature variation of
crystalline fraction confirms a strong infiuence of gram sizes and gram boundaries on
electrical resistivity of the nanocrystallizing alloy. This first thermoelectric study of
Nanoperm alloys exhibits a weak dependence of TEP on temperature and chemical
composition of amorphous alloys. TEP of the nanocrystalline alloy is determined
mainly by the Cl! - Fe phase.

1 Introduction

The recently studied Nanoperm alloys are distinguished by excellent magnetiC

properties, such as high saturation magnetization, small coercive field, high
magnetic permeability. Numerous experimental approaches, mainly mag-
netic, were applied in order to unfold the microscopic structure which entails
such specific physical properties. The Moessbauer spectroscopy allowed to
separate the fine processes of surface crystallization [1]. Transport investiga-
tions of Nanoperm alloys are very limited [2,3]. The present work is aimed
at studying the unresolved change in temperature variation of resistivity ap-
pearing at 500 K. Electrical resistivity is supplemented for a first time by
thermoelectric power, which is also a sensitive probe of structural transfor-
mations.

2 Experiments

Alloys with compositions FeS9Zr7B4, FeS7Zr7B6, FeSSZr7BlO' FeS7Zr6

Cu 1 B 6 and FeS5Zr7Cu2B6 were produced in argon atmosphere by melt spin-
ning technique. Details are described separately [4]. Electrical resistivity was
measured using the dc four probe configuration with the accuracy of 0.1%.
Temperature variation of thermoelectric power was recorded by a differential
method. The error was smaller than 0.2 mV / K.

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
76 K. Pekala

3 Results and Discussion

3.1 Electrical Resistivity

All our alloys exhibit a broad minimum of electrieal resistivity located around
350 K, which is close to the Curie temperature of the alloys (Fig. 1 and 2).
A similar temperature variation of resistivity in the vicinity of the para- to
ferromagnetic transition is reported for Fe - Zr alloys [2,3] and also for other
alloys, such as e.g. Co - Si - B [2,5] . Above this minimum, the temperature
variation of resistivity exhibits two ranges of approximately linear variation
with a slope change of 500 K - 630 K depending on composition. An abrupt
drop of resistivity is observed above the threshold temperature of 800 K,
when crystallization processes start. The decrease of resistivity above 900 K
reveals further crystallization stages.
The observed temperature variation of electrical resistivity is buHt by a
superposition of the structural and magnetic eomponents. Aeeording to the
Ziman model the structural part above the Debye temperature varies pro-
portionally to the temperature [6]. The magnetic part related to the electron
seattering on disordered spins beeomes constant above the Curie temperature
[7].
Analysis of the resistivity data allows us to distinguish subtle phenomena in
erystallization. It is weIl known that due to additional topologieal and chem-
ical disorder , experieneed by atoms loeated at surfaces, the threshold surface
erystallization temperature is lower as eompared to the bulk erystallization
threshold. Thus a change in the slope of resistivity versus temperature ob-
served at 500 K may be ascribed to the surface erystallization affecting a
small part of the sampie volume (Figs. 1 and 2). The structural rearrange-
ments at the surface were also detected at this temperature by Moessbauer
spectroseopy [8]. An abrupt resistivity change around 800 K refiects the crys-
tallization proeesses oecurring in a bulk. They result in a precipitation of
nanoerystals a - Fe phase. This iron phase was eonfirmed by the X-ray and
Moessbauer data. It is noteworthy that the relative change in resistivity dur-
ing erystallization is greater than in Finemet and aluminium based alloys
[9,10]. Moreover, erystallization oecurs in a relatively narrow temperature
interval.
Alloys with higher boron eontent exhibit higher threshold temperatures both
for the surface and for bulk proeesses. Fig. 1 eonfirms the stabilizing role of
boron atoms which introduee eovalent bonds Fe-B. It was checked that the
loeal surface erystallization temperature is suppressed when boron atoms are
displaced from a surface layer. On the other hand, the eopper atoms affect
only the bulk erystallization but not the surface erystallization. Previous ob-
servations lend a firm support to the idea that eopper atoms form clusters
being nucleation centres [11]. This in turn explains that the raising copper
content diminishes the threshold bulk crystallization temperature.
 Electron Transport of Nanoperm Alloys 77

1.10 - , - - - - - - - - - - - - - - ,

1.08

1.08

1.02

~ 0.98
0.96

0.94

0.112

0.90

0.88 +---.,.--.---.,.--.---.,.-,...--.--,--'
300 400 600 600 700 800 QOO 1000 1100
T(K)

Fig.1. Temperature variation of normalized electrical resistivity of Fe - Zr - B

alloys for various boron contents.

1.08

1.08

1.04

1.02

J 1.00

"' 0.98
0.96

0 .94

0.92

O.9()

0.88 +--,----,r--,--r-r--.-~-,__~
300 400 500 600 700 IlOO goo 1000 1100 1200
T(K)

Fig. 2. Temperature variation ofnormalized electrical resistivity of Fe-Zr-Cu-B

alloys for various copper contents.

3.2 Crystalline Fraction

Kinetics of crystallization may be quantitatively evaluated, when evolution
of a crystalline fraction X is determined as a function of temperature or time.
Transport studies are highly sensitive to structural rearrangements and of-
fer electrical resistivity measurements as the precise probe for crystallization
processes . The so called Maxwell - Garnett model is applied for numerical
calculations ofthe crystalline fraction X, [12]. This model describes electrical
resistivity p(T) considering an alloy as a mixture of the crystalline and amor-
phous phases characterized by different Pe (T), Pa (T) electrical resistivities,
78 K. Pekala

respectively. Then

1 + 3Xß]
p(T) = Pa(T) [ 1- Xß '

where ß = [Pa(T) - pc(T)l![Pa(T) + 2pc(T)]. It was checked that an as-

sumption of the linear temperature dependence of electrical resistivity of the
amorphous phase is weIl substantiated. Electrical resistivities of the amor-
phous and crystalline phases are used as experimental parameters in calcula-
tion. The temperature dependence of electrical resistivity for the crystalline
0: - Fe phase taken from [15] is used only in the relatively narrower temper-
ature interval, in which a crystallization occurs. The calculated evolution of
the crystalline fraction with temperature is plotted in Fig. 3. It shows that
during the first crystallization stage below 830 K only 5 % of the sampIe
volume transforms to the nanocrystalline phase. Keeping in mind that the
surface crystallization starts at 500 K one may conc1ude that this initial 5 %
nanocrystalline volume results from surface crystallization. It is very plausi-
ble that in a temperature interval just below 830 K also the intense nuc1eation
processes occurs. The nuc1eation centers created inside the whole sampie in
this interval facilitate bulk crystallization which proceeds rapidly above 830
K (Fig. 1 and 2).
CRYST

1.0

0.8 I
If
~
w
~ ~
i
0.4

~
:3 §
.J
0.2

0,0

500 600 700 800 900 1000

T(Kl

Fig. 3. Crystalline fraction X as a function of temperature calculated from electrical

resistivity data for FeS9Zr7B4 alloy.

During the second crystallization stage the nanocrystalline grains devel-

oping around the nuc1ei grow up quickly - leading to final 87% fraction of
 Electron Transport of Nanoperm Alloys 79

nanocrystalline phase in the alloy. This is in agreement with the crystalline

fraction determined from magnetization data by Malkinski et al. [13) Electron
scattering on grain boundaries is wen known for enhancing strongly electrical
resistivity when grain sizes are reduced [15,16 ). It is found that calculations
of the crystalline fraction optimally reproduce the measured electrical resis-
tivity when applying the electrical resistivity of nanocrystalline phase is 2.3
times higher than the resistivity of the polycrystalline Cl - Fe phase with mi-
crometer size grains. Such a high ratio of electrical resistivities in the nano-
and microcrystalline phases is not exceptional and was reported e.g. for the
nanocrystalline Cu, Ni and Fe [17,18).

3.3 Thermoelectric Power

Thermoelectric power (TEP) is negative for the amorphous alloys studied, as
shown in Fig. 4 and 5. It varies slowly and proportionally to the temperature.
Per analogiam to electrical resistivity, a change in TEP slope is found at 500-
600 K due to the surface crystallization. TEP is only insignificantly influenced
by chemical composition of amorphous alloys. The abrupt deviations appear
at ab out 800 K, when the bulk crystallization starts. After crystallization,
the TEP shifts towards positive values (Fig. 4 and 5). For a comparison,
TEP of pure polycrystalline iron is plotted in Fig. 4 [14). A similarity in
TEP temperature variations of iron and crystalline alloys proves that TEP
of these alloys is overwhelmingly determined by the atomic structure of the
Cl - Fe phase. A similar behavior of TEP was reported for AI-Y-TM alloys [4).
The values of TEP in nanocrystalline alloys are relatively smaller compared
to polycrystalline alloys of the same composition.The same ratio of TEP
magnitudes in polycrystalline and in nanocrystalline sampies was reported
for the Cu, Ni, Fe sampies [18].

4 Conclusion
Electron transport parameters are strongly correlated to the structural evolu-
tion of solids. The advanced analysis of temperature variations of resistivity of
the alloys allows to distinguish intervals of the surface crystallization starting
from 500 to 630 K, depending on composition, and of the bulk crystallization
above 800 K. The threshold temperatures of both crystallization processes
reveal the stabilizing role of the boron atoms creating covalent Fe - B bonds.
The negligible influence of copper atoms on the surface crystallization, along
with the noticeable suppression of the bulk crystallization temperature, sug-
gests specific nucleation mechanisms for both types of crystallization. The
reduced boron content in a surface layer suppresses the local threshold crys-
tallization temperature [7]. On the other hand, the copper clusters form bulk
nucleation centers [11). Temperature variation of crystalline fraction, as wen
as a comparison to magnetic studies of a crystallization kinetics [13), con-
firm a strong influence of the grain sizes and grain boundaries on electron
80 K. Pekala

14

- ·
~ 12
><: Fe.zt18,.trnOf'Ph
~

·
10 - - Fe.,z,,s•• aytI
~ ' •.,zr,oe . ....."..

·
0: 8 - "AB.' <IysI
...zr,s•• -
0
~ 6
~~,s •• cry1I

Q. o "-cry1I
U ClYstalline <>
Ci: 4
I-
U
w 2
-'
w
0 0
==
0:
W
:I: ·2
I-

-4

-8
300 400 500 600 700 800 900 1000 1100
T( K )

Fig.4. Temperature variation of thermoelectric power of Fe - Zr - B alloys for

various boron contents.

4 .-~--------------------------~

:: 2 o f..,z.,s,·_
~ F..,z.,s.· ""'"
~ 1 • Fo"z..c-.e••....."..
0: '..,vP,8.·cry1I
F."pf:.ol'•• omotph
~ 0 Fe",ZI,co,B.-cry1I
2
o -1
~
[;l ·2
-'
w
o .3
~
~ -4
I-

·5

300 400 500 600 700 800 900 1 000 1100

T ( K)

Fig. 5. Temperature variation of thermoelectric power of Fe - Zr - Cu - B alloys

for various copper contents.

seattering in alloys eomposed of the bee - Fe nanoerystals embedded in the

amorphous matrix. Temperature variations of thermoeleetrie power of the
amorphous alloys depend weakly on temperature and on alloy eomposition.
Thermoeleetric power is less affeeted by the surface and bulk crystallizations,
 Electron Transport of Nanoperm Alloys 81

as compared to the electrical resistivity. TEP of the nanocrystalline alloys is

overwhelmingly determined by the atomic structure of the Q- Fe phase.

Acknowledgement
This work is supported in parts by grants No 7 T08A02121 of KBN and 503
G022 of Faculty of Physics.

References
1. M. Kopcewicz, A. Grabias, J. Appl. Phys. 80 (1996) 3422.
2. J.M. Barandiaran, L. Fernandez Barquin, , J.C. Gomez Sal, P.Gorria, A. Her-
nanando, Sol. St. Commun, 88 (1993) 75.
3. P. Duhaj, I. Matko, P.Svec, J. Sitek, D. Janickovie, Mater.Sei. Eng. B39 (1996)
208.
4. K. Pekala, J. Non-Crystalline Solids 287 (2001) 183.
5. U. Mizutani, Progress in Mater. Sei. 2 (1983) 97.
6. P.J. Cote, L. Meisel, Phys. Rev. Lett. 40 (1978) 1586.
7. G. Bergmann, P. Marquardt, Phys. Rev. 17 (1978) 1355.
8. M. Kopeewiez, A. Grabias, J. Appl. Phys. 80 (1996) 3422.
9. T. Kulik, T. Horubala, H. Matyja, Mater. Sei. Eng. A157 (1992) 107.
10. P. Jalkiewiez, K. Pekala, J. Latueh, Nano-Structured Materials 11 (1999) 733.
11. Y. Zhang, K. Hono, A. Inoue, A. Makino, T. Sakurai, Acta Mater. 44 (1996)
1497.
12. R. Landauer, in Eleetrieal Transport and Optieal Properties of Inhomogeneous
Media, AlP Conf. Proe. No. 40, ed. J.C. Garland, D.B. Tanner (AlP, New York,
1978), 1.
13. L. Malkiwski, A.Slawska -Waniewska, Mater. Sei. Eng. 226 (1997) 716.
14. W. Fulkerson, J.P. Moore, D.L. MeElroy, J. Applied Physics 37 (1966) 9644.
15. A.F. Mayadas, M. Shatzkes, Appl.Phys.Lett 14 (1969) 345.
16. J. Ederth, L.B. Kish, E. Olsson, C.G. Granqvist, J. Applied Physics 88 (2000)
6578.
17. R.K. Islamgaliev, K. Pekala, M. Pekala, R.Z. Valiev, Phys. Status Sol (a)162
(1997) 559.
18. K. Pekala, M. Pekala, Nanostruetured Materials 6 (1995) 819.
Self-Organized Formation of Fractal and
Regular Pores in Semiconductors

Jens Ohristian Olaussenl ,2, Jürgen Oarstensenl , Marc Ohristophersenl ,

Sergiu Langal, and Helmut FöHl

1 Chair for general materials science, CAU University of Kiel, Kaiserstr. 2, 24143
Kiel, Germany http://www.tf .uni-kieLde/matwis/amat/
2 Theoretische Physik und Astrophysik, Universität Kiel, Leibnizstr. 15, 24098
Kiel, Germany http://www . theo-physik. uni-kieL der claussen/

Abstract. Electrochemical etching of semiconductors, beside technical applica-

tions, provides an interesting experimental setup for self-organized structure forma-
tion capable of regular, diameter-modulated, and branching pores. The underlying
dynamical processes governing current transfer and structure formation are de-
scribed by the Current-Burst-Model: all dissolution processes are assumed to occur
inhomogeneously in time and space as a Current Burst (CB); properties and inter-
actions between CB 's are described by a number of material- and chemistry- depen-
dent ingredients, like passivation and aging of surfaces in different crystallographic
orientations, giving a qualitative understanding of resulting pore morphologies.

1 Electrochemical Etching: Basic Experimental Setup

The solid - liquid junction of Silicon and HF - containing electrolytes ex-
hibits a number of peculiar features, e.g. a very low density of surface states,
Le. an extremely weH "passivated" interface [1]. If the junction is biased, the
N - characteristics (Fig. 1) in diluted HF is quite complicated and exhibits
two current peaks and strong current- or voltage oscillations at large cur-
rent densities (for reviews see [2,3]). These oscillations have been decribed
quantitatively by the Ourrent-Burst-Model [4-7].
The PO controHed experimental setup is shown in Fig. 1. Using a four
electrode arrangement, a potentiostatjgalvanostat is contacting the sampIe
and the electrolyte, allowing for a weH defind potential resp. current for the
electrochemical dissolution reactions. Backside contact, front- andjor back-
side illumination and electrolyte pumping can be varied as weH as ceH size
(from under 0.3 cm up to wafers of 6 in), semiconductor material (Si, InP,
GaAs, GaP) including various doping levels and crystallographic orientations.
In addition, the electrolytes (e.g. HF, HOl, H2 S04 ) and their concentrationS
and temperature can be varied. With this setup a rich variety of porous
semiconductors [8,9] can be generated, from nanopores to mesopores and
macropores, and with various regular and branching structures.
Most of the phenomenonology can be weH understood within the frame-
work of the Ourrent-Burst-Model which seems to refiect a number of quite
general properties of semiconductor electrochemistry.

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 Self-Organized Pore Formation in Semiconductors 83

2 The Current-Burst Model

The Current Burst Model [4-6,10] states that the dissolution mainly takes
place on small spots in short events, starting with a direct Si-dissolution, and
possibly followed by an oxidizing reaction (see Fig. 2).
After these two short processes, the oxide hump undergoes dissolution, a
time-consuming process which ensures at the location of the current burst a
dead-time of fixed length during which no new burst can start. However, im-
mediately after dissolution the Si surface has the highest reactivity, resulting
in a maximal probability of annother current burst. Due to H-termination
the sudace becomes passivated, and the probability for bursts decays until it
reaches the properties of a completely passivated Si sudace, comparable to
the situation before the nucleation of the first Current Burst.
Smoothing Effect of Oxide Layers and Global Current Oscillations
Depending on the regime in the IV-curve current bursts and their interaction
play different roles: For extremely high currents, in the oscillation regime, one
has a permanent oxide coverage, and due to the high forcing of the system
Current Bursts are started at alilocations where the oxide layer is sufficiently
thin for a breakthrough. In this regime, a detailed Monte Carlo study [5~1]
including the lateral interaction between CBs (overlap of the oxide humps of
neighboring CBs) has shown that the CB Model quantitatively can explain
the experimental observations of globally oscillating etching current. Due to
aphase synchronization of neighboring CBs oxide domains are formed. The
size of this domains increases with increasing oxide generationjreduced oxide
dissolution. At a percolation point only one oxide domain exists on the sampie
sudace with a synchronized cycle of oxide growth and dissolution, resulting
in a macroscoping oscillation of the external current. The size of the domains
as weIl as the oscillation time can be controlled by the chemical parameters.
Even without global oscillations for all regions of the IV-curve where oxide
is formed one finds domains of synchronized oxide growth which define the
length scale for the roughness of the electrochemically polished sampie sudace
and thus lead to a smoothing of the surface.
Passivation Effects: The Aging Concept While at high current densities
the semiconductor surface is completely covered with oxide, at low current
densities most of the semiconductor sudace will be in direct contact to the
electrolyte. It is weIl known [1] that after chemical dissolution the free sudace
is passivated, i.e. the density of surface states reduces as a function of time
which increases the stability of the sudace against further electrochemical at-
tack. Schematically the pedection of the sudace passivation and the resulting
reduction of the probability for a chemical attack as a function of time are
plotted in Fig. 3. In silicon, the speed and the pedection of passivation of the
(111) crystallographic sudace is larger than for the (100) sudace. This selec-
tive aging of sudaces leads to a self amplifying dissolution of (100) sudaces
(which will become pore tips) and a preferential passivation of (111) sudaces
84 J. C. Claussen et al.

(which will become pore walls). Under optimized chemieal conditions with
an extremely large passivation difference between (111) and (100) surfaces a
self organized growth of octahedral cavities occurs. The octahedra consists
of (111) pore walls. As soon as the complete surface of the octahedra reaches
a critieal value, it is easier to start a new cavity at a (100) tip of the old
cavity, since the current density in the new, small cavity is larger and no
surface passivation will occur until the surface again becomes to large. This
growing mechanism leads to an oscillation of both the current through each
pore and the diameter of each pore as a nmction of time. As in the case of
oxide dissolution an internal time constant is related to the pore growth.
H a global current density smaller than the average current density in a
current burst is applied, the aging concept becomes essential to explain the
generation of macropores: At passivated surfaces, the nucIeation probabil-
ity of new current bursts is lower than at sites where a current burst has
taken place before. Therefore the situation shown in Fig. 2(c) is much more
unlikely than the situation in Fig. 2(d), so that the whole surface separates
in two phases with current-carrying pores and passivated surface without
contributions to the total current.

3 Consequences on Pore Geometries

As described above, depending on the electrochemieal reactions the interac-

tion of Current Bursts, either surface smoothing due to the dissolution of
an isotropic oxide layer or strongly anisotropie dissolution of the semieon-
ductor (due to the crystallography dependent aging of surfaces which is one
of the most important reasons for pore formation) will dominate. Thus one
obtains completely different surface morphologies when changing the elec-
trochemical etching conditions; e.g. in a Si-HF-organie electrolyte system
only by increasing the HF concentration (Le. faster oxide dissolution and
thus reduced influence of oxide) the electrochemieal etching changes from
electropolishing (strong oxide smoothing) over macropore formation (pores
with diameters of several ILm and smooth pore walls) to mesopore formation
(strongly anisotropie, narrow pores with diameters of less than 400nm) [11].
Silicon has one of the most stable oxides of all semieonductors. So gener-
ally electrochemieal etching of III/V compounds leads to rougher surfaces and
smaller pore diameters. In addition the surface aging of III-V compounds is
more complicated, since there exist two different (111) surfaces; e.g. in GaAs
only the {I11}A planes (Ga-rich planes) appear as stopping planes. So in most
III-V compounds not octahedra (eight (111) surfaces as stopping planes) as
in silicon are etched but tetrahedra with only four (Il1)A surfaces as stop-
ping planes, and the {l1I}B planes (As-rieh) serve as preferential growing
directions like the (100) directions in Si (Fig. 4).
 Self-Organized Pore Formation in Semiconductors 85

25

~20
"'8
~ 15
.!! 10
.... 5

o L-~ __ ~~ __ ~~

0.0 2.5 5.0

Uvs. SeE [Volt]

Fig. 1. Left: The IV- characteristics of the silicon-hydroßuoric acid contact shows
different phenomena from generation of a porous silicon layer (PSL), oxidation and
electropolishing (OX) and electrochemical oscillations at higher anodic bias. Right:
The experimental set-up used for electrochemical anodization of semiconductors.

Fig. 2. Applying a global current density smaller than the average current density in
a current burst will either require very long time constants to keep the mean charge
density passed in a burst small and to cover the surface completely with current
lines (a), a statistical arrangement of current lines with the optimized (smalier) time
constant of the system, leaving parts of the surface without current and creating
nanopores (b, c), or induces a phase separation by rearranging the current lines in
areas corresponding to macropores (d).

IU} {lOG} Probabllity per Tune

.r storting Curreot
Fig. 3. Time constant and perfection of H-termination differ strongly for (100) and
(111) Si surface orientation with higher nucleation probability on (100) surface.
86 J. C. Claussen et al.

Fig. 4. Tetrahedron-like pores oriented along < 111 > directions obtained in (100)-
oriented GaAs at high current densities in HCI electrolytes (n = 1017 cm- S ).

100 IJm <100>

]<810>

Fig. 5. Left: Without crystallographic preferentiation, surface fractals similar to

dendritic structures occur. Right: Under certain etching conditions lateral growth
along lines in crystallographic direction occurs which also seems to generate some
self-similar structure.

Fig.6. Branching of pores (left, for DMF, 4wt-% HF, constant current) can be
suppressed by an open-Ioop control method triggering to the system-inherent time
scale (right, period 0.5 min).
 Self-Organized Pore Formation in Semiconductors 87

Lateral Growth of Current Bursts: Generation of Surface Fractals

For extremely low current densities and low HF concentration, the total num-
ber of active current bursts is so low that most of the surface is passivated
or covered with oxide. CB nucleation becomes a severe problem. The neigh-
borhood of a just terminated CB is therefore a preferential site for new CB
nucleation, leading to the formation of surface fractals as shown in Fig. 5
(left). A further decrease of the current density allows for a partial passiva-
tion of the surface area of a terminated CB. According to Fig. 3 nucleation
in the 100 direction of the surfaee near CBs will be preferential (Fig. 5).
Sidebranching Pores and Suppression by Open-Loop-Control For
technical applications, it would be desirable to generate pores with a diameter
ofless than 500nm. However, at all etching eonditions investigated up to now,
in this diameter range always pores always show intensive side-branching. By
modulation of the intensity of illumination or by modulation of the etching
current with a modulation frequeney meet the intrinsic time seale of the pore
formation, the side pore formation is hindered by the reduction of the etching
eurrent at the right phase. (Fig. 6).

4 Conclusions
The Current Burst model, a loeal stochastie and highly nonlinear model,
and the Aging coneept describing the time- and orientation-dependenee of
the passivation behavior, provide on an intermediate level of abstraction a
general approach to explain pore geometries, oscillations and synchroniza-
tion with a minimum of material-dependent, but experimentally aceessible,
parameters. The lateral interaetion of Current Bursts gives rise to synchro-
nization phenomena, and a pereolation transition to global ordering. Under
special eonditions lateral growth of fractal struetures is obtained. An open-
loop control can be suecessfully applied to suppress sidebranching of pores.
To clarify quantitatively the dynamical proeesses in this system, further ex-
perimental and theoretical efforts have to be made.

References
1. Yablonovich, E., Allara, D. L., Chang, C. C., Gmitter,T., Bright, T. B.: Phys.
Rev. Lett., 57, 249 (1986)
2. Föll, H., Appl. Phys. A 53, 8 (1991)
3. Smith, R. L., Collins, S. D., J. Appl. Phys., 71, R1 (1992)
4. Carstensen, J., Christophersen, M., Föll, H.: Mat. Sci. Eng. B, 69, 23 (2000).
5. Carstensen, J., Prange, R., Popkirov, G., Föll, H.: Appl. Phys. A 67, 459 (1998)
6. Carstensen, J., Prange, R., Föll, H.: in: Proc ECS San Diego 1998, 98 148
7. Carstensen, J., Prange, R., Föll, H.: J. Electrochem. Soc. 146(3), (1999) 1134
8. Propst, E. K., Kohl, P. A.: J. Electrochem. Soc., 141, (1994) 1006
9. Ponomarev, E. A., Levy-Clement, C.: J. Electrochem. Soc. Lett., 1, (1998) 1002
10. Föll, H., Carstensen, J., Christophersen, M., Hasse, G.: in: Proc. ECS 2001
11. Christophersen, M., Carstensen, J., Föll, H.: Phys. Stat. Sol. (a) 182, 45 (2000)
Evolution and Shapes of Dunes

Hans J. Herrmann and Gerd Sauermann

Institute for Computer Applications 1, University of Stuttgart Pfaffenwaldring 27,

D-70569 Stuttgart, Germany

Abstract. We give a review on our field measurements on Barchan dunes. Then we

present a model for dune formation that takes saturation transients in the saltation
layer into account. We show that these transients introduce a characteristic length
scale that is crucial for dune formation and the macroscopic shape of dunes, albeit
determined by intrinsic properties of the saltation kinetics. Finally, we demonstrate
that our model can explain the scale dependence of the shape of barchan dunes
revealed in recent field

1 Introduction
We have all seen the majestic wave-like shapes of dunes in the desert.
Sand dunes develop wherever loose sand is driven by a fluid (air, water ... )
that lifts grains from the ground and entrains them into a surface flow. The
diverse conditions of wind and sand supply in different regions on Earth give
rise to a large variety of different shapes of aeolian dunes [1-3]. Moreover,
dunes have been found on the sea-bottom and even on Mars. Despite the long
history of the subject, the underlying physical mechanisms of dune formation
are still not very weIl understood. How are aerodynamics (hydrodynamics)
and the particular properties of granular matter acting together to create
dunes? How is the shape of a dune maintained when it grows and moves?
What determines the size of dunes? Due to the fact that neither now nor
in the near future we will be able to simulate dunes on the grain scale (an
average barchan comprises 10 15 grains), we concentrate in the following on
an effective continuum model that can be applied to sand dunes or other
geomorphological problems on a large scale. Due to the highly complicated
physical processes involved (saltation, turbulent wind) and the wide range
of length and time scales to be covered - from the dynamics of single sand
grains, the formation of ripples, to the genensis and migration of dune fields,
the time and length scales span over more than seven orders of magnitude
- the derivation of the model will not be given here. We will instead after
reviewing our experimental measurements present the basic elements of the
model and finally compare the two.
The simplest and best known type of dune is the barchan dune shown in
Fig. 1, shaped in acrescent, which occurs if the wind comes steadily from
the same direction throughout the year and if there is not enough sand to
cover the entire surface. Barchan dunes move proportionally to the wind ve-
locity and inversely proportionally to their height. They are encountered for

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 Evolution and Shapes of Dunes 89

instance in Peru [4-7], in Namibia [8] and Morocco [9] . On these dune fields,
hundreds of barchans can be found, generally all of the same size. The dunes
have heights between 1.5 and 10 m, while their bases are typically 40 to 150 m
long and 30 to 100 m wide. The windward or stoss-side of the dune has typieal
slopes between 8° and 20° and is limited by a sharp edge, called the brink.
The brink coincides in many cases with the crest of the dune and separates
the slip face from the dune's windward side. Roughly speaking, its section is
a parabola-like curve reaching from the tip of one horn to the point of max-
imum slip face height and back to the tip of the other horn. Despite the fact
that for more than 50 years geologists and geographers have been measuring
dunes in the field and have obtained data on height, width, length, volume
and dune velo city, very little is yet known [10,11] ab out the exact quantita-
tive shape of barchans. From a mathematical point of view, the barchan dune
is a symmetrieal object in the wind direction, but in nature there are many
factors, like non-steady winds or inclined ground surfaces leading to asym-
metrical shapes. Numerical simulations to prediet the evolution of barchan
dunes and their exact shape have been performed by [12,13].

Fig. 1. Barchan dunes near LaAyoune, Morocco. The dune in the front, on the left
side was measured in detail during our field trip in May 1999, see [14].

2 Experimental Measurements of a Barchan Dune

2.1 Dune Morphology

Qualitatively the crescent-like shape of the barchan is weH known. The first
measurements concerning barchan dunes and their morphologie relationships
were performed by Coursin (1964) [15] in Mauritania and Finkel (1959) [4]
in the Pampa de La Joya in southern Peru. Hastenrath (1967,1987) [5,6]
analysed barchans in the same area and revisited the site 20 years later.
Additional investigations in the same field were undertaken by Lettau and
Lettau (1969) . [7] Slattery (1990) [8] measured barchan dunes in Namibia.
90 H. J. Herrmann, G. Sauermann

However, they did not measure the entire shape of the dune, but only the typ-
ieallengths. Our own field measurements [14] were performed in a dune field
in the Sahara desert, loeated in southern Moroeeo (former Spanish Sahara)
near the city of Laäyoune.
We define for each horn, the lengths La, Lb and the widths W a, Wb inde-
pendently, as did Finkel (1959) [4]. The orientation of the measuring axis is
chosen aceording to the wind direetion, which eoincides with the symmetry
line for a totally symmetrie dune. Furthermore, we introduee the length of
the slip face L s and the length L o from the dune's toe on the windward side
to the brink. Finally, the height of the slip face H is defined at the highest
point of the brink, which is the interseetion of the brink and the longitudinal
eenterline of the dune.
The relationship between the width of the horns W = W a + Wb and the
height H of the slip face has been studied many times. An overview can
be found by Hesp and Hastings (1998). [10] A linear relationship was found
between the height H and the width of the horns W.

W=awH+bw (1)

For the management of dune movement the volume V of a dune is one of

the most interesting features, apart from the rate of movement Vd. Together
they determine the bulk fiux Qb of sand transported by the dunes.

(2)

To obtain the total flux Qt, in addition to the bulk fiux the inter-dune-flux
Qi, has to be taken into aceount.

(3)

A study of these fiuxes has been performed by Sarnthein and Walger (1974)
[16]. In the following we eoneentrate on the bulk fiux and the dune volume.
Aecording to Oulehri (1992) [9], the rate of movement of the dunes in the
area of Laäyoune is 32 m yr- l for a dune of 9 m height. The bulk density
of the dune sand is 1670 kg m- l , on average. Using these data and the
ealeulated volume of 23 000 m 3 for dune 7, we obtain a fiux of 1.2 million
t m- l yr- l or 736 000 m4 yr-l. Lettau and Lettau (1969) [7] estimated a
bulk fiux of 20 550 m4 yr-l for an average barchan of the Pampa de La Joya
whose height was 3 m. A large barchan with a height of 5.2 m gave a bulk
fiux of 60 000 m4 yr- l .
The overall length L of a barchan is the sum of the length Lo from the
windward foot of the dune to its erest, the length ofthe slip face L s , and the
average of the horn lengths Lh = (La + Lb)/2

L-L b
- 0+ L s+ La+L
2 (4)
 Evolution and Shapes of Dunes 91

This definition uses four lengths, of which only the length of the slip face L s
has an obvious dependence on the height H,
H
Ls=-,e
tan
(5)

where e is the angle of repose with typical values between 31 0 and 350 • This
angle is an intrinsic property of the sand and is therefore independent on the
aeolian processes.
Finkel [4] reported for the horn length.

(6)
Further the ratio of the horn length Lh to the stoss-side length L o also de-
pends on the dune's height. This shows clearly that the relative position of
the slip face within the whole dune varies from small to large dunes and that
the ratio of horn length to totallength increases with the height as shown in
Fig. 3. This disproves scaling invariance of barchan dunes. The size scaling
of Barchans can be used to put them all on top of each other by rescaling
the axis for each dune and using dimensionless variables. This is done for
the longitudinal cross section in Fig. 2. We see that on the windward part
the shape is a parabolloid and the brink position bo moves to the right for
larger dunes. In fact bo is a linear function of the height. Therefore we have a
deviation from perfect size scaling as shown in Fig. 3. Looking from the top,
the brink line has also the shape of a parabola.

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

o
-0.5 o 0.5 1.5
normalized length x

Fig. 2. Profile along the symmetry plane of the dunes (thin lines) using normalized
variables and the standard parabola (thick line) from [14].
92 H. J. Herrmann, G. Sauermann

large small

Fig. 3. Sketch of a small and a large dune.

2.2 Wind Velo city and Sand Flux

Correlated measurements of the wind velo city v and the sand flux q have been
performed [17] on a large barchan dune near the beach of Jericoacoara (see
Fig. 4) during the first week of December 2000, at the end of the dry season.
The dune has approximately 34 m in height, a width of 600 m, and the length
of its wind ward side is 200 m. When a fully turbulent atmospheric boundary

Fig.4. Aerial photography of the coastal dune field near Jericoacoara.

 Evolution and Shapes of Dunes 93

layer develops over a Hat surface it gives rise to a logarithmic velocity profile
v(z) [18],
u* z
v(z ) = -ln-,
zo
/'i,
(7)

where U* denotes the shear velo city, zo the roughness length of the surface,
and /'i, = 0.4 the von Karman constant. The shear velocity u* has dimensions
of velocity but is defined in terms of the shear stress T = PairU; and density
Pair of the air. According to Hunt et al. [19] the height 1 of the shear stress
layer can be obtained implicitely through
2/'i,2L
1=---, (8)
In 1/ zo
where L is the characteristic length of the dune which is measured from the
half height of the windward side to the crest, according to the definition of
[19]. From (8), we obtain for L = 100 m and a roughness length zo = 10-4 m,
a height of this layer of 1 ~ 3.2m (1 ~ 4m for zo = 1O-3 m). Hence, we
placed the anemometers at a height of 1 m that is weH inside the shear stress
layer. One should note that it is very difficult to measure the wind velo city
with standard anemometers within this layer for small dunes. This is the
main reason for choosing a large dune. A reference anemometer has been
placed approximately 300 m upwind of the dune's foot and has been kept
there during all the measurements. With a second anemometer, we measured
the average velo city Vi every 24 m on the central profile during 10 minutes.
Finally, we normalized the average velo city (Vi) by the average velocity in the
same period obtained from the reference anemometer, (Vr,i). By doing so, we
could get rid of the long term (> 10minutes) variations in the wind speed
time series throughout the day and obtain the shear velo city U*,i through

1+ U*,i _ (Vi)
(9)
A • _

U*,t - - ,
U*o (Vr,i)

where U*,i is the dimensionless shear velo city perturbation of the air caused
by the dune and U*o is the undisturbed shear velo city far upwind of the dune.
U*o can be calculated assuming a typical logarithmic profile from turbulent
How (7). During our measurements, the wind was blowing quite constantly
and had an average value of 7.5 m S-l at the reference station. Assuming a
roughness length zo = 2.5 X 10-4 m, we obtain from (7) the undisturbed
shear velo city over the plane U*o = 0.36 m S-l. The averaged and normalized
measured shear velocities are plotted in Fig. 5.
We also measured the sand flux on the central slice of the dune using
cylindrical traps with a diameter of 5 cm and an opening of 1 cm at the front.
The back-side of the traps have an opening of 2 cm covered by a fabric with
pores smaller than the grain diameter. The traps were placed at the same
positions where the wind speed has been measured. From the mass m of the
94 H. J. Herrmann, G. Sauermann

collected sand, the collection time T, and the width w of the opening, we
calculated the sand flux q = mj(Tw). The measured sand fluxes are shown
in Fig. 6.

1.6 .---.--- --.--- - - r - - r - - --,-----.-----,.- ---,

·300 ·250 ·200 ·150 ·1 00 ·50 o 50 100

X (m)

Fig. 5. The circ1es show the measured shear velocity u. and wind speed v(z = 1m)
normalized by their reference values u.o and vo(z = 1m), respectively. The solid
line depicts the prediction of (11) using the measured height proille h(x) shown in
the bottom curve (crosses). The depression at the dune's foot is about 0.8 and the
maximum speed-up at the brink is approximately 1.4 from [17].

3 The Model

Since we are interested in the formation and movement of dunes, the

important time scale of our problem is defined by the erosion and deposition
processes that changes the surface. A significant change of the surface hap-
pens within some hours or even days. In contrast to this, the time scale of the
wind and the saltation process is of the order of seconds and therefore sev-
eral orders of magnitude faster. Hence, we will use in the following stationary
solutions for the wind field and the sand flux. Similarly we neglect the finite
life time of avalanches (a few seconds) and consider them as instantaneous
compared to the movement of the dunes. The separation of the different time
scales and the resulting approximations lead to an enormous simplification,
because it decouples the different physical processes. The entire model can be
thought of as four (almost ) independent parts: the stationary wind field over
a complex terrain, the stationary aeolian sand transport, the time evolution
of the surface due to erosion, and avalanches.
 Evolution and Shapes of Dunes 95

x (m)

Fig. 6. Measured and calculated sand :fI.ux onto the central slice of a barchan dune.
The circles denote measurements and the solid line the prediction of the non-
equilibrium sand :fI.ux model, (13) (solid line). The two dashed lines correspond to
the classical empirical relation for saturated :fI.ux: ref. [19] and ref. [27] is depicted.
Yet, these saturated sand:fl.ux relations cannot correctly predict the sand :fI.ux near
the dune's foot and show clearly that the assumption of saturation breaks down
from ref. [17].

3.1 The Wind Shear Stress

2.5,----.,--,---,---...,---,--,----,---,---,-,- - ,
/. !
i. -..-..-i-·-·-·r··-··-..t-·7i~:~-··-\·ti: . ---..1·-.: .-.. -·!-.:. --_··t··_: . _··
i

2 •••• ..
~ :: I :.':
~ 1.5 ____~L _ .. _.._l _____ .. L _____ ~~ ____l_ .. ____ il ____.. _L_.. ___.L _____.l.. ___ ..
~ ~!! ./1 ! U ! ! !
~ -'-l _CC:~"~:~=-"+'----}.)'/"+-"---~- '-' ~: --_·:;~+- . _·_·l------t·-.._.
~ ! "" ,;...'/! i i ~ i"" ! i
·----·'i-----··-t-··-··--t--··---! -------i-----t-- --··i·-··-··-~r~::--t----··
0.5
i i . i j i . i .. ~
~ i .. ----L-.----l.-------L------L- : :..... ::_~
-3 -2.5 -2 ·1.5 -1 -0.5 0 0.5 1.5 2
x/L

Fig.1. The envelope h(x) of the windward profile of a dune h(x) (solid line) and
the separating streamline s(x) (dashed line) form together a smooth object which
is used to calculate the air shear stress T( x) (dash-dotted line) on the windward
side. In the region of recirculation the air shear stress T is set to zero.

The fully turbulent atmospheric boundary layer develops over a Hat

surface the logarithmic velo city profile v(z) of 7 [18]. Aperturbation of the
96 H. J. Herrmann, G. Sauermann

ground hex) sueh as a dune or hill gives rise to a non-Ioeal perturbation fex)
of the undisturbed air shear stress TO,

T(X) = TO [1 + fex)]. (10)

The functional dependenee of the air shear stress perturbation is erucial for
the understanding of the stability of dunes and to predict the sand flux onto
the windward side of a dune. Analytical ealeulations of the flow over a gen-
tle hill yield an analytieal expression for the shear stress perturbation fex)
[19-21]. We performed further simplifieations in order to obtain a minimal ex-
pression that eaptures the erucial features (and only those) and is applieable
for sand dunes [22]. The resulting formula for the air shear stress perturbation
fis,

(11)

where h' denotes the spatial derivative of the dune's profile hex) in wind di-
reetion. The eoefficients A(LI zo) and B(LI zo) depend only logarithmieally
on the ratio between the eharaeteristic length L of the dune and the roughness
length Zo of the surface. For a dune with a length and width ratio W I L ~ 1
and Llzo = 4.0105 we obtain A ~ 3.2 and B ~ 0.3 from Ref. [14,22]. Equa-
tion (11) has several features that are important for dune formation. First,
the air shear stress is eompletely seale invariant and leads to the same speed-
up for small and large dunes. This is expeeted in the fully turbulent regime
where no characteristic length exists. Secondly, the shear stress perturba-
tion fex), Equation (11), seales with the height Hand inversely with the
eharaeteristic length L of the dune and thus with the average slope of the
dune's windward side, f cx: HIL. Thirdly, adepression of T(X) in front of
the hill oeeurs as a eonsequenee of the strongly non-Ioeal eontribution in
Equation (11). Finally, the shear stress perturbation fex) for the windward
side of the dune is ealculated using (11), the profile hex) on the windward
side, and the separating streamline sex) on the lee side. The result is shown
in Fig. 5 together with the measured mean values (averages over 10 minutes
intervals) normalized aeeording to 9. The agreement between model results
and measurements is good. From this, we ean eonclude that the heuristic
model of the separation bubble eombined with the analytic expression, (11),
provide a reasonable approximation for the wind field above the dune. This
strategy enormously reduces the eomputational effort, eompared to the nu-
merical solution of turbulence models and the averaged three dimensional
Navier-Stokes equation.
Equation (11) is based on aperturbation theory and ean only be applied
to smooth hills. Jackson and Hunt [20] assumed HIL < 0.05, whereas Car-
ruthers et al. [23] showed that mean slopes up to H I L ~ 0.3 give reasonable
results. The windward side of a barehan dune is always below the latter value
and the formula should be applicable. However, flow separation occurs at the
 Evolution and Shapes of Dunes 97

brink, which is out of the scope of the linear perturbation theory. A heuris-
tic solution to solve this problem has been suggested by Zeman and Jensen
[24]. They introduced aseparation bubble that comprises the recirculating
flow (the large eddy in the wake of the dune), which reaches from the brink
(the point of detachment) to the bottom (to the point of reattachment) see
fig 3.1. We model the separating streamline by a third order polynomial that
is a smooth continuation of the profile h(x) at the brink Xbrink and at the
reattachment point Xbrink + L r , Le. h(Xbrink) = 8(0), h'(Xbrink) = 8'(0),
8(L r ) = 0, 8'(L r ) = 0, where L r ~ 6H is the downwind distance of the
reattachment point from the brink. The shear stress perturbation f(x) for
the windward side of the dune is finally calculated using equation (10), the
profile h(x) on the windward side, and the separating streamline 8(X) on the
lee side. An example is depicted in fig. 3.1.

3.2 The Sand Flux

Sand transport has been studied already by Bagnold [25] and it was also
him who proposed the first phenomenologicallaw that predicted the sand
transport from the shear stress of the air. Improved laws have been proposed
by several authors in the meantime [26,7,27]. However, all these relations
assume that the sand flux q is in equilibrium and can be written as a function
of the shear stress r, q (r(x)). Temporal or spatial transients are completely
neglected. In the following we will call such a relation saturated, because it
predicts the amount of sand that can be maintained in the saltation layer at
a certain air shear stress r.
This condition is hardly fulfilled at the windward foot of an isolated dune
[21], e.g. a barchan, where the bed changes rapidly from bedrock or vegeta-
tion to sand. Besides the particular conditions at the dune's foot, the sand
flux may never reach saturation [28] on the entire windward side, where the
shear velocity increases gradually from the foot to the crest. Wind tunnel
measurements indicate that the typical time to reach saturation in saltation
is approximately two seconds [29], which corresponds to a saturation length
of the order of 10 m. This length is of the order of the dune size and can
not be neglected if the sand flux on the entire windward side is significant.
Furthermore, it has been observed that the time to reach saturation increases
for shear velocities dose to the threshold [29]. In this situation, the sand flux
may never reach saturation on the entire windward side and should increase
exponentially with distance from the dune's foot [28]. In recent years, sev-
eral models to calculate the wind field have been developed, from analytic
boundary layer approximations to numerical solutions of the Navier-Stokes
equation with an enormous computational effort. Although some previous
studies have discussed the limits of the saturation approximation in detail
[21], much less effort has been dedicated to the development of sand flux
relations that effectively incorporate non-saturation effects [30].
98 H. J. Herrmann, G. Sauermann

For the saturated flux many different functional forms of these sand trans-
port laws exist and have been used in the past. For high shear stresses, how-
ever, they all converge to the simple relation proposed by [1],

(12)

All other more elaborate relations add higher order corrections to the Bag-
nold formula that become important elose to the air shear stress threshold.
To overcome the limitation of saturation and to obtain information about the
dynamics of the saltation process, numerical simulations on the grain scale
have been performed in the last years [31-33]. Still, concerning the modeling
of dune formation, both approaches had to be discarded. The microscopic
models are computationally too expensive and the equilibrium assumption
that is inherent in the simple flux relations does not hold on the entire wind-
ward side of a dune [11,12,14,21,34]. Since both known approaches cannot be
used to model dune formation we developed a new phenomenological contin-
uum saltation model that is computationally very efficient on the one side and
on the other side incorporates the dynamics of the saltation layer and thus
allows for saturation transients [34]. In this model the sand flux is defined by
a differential equation of the form

!...-q = .!.q (1 - 1.) , (13)

8x l8 qs
where qs(r) is the saturated sand flux and l8(r) the characteristic length
of the saturation transients, called saturation length. The saturation length
ls(r) depends on the air shear stress, but converges towards a constant value
for r » rt [34]. A comparison between the saturated sand flux, our model,
Equation (13), and the field measurements from Jericoacuara can be seen in
Figure 6. One observes that Equation (13) gives the right behaviour at the
foot of the Barchan while the neglect of the transient gives an unphysical dip.

3.3 The Surface Evolution

A spatial change in sand flux implies that erosion or deposition takes

place and the surface changes in height. The time evolution of the surface
can be calculated from the conservation of mass,

8h 1 8q
-=----, (14)
8t Psand 8x
where Psand is the bulk density of dune sand. Finally, we note that Equa-
tion (14) is the only remaining time dependent equation and thus defines the
time scale of the model.
The full dune model can be sketched as follows. An initial surface h is
used to start the time evolution. If flow separation has to be modeled the
 Evolution and Shapes of Dunes 99

S 6
.S
~

x in m
Fig. 8. The solutions for large volumes - above a critical heigth - are dunes
inciuding a slip face, whereas for small volumes heaps develop. An important fact
is that the steepest lee side of a heap (dashed lines) is approximately 15°, which is
weil below the angle of repose of 340 .

1 ........ -;............ ..

0.8 . ................... .. . .

0.6 ..... .... ; .........:.

:x:
0.4 ........ ~

o
-2 -1.5 -1 -0.5 o 0.5
L

Fig. 9. Profiles along the symmetry plane of 3d Barchans of different size normalized
on a single height as obtained from the model (from [35)).

separating streamline s(x) is calculated. Next, the air shear stress T(X) onto
the given surface h (or h and s) is calculated using Equation (10). From the
air shear stress T(X) the sand flux can be determined using Equation (13).
Then, the integration forward in time of the surface is calculated from the
mass conservation, Equation (14). Finally, sand is eroded and transported
downhill if the local angle Ox h exceeds the angle of repose. This redistribution
of mass (avalanches) is performed until the surface slope has relaxed below
the critical angle. The time integration is calculated until the final shape-
invariantly moving solution is obtained.
100 H. J. Herrmann, G. Sauermann

Fig.l0. Complex dune pattern, calculated with the fuH three dimensional model.
Wind is blowing from the left to the right. When Barchan dunes are too elose
they interact, get eventually connected, and form complex dune structures. The
large dunes are shielding the small dunes from the arriving sand flux which then
constantly loose volume [35].

4 The Shape of the Dune and Outlook

In order to analyze the properties ofthe shape-invariantly moving solution
of our model, we performed aseries of calculations varying the volumes of the
Gaussian hills that have been used as initial configuration. The final shape
invariantly moving solutions are displayed in Fig. 8. For small volumes we
obtained heaps without a slip face, whereas for large volumes dunes with a
slip face developed. Hence, there is a minimal height for dune formation or,
more precisely, a minimal height for the formation of a slip face. EmpiricaHy,
this was observed many times in nature.
The simulation also showed that the Barchan shape is a steady state so-
lution. Starting from different initial configurations having the same volume,
one always obtains after a certain transient the same crescent shape dune.
This dune moves with constant velo city inversely proportional to its height
and quantitatively agrees with the ones measured in the field for correspond-
ing volume.
Simulating 3d Barchans by starting from an initial Gaussian heap can
be done under essentially two boundary conditions: an absorbing bedrock
[q = 0] for coastal dunes or a finite influx and q =I 0 on the bedrock for desert
dune fields. In Fig. 9 we see longitudinal profiles along the symmetry plane
of desert Barchans for different initial volumes of dunes, normalized such
that their heights fall on top of each other. One observes that the measured
profiles from Fig. 2 agree very weH with this prediction. The numerical study
of Fig. 9 also shows that the tail on the windward side is not parabolic but
rather logarithmic.
 Evolution and Shapes of Dunes 101

One can also using our programme construct virtual dunes and produce
virtual desert landscapes. One example is shown in Fig. 10 where one sees
the resulting coalescence of Barchans after they have developed from a set
of several Gaussian heaps. In this way one can, starting from measured to-
pographies predict the future evolution and therefore planify in advance to
protect cities and fields against moving sand masses in the Sahara. Another
perspective of the use of our equations of motion is the possibility to study
techniques used to stop or destroy dunes, like Bofix, as introduced by Meunier
in Nouakchott.

Acknowledgement. We want to thank our collaborators P. Rognon, J .S. de

Andrade, A. Poliakov, L.P. Maia and K. Kroy.

References
1. Bagnold, R. A.: The physics 0/ blown sand and desert dunes. Methuen, London,
1941.
2. Pye, K. and Tsoar, H.: Aeolian sand and sand dunes. Unwin Hyman, London,
1990.
3. Lancaster, N.: Geomorphology 0/ desert dunes. Routledge, London, 1995.
4. Finkei, H. J.: The Barchans of Southern Peru. Journal of Geology 67, 614-647
(1959)
5. Hastenrath, S.: The Barchans of the Arequipa Region, Southern Peru, (1967).
6. Hastenrath, S.: The barchan dunes of Southern Peru revisited. Zeitschrift für
Geomorphologie 31-2, 167-178 (1987)
7. Lettau, K., Lettau, H.: Bulk transport of sand by the barchans of the Pampa de
La Joya in Southern Peru. Zeitschrift für Geomorphologie N.F. 13-2, 182-195
(1969)
8. Slattery, M. C.: Barchan migration on the Kuiseb river delta, Namibia. South
African Geographical Journal 72, 5-10 (1990).
9. Oulehri, T.: Etude geodynamique des migrations de sables eoliens dans la region
de LaAyoune (Nord du Sahara marocain). Ph.D. thesis, l'Universite Paris 6,
Paris. no. 92-12 (1992)
10. Hesp, P. A., Hastings, K.: Width, height and slope relationships and aerody-
namic maintenance of barchans. Geomorphology 22, 193-204 (1998).
11. Wiggs, G. F. S., Livingstone, 1., Warren, A.: The role of streamline curvature
in sand dune dynamics: evidence from field and wind tunnel measurements.
Geomorphology 17, 29-46 (1996).
12. Wippermann, F. K., Gross, G.: The wind-induced shaping and migration of
an isolated dune: A numerical experiment. Boundary Layer Meteorology 36,
319-334 (1986).
13. Howard, A. D., Morton, J. B.: Sand transport model of barchan dune equilib-
rium. Sedimentology 25, 307-338 (1978).
14. Sauermann G., Rognon P., Poliakov A., Herrmann H.J.: The shape of the
barchan dunes of Southern Morocco. Geomorphology 36,47-62 (2001).
15. Coursin, A.: Observations et experiences faites en avril et mai 1956 sur les
barkhans du Souehel el Abiodh (region est de Port-Etienne). Bulletin de l' I.
F. A. N. 22A, no. 3, 989-1022 (1964).
102 H. J. Herrmann, G. Sauermann

16. Sarnthein, M., Walger, E.: Der äolische Sandstrom aus der W-Sahara zur
Atlantikküste. Geologische Rundschau 63, 1065-1087 (1974).
17. Sauermann, G., Andrade Jr., J.S., Maia, L.P., Costa, U.M.S., AraUjo, A.D;,
Herrmann, H.J.: Wind Velocity and Sand Transport on a Barchan Dune (2001).
18. Landau and Lifshitz.: Fluid Mechanics, volume 6 of Course 0/ Theoretical
Physics. Pergamon Press, London, 1963.
19. Hunt, J. C. R., Leibovich S., Richards K J.: Turbulent wind flow over smooth
hills Q.J.R. Meteorol. Soc. 114, 1435-1470 (1988).
20. Jackson, P. S., Hunt J. C. R.: Turbulent wind flow over a low hill. Q.J.R.
Meteorol. Soc. 101, 929 (1975)
21. Weng, W. S., Hunt J. C. R., Carruthers, D. J., Warren A., Wiggs, G. F. S.,
Livingstone, 1, Castro, 1: An experimental study of Froud number effect on
wind-tunnel saltation. Acta Mechanica Supppl. 1, 145-157 (1991).
22. Kroy, K, Sauermann, G., Herrmann, H. J.: A minimal model for sand dunes,
Phys. Rev. Lett., 88:054301, 2002. Preprint cond-mat/0203040.
23. Carruthers, D. J., Hunt J. C. R.: Atmospheric Processes over Complex Terrain.
Vol. 23, Chapter Fluiid Mechanics of Airflow over Hills: Turbulence, Fluxes, and
Waves in the Boundary Layer. Am. Meteorological Soc.
24. Zeman, 0., Jensen N. 0.: Progress report on modeling permanent form sand
dunes. Ris0Natinal Laboratory M-2738 (1988).
25. Bagnold R. A.: The movement of desert sand. Proc. R. Soc. London, Sero A,
157:594-620, 1936.
26. Owen P.R.: Saltation of uniformed sand grains in air. J. Fluid. Mech., 20:225-
242,1964.
27. S!Ilrensen M.: An analytic model of wind-blown sand transport. Acta Mechanica
(Suppl.) 1, 67-81 (1991).
28. Lancaster, N., Nickling W. G., McKenna Neumann C. K, Wyatt V. E.: Sed~
iment flux and airflow on the StOBS slope of a barchan dune. Geomorphology
17, 55-62 (1996).
29. Butterfieid, G. R.: Sand transport response to fluctuating wind velocity. In
N.J. Cli:fford, J.R. French, and J. Hardisty (eds.) Turbulence: Perspectives on
Flow and Sediment Transport, Chapter 13, 305-335, John Wiley & Sons Ltd.
(1993).
30. van Dijk, P. M., Arens, S. M., van Boxel J. H.: Aeolian processes across trans-
verse dunes ü: Modelling the sediment transport and profile development. Earth
Surf. Process. Landforms 24, 319-333 (1999).
31. Robert S. Anderson and Peter K Haff.: Simulation of eolian saltation. Science,
241:820, 1988.
32. Anderson, R. S.: Wind modification and bed response during saltation of sand
in air. Acta Mechanica (Suppl.), 1:21-51, 1991.
33. McEwan, I. K and Willetts, B. B.: Numerical model of the saltation cloud.
Acta Mechanica (Suppl.), 1:53-66, 1991.
34. Sauermann, G., Kroy, K and Herrmann, H. J. : A phenomenological dynamic
saltation model for dune formation. Phys.Rev.E, 64:31305, 2001.
35. Gerd Sauermann: Modeling of wind blown sand and desert dunes. Universität
Stuttgart, 2001. PhD Thesis.
 Morphogenesis of Growing Amorphous Films

Stefan J. Linz, Martin Raible and Peter Hänggi

Theoretische Physik I, Institut für Physik,

Universität Augsburg, 86135 Augsburg, Germany

Abstract. In a first part of this paper, we survey the conceptional and method-
ological background of the stochastic field equations approach to model surface
growth processes. In the second part, we focus on recent progress in the modeling
of such equations for the specific case of vapor deposited amorphous thin films that
allow for a quantitative validation with experimental data.

1 Introduction
The fundamental physical laws governing the macroscopic and the micro-
scopic world are, in principle, rather simple. Nevertheless, nature is able to
create a highly complicated and structured world on the basis of these laws.
Unraveling the hidden rules of nature how to fabricate such complex systems,
Le. systems being built up of many interacting constituents and exhibiting a
complicated overall behavior that is not at all evident from the known un-
derlying interaction between the individual constituents, has developed into
a central area of current research in physics.
The morphogenesis or structure formation of growing surfaces and in-
terfaces due to deposition processes constitutes a specific and nonetheless
pararugmatic example of such a complex system [1,2]. Focussing on atomic
deposition processes, the interplay of two individual components, Le. atoms
or moleeules of the typical size of lO-lnm, is basically determined by their
electromagnetic interaction. This, however, does not give us immediate in~
sights into the deposition processes of particles when they arrive at a surface
profile of already collectively condensed particles: It is not at all obvious how
to theoretically understand the plethora of different, experimentally observed
surface structures that can be created by the variety of deposition methods,
deposition conditions and specific type of particles. In this context, vapor
deposited amorphous films exhibit the spectacular fact of building up surface
patterns with some intrinsic regularity or periodicity on scales of the order
of lOnm if the film thickness reaches values of several hundred nanometers
[3-5].
Even if one were able to simulate the corresponding full ab-initio many-
particle problem on square areas of the corresponding size and for correspond-
ingly long deposition times, this would not directly lead to physical insights
into the underlying physical phenomena happening on the mesoscale. In the
last fifteen years, an alternative approach to study surface growth phenomena
[1,2,6,7] basically pioneered by the seminal work of Kardar, Parisi and Zhang

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
104 S. J. Linz, M. Raible, P. Hänggi

jjjjjjjjjjj j I II II jj j

Fig. 1. Sketch of a representative deposition process: Particles from the incoming

ßux (denoted by the arrows) are deposited on the rugged surface and contribute to
the spatio-temporal evolution of the surface.

[8] has attracted considerable interest and developed into a significant branch
of surface physics. Here, the primary ideas are to consider the surface struc-
ture as a continuous two-dimensional manifold on a superatomic level as, for
instance, represented by scanning tunneling microscopy images and then to
model its spatio-temporal evolution by means of stochastic field equations
(SFEs) that incorporate the relevant mesoscopic relaxation mechanisms of
the deposited particles.
The focus of this work is twofold. In section 2, we critically review the
basic strategies of the SFE approach for spatio-temporally evolving surface
morphologies and put some emphasis on the pattern forming aspect of such
equations. In section 3, we specifically apply this method to the problem of
amorphous thin film growth by physical vapor deposition. We review and ex-
tend recent investigations that lead, by comparison with experimental data,
to a rather complete picture of the significant relaxation phenomena and,
most importantly, to an elementary description of the vapor deposition pro-
cess in form of a minimal deposition equation.

2 Basic Concepts
General strategy. As starting point, we present here a synoptic account
of the framework of stochastic field equations (SFEs) to model surface mor-
phologies (for a thorough overview see Refs. [2,7]) including the clarification of
some misconceptions in the literature. The general idea of the SFE approach
is comparatively straightforward: A generally spatio-temporally varying fiux
of particles given by lex, t) reaches the surface (cf. also Fig. 1), the particles
from the beam are deposited at the surface and then undergo various surface
diffusion processes until they arrive at their final position. The growing layer
built up by the deposited particles forms a spatio-temporally evolving free
surface that is characterized by its height or morphology H(x, t) at time t and
the location x = (x, y) measured with respect to coordinates of the initially
 Morphogenesis of Growing Amorphous Films 105

flat surface H(x, 0) = O. The SFE approach disregards the mieroseopie details
of the particle arrangement and interaction at the surfaee and eonsiders the
growth proeess on a slighty larger length seale, the meso- or nanoseale, where
the (eoarse-grained) surface morphology H(x, t) ean be regarded as a field
variable evolving eontinuously in spaee and time. Then, the rate of change
of the surface height H(x, t) ean be expressed in form a partial differential
equation

8t H (x, t) = G[V'H, ...] + I(x, t) (1)

where the funetional G[V' H, ...] depends only on spatial derivatives of H and
their nonlinear eombinations and eomprises all physieal mechanisms leading
to growth and relaxational proeesses at the surface. Writing down (1), three
fundamental symmetry requirements [2] for surface growth proeesses have
been already ineorporated: (i) no dependenee of (1) on the speeifie choice
of the origin of time implying invariance under translation in time, (ii) no
dependenee of (1) on the specifie choice of the origin of the eoordinate sys-
tem implying invariance under translation in the direction perpendicular to
the growth direction, and (iii) no dependenee of (1) on the speeifie choice of
the origin of the H -axis implying invariance under translation in growth di~
rection. These symmetry requirements on the evolution equation (1) exclude
anyexplicit dependenee of the functional G[ .. ] on the time t, the spatial po-
sition x, and the height H, respeetively. Note, however, that these symmetry
requirements do not neeessarily apply to the solutions H(x, t) of (1), too.
It is worthwhile noting that (1) eontains several implicit assumptions:
(i) the relaxation proeesses are loeal in space, (ii) no, in principle possible
changes in bulk of the already built-up layer are taken into aceount, (iii) in
order to guarantee single-valuedness of H(x, t), no overhangs in the evolv-
ing surface strueture are allowed, and (iv) to fulfill moderate existenee and
uniqueness requirements for H(x, t), the (eoarse-grained) surface profile can
have at most some eusps with a non-zero opening angle.

How does stochasticity enter into (I)? In many physieal applieations

such as vapor deposition experiments [3-5], the deposition flux is basically
eonstant with some small superimposed spatio-temporal variations resulting
from the particle souree. As a eonsequenee, the deposition flux ean be split
into a spatio-temporally eonstant mean deposition flux F and a fluctuating
part I(x, t) = F + 1J(x, t). As a simplest model for the generally not well-
known fluetuations 1J(x, t) one usually uses spatio-temporal Gaussian white
noise determined by (1J(x, t)}1/ = 0 and (1J(x, t) 1J(x', t')}1/ = 2D o(x-x') o(t-
t'). Here, ("'}1/ denotes the ensemble average, and D the fluctuation strength.
Sinee the mean deposition flux F is constant it also proofs useful to introduee
the height profile h(x, t) = H(x, t) - Ft in the frame eomoving with the
velo city F. Then, (1) simplifies to

8t h = G[V'h] + 1J(x, t). (2)

106 S. J. Linz, M. Raible, P. Hänggi

As an aside, we note that the method leading to (2) also works if the mean
deposition flux F(t) contains an experimentally predetermined dependenee
on time t. Then one straightforwardly arrives at (2) if the definition hex, t) =
H(x, t) - J~ F(t') dt' is used.

How to specify the functional G[.. ]? As explained before, all physics

of the growth proeess is hidden in the funetional G[.. ]. So, the variety of
different experimentally observable surfaee structures must be direetly re-
lated to the specifie funetional form of G[.. ]. Basically two approaches are
used in the literature. One way is to select known physical relaxation meeh-
anisms (for eolleetion of such mechanisms see e.g.[2]) that are eonsidered to
be relevant for the speeifie system, eombines them and tries to eompare the
outeome with experimental data. This method has the drawback of being non-
systematic. Another way is to start from the guiding principle of simplicity
of the funetional form of G[ .. ] eombined with further symmetry requirements
for the specifieally eonsidered system and aseries expansion of G[ .. ] in small
gradients of hup to some given order. In this ease, one obtains a systematic
skeleton ofthe funetional form ofthe SFEj it is, however, not direetly evident
how to relate all the terms to underlying growth and relaxation processes.
In the remainder of this section, we foeus on deposition processes that also
possess invariance under rotation and reftection in the plane perpendieular
to the growth direetion. An example for such processes is amorphous growth
where the isotropy of the amorphous phase implies such an invarianee. This
symmetry immediately excludes any odd derivatives of h in G and implies
that V' -operators entering the various contributions in G must be multiplied
in eouples by sealar multiplieation. Assuming that all surfaee relaxation pro-
cesses are loeal, we finally expand the funetional G in apower series in all
possible spatial derivatives of h and keep only the terms that are linear or
quadratic in h and only possess a maximum of four V'-operators. As a result
of the afore-mentioned symmetries, the deterministic part of (2) ean only eon-
sist of the terms V' 2h, (V'h)2, V' 4h, V'2(V'h)2, (V' 2h)2, and V' . [(V'h)(V' 2h)].
The last term ean be slightly rearranged in the form

(3)

with

(4)

Consequently, a systematic expansion of the funetional form of the growth

equation (1) that takes into aceount (i) all afore-mentioned symmetries and
(ii) all admissible eombinations of terms being linear or quadratie in hex, t)
and eontaining up to a maximum of four V'-operators is given explicitly by
 Morphogenesis of Growing Amorphous Films 107

[9]
8t h = alV2h+a2V4h+a3V2(Vh)2
+ as(V 2h)2 + a6M + 'fJ.
+a4(Vh)2 (5)
Eq.(5) constitutes the main result of this section.

Some general remarks. Equation (5) consists of two linear terms and
four nonlinear terms in h. The term being proportional to a6 becomes zero
in the one-dimensional limit. This shows the principal problem that one-
dimensionally motivated surface growth equations cannot be carried over
to the two-dimensional case by simply replacing 8z -t V. Moreover, it is
interesting to note that Lai and Das Sarma [10] have also attempted to derive
the leading order functional form of a growth equation using isotropy and the
fact that the functional G in (2) is a scalar. Their result, however, significantly
differs from (5) since the terms as(V 2h)2 and a6M are missing. Therefore,
we conclude that Lai and Das Sarma's growth equation [10] represents an
inconsistent systematic expansion since the terms (8;h) (8;h) - (8z 8yh)2 and
(8;h)2 + (8;h)2 + 2(8;h) (8;h) are not properly taken into account.
The growth equation (5) contains several known limiting cases. The limit
ai = 0 for i = 1, .. , 6, 8t h = 'fJ, is considered as an appropriate model for
random deposition [2]. Setting ai = 0 for i = 2, .. ,6, 8t h = al V 2h + 'fJ, yields
the Edwards-Wilkinson (EW) equation originally motivated in the context
of granular systems [11]. The limit ai = 0 for i = 2,3,5,6, determines the
Kardar-Parisi-Zhang (KPZ) equation, 8t h = al V 2h+a4(Vh)2 +1], being the
paradigm for a stochastic roughening process [8]. Finally, the limit ai = 0
for i = 3,5,6, 8t h = al V 2h + a2V 4h + a4(Vh)2 + 'fJ, leads to the stochastic
version of the Kuramoto-Sivashinsky (KS) equation [12].

Pattern forming aspects. Prom the viewpoint of nonlinear dynam,.

ics, the KPZ equation and the KS equation can be considered as an-
tipodal paradigms although they only differ by the term V 4 h. This can
seen from the linearized version of (5) that dominates the initial stages
of the growth process. Neglecting any stochasticity for the moment, in-
serting a solution or mode ansatz hex, t) = ho exp(ik . x) exp(at) into
8t h = al V 2 h + a2 V 4 h directly leads to the dispersion relation for the growth
rate, a(k) = -a1 k 2 +a2k4, with k = Ikl being the modulus ofthe wave vector
k. Depending on the sign of the coefficients al and a2, four main types of
behavior can be distinguished. For al > 0 and a2 ::; 0, all modes are damped
with time, whereas for al < 0 and a2 ~ 0 all modes grow with time and they
grow the faster, the larger the wave number iso If al > 0 and a2 > 0, then all
mo des beyond the threshold kT = JaI/a2 grow. The most important case,
however, occurs if al < 0 and a2 < O. Then, only wave numbers k in the
range 0 ::; k ::; J al / a2 can grow with time and, moreover, there is a fastest
growing mode km = J al /2a2 that dominates the evolution of h. The latter
result teIls us that, at least for stages where the linearized version of the KS
108 S. J. Linz, M. Raible, P. Hänggi

equation is sufficient, a pattern with an underlying periodicity dominated by

km develops. This growth mechanism disappears in the KPZ-type limit when
a2 is set to zero. Physically interesting behavior of the KPZ equation appears
when al > O. Then, the col1aborate effect of the noise and the nonlinearity
leads to a stochastic roughening of the surface with self-affine character [2].
The noisy KS equation for al > 0 and a2 < 0 leads to a similar behavior. H,
however, al < 0 and a2 < 0, the pattern forming mechanism dominates (for
smal1 enough noise amplitudes) the early stages of the growth process. The
role of the entering nonlinearity is to modify the pattern at later stages of
the growth process.
Besides the invariances already invoked for its derivation, the growth equa-
tion (5) possesses an interesting additional symmetry: It remains invariant
under the combined transformation
(6)
As a consequence, one has to expect that a simultaneous change of the sign of
the coefficients a3, a4, a5, a6 belonging to the nonlinear terms in (5) only leads
to an inversion of the surface profile hex, t) about h = O. Note, however, that
(5) does not possess mirror symmetry about h = 0, i.e. it does not fulfill the
up/down invariance h -+ -h (without inversion of the signs of the nonlinear
coefficients). This already implies some degree of asymmetry of the resulting
surface profile hex, t).

Conservative growth processes. A frequently invoked further requir&-

ment on surface growth equations [2,7] is that the functional G should be
represented by the divergence of a surface current, G = -\7 . ja (\7h) if no
desorption of particles can occur. Such an assumption directly rules out the
appearance of a KPZ term (\7h)2 and the term (\72h)2 in G. It also implies
that the spatial and ensemble averaged height is related to the deposition
flux by ((H(x,t)}'l}x = Ft or, equivalently, that ((h(x,t))'l}x = 0 holds.
The afor&-mentioned assumption, however, also implicitly implies that nO
coars&-grained density variations can occur. In the presence of local den-
sity variations, a discussion starting from the condition that nO incom-
ing particles are lost (cf. the following section) can lead to a KPZ term
(\7h)2 and a term (\7 2h)2 and, therefore, also to a nonzero excess velocity
v = ((8t h}'l}x = ((a4(\7h)2 + a5(\7 2h)2}'l}x of the surface profile hex, t).
How to make contact with experimental results? Modern experi-
mental investigation tools such as scanning tunneling microscopy combined
with image processing allow for a detailed resolution of the surface morphol-
ogy and its spatio-temporal evolution [3-5]. Since the obtained data set is too
immense and the data also contain some degree of stochasticity due to the
small deposition noise resulting from the particle source, the height-height-
correlation function
C(r, t) = (([H(x + r, t) - (H}x][H(x, t) - (H}x]}'l}x.lrl=r (7)
 Morphogenesis of Growing Amorphous Films 109

determines an appropriate quantitative statistieal measure for the informa-

tion on height variations and lateral correlations. In (7), ( ... }7J represents
an average over different sampies (ensemble average), ( ... )x = L-2 JoL d?x ...
the spatial average over a sampie area of size L2, and (H}x = (H}x(t) =
(H(x, t)}x the spatially averaged surface profile at time t. The height-height-
correlation function C(r, t) contains the two most important global quan-
tities that characterize the surface morphology: (i) The correlation length
Re(t) that is given by the first maximum of C(r, t) for non-zero r, Le. by
Re(t) = min{r > 018r C(r, t) = 0, 8;C(r, t) < O}, and, therefore, determines
the typieallength scale over whieh height fluctuations are correlated, and (ii)
the surface roughness w( t) or root mean square deviation ofthe relative height
fluctuations that is determined by the r = O-limit of C(r, t), w2 (t) = C(O, t).
Another often used statistieal measure of the surface morphology is the spec-
tral power density that is determined by

C(k, t) = C(lkl, t) = F[C(r, t)] = F[C(lrl, t)] (8)

where F[ ... ] represents the two-dimensional Fourier transform with respect

to the wave vector k. As a minimum requirement for a successful model-
ing attempt of the spatio-temporal evolution of H(x, t), the validation of
the temporal evolution of Re(t) and w(t) in comparison with the available
experimental data needs to be achieved.

3 Deposition Equation for Thin Film Growth

In this section, we specifically focus on the growth of solid amorphous films

generated by physical vapor deposition under normal incidence of the particle
flux that is important e.g. in the context of coating and the manufacturing of
thin glassy ZrA1Cu films and has recently attracted interest [3-5,13] in ma-
terials science. We review and partly extend some major results obtained in
recent works [13-16] on the development and detailed analysis of a minimal
model in form of a SFE that (i) appropriately describes the spatio-temporal
evolution of such amorphous surface growth processes and (ii) stands the test
of a quantitative comparison with available experimental data [3-5,13]. From
the theoretieal point of view, amorphous film growth constitutes a partieu-
larly attractive testing ground for a quantitative comparison of experimental
data and theoretieal approaches since (i) there are not any long range or-
dering phenomena (as in epitaxial growth processes) to be expected, (ii) the
effect of terrace formation and, therefore, the Ehrlich-Schwoebel effect being
significant for epitaxial growth processes are absent, and (iii) the growing
film should be spatially isotropie. As an appropriate starting point, we can
directly use the general form (5) that contains all basie symmetries relevant
for the vapor deposition process.
110 S. J. Linz, M. Raible, P. Hänggi

11111 I I 11
\.

amorphous
film

Fig.2. Microscopic effects of amorphous surface growth. Left part: InHection of

particles due to interatomic interaction. Middle part: Surface diffusion of deposited
particles due to surface relaxation. Right part: Equilibration of the inhomogeneous
particle concentration due to the geometry of the surface.

Physics behind the growth equation. Guided by the principle that

any mathematically admissible term might have some physieal significanee,
we next relate all terms appearing in the growth equation (5) to the four
competing mieroseopic mechanisms
- surface tension [18]
- eoneentration equilibration of deposited partieies [19,20]
- steering of arriving partieies [14]
- inhomogeneous density distribution [14,15]
that, as we shall see in the next section, seem to dominate physieal vapor de-
position and are all, at least at some stages of the growth proeess, important.
Also the signs and the order of magnitude estimates of some eoefficients in
(5), as weH as a physically motivated simplifieation of (5) are obtained.
The linear term proportional to a2 in (5) ean be interpreted as the result
of a type of a mieroseopie surface tension effeet as originally suggested by
MuHins [18]. The basie idea behind this effect (cf. also the middle part of
Fig. 2) is that the just deposited partieles favoritely move to positions at
the surfaee that have positive eurvature '\72 h > 0 sinee there, the already
eondensed surface partieies form a loeal vieinity with higher binding energy.
This gives rise to a diffusion eurrent jm oe '\7('\7 2h) that, depending on the
loeal eurvature, ean be uphiH or downhill. The divergenee of this eurrent,
-'\7 ·jm = a2'\74h, contributes to the surface evolution in (5) with a2 being
neeessarily negative. This term basica1ly tries to minimize the area of the
surface and, as a consequenee, to smooth the surface morphology. .
The nonlinear term proportional to a3 can be related to the tendeney of
equilibrating the non-homogeneous eoneentration C of the deposited partieies
just after arriving at the surface. This effect has originally been suggested
by Villain [19] (cf. also [20)). The underlying reason is of purely geometrie
nature. Although the deposition flux is basieally homogeneous, more partieies
per surface area arrive at positions with a small or zero modulus of slope '\7 h
 Morphogenesis of Growing Amorphous Films 111

than at positions being strongly inelined with respect to the partiele beam, cf.
also the right part of Fig. 2. Therefore, the local concentration of the diffusing
particles right after the deposition is not constant, but is weighted by the local
slope of the surface, C (X 1/VI + ('\1h)2, or in a small gradient expansion, C (X
1- !('\1h)2. Then, the tendency to equilibrate the concentration is reflected
by a diffusion current je (X -'\1 C (X '\1 ('\1 h)2, or, after taking the divergence,
by the term -'\1 . je = a3 '\12 ('\1 h)2 that contributes to the height changes
in (5). Obviously, concentration equilibration requires that the coefficient
a3 is negative and also tries to smooth the surface morphology. A simple
dimensional argument leads to an estimate for a3. Equation (5) implies that
the coefficient a3 has the dimension of length3 /time. The magnitude of a3
necessarily depends on the deposition flux F that possesses the dimension of
length/time and the mean diffusion length 1 which is the only relevant length
scale determining this process. The only combination of F and 1 leading
to the correct dimension of a3 is Fl2. Therefore, one expects a3 (X - Fl 2.
A thorough discussion of the concentration equilibration [14] supports this
argument and yields the explicit relation a3 = -~FI2. Moreover, one expects
that the typical magnitude of I is of the order of several atom diameters.

The two terms in (5) that are proportional to al and a6 can microscop-
ically be related to the steering of the arriving partieles. Here, the basic
idea [14] is that the particles from the beam experience elose to the growing
surface a deflection due to the interatomic attractive interaction with the
already condensed surface partieles. As a consequence, the particles do not
hit the surface perpendicular to the substrate orientation, but perpendicular
to the surface itself. This implies that more particles arrive at positions at
the surface with negative curvature, '\12 h < 0, than at positions with posi-
tive curvature '\12 h > O. Effectively, this leads to a tendency to roughen the
surface morphology. We refer to Ref. [21] for experimental indications of the
relevance of this effect. To model this scenario in a dynamical way [14], we
use the idealization that the particles undergo a change of direction only af-
ter reaching a critical distance b, the effective range of the interaction, from
the surface and are then attracted such that they arrive perpendicular to the
surface, cf. the left part of Fig. 2. A detailed mathematical derivation [14]
using a reparametrization in the coordinates of the imaginary surface where
the interaction becomes effective (cf. the dotted line in the left part of Fig. 2)
and a small gradient expansion in h in fact shows that this scenario gives
simultaneously rise to the two contributions al '\12h and a6M in (5). More-
over, the coefficients al and a6 can be related to the mean deposition flux
F and the effective range b of the interatomic interaction yielding al = - Fb
and a6 = Fb2 [14]. Although b cannot be directly measured its magnitude
should be typically of the order of one atomic diameter and, therefore, much
smaller than the radius of the surface curvature. This implies that the term
proportional to a6 is of minor relevance in comparison to the al-term and
can be neglected. Moreover, the sign of al is negative.
112 S. J. Linz, M. Raible, P. Hänggi

The physical origin of the nonlinear terms proportional to a4 and a5 is

determined by the potential variations of the coarse-grained density [13,14].
These terms cannot result from particle desorption since the substrate is held
at room temperature and the particle energy in the vapor beam is rather
low (typically of the order O.leV). Therefore, all arriving particles finally
contribute to the surface growth. As a consequence, any term that cannot be
recast in form of the divergence of a current in (5) arises from changes of the
coarse-grained density. Assuming for the moment that the deposition noise
is zero (1] = 0), particle conservation implies that the rate of change of the
number of particles per substrate area above a given substrate location, C, is
determined by a balance equation 8t C = -V -je+PoF. Here the divergence of
the current je is given by the combination of all surface relaxation processes
(cf. the afore-mentioned arguments), Le. by -V ·je = Po [al V 2H +a2V 4H +
a3 V 2 (V H)2 + a6M], and Po represents the density of the growing film in the
case of a horizontal surface. Allowing for density variations at the growing
surface, the rate of change of C is related to the rate of change of the height
H by 8t C = p(VH)8t H. Here p(VH) denotes the density at the surface.
Without the incorporation of density changes (p = Po = const.), there is a
direct proportionality 8t C = Po8t H. H small density variations are taken into
account, p(V H) can be expanded in the derivatives of H yielding p(V H) =
po[1 + ql(VH)2 + q2V 2H] in lowest order approximation. Therefore, 8t H =
Pr;t [1 - ql (V H)2 - q2 V 2 H]8t C holds. Inserting this in the balance equation
from above, explains the presence of the two terms -ql F(VH)2 = a4(Vh)2
and -Q2al(V2H)2 = a5(V 2h)2 appearing in (5). From the physical point
of view, however, density changes are primarily connected to the gradients
of the surface profile reflecting the local arrangement of the particles at the
surface and not so much to the surface curvature. Therefore, it is plausible to
disregard the term a5 (V 2 h)2 in a minimal description of the growth evolution.
Since the density variations result from a widening of the mean inter-particle
distances at the surface or an enlarged number of microscopic vacancies in
the growing material one has to expect that they locally decrease the density
implying that a4 > 0 holds.
Taking into account the afore-mentioned physical arguments, the terms
a5(V 2h)2 and a6M are negligible in leading order and, as a final result, we
obtain the model equation for amorphous film growth [9,13-15],
8t h = al V 2h + a2 V 4h + a3 V 2(Vh)2 + a4(Vh)2 + 1] (9)
with al, a2, a3 being negative and a4 being positive.
Using stochastic numerical simulations of the surface growth equation
(9) starting from a flat substrate (for details of the different methods see
Ref. [17]), we investigate in the remainder of this contribution the evolution
of the correlation length Re and surface roughness w as a function of the
experimentally measurable layer thickness H. This quantity is determined by
H = ((H(x, t)},,)x = Ft+((h(x, t)},,)x and is, in general, implicitly connected
to the time t via the solution of (9). The latter results from the fact that
 Morphogenesis of Growing Amorphous Films 113

3.0

a) b)
E E
.s .s
rr.
U
20
~
2.0
!ZCD
~ • c:
.z::
j! CI
c: :::J
o 2
1; 15 1.0
~
ca
~ • ZrAICu 't: .ZrAiCu
8 :::J
U)

10 0~--=100::-------'2=00::--""--:300==-~-::4OO=-----:! 0.0
soo 0 100 200 300 400 SOO
layer thickness [nm] layer thickness [nm]

Fig.3. Solid lines (dashed lines): Correlation length Re and surface roughness w
for the experimentally investigated thickness interval 0 ~ H ~ 480nm calculated
from the nonlinear growth equation (9) using the parameters al = -O.0826nm2 /s,
a2 = -O.319nm4 /s, a4 = O.055nm/s, D = O.0174nm4 /s and a3 = -O.10nm3 /s (a3 =
Onm3 /s). Diamonds represent the corresponding experimental results previously
published in Ref.[3,22]. This figure is taken from Ref.[13].

the surface profile generated by (9) possesses a finite excess velocity, v =

«8 t h}'1}x = ((a4{VTh)2}'1}x. Since a4 is positive the average of the surface
morphology H{x, t) = Ft + hex, t) grows with a faster speed than F as a
result of the inhomogeneous density distribution.

Selected results. Here, we show that the model equation (9) is indeed
able to quantitatively reproduce experimental data on the correlation length
Re and surface roughness w if the coefficients al, a2, a3, a4 and D are ap-
propriately chosen. For the specific example ofthe growth of Zr65AI7.5Cu27.5
films [3-5,13], a parameter estimation procedure discussed in detail in [13]
yields for the coefficients in (9) al = -0.0826nm2js, a2 = -0.319nm4 js,
a3 = -0.lOnm3 js, and a4 = 0.055nmjs and for the strength of the depo-
sition noise D = 0.0174nm4 js. The experimentally determined mean depo-
sition flux is given by F = 0.79nmjs. For this set of parameter values, we
show the dependence of the correlation length Re and surface roughness w
(solid lines) on the thickness of the amorphous film in Fig. 3 and infer a very
good agreement with the corresponding experimental data. For comparison,
the corresponding results of the Kuramoto-Sivashinsky limit (a3=0, dashed
lines) are given. Since the correlation length ceases to exist at a film thick-
ness of about 300nm in this limit we also conclude that both nonHnear terms
proportional to a3 and a4 are necessary to reproduce the experimental data.
Leaving off the term that describes the effect of density inhomogeneities,
a4 = 0, the surface roughness increases strongly with time and does not show
the cross-over to a saturation at layer thicknesses of about 480nm {for more
114 S. J. Linz, M. Raible, P. Hänggi

details ofthis limit cf. Ref.[14,15]). Moreover, both linear terms proportional
to al and a2 are necessary to excite the growth instability at the initial stages
ofthe growth process [14]. Consequently, (9) must be considered as a minimal
model for the growth of amorphous Zr65Ah.5Cu27.5 films.
The extrapolated parameters al, a2, a3, a4, and D also allow for mi-
croscopic estimates [13]. (i) Since al = -Fb, the typical range b of the
interaction between the surface atoms and the particles to be deposited is
about O.lnm, Le. of the size of the radii (0.2nm) of the surface atoms. (ii)
Since a3 = - Fl 2 /8, the diffusion length l is ab out 1.0nm. Consequently, the
deposited particles experience a surface diffusion on a nanometer scale and
do not just stick at the places where they hit the surface. (iii) If the par-
ticles arrive independentlyon the surface, the deposition noise is related to
the particle volume n and the mean deposition rate F by 2D = Fn [14],
yielding n = 0.04nm 3 . This is up to a factor of two the averaged particle
volume of ZrAICu. (iv) The local density of the growing film varies with the
surface slope: On an inclined surface area the local density is decreased by
p('Vh) = poh with I = 1 + (a4/F)('Vh)2 (where a4/F is about 0.07). These
finite density variations are physically compatible with the small diffusion
length l of two to three atom diameters. At the layer thickness 480nm, this
local density reduction I (averaged over the surface) possesses a mean 1.021
and a standard deviation 0.017.
Finally, we present in Fig. 4 the dependence of the Fourier transform of
the height-height correlation function, eCk, t) (see (8», on the modulus ofthe
wave vector for five subsequent deposition times or layer thicknesses obtained
from the numerical solution of (9). Also here, our numerical simulations show
a striking agreement with previously published experimental data (cf. Fig. 4
[right part] in Ref.[22] and also Fig. 3 in Ref.[23]). As in these experiments,
one recovers (i) the characteristic decay proportional to k- 4 that gives clear
evidence of the importance of the Mullins term a2 'V 4h in (9) and (ii) the
slight buckling of eCk) for the layer thickness of 480nm and wave numbers
k somewhat larger than 1O-lnm- 1 where in a narrow range of k a decay
proportional to k- 1 can be fitted. A detailed numerical comparison with and
without the density variation term, a4('Vh)2, in (9) shows that this effect
is directly related to the inclusion of density variations. Based on their ex-
perimental data (cf. Fig. 3 in Ref.[23]), Mayr and Samwer [23] have recently
suggested an alternative explanation for this effect. Using the old idea tracing
back to Mullins [18] that a linear term leading to a decay proportional to k- 1
might be related to viscous flow in the bulk, these authors have designed a
qualitative model of viscous hill coalescence to explain this feature. Since the
decay of the power spectral density eCk, t) proportional to k- 1 in a small
range of k is in our analysis a direct consequence of the nonlinear KPZ term,
it does not seem to substantiate the presence of viscous flow in the bulk.
Moreover, the remark in [23] that the KPZ term, a4('Vh)2, in (9) might be
interpreted as the lowest order mathematical representation of viscous hill
coalescence is not convincing. If no desorption takes place and a KPZ term,
 Morphogenesis of Growing Amorphous Films 115

10'

103
480nm
360nm
10'
240nm
,.....,
.... 101

--
E 100nm
c:
....... 30nm
10°

.::Z
-10-1
Ü
10-.2

10-3
,,
,,
,,

Fig.4. Dependence of the Fourier transform of the height-height correlation func-

tion or power spectral density, G(k, t), on the modulus of the wave vector for
five subsequent deposition times or layer thicknesses calculated from the nonlinear
growth equation (9) using the parameters Ul = -O.0826nm2 /s, U2 = -O.319nm 4 /s,
U3 = -O.lOnm3 /s, U4 = O.055nm/s and D = O.0174nm4 /s.

a4(V'h)2, is invoked to explain experimental data, then it must be related to

density variations. All other effects that are based on transport mechanisms
should be expressible in terms of divergences of currents.
Next, we explore some properties of the growth process in a layer thick-
ness range up to 2000nm that has so far not yet experimentally investigated.
The results in Fig. 3 up to a layer thickness of 480nm suggest that the growth
process has not yet reached a final, not necessarily stationary state. Using
again the afore-mentioned parameter values, the dependence of the corre-
lation length Re and surface roughness w (solid lines) on the thickness of
the amorphous film is shown in Fig. 5 (the dashed lines refer to the special
case a3=O). Obviously, the surface roughness has reached an almost constant
value for a layer thickness larger than 600nm that increases only very weakly
as the growth process proceeds. In contrast to that, the correlation length
steeply decays after reaching a maximum and then saturates in an almost
constant value for a layer thickness larger than 600nm as the growth process
proceeds. For further results, in particular the properties of the correlation
function and the related height difference correlation function as well as vi-
sualizations of the surface morphology and a theoretical interpretation of the
various stages of the growth process, we refer to Ref.[15].
116 S. J. Linz, M. Raible, P. Hänggi

2t>

a) b)
E E
oS 20 oS 3

i
, , ,-
a:.
~
0
I/)
.s::.
"
a~c:
Öl
r:: 15
~ 2 '
r::
.Q
a;
,~r-----------------1
ä; 10
t:
0
<.>

5
0 600 1000 lSOO 2000 °O~--~SOO
~----~
,ooo
~--~l600
~--~2000
~
layer thickness [nm) layer thlckness [nm]

Fig.5. Solid lines (dashed lines): Correlation length Re and surface roughness w
for the thickness interval 0::; H::; 2000nm calculated from the nonlinear growth
equation (9) using the parameters al = -O.0826nm2 /s, a2 = -O.319nm 4 /s, a3 =
-O.lOnm3/s (a3 = Onm 3 / s), a4 = O.055nm/s, and D = O.0174nm 4 /s.

2600 r----~---~----~----__,

E2000

:eE2000~~
oS oS
~ ~
1\ 1600 1~

•••••
I
v
~
+
V
.;

~~1OOOIilIl
~1000
1\
I
v

~~
;!;500
8
o
50 100 150 200 0 50 100 150 200
x [nm] x [nm)

Fig. 6. Visualization of the cross-section of the spatio-temporal evolution of the

growing surface of the film calculated from the nonlinear growth equation (9). Left
part: realistic parameters for ZrAICu al = -O.0826nm 2 /s, a2 = -O.319nm 4 /s, a3 =
-O.10nm3 /s, a4 = O.055nm/s, and D = O.0174nm 4 /s; right part: same parameters
except of a sign change of al, Le. al = O.0826nm 2 /s.

To obtain further insight into the spatio-temporal evolution of the sur-

face morphology, we present in the left part of Fig. 6 a representative one-
dimensional cross-section of the growth of the surface profile (for y = 0) with
increasing time or layer thickness. For demonstration purposes, the relative
height fluctuations have been weighted by a factor of 20 relative to the mean
thickness (H)x = (H)(x,y). From the left part of Fig. 6, three remarkable fea-
tures can be read off. First, as the time proceeds and the layer builds up, the
surface morphology develops into a predominantly almost periodic structure
 Morphogenesis of Growing Amorphous Films 117

with an averaged periodicity length given by the eorrelation length Re .and

some superimposed stoehastic variations. Seeond, the evolving mound and
dip strueture is asymmetrie in the sense that the dips are eomparatively nar-
row in eontrast to the wide mounds. Third and most remarkably, the surfaee
morphology does not approach a stationary profile in the thickness range
lOOOnm ::; (H)(z,y) ::; 2000nm. Despite the fact that statistical quantities
such as the correlation length and the surfaee roughness are almost eonstant
in this thickness interval, the surface profile still varies signifieantly with time.
The right part of Fig. 6 depicts (up to some scaling of the amplitudes) the
evolution of the surface morphology for the same parameters except that the
sign of the eoefficient al has been inverted. Although this parameter set is
not really physical, it shows the remarkable fact that, as a eonsequenee of the
inversion of the sign of al, the surfaee strueture changes from a pattern with
intrinsie regularity (depicted on the left panel of Fig. 6 with negative al) to a
stochastically varying profile without any regularity when al ehanges its sign
(see right panel of Fig. 6). Beyond that, the roughness of the surface itself is
largely redueed in comparison to the ease al < O. The distinet surface evolu-
tion for positive and negative values of al supports our previous statements
about pattern forming aspeets in section 2.

4 Conclusions and Perspectives

Based on (i) a systematically derived minimal functional form of a growth
equation being appropriate for the understanding of amorphous thin film
growth, cf. equation (5), and (ii) relations of the terms oeeuring in the fune-
tional form (5) to underlying mieroseopie surfaee relaxation meehanisms, a
quantitative agreement of the temporal evolution of the eorrelation length
and the surface roughness of the surfaee morphology with experimental data
ean be aehieved. We expect that similar quantitative agreement of experimen-
tal data and appropriately modeled SFEs for the spatio-temporal evolution
of surface morphologies ean also be obtained for different systems such as for
erystalline growth processes or sputter deposition.

Acknowledgement: This work has been supported by Sonderforschungsbereich

438 (TU München/Univ. Augsburg), Projeet Al.

References
1. Tong, W. M., Williams, R. S.: Kinetics of surface growth. Annu. Rev. Phys.
Chem. 45 (1994) 401-438
2. Barabasi A.-L., Stanley, H. E. : Fractal concepts in sur/ace growth (Cambridge
University Press, Cambridge, 1995)
3. Reinker, B., Moske, M., Samwer, K.: Kinetic roughening of amorphous ZrAICu
films investigated in situ with scanning tunneling microscopy. Phys. Rev. B 56
(1997) 9887-9893
118 S. J. Linz, M. Raible, P. Hänggi

4. Mayr, S. G., Moske, M., Samwer, K.: Early stages in amorphous Zr65Ah.5CU27.5
film growth on HOPG. Europhys. Lett. 44 (1998) 465-470
5. Mayr, S. G., Moske, M., Samwer, K.: Identification of key parameters by
comparing experimental and simulated growth of vapor deposited amorphous
Zr65 AI7.5Cu27.5 films. Phys. Rev. B 60 (1999) 16950-16955
6. Krug, J.: Origins of scale invariances in growth processes. Adv. Phys. 46 (1997)
139-282
7. Marsili, M., Maritan, A., Toigo, F., Banavar, J. R.: Stochastic growth equations
and reparametrization invariance. Rev. Mod. Phys. 68 (1996) 963-983
8. Kardar, M., Parisi, G., Zhang, Y. C.: Dynamic scaling of growing interfaces.
Phys. Rev. Lett. 56 (1986) 889-892
9. Linz, S. J., Raible, M., Hänggi, P.: Stochastic field equation for amorphous sur-
face growth. Lecture Notes in Physics 551 (2000) 473-483
10. Lai Z.-W., Das Sarma S.: Kinetic growth with surface relaxation: continuum
versus atomistic models. Phys. Rev. Lett. 66 (1991) 2348-2351
11. Edwards S., Wilkinson D.R.: The surface statistics of a granular aggregate.
Proc. Roy. Soc. London A 381 (1982) 17-31
12. Drotar, J. T., Zhao, Y.-P., Lu, T.-M., Wang, G.-C.: Numerical analysis of the
noisy Kuramoto-Sivashinsky equation in 2 + 1 dimensions. Phys. Rev. E 59
(1999) 177-185
13. Raible, M., Mayr, S. G., Linz, S. J., Moske, M., Hänggi, P., Samwer, K.: Amor-
phous thin film growth: theory compared with experiment. Europhys. Lett. 50
(2000) 61-67
14. Raible, M., Linz, S. J., Hänggi, P.: Amorphous thin film growth: minimal de-
position equation. Phys. Rev. E 62 (2000) 1691-1705
15. Raible, M., Linz, S. J., Hänggi, P.: Amorphous thin film growth: effects of
density inhomogeneities. Phys. Rev. E 64 (2001) 031506 1-11
16. Linz, S. J., Raible, M., Hänggi, P.: Amorphous thin film growth: modeling and
pattern formation. Adv. Solid State Phys. 41 (2001) 391-403
17. Raible, M., Linz, S. J., Hänggi, P.: Amorphous thin film growth: simulation
methods for stochastic deposition equations. Acta Phys. Pol. B 33 (2002) 1049~
1061
18. Mullins, W. W.: Theory ofthermal grooving. J. Appl. Phys. 28 (1957) 333-339;
Flattening of a nearly planar solid surface due to capillarity. J. Appl. Phys. 30
(1959) 77-83
19. Villain, J.: Continuum models of crystal growth from atomic beams with and
without desorption. J. Physique I 1 (1991) 19-42
20. Moske, M.: Mechanische Spannungen als Sonde für Schichtwachstum und
Schichtreaktionen (Habilitation thesis, Universität Augsburg, 1997)
21. van Dijken, S., Jorritsma, 1. C., Poelsema, B.: Steering-enhanced roughening
during metal deposition at grazing incidence. Phys. Rev. Lett. 82 (1999) 4038-
4041
22. Mayr, S. G., Moske, M., Samwer, K.:ldentification of key surface processes for
vapor deposited amorphous metallic film growth. Mater. Sei. Forum 343-346
(2000) 221-230
23. Mayr, S. G., Samwer, K.: Model for intrinsic stress formation in amorphous
thin films. Phys. Rev. Lett. 81 (2001) 036105 1-4
 Part 11

INTERFACE DYNAMICS:
Modelling and Simulation
 Density Effects and Fluid Flow in Phase-field
Models

Massimo Conti

Dipartimento di Matematica e Fisica, Universita' di Camerino, and Istituto

Nazionale di Fisica della Materia, Via Madonna delle Carceri,I-62032 , Camerino,
Italy

Abstract. In this paper we discuss a method to incorporate hydrodynamic effects

into phase-field models for the solidification of both pure substances and binary
alloys. We start from a generalised thermodynamic potential with squared gradient
terms for the associated fields; the condition of local positive entropy production is
then utilised to derive a set of equations which drive the system towards equilibrium.
The models are numerically solved in one dimension; the effects of the flow field on
the growth dynamics are presented and discussed.

1 Introduction

The change of density upon freezing is a common feature of both pure ma-
terials and alloys. The solid is generally denser than the liquid; the change
ranges from a few percent for simple metals to more than 20% for some eutec-
tic mixtures. The shrinking (or, in some cases, the dilatation) ofthe system,
causes a fluid flow in the liquid phase towards the interface (or away from
it). Then, even in absence of natural convection, the purely diffusive picture
for the heat (and solute, for alloy solidification) released at the growing inter-
face should be extended, to incorporate an accurate analysis of the flow field.
Recently, this subject attracted an increased attention for practical reasons.
The reduction in volume in metal solidification is a source of stress at the
growing interface, and may induce cavitation, with formation of defects or
micropores in cast products [1,2]. A further effect results from the dynamic
pressure drop across the melt: at large growth rates the change of the melt-
ing temperature with pressure may significantly alter the undercooling of the
liquid phase.
A stimulating approach to the density effects in solidification is based on
the diffuse interface picture, through an extension of the phase-field model.
In the classic formulation of the model, a non-conserved order parameter
fjY(x, t) characterises the phase of the system at each point. A suitable free
energy (or entropy) functional is then constructed, that depends on the order
parameter as weH as on the associate (conserved) fields and their gradients.
The extremization of the functional with respect to these variables results
in the dynamic equations for the evolution of the process. Studies conducted
on solidification of both pure substances and binary alloys [3-11] pointed

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
122 M. Conti

out that the model incorporates in a natural fashion the effects of the in-
terface curvature and non-equilibrium phenomena as the trapping of solute
into the solid phase and the kinetic undercooling of the solid-liquid interface.
An extension of the phase-field model to incorporate flow effects was first
proposed by Caginalp and Jones [12]. They derived a system of differential
equations for the variables temperature, order parameter, fluid velo city, den-
sity and pressure. In the momentum equation capillary and viscosity effects
were neglected. Anderson et al. [13] developed a phase-field model with con-
vection in the melt, deriving an expression for the stress tensor within the
arguments of the extended irreversible thermodynamics. The problem was
treated in the quasi-incompressibility approximation, neglecting the pressure
dependence of the local density, which is assumed uniform in the two bulk
phases. The condition of positive entropy production was employed to de-
rive the governing equations of the model. In a subsequent study the same
authors [14] examined the sharp interface limit of the model. Their analysis
recovered the standard interfacial conditions, including the Young-Laplace
and Clausius-Clapeyron equations for the mechanical and chemical equilib-
rium at the interface. Conti [15] foHowed a different approach, starting from
agrand canonical potential which includes the local density as a dynamic
variable. Square gradient terms were allowed for both the structural order
parameter and the density field. The stress tensor was derived assuming that
any dissipation of tensorial order should be ascribed to the fluid viscosity. In
this formulation, the pressure field is consistently related to the local density
via an equation of state. The model equations for the order parameter and
the local density, moment um and energy allow to describe the solidification
process and the flow field, including the propagation of elastic waves. The
model was numerically solved in one dimension to assess the effects of the
fluid convection on the growth dynamics.

In the present paper we present a model which incorporates flow effects

due to density change. The solid phase is modeled as an isotropie fluid with a
viscosity much larger than that of the liquid. The entropy production equa-
tion, coupled with the balance of mass, momentum and energy, is used to
derive governing equations that drive the system towards equilibrium. The
model, derived for a pure substance, is then extended to describe alloy solid-
ification. The equations of the model have been solved numerically in one di-
mension. As solidification starts, the sudden contraction of the liquid in front
of the interface originates apressure (and density) wave that propagates into
the sampie. The results of our simulations show that this phenomenon, as
weH as the other coupled effects of the chemical and mechanical relaxation,
is properly described. In alloy solidification, trapping of solute is observed
at large growth rates. Due to the flow effects on the melting point, the bor-
der between steady and diffusive growth, in the dynamical phase diagram, is
slightly shifted.
 Density Effeets and Fluid Flow in Phase-field Models 123

2 Phase-field Model for a Pure Material

2.1 The Entropy Balance
Let us first eonsider a pure substance which undergoes a solid-liquid transi-
tion. The Ioeal state of the system is characterised by a coarse grained density
p(x, t), the Iocal temperature T and a nonconserved order parameter </J(x, t),
which is assumed to take the values </J = 0 in the solid and </J = 1 in the liquid.
Let us denote the velo city field by v, the specific energy by e' and the spe-
cific entropy by s'. The two latter quantities are determined by the specific
free energy f' (p, </J, T, \1 p, \1 </J) and involve gradient contributions. The stress
tensor will be denoted by P; JE, Js stand for the energy and entropy fiux
vectors respectively, and u is the entropy production rate. Finally, gstands
for a specific body force field. In terms of these variables the classical balance
laws read:

-dp
dt
= -p\1·v (1)

dv
p - =pg- \1.p (2)
dt

de'
p - = -\1. JE - P : \1v (3)
dt

ds'
p-=-\1·Js+u (4)
dt
The constitutive relations and the explicit form of the fluxes will follow
from the Courie principle and from the local form of the second law of thermo-
dynamics, which implies u ~ O. We postulate a generalised specific Helmoltz
free energy of the form

(5)

where f(p, </J, T) is the specific bulk free energy, and the the gradient terms
account for nonlocal contributions in the interfacial region. We assume that
EF and t>F depend only on temperature. The non-gradient part of the specific
energy and entropy are defined by

8f e(p, </J, s) = f(p, </J, T) + T s(p, </J, e) (6)

s(p, </J, e) = - 8T
124 M. Conti

Similar relations are postulated for the corresponding quantities incorpo-

rating gradient terms. Denoting

2 _ d4 (7)
ES = dT

we obtain from (5)

(8)

e'(p, rjJ, 8, V' p, V'rjJ) = e(p, rjJ, 8) + 2~ [<>~(\7 p)2 + E~(\7rjJ)2] (9)

Under the above assumptions the differential form of the second law of
thermodynamics reads:

T ds , = de, - Ö f drjJ - p21 [P - :2<>F

ärjJ 1 2(\7 P)2 - :2EF(V'rjJ
1 2 )2] dp

1
-- [c5~(V'p)d(V'p) + Ei(V'rjJ)d(\7rjJ)] (10)
P
Combining (10) with the balance equations (1,2,3) yields, after some ma-
nipulations:

(11)

where we denote Öi == Ö/ÖXi, and Tik is the capillary stress tensor, defined
by

Notice that this form of the stress tensor extends to a solid-liquid phase
transition with density change a weH known result obtained either for fluid-
fluid interfaces or for solidification without density change. We can rearrange
(11) according to the entropy balance equation (4); the entropy production
rate (1 results as
 Density Effects and Fluid Flow in Phase-field Models 125

(13)

To simplify the discussion, extending the choice of Wang et al. in their

model for solidification at constant density [6], we assume that €fu = dfu = 0;
€§,8§ = const. Moreover in the sequel we shall neglect the spatial variation
of both €} and 8~, which amounts to neglect the thermal gradient across the
interface. Then, the eonstraint of loeal positive entropy production reduces
to

(14)

(15)
where r is a positive constant and K is the thermal eonductivity. Moreover,
assuming that the tensor contribution is only amenable to viscous dissipation,
we obtain

P = -T-II (16)
with II indicating the standard stress tensor for viscous fluids.

2.2 The Thermodynamic Potential

To proceed further we need an explicit expression for the specific free energy
f. At equilibrium the latter should take the form of a double weH over the p, rp
plane, with the two minima centered at the bulk solid (p = PsO, rp = 0) and
liquid (p = PIO, rp = 1), where we denoted as Pso, PIO the equilibrium densities
in the two phases at some reference temperature To and pressure Po. We start
from a linearised equation of state which describes the bulk density shift out
of equilibrium as

P - Po = -ß Po (T - To) + kpo (p + P - Po) (17)

where ß is the thermal expansion coefficient and k the isothermal eompress-
ibility; p is the thermodynamic pressure given by p = p2 (8f /8p) and P is the
exeess pressure due to the capillary stress. In the case of a planar interface
normal to z, the latter is written as P = (4 <p~ + 8~ p~)/2 - pd~ pzz
126 M. Conti

An expression of the free energy, consistent with (17) may be written as

f(p,</J,T) = ag(</J) + ~(T _ To) [p - pO(</J)] + (Po _ P) [p - pO(</J)] -

P k PPo(</J) PPo(</J)
T T
-CTln To + C(T - To) + p(</J)Lo(l- To ) -

_ (po - P) + po(</J) [p - Po(</J)] 2 (18)

Po(</J) 2k PPo(</J)

with g(</J) = (1/4)</J2(1- </J)2. In Equation (18) Cis the specific heat and L o
the latent heat per unit mass in the reference state. We observe a Landau-
Ginzburg contribution for the order parameter </J (first term in the r.h.s of
the equation), and the elastic contribution of the last term. Along the way
followed by [12] we shall neglect in the pressure field the contribution due to
the density dependence of the weIl height. The function p( </J) is monotonie
and increasing with </J, assuming the values p(O) = 0, p(1) = 1, and describes
the transition of the free energy from the solid to the liquid phase. Choosing
p( </J) = </J3 (10 - 15</J + 6</J2) fixes the bulk solid and liquid to the values </J = 0
and </J = 1, respectively, for every value of temperature. The equilibrium
density Po(</J) is assumed to change in the interfacial region as Po == Po(</J) =
PsO + p(</J)(Plo - Pso). In the sequel we assume that all the bulk properties
of the system change in the same fashion along the solid-liquid interface, for
example we represent the loeal viscosity as '" = "'8+p(</J)(",,-1Js), with "'8 » "".
Then, Equations (1,2,3,14) along with specifications (15,16,18) represent the
evolution equations for the system.

2.3 Nondimensional Equations in one Dimension

We assume equal values of the thermal diffusivity D and the specific heat
(at constant pressure) Cp in both phases. A nondimensional form of the
model equations is obtained adopting a reference length ~ and scaling time
to T = e/
D. Density is scaled as P/ Plo and a nondimensional temperature is
introduced as u = Cp (T - To)/ Lo, with To being the melting temperature in
the reference state. Specific energies will be sealed to v~, where Vo = ~/T is the
natural reference for velocities, and the scale for the components of the stress
tensor is PIOV~, The equilibrium density takes the form Po = S + p(</J)(1- S),
with S = Pso/Plo. Notice that in the following we neglect thermal expansion
effects and we assume equal compressibilities in both phases. Retaining for
simplicity the same simbols for the scaled (nondimensional) quantities, and
in absence of body forces, the model equations in one dimension read:

op op ov
-+v-=-P- (19)
ot oz oz
 Density Effects and Fluid Flow in Phase-field Models 127

(20)

p8v
- 8v _-
+ PV- - r7
v·
[Rl
--
R3 (p - Po)] + R 48- (,"I (A.)8V)
'I' - (21)
8t 8z m p 8z 8z

8u +v 8u = -p'(</J) (8</J + v 8 </J) + 8 2u2 + R 4 (8v)2 (22)

8t 8z 8t 8z 8z R5 8z
with Al(</J) = 1 + [1 - p(</J)][(1]s/1]I) - 1]. The parameters appearing in the
above equations are defined as:

rE 2 r
m=_F_. R1 = - ;
rr
~2 ' k
R _ 8~PIO. Lo
2- ."C2 V 2' R5=2 (23)
o Vo

Notice that imposing Po = S = 1 the model collapses on the the classical

phase-field description of the the solidification of a pure substance. In this
case (see for example [6,16]) the model parameters m, €, a, EF, a can be related
to the material properties through

J.L(7To _ h
m = DpoL; E =-; a=-~-·
~ 6v'2do '
E~ = 6v'2 (7 h; a= 6v'2~ (24)

where h is the interface thickness, and do = «(7CTo)/(PoLo2) the capil-

lary length (with (7 being the surface tension); J.t is the kinetic undercooling
coefficient that relates the interface undercooling to the interface velocity VI
through VI = J.t(To - T). We assume that the above equations still represent a
reasonable estimation of the model parameters in terms of the thermophysical
properties of the material.

3 Phase-field Model for a Binary Alloy

We are concerned with an ideal binary solution with components A (solvent)
and B (solute). To simplify the discussion we neglect the difference of the
128 M. Conti

molar mass M of the two components, assuming MA MB _ M, that

means to identify the molar concentration with the mass fraction. A further
balance equation, in addition to (1,2,3,4), must be supplemented for the solute
concentration c. The latter reads

dc
p- = -V ·JD (25)
dt
where JD is the diffusive solute flux. mternal energy, entropy and free energy,
exhibit an additional gradient term for the local concentration, with coeffi-
cients 1]~, 1]§, 1]~, respectively. The entropy production rate reads, in this case

CI = _!..T d<jJdt (p {)f{)cp _ €~ V cp) +

2 JE' V (!..)
T
+

dcp (€2) dp 2)
(8..E. dc (1]2
+-V<jJ·V ..E. +-Vp.V +-Vc.V ..E ) +
dt T dt T dt T
+JD'V [(PA; /l-B) + 1]~p;2c] _ ~ (P + T) : Vv (26)

with the capillary stress tensor T given by

Tik = 8ik [-p + p8F2V2P + iF(V

1 2 2
p) +
1 2 2 1
"2€F(VCP) + "21]F(VC)
2 2] -

-8~PiPk - €~CPiCPk - ~CiCk (27)

Then, the constraint of positive local entropy production pro duces a fur-
ther condition, in addition to (14,15,16), Le.

J = MV [(/l-A -/l-B) + 1]~V2c] (28)

D e T pT

where Me is a positive constant. The specific free energy is written as

(29)

where /A, fB represent the free energy of pure A and B, obtained replacing
in (18) the material parameters with the ones of the solvent and the solute,
respectively, and the equilibrium density with

PO(cp) = POA POB (30)

(1 - C)pOB + CPOA
 Density Effects and Fluid Flow in Phase-field Models 129

Then, the solute flux may be expressed as

McR ('T}~
JD=-Me(l_e)V'e+McV' pTV'e
2)
+McV' { (aA-aB)
pT g(</J) p(</J)
+T [LOA ( T
1- ) - LOB (1-
TOA TTOB
)]}

We recover the standard definition of the solute diffusivity D c taking

Dc = McR (31)
pMe(l- e)

As the solute diffusivity is quite different in the solid and liquid phases,in
the following D c will be taken as D c = Ds + p( </J) (Dl - Ds), Dl and Ds being
the diffusivities in the liquid and in the solid, respectively. We now adopt
a reference scale for time as T = e/ Dlj the nondimensional temperature is
defined as u = Cp (T - TOA) / LOA and density is scaled to PIOA.
In nondimensional form the governing equations become

op op ov
-+v-=-p- (32)
ot oz oz

(33)

OV ov
p~+pv-=-V'· [(P-PO)]
ß3 0 ( Al(</J)-
+ß4- Ov) (34)
ot oz P OZ OZ

oe oe 0 0 { 1 - e TOA
P-
ot
+ pv-
oz
= --A2(</J)pe(1-
OZ
e)- In -
OZ e
+ -ß5g(</J)(1-
T
Wl)

TOA
+-p(</J)
T
( ßIß5-
m
f2) [(1- -TOAT ) - W2 ( 1- TOB
- T )] + -
TOA - -e }
ß6 02
T P OZ2
130 M. Conti

8u 8u (8c/> 2
I
-+v-=-p(c/»[(I-c)+cw2] - + v8c/»
- +DT-8 -
u+
2
ß2ß4
- (8V)
- 2
8t 8z 8t 8z Dl 8z ßIß3 8z
(35)
with )..2(c/» = (Ds/D l ) +p(c/>)[I- (DsjD I )], and the parameters defined as

aB LOB
Wl =-aA
j W2 = - - j
LOA
SB = PsOB j E = PIOB (36)
PIOA PIOA

4 N umerical Solutions
4.1 Pure Substance
The equations of the model have been solved on the computational domain
o~ z ~ Zm. Initially aphase boundary at z = Zo separates a solid region
(z < Zo, c/> = 0, P = S) from the liquid region (z > Zo, c/> = 1, P = 1). The
system is initially at rest and the liquid is undercooled, i.e. v(z, 0) = 0, u(z ~
zo,O) = 0, u(z > Zo, 0) < O. For the phase, density and temperature fields we
imposed Neumann boundary conditionsj the velo city is fixed as v(O, t) = 0 at
the left end ofthe solid, while we chose (8vj8z) = 0 at z = Zm. To advance the
solution forward in time, we employed an explicit Euler integration scheme for
the density, temperature and phase-field equations. To allow for the condition
'Tls » 'Tll, which originates a large diffusion term in the momentum equation,
the latter was integrated with a Crank-Nicolson implicit scheme. Second order
central differences were used to discretize the Laplace operator, and upwind
differences for the convective terms. To ensure an accurate resolution of the
solid-liquid interface the grid spacing was selected as L1z = 0.25 ·lj the time
step required for numerical stability is L1t = 0.25 . 10- 7 • We referred to
the thermophysical properties of Nickel, fixing the parameters as m = 0.1,
a = 265, l = 0.001, S = 1.05, R 1 = 5.29.106 , R 2 = R 3 = 7.12'10- 3 ,
R4 = 0.105 and R5 = 4940.
We first checked whether the model gives a consistent description of the
mechanical effects due to the density change in solidification. To this aim we
solved the model equations at constant temperature (u = -0.005), fixing our
attention to the mechanical relaxation of the system. The contraction of the
liquid in front of the interface originates apressure wave which propagates
 Density Effects and Fluid Flow in Phase-field Models 131

both into the solid and into the liquid. This effect is illustrated in Fig. 1, where
only the liquid portion of the system is shown. The solid-liquid interface is
located near z = 0.3 (here and in the following all the numerical results will
be presented in nondimensional form), and the pressure field is represented at
different times. The initial pressure was initialised as p(z,O) = 0, and we see
the negative pressure front which propagates into the liquid. The velo city of
the wave, estimated tracking the position of the wave front, results Vs = 615.5;
this value is in excellent agreement with the theoretical one for the liquid in
equilibrium, (ljJ = 1,PI = 1), assumed as a pure elastic medium: the latter is
Vap/ap'= VRl R3 /m = 613.7.

20

)~~,~~~,V\ (
V
V

-40

.(,()

0.30 0.40 0.50 0.60

z

Fig. 1. The pressure wave originated at the solid-liquid interface. The interface is
located near z = 0.30, an the different curves represent the pressure field at times
1.03 . 10- 4 , 1.69 .10- 4 , 2.34 · 10- 4 ,3.00 . 10- 4 and 3.66· 10- 4 from left to right. The
wave speed is vs = 616.

After a short transient, growth at fixed temperature results in a steady

advancement of the solid-liquid interface. In Figs. 2(a)- 2(c) we show the
phase, density and velo city profiles obtained at different times, with u =
-0.004. We observe that the solid is at rest, while the liquid is advected
towards the interface with a velocity v = -0.0356. The interface velocity, as
resulting from the numerical data, is VI = 0.7123. Notice that this is the same
value fixed by the mass conservation law through the relation v = (1- S)VI.
132 M. Conti

The conditions for mechanical equilibrium at a curved interface force a

pressure jump across the phase boundary. This effect is shown in Fig. 3
for the growth of a circular germ (the initial undercooling is Ll == u(z >
zo,O) = -0.4). The normal stress is represented versus the radial coordinate
at different times. We observe the difference Llp of the pressure in the two
phases, due to the Young-Laplace effect. Notice that the jump decreases with
increasing the radius R of the germ. The product R Llp remains constant
during the process, and should correspond to the (nondimensional) surface
tension. Indeed, the latter, given by R3 / (i..;72) is in excellent agreement with
the observed value R Llp = 0.84 .

.-.
t
(
f {
:
! -
-- -
-
J j j
on
0 , 10 O,U .~o 0.2> .....

!I,I,
,
\
\
1.
i
\
:I

.). 0$ r - - -- - - - - - - - - - - ,

1
f
...
0_00
~ ~
,
i i
..0 ,0$
O •.IU 1,) , 1:.'0 0.,. 0= 0.>.

Fig.2. The phase (a), density (b) and velo city (c) fields at different times. The
curves are taken at t = 0.06, t = 0.12, t = 0.18 from left to right.
 Density Effects and Fluid Flow in Phase-field Models 133

...0 .0 . - - - - - - - - - - - -- - - - -_ _ _- ,

~ 20 .0
f----'=--_
"\
- ,
: ~ \
0.0

0 .00
i o,os
i \\.
' - - _-.a..._..........._ _ _ _ _ _ _ _ _ _ _-'
0 . 10 0 . 15 0 .20

Fig. 3. Profile of the negative of the normal stress in circular growth. The curves
are taken at times 0.875.10- 3 ,2.250.10- 3 ,3.625.10- 3 ,5.000 .10- 3 from left to right.
We can observe the pressure jump across the interface due to the Young-Laplace
effect.

4.2 Binary Alloy

In general we referred to phase diagram of an ideal solution of nickel (sol-

vent) and copper (solute), fixing the model parameters as m = 466, ßl =
8.484 . 108 , ß2 = 6.681 . 1010 , ß3 = 9.231 . 1013 , ß4 = 1.282· 103 , ßs =
1.071, ß6 = 6.853· = 5.727·
10-7 , Wl = 6.910 · 10- 1 , SA =
1O- 1 ,W2
1.051, SB = 1.026, E = 0.983. The interface width was chosen as l = 8.10- 4 ,
and the initial concentration of the solution is c = 0.07214, that belongs to
the solidus line at T = 1700K. We recovered all the expected phenomenol-
ogy. For example, we observed the transition from the kinetic to the diffusive
regime in isothermal growth, as the temperature approaches the To line from
below, and solute trapping occurs at large growth rates. We want now to
draw the attention on an interesting point. The flow effects result in a slower
interface dynamics (see for example [1]) . Then, we may expect that in the
dynamic phase diagram (in the T, c plane) steady state solutions, due to den-
sity change, could be pushed into the region of diffusive regimes. This effect
is illustrated in Fig. 4, where the growth rate is represented versus time, at
T = 1703K. The upper set of data shows the kinetic (steady) growth obtained
without density change (Le. fixing SA = SB = E = 1). In the lower curve we
accounted for the density effects, fixing SAY = SB = 0.8, E = 1. We see in
this case the transition from the kinetic to the diffusive regime.
134 M. Conti

5 Conclusions
The classical phase-field model is a weH established tool to describe solidifi-
cation far from equilibrium. In this paper we derived the governing equations
of the model in the framework of the extended irreversible thermodynamics,
considering different densities of the solid and liquid phases and taking into
account the effects of capillary stresses. The equations reduce to the classical
formulation for equal solid and liquid densities. The numerical solutions of
the model show that the sound wave propagation, the interfacial dynamics
and the ßow field are properly described. Then, the model has been extended
to describe alloy solidification. In future work we plan to study the effects of
the mechanical relaxation at the early stage of the growth. In areal process,
as the nucleation starts, the homogeneity of the melt is rapidly lost and all
the growing nuclei become the source of (primary or reßected) elastic waves.
This phenomenon results in a new interaction mechanism between the grow-
ing germs, and could alter in a significant way the first stage of the crystal
growth.

10000

1000
--
-'.
....1$~ 100

ä
'0

0 . 00001 0.00010 0 . 00'00 0 .0.1 000 0 . 10000

Fig. 4. The growth rate for T = 1703K. The upper curve shows the steady regime
when density effects are neglected. The ßow field shifts the growth regime into the
diffusive region (lower curve)

References
1. Caroli, B., Caroli, C. and Roulet, B. in: Bolids lar /rom equilibrium, ed. Go-
dreche, (Cambridge University Press, Cambridge, 1992).
2. Kurz, W. and Fisher, D. J .: Fundamentals 01 solidification (Trans Tech Publi-
cations, Aedermannsdorf, 1992).
 Density Effects and Fluid Flow in Phase-field Models 135

3. Caginalp, G.: Arch. Ration. Mech. Anal. 92, 205 (1986).

4. Caginalp, G.: Phys. Rev. A 39, 5887 (1989).
5. Penrose, O. and Fife, P. C.: Physica D 43, 44 (1990).
6. Wang, S. L., Sekerka, R. F., Wheeler, A. A., Murray, B. T., Coriell, S. R.,
Braun, R. J. and McFadden, G. B.: Physica D 69, 189 (1993).
7. Caginalp, G. and Jones, J.: Annals of Physics 237, 66 (1995).
8. Wheeler, A. A., Boettinger, W. J. and McFadden, G. B.: Phys. Rev. E 47, 1893
(1993).
9. Conti, M.: Phys. Rev. E 56, 3197 (1997).
10. Ahmad, N. A., Wheeler, A. A., Boettinger, W. J. and McFadden, G. B.: Phys.
Rev. E 58, 3436 (1998).
11. Zhiqiang Bi, Robert F. Sekerka, Physica A 261, 95 (1998).
12. Caginalp, G. and Jones, J.: in IMA Volumes on Mathematics and its applica-
tions, edited by Gurtin, M. E. and McFadden, G. B. 29-50, (1991).
13. Anderson, D. M., McFadden, G. B. and Wheeler, A. A.: Physica D 135, 175
(2000).
14. Anderson, D. M., McFadden, G. B. and Wheeler, A. A.: Physica D 151, 305
(2001).
15. Conti, M.: Phys. Rev. E 64, 051601 (2001).
16. Warren, J. A. and Boettinger, W. J.: Acta Metall. Mater. 43, 689 (1995).
A Lattiee Boltzmann Method for the
Mesoseopie Calculation of Anisotropie Crystal
Growth

Wolfram Miller1 and Igor Rasin12

1 Institut für Kristallzüchtung (IKZ), Max-Born-Str.2, 12489 Berlin, Germany

2 Brandburgische Technische Universität Cottbus

Abstract. We developed a new kind of a phase-field model in the framework of the

lattice Boltzmann method. The latter is used to calculate the convection und heat
transport. In the continuum limit of second order the new lattice phase-field model
converges towards the ordinary continuum phase-field model. The calculation of
dendritic growth in two dimensions shows the amenibility of the new method.

1 Introduction

In industry as well as in nature, erystal growth proeesses are often strongly

infiueneed by eonveetion. In addition, the eonvection ean exhibit patterns on
rather small seale, sometimes eonvection is even turbulent. During the last
years the lattice Boltzmann method has been evolved as a powerful numerical
tool for fiow ealeulations [12,16], especially regarding its eflicieney on parallel
platforms [8,11]. This was the motivation to develop a new kind of lattiee
phase-field model for the lattiee Boltzmann method [3,10]. In this paper we
will foeus on the deseription of the lattice phase-field model and its appli-
eation to dendritic growth. Details of the lattice Boltzmann method ean be
found e.g. in [9,12,7].

2 Lattice Phase-field Model

In analogy to the eontinuum models for the phase transition we define a

eontinous order parameter -1 ::; cl> ::; +1, which takes the value +1 for melt
and -1 for solid. The order parameter is defined on a lattice point r * at a
diserete time t*. Its evolution is given by a simple rate equation:

(1)

The rate 'R is given by the probability for melting and solidifeation, K+
and K-, respeetively. The timeseale for the phase transition is defined by a
frequeney factor f.

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 A Lattice Boltzmann Method for Mesoscopic Calculation 137

For the probabilities we make the following approachs:

K± = ~ (1 ± tanh {(T'~ Ts )}) ~(1 ± <p(r* + cD)2 (3)

t

for model I and

(4)

for model Ir. In model 11 we have an additional non-Iocal term in (2):

(5)

It can be shown that in the continuum limit of second order modell 11 con-
verges towards the continuum phase-field equation

whereas model I includes additional first order derivatives in the phase-field

<p [9]. <Ps = 1/2(1 - <p) is the solid fraction, J.tk the kinetic coefficient, r
the Gibbs-Thomson coefficient, and 68 is the width of the transition region
(-0.9 ~ <p ~ 0.9).
The release of latent heat is included by a temperature change at the
nodes:

(7)
where St is the Stefan number or undercooling St = cpiJ.T / L, iJ.T* and iJ.T
are the characteristic temperature differences of the system in lattice and
physical units, cp and L are the specific and latent heat, respectively.
We want to mention that all entities are measured in lattice units (l.u.). The
length scale is measured in units of lattice distances, the time scale in number
of iterations, the temperature scale can be chosen arbitrarly with the restric-
tion ITI < 0.2. The relevant energy scale is defined via 7.
Often, the surface energy of a crystal differs in the various spatial directions.
This anisotropy can be considered in the phasefield model by introducing a
dependency on the direction into the length and time scale, 8 and f, respec-
tively. The scales are modified in the following way:

8 = 80 (1 + Ecosn8),
8 is the angle between the crystal axes and the growth direction and n defines
the symmetry of the crystal. In this article we consider a fourfold symmetry,
i.e. n = 4. Some additional modifications we have to made to our models in
order to compute the growth direction. For model I these modifications are
described in [9], for model 11 they will presented in a forthcoming article.
138 W. Miller, I. Rasin

3 Results

We have studied the growth of a small circular seed into an undercooled

melt. The problem is two-dimensional with a fixed temperature at the bound-
aries. The problem is characterized by a length, time and energy scale, rep-
resented by the capillary length number Ca, the Damkähler number Da and
the Stefan number St. The definitions are Ca := do/(V28), Da := rpk/D,
and St = i1Tcp / L. The relations in terms of the lattiee parameters are:
Ca = T w St/V2i1T* and Da = 82 f /2D* . The width of the transition region
8 is an input parameter in model 11 but in model I it can be only measured
after a run. In the case of anisotropie growth, the strength of anisotropy € en-
ters the list of characteristic parameters. Firstly, we present a result obtained
with model 1. The capillary length, the Damkähler, and the Stefan number
were Ca = 3.5 x 10- 3 , Da = 0.2, and St = 0.1, respectively. Choosing an
anisotropy of € = 0.22 we obtained a dendritie growth. In the first run the
crystal axes coincided with the lattiee directions, in the second run they were
tilted by 18°. The resulting dendrite was the same in both cases (see. Fig. 1)
[10]. If we switch on thermal convection and define the gravity in y-direction,

Fig.1. Dendritic growth computed with modell. Interfaces (</> = 0) are shown for
different timesteps (4000, 8000, 12000, 16000, and 20000 iterations). In the figure
on the right the crystal axes were tilted by 18° according to the lattice.

the symmetry is broken and the resulting crystals will not be the same for
the tilted and the untilted case. In Fig. 2 the shape of the dendrite and the
convection field is shown for the case of a tilted crystal. The runs were per-
formed on a 200x200 lattiee. The melt had a Prandtl number of Pr = 0.1
and the Rayleigh number was Ra = 1 X 105 • The Prandtl number is the
ratio of viscous to thermal diffusivity. Due to the better heat transport, the
dendrite is growing is faster in the direction of upstreaming flow. Our obser-
vations are consistent with those of Bänsch and Schmidt [1]. Other authors
 A Lattice Boltzmann Method for Mesoscopic Calculation 139

have considered cases of dendritic growth in a forced flow, either with a sin-
gle dendrite [14] or an array of dendrites on a surface [15]. With model 11 we

Fig. 2. Dendritic growth computed with modell. The thermal convection is induced
by the heated crystal. On the lefthand side the interface (</J = 0) is shown for
different timesteps (4000, 8000, 12000, 16000, and 20000 iterations) and righthand
side the velocity field in the vicinity of the crystal (iteration 15000) is represented
byarrows.

have calculated the dendritic with a parameter set of the same order as in the
paper of Tong et al. [13] . In particular, we have chosen Ca = 0.1, Da = 0.063,
St = 0.5, and € = 0.05. The resulting dendrite at iteration 5000 is shown in
Fig. 3. The computational domain was 400x400. H we increase the Damköh-
ler number by increasing the frequency for the phase transition the dendrite
is growing faster with a smaller tip radius (Fig. 4). The tip radius is evaluated
numerically by interpolation with a parabola. On the lefthand side of Fig. 4
the accuracy of this approach is chown in comparison with the real shape of
the tip. We measured a tip radius of Ptip=2.2 l.u. and a tip velocity of Vtip=
0.033l.u., which yield a tip Peclet number of Pe=0.145. The tip Peclet num-
ber is defined via Pe = PtipVtip/2D* and according to the linear solvability
theory of Ivantsov the Peclet number depends on the undercooling only [6,41.
Nevertherless, more eleborate solvability theories and numerical calculations
show an decreasing Peclet number with increasing isotropy [5]. The Peclet
numbers of our computations are in good agreement with the oberservations
of Karma and Rappel [5]. Another point of discussion is the shape of the
tip. In an experiment with xenon Bisang and Bilgram observed a tip shape,
which deviates significantly from a parabola. Instead of an exponent of 2 the
shape was fitted weH by apower law with an exponent of 1.67 [2]. We did
not find such an deviation in our calculations. The parabola fits the shape of
the tip quite weH as it can be seen in Fig. 4.
140 W . Miller, I. Rasin

3M ~ .- __- -__- -__- -______- ,

300 300

2M 2W

200 200

100 100

M50 100 1M 200 250 300 ~

Fig. 3. Dendritic growth computed with model 11. Isolines of phase-field (cjJ =
-0.9, 0.0,0.9) and temperature isolines for iteration 5000. The righthand figure
presents the result, when crystal and lattice axes are tilted by 17.1° .

400 T =-O,OI
f

200

8~8'
5 ~- 200 215

Fig. 4 . Dendritic growth computed with model 11. Isolines of phase-field (cjJ =
-0.9,0.0,0.9) and temperature isolines for iteration 3000 (left). Magnification of
the lower tip region and comparison of the tip interface with a theoretical parabola
curve (right).

4 Conclusions

We have presented a lattice version of the phasefield model, which is embed-

ded in the lattice Boltzmann method for the calculation of convection and
heat transport. We showed its amenability to calculate dendritic growth in
two dimensions without exhausting much computer time. This result is en-
couraging to develope an efficient parallel version of the programm for large
scale simulations in the future. This will also include calculations of growth
and flow in three dimensions.
 A Lattice Boltzmann Method for Mesoscopic Calculation 141

Acknowlegement
Part of the model has been jointly developed with Sauro Succi. Financial
support from ESF-PESC Scientific Programme "Chalienges in molecular sim-
ulations: bridging the length and time scale gap" is kindly acknowledged.

References
1. Bänsch, E. and Schmidt, A.: Simulation of dendritic crystal growth with ther-
mal convection, Interfaces and Free Boundaries 2 (2000), 95-115.
2. Bisang, U. and Bilgram, J. H.:, Dendritic solidification of xenon, J. Crystal
Growth 166 (1996), 212-216.
3. de Fabritiis, G., Mancini, A., Mansutti, D. and Succi, S.: Mesoscopic models
of liquid/solid phase transitions, Int. J. Mod. Phys. C 9 (1998), 1405-1415.
4. Ivantsov, G. P.: Dokl. Akad. Nauk SSSR 58 (1947).
5. Karma, A. and Rappel, W. -J.: Quantitative phase-field modeling of dendritic
growth in two and three dimensions, Phys. Rev. E 57 (1998), 4323-4349.
6. Langer, J. S.: Instabilities and pattern formation in crystal growth, Rev. Mod.
Phys. 52 (1980), 1-28.
7. Massaioli, F., Succi, S. and Benzi, R.: Exponential tails in Rayleigh-Benard
convection, Europhys. Lett. 21 (1993), 305.
8. Miller, W. and Succi, S.: Parallel three-dimensionallattiee Boltzmann hydro-
dynamies on the IBM 9076-SPl and SP2 scalable parallel computers, in Com-
putational Fluid Dynamies '96, Desideri, J. -A., Hirsch, C., Tallee, P., Pandolfi,
M. and Periaux, J.: eds., ECCOMAS, John Wiley & Sons, 1996, 1052-1058.
9. - - , A lattice Boltzmann model for anisotropie crystal growth from melt, J.
Stat. Phys. 107 (2002), 173-186.
10. Miller, W., Succi, S. and Mansutti, D.:, A lattice Boltzmann model for
anisotropie liquid/solid phase transition, Phys. Rev. Lett. 86 (2001), 3578-
3581.
11. Punzo, G., Massaioli, F. and Succi, S.: High-resolution lattice-Boltzmann com-
puting: on the SPI scalable parallel computer, Comput. Phys. 5 (1994), 1-7.
12. S. Succi The lattice Boltzmann Equation, Numerical Mathematics and Scien-
tific Computation, Clarendon Press, Oxford, 2001.
13. Tong, X., Beckermann, C. and Karma, A.: Velocity and shape selection of
dendritic crystals in a forced ßow, Phys. Rev. E 61 (2000), R49-R52.
14. Tong, X., Beckermann, C., Karma, A. and Li, Q.:, Phase-field simulations of
dendritie growth in a forced ßow, Phys. Rev. E 63 (2001), p. 061601.
15. Tönhardt, R. and Amberg, G.:, Dendritic growth of randomly oriented nuclei
in a shear ßow, J. Crystal Growth 213 (2000), 161-187.
16. Wolf-Gladrow, D. A.:, Lattice-Gas cellular automata and lattice Boltzmann
models, vol. 1725 of Lecture Notes in Mathematics, Springer-Verlag, Berlin
Heidelberg New York, 2000.
 A Phase-field Model for the Solidification
Process in Multicomponent Alloys

H. Garcke 1 , B. Nestler 2 and B. Stinner1

1 Naturwissenschaftliche Fakultät I - Mathematik, Universiät Regensburg

2 Fachbereich Informatik, FH Karlsruhe

Abstract. Our aim is to describe phase transitions in a system of an arbitrary

number of components and phases. Based on a gradient ßow for the entropy (in-
cluding surface anisotropy) we propose aphase field model that can be regarded
as an extension of the Penrose-Fife model and that is thermodynamically consis-
tent. By formal asymptotic expansions we see that the considered domain splits
into several phases. We define the surface entropies on the phase bouridaries and
then we can show that in the limit the model satisfies the Gibbs-Thomson relation
and other conditions known from classical sharp interface models. Finally, some
possibilities to linearize the equations are outlined.

1 Solidification Effects and Length Scales

The solidification process in multicomponent alloys is charactarized by effects
that occur on severallength and time scales and that inßuence each other. On
a large length scale castings reveal aseparation into several grains involving
various kinds of grain boundaries. This separation into grains is combined
with a complex microstructure. Examples are given in Fig. 1.

Fig. 1. Examples for microstructures occuring during the solidification process of

alloys

On the first image we can see a dendrite combined with lamellar eutec~
tic structures growing into a liquid melt. The second image shows growing

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 A Multicomponent Phase-field Model 143

eutectic grains/cells, so called eutectic colonies. The effective growth laws of

the solid-liquid front which is influenced by''macroscopic'' variables (as e.g.
the temperature) are ofinterest as well as the infiuence ofthe microstructure
formation on such macroscopic variables. Modelling a solidification process
means therefore that one has to take into consideration the different scales
as weH as their interactions.
Mathematically, phase field models can be used to avoid topological diffi-
culties with the motion of free boundaries in sharp interface models; by the
technique of matched asymptotic expansions such a diffuse interface model
can be related to sharp interface models. Our goal is to develop aphase
field model for alloy systems with multiple components and phases with the
following properties:
- consistency with the second law of thermodynamics, i.e. there is an en-
tropy inequality and the total entropy grows,
- the model is related to classical sharp interface models for systems with
multiple components and phases,
- only the knowledge of the bulk free energies of the occuring phases, the
surface free energies of the phase transitions and some mobility and dif-
fusion coefficients is necessary to derive the governing equations,
- earlier results obtained for phase field models are generalized, e.g.
• by Caginalp ([2] and [3]) for the limiting free boundary problem,
• by Penrose, Fife [14], Alt, Pawlow [1] and Wang, Sekerka, Wheeler,
Murray, Coriell, Braun, McFadden [17] for thermodynamically con-
sistent models,
• by Wheeler, Boettinger, McFadden [18] and Caginalp, Xie [4] for
binary alloys,
• by Steinbach, Pezolla, Nestler, Seeßelberg, Prieler, Schmitz, Rezende
[15] and Garcke, Nestler, Stoth [7] for systems with multiple phases,
• by Wheeler, Boettinger, McFadden [19] for eutectic systems, by Lo,
Karma, Plapp [10] for peritectic systems and by Nestler, Wheeler [11]
for both.
In the fOllOwing section we present the model in its general form. The
corresponding sharp interface model is described in the third section. In the
last section we will present some linearized models.

2 Description of the Model

The model is based on an entropy functional of the form

S(e,c,cjJ) = L
(s(e,c,cjJ) - (ea(cjJ, VcjJ) + ~W(cjJ))) dx. (1)

The unknown variables are the internal energy density e, the concentrations
ofthe N components Ci, 1 ~ i ~ N, and the phase field variables cjJ = (cjJa)~=l
144 H. Garcke, B. Nestler, B .. Stinner

(tjJo: gives the fraction of phase a). For the concentrations and the phase field
variables the following constraints have to be fulfilled:

(2)
i=l 0:=1
Le. cE E N and tjJ E E M where E K := {d E lRK : Ef=1 dk = I}.
The entropy density and the internal energy density are given in terms
of the free energy density J by the fundamental thermodynamic relations
s = - J,T and e = J + Ts respectively where T is the temperature. Knowing
the free energies r(T, c) ofthe pure phases we obtain the total free energy as
a suitable interpolation, Le. J(T, c, eo:) = r(T, c) where eo: is the a'th unit
coordinate vector. As an example we obtain the ideal solution free energy
density

M
J(T, c, </J) := ~ ~ ciLi
N (T Tt-Trt) h(</Jo:)
t

+ t, (kTt,(Ciln(Ci)) -cv T(1n(T) -1)) tjJo:.

The terms a and W are introduced to model surface contribution to the

entropYj possible choices are:

a(</J, \7tjJ) =L ?o:,ß ItjJo: \7tjJß - tjJß\7</J0:1 2 , (3)

a<ßma,ß
Wst(</J) =9 L
mo:,ßia,ß tjJ~tjJ~. (4)
a<ß
In [8], an obstacle potential was introduced which allows for a simple calibra-
tion between coefficients in the phase field model on the one hand and the
surface free energies and mobility coefficients of the sharp interface model on
the other hand.
The governing set of equations consists of energy balance, mass balance
and a diffuse Gibbs-Thomson equation:
8t e = -\7. Jo, (5)
8t Ci = -\7. Ji , 1~i ~ N, (6)
88
wc8ttjJo: = 8</Ja -.A, 1 ~ a ~ M, (7)

where w is a constant and .A is an appropriate Lagrange multiplier such that

the constraint (2) for the phase field variable is satisfied. The variation of the
entropy with respect to tjJ is
88
8tjJ = c (\7. a,x(tjJ, \7tjJ) - a,t/>(tjJ, \7tjJ)) - e1 w,t/>(tjJ) - ft/>(T,c,</J)
' T (8)
 A Multicomponent Phase-field Model 145

and the fluxes Jo, ... , JN for the eonserved quantities are given by the phe-
nomenologieal equations (see [9])
N
1
Jo = Loo(T,c,4J) '\1'f + "L..JLoj (T,c,4J) '\1 -J.t,
T3, (9)
j=1
N
1 "
Ji = LiO(T, c, 4J) '\1 T + L..J Lij(T, c, 4J) '\1 -J.t,
T3• (10)
j=1
Observe that there is the thermodynamic relation

where J.tj = i,cj is the chemieal potential of the eomponent j E {I, ... , N}.
By Onsager's postulate (see [12] and [13]) every thermodynamie flux is given
by a linear eombination of the thermodynamie forees whieh are here given
bY t"7 1 t"7 =.1&
V T' v
t"7 -I"N
T , .• " v T .
The matrix of mobility eoefficients L = (Lij)f,;=o is assumed to positiv
semi-definite and symmetrie (Onsager relations). For the first eonstraint of
(2) to be satisfied we assume
N
L Lij = 0, 0 ~ j ~ N. (11)
i=1

In [5] we derive an entropy inequality for this model, henee it is eonsistent

with the seeond law of thermodynamics, For closed systems we show that the
total entropy eannot deerease in time.

3 The Related Sharp Interface Model

By formally matched asymptotie expansions it is shown in [5] that the above
set of equations is related to the following sharp interface model:
In every phase q E {I, ... , M} there are the evolution equations for energy
and mass:

8t eq = -'\1. (Lgo(Tq'C q) '\1~ + tLgj(Tq,cq) '\1 ;~) , (12)

3=1

8t c1 = -'\1. (L1o(Tq,C q) '\1~ + "tL1j(Tq,Cq) '\1 ; } ) Vi; (13)

3=1
we obtain the terms with the upper index q by setting 4J = e q , the q'th
prinzipal eoordinate veetor, in the eorrespondings terms of the phase field
model.
146 H. Garcke, B. Nestler, B. Stinner

On a (smooth) boundary r between two phases a and ß the tempera-

ture and the generalized chemical potential differences are continuous, there
are jump conditions for the energy and the mass, and the Gibbs-Thomson
relation must be satisfied:
Ta. =Tß =:T, (14)
flf = flf =: fli Vi, (15)
[e]~ v = [Jo]~ . v, (16)
[Ci]~ v = [Ji]~ . v Vi, (17)
m v = '" K + [f]~ - ~i fldci]~ . (18)
a.,ß 1a.,ß T

Here, v = va.,ß is the unit normal pointing into ß, v is the speed of r in this
direction, K the m~@. curvature and

are the generalized chemical potential differences. 'Ya.,ß is the surface entropYj
as observed in [16] in the isotripie case and later more generally in [7] the
relation between the surface entropy and the phase field model is given by

(20)

where the infimum is taken over all Lipschitz continuous functions p connect-
ing the mimima of w which correspond to the phases adjacent to the interface,
i.e. p( -1) = ea. and p(l) = eß. A function p : IR ~ EM is a solution of the
(suitably reparametrized) Euler-Lagrange equation of (20) if and only if

0= _pM (8z (a,x (p, 8z p ® v))v + a,,p(p, 8z p ® v) + w,,p(P)) . (21)

Here, pM is the projection of lRM onto the tangential space TEM = {d' E
IRM : ~~1 dA: = O} of EM. The boundary conditions for p are

_( ) z-tqo _( ) z-t-oo
Pz ---t eß' pz -t ea.. (22)

The above Euler-Lagrange equation is exactly the equation to lowest order

obtained from the lfo-equation of the phase field system after expanding in
the (diffuse) phase transition, and the boundary conditions are the matching
conditions with the outer expansion in the adjacent bulk phases. Another

I:
representation of the surface entropy is

'Ya.,ß = (a(p, 8z p ® v) + w(P))dz. (23)

 A Multicomponent Phase-field Model 147

FinaHy, ma,ß is a mobility coeflicient which is also given in terms of that

ma,ß = i:
mimimizer p respectivly p:

18zpl2dz. (24)

4 Examples
It can easily be verified that the above model reduces to the Penrose-Fife
=
model [14] if there is only a pure substance (i.e. N 1) and if there are two
phases (i.e. M = 2). In this case the system for (T, cp) := (T, CP2 - cPt) consists
of the equations derived by Penrose and Fife.
Various simplifications of the model are possible by partially linearizing
the fuH set of equations. We will give an example for a binary alloy with three
phases (i.e. M = 3, N = 2). We set c = Cl, then C2 = 1 - c holds.
By a suitable choice of the diffusion coeflicients it is possible to linearize
the balance equations. Set
T T T
L oo := (e,e)2 Dj - + kT 2, L Ol = L lO := e,eD-j , L u := Dj -.
,ce ,ce ,ce

This yields the equations

8t e = V· (kVT + e,eDVc+ e,eD;'::VCP) , (25)

8t c = V . (DV c + D ;,:: V cp) . (26)

If e,e and Cv = e,T are constant then in a pure phase (i.e. V cP = 0) we

obtain the simple equations
ev8t T = kLlT (Fourier's law),
8t c = DLlc (Fick's law).
We remark that (26) becomes linear even across a phase transition (where
V cP i:- 0) if we choose J such that J,e</JI J,ee is constant.
FinaHy let us consider a case in which the internal energy does not depend
on c. More precisely we set

J(T, c, cp) := L.,; La

~ (T-Tf
C Ta + (1 - c) T-Tf)
TP CPa
a=l 1 2
+kT(cIn(c) + (1 - C) In(1 - C)) - evT(In(T) - 1).
In addition we take the gradient term a( cP, V cp) := ~ E a IV CPal 2 , we choose
w to be the standard multi weH potential, and let w = 1, L oo := T2, L lO =
LOl := 0, Lu := D c(1 - c). Approximating the nonlinear term ~ by

1
- R:l - 1 - (1)2
- (T - Tf)
T Tf T 0l
2 •
148 H. Garcke, B. Nestler, B. Stinner

the governing equations become

Ote = ßt(c'IIT-- L:Vl<cPa) = .6.T,

Ot C = kD .6.c+ V ·Dc(l- c)V (~(~a - ~a) LacPa) ,

1
cOtcPa = c.6.cPa - -gw,<P a (cP)

-Lac (~a - ~a) - La (~a) 2 (T - T a) -

2 A.

This set of equations can be interpreted as a generalization of the phase model

studied by Caginalp in [3] which can be obtained from the Penrose-Fife model
by a similar linearization as already done.

5 Further Generalizations
Observe that in this paper the internal energy contains no constributions by
surface terms. The surface entropy density of an interface is independent of
the temperature, hence its surface free energy density is linear in T so that
the internal energy density of the interface vanishes. In a forthcoming paper
we will study more general dependences of the surface free energy density on
the temperature.
Furthermore, the infiuence of the components will be taken into consider-
ation. This will be done via a dependence of the surface free energy density
on the generalized chemical potential differences j],i which are continuous and
hence weIl defined on an interface.

References
1. H. W. Alt and I. Pawlow: Models of non-isothermal phase transitions in
multicomponent systems. Part I: Theory, SFB256 University Bonn, Preprint
222 (1992)
2. G. Caginalp: Stefan and Hele-Shaw type models as asymptotic limits of the
phase-field equations, Phys. Rev. A 39, 5997 (1989)
3. G. Caginalp: An analysis of a phase field model of a free boundary, Arch.
Rational Mech. Anal., 92, 205-245 (1986)
4. G. Caginalp and W. Xie: Phase-field and sharp-interface alloy models, Phys.
Rev. E 48, 1897 (1993)
5. H. Garcke, B. Nestler and B. Stinner: A diffuse interface model for alloys
with multiple eomponents and phases, in preparation for submission to Phys.
Rev.Lett.
6. H. Garcke, B. Nestler and B. Stinner: ... , in preparation for submission to ...
7. H. Garcke, B. Nestler and B. Stoth: On anisotropie order parameter models
for multi-phase systems and their sharp interface limit, Physiea D 115, 87-
108 (1998)
 A Multieomponent Phase-field Model 149

8. H. Garcke, B. Nestler and B. Stoth: A multi phase field eoneept: numerieal

simulations of moving phase boundaries and multiple junetions, SIAM J.
Appl. Math. 60, No.1, 295-315 (1999)
9. J. S. Kirkaldy and D. J. Young: Diffusion in eondensed state, The Institute
of Metals, London (1987)
10. T. Lo, A. Karma, M. Plapp: Phase-field modeling of microstructural pat-
tern formation during direetional solidifieation of periteetie alloys without
morphologieal instability, Phys. Rev. E. 63, 031504 (2001)
11. B. Nestler and A. A. Wheeler: A multi phase field model of eutectie and
periteetie alloys: numerieal simulation of growth structures, Physiea D 138,
114 (2000)
12. L. Onsager: Reciproeal relations in irreversible proeesses I, Phys. Rev. 31,
405 (1931)
13. L. Onsager: Reciproeal relations in irreversible proeesses II, Phys. Rev. 38,
2265 (1931)
14. O. Penrose and P. C. Fife: Thermodynamieally eonsistent models of phase-
field type for the kineties of phase transitions, Physica D 43, 44 (1990)
15. I. Steinbach, F. Pezolla, B. Nestler, M. Seeßelberg, R. Prieler, G. Schmitz,
J. Rezende: Aphase field eoneept for multiphase systems, Physiea D 94,
135-147 (1996)
16. P. Sternberg: Vector-valued loeal minimizers of noneonvex variational prob-
lems, Rocky Mt. J. Math. 21 No. 2, 799-807 (1991) .
17. S-L. Wang, R. F. Sekerka, A. A. Wheeler, B. T. Murray, S. R. Coriell, R. J.
Braun, G. B. MeFadden: Thermodynamieally-consistent phase-field models
for solidifieation, Physiea D 69, 189-200 (1993)
18. A. A. Wheeler, W. J. Boettinger and G. B. MeFadden: Phase-field model for
isothermal phase transitions in binary alloys, Phys. Rev. A 45, 7424 (1992)
19. A. A. Wheeler, W. J. Boettinger and G. B. MeFadden: Phase-field model for
solidifieation of a eutectie alloy, Proe. R. Soe. Lond. A 452, 495-525 (1996)
Planar Solidification from Undercooled Me1t:
An Approximation of a Dilute Binary Alloy for
a Phase-field Model

Denis Danilov

Laboratory of Condensed Matter Physics, Physical Faculty, Udmurt State

University, 1 Universitetskaya Street, Izhevsk 426034, Russia
Email: [email protected]

Abstract. Planar solidification from an undercooled melt has been considered

using the phase-field model. The solute and the phase fields have been found in
the limit of small impurity concentration. These solutions in the limit of vanishing
velocity of the interface motion give the equilibrium partition coefficient and the
liquidus slope. Asymptotic expansions for the solute and for the phase fields and
the relation between the diffusive speed and the parameters of the phase field model
have been found at high growth velocity. A comparison with numerical calculations
is presented.

1 Introduction
Classical macroscopic models of solidification are based on the assumption
that the interface between regions of different phases have zero thickness [1].
Sharp-interface descriptions require the introduction of separately derived
nonequilibrium models for the behavior of the interfacial temperature and
of the solute concentrations. In contrast to the sharp-interface models, the
phase-field models [2,3] describe the bulk phases as weH as the interface, Le.
these models treat the system as a whole and eliminate the need to specify the
interfacial conditions separately. In phase-fields models, equilibrium behav-
ior is recovered at low growth velocities and nonequilibrium effects naturally
emerge at high growth velocities. From a computational point of view, the
advantage of the phase-field formulation is that the interface is not tracked
but is given implicitly by the value of the phase-field variable. However, the-
oretical analysis solutions were obtained by matching of separately derived
''inner'' (near the interface) and "outer" (far from the interface) solutions in an
intermediate region [4,5]. In the work [6], phase-fields and solute profiles have
been obtained without using the multiply-variable expansion, which leads to
''inner'' and "outer" solutions, but the profiles were obtained from a simplified
set of equations in which the interface kinetics was eliminated.
In this work an analytical solution of the phase-filed model for planar
solidification from an undercooled melt as an expansion in terms of the solute
concentration far from the interface is presented. This solution is valid for the
bulk phases as wen as for the interface and includes the interface kinetics. To
verify the expansion, numerical simulations of the problem were carried out.

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 Planar Solidification from Undercooled MeIt 151

2 The Model
Let us consider the motion of a planar solid-liquid interface during the solid-
ification of a binary alloy. The governing equations for the phase-field 4J(x, t)
and for the solute concentration c(x, t) are given by the following set of equa-
tions [2]

-8c =
8t 8x
8 8 8!
-M2 c(1-c)--
8x 8c'
(1)

which satisfy the conservation of solute and which have been derived under
the assumption that the total free-energy decreases monotonically in time.
The parameters MI, M 2 , and € are related to the growth kinetics, to the
diffusion coefficient, and to the surface energy respectively.
In the approach of ideal solution, the free-energy density !(4J, c, T) at a
temperature T is given by [2]
RT
!(4J,c, T) = C!B(4J, T) + (1- c)!A(4J,T) + -[clnc+
Vm
(1- c) ln(l-c)], (2)

where R is the universal gas constant, V m is the molar volume, which is

assumed to be constant. Free-energy densities of the pure materials A and B
are assumed to be of the form [2]

h(4J, T) = Wi Jr<l>
o p(p-1)
(1 )
P - "2 - ßi(T) dp = -fg(4J)+
W' W-ß'(T)
'~
.
p(4J) (3)

where

,I.)
9 ('I' = 4J4 - + '1',
2'1',1.3,1.2 P (,I.)
'I' = 4J2(3 - 24J) , Wi~i(T) = L;T
• :;Ti .
.L •
(4)

Here Wi is a constant, Li is the latent heat per unit volume, Ti is the melting
point of material i = A, B.
To investigate steady-state interface motion, the boundary conditions will
be taken in the form
8c
8x 13:-1--00 = 0, 4J13:-I--00 = 1, (5)

and the moving frame z will be adopted according to the equation

z=x- Vt, (6)

so that the interface (given by 4J = 1/2) corresponds to z = 0 at this frame.

In the moving frame, equations (1) may be written by using (2) as

2 82 4J ~ 84J _ [8fA (8!B _ 8 f A )] =0 (7)

€ 8z 2 + MI 8z 84J +C 84J 84J '
152 D. Danilov

M 2 RT 82 c V 8c M ~ [ (1- ) (8 1B _ 8 iA ) 8f/J] = 0. (8)

vm 8z 2 + 8z + 2 8z C C 8f/J 8f/J 8z
For further analysis we introduce the notation for the diffusion coefficient
D = M 2 RT /v m and integrate equation (8) with respect to z over an interval
[z, 00). Employing the boundary conditions (5) gives

8c (81B 81A ) 8<jJ

D 8z + V(c - coo ) + M2 c(1 - c) 8f/J - 8<jJ 8z = o. (9)

3 An Approximation of a Small Impurity

Concentration

Let us examine a dilute binary alloy with initial impurity concentration Coo «
1. For this purpose we shall consider the expansion in terms Coo for the phase-
field, the solute concentration, and the temperature

<jJ(z) = f/Jo(z) + coo f/Jl(Z) + 0(2

00 ), (10)
c(z) = Coo Cl(Z) + O(Coo ) , (11)
T = To + cooTt + 0(2 00 ), (12)

As a result, in the zeroth-order approximation, from (7) by using (3) we

obtain the equation for f/Jo(z)

(13)

The solution of (13), which satisfies the boundary conditions (5), describes
the phase field during the solidification of a pure material A and has the form

(14)

where lA is the interface thickness and the interface velocity V is related to

the temperature To by the relationship

(15)

which determines the interface kinetics for a pure material. The relationship
(15) corresponds to the model of collision limited growth [7] with a kinetic
coefficient

(16)
 Planar Solidification from Undercooled Melt 153

In the first-order approximation, the equations (7) and (9) give

(17)

OCI V of1F
OZ + D (Cl - 1) + Cl Tz = 0, (18)

where function f1F(z) is defined by expression

f1F(z) = ~;o [!B(<Po(z),To) - !A(<Po(z),To)]. (19)

Taking into account the boundary conditions (5), the solution of (18) is

CI(Z) = ~ [Zoo eV(z'-z)/D+LlF(z')-LlF(z)dz'. (20)

The point z = 0 corresponds to the solid-liquid interface, i. e. <p(0) =

= 1/2, and to the inflection of the phase field, therefore
<Po(O)

02<PI I = O. (21)
oz2 z=o
As for metallic systems MI rv 108 cm3 /J·s [2], then the relation V/MI « 1
will be satisfied. Taking into account the conditions (21) at z = 0, one can
find from (17) the relationship between Tl and Cl

( O!B _ O!A) I
o<p o<p
I
<Po, To
Tl = - 02!A CI!Z=O.
(22)

oTo<p <Po,To

Equations (11), (12), (16) and (22) give the expression for the temperature
V
T = TA - - - m(V)c!z=o, (23)
{t

where the function m(V) is defined by

154 D. Danilov

4 Stationary Interface
Now we consider the solute field due to the stationary solid-liquid interface,
at V = 0, which corresponds to the equilibrium state. Let us introduce the
definitions for the bulk concentrations in the liquid CL = c(z -t (0) and in
the solid Cs = c(z -t -(0). For a stationary interface, first it follows from
(16) that To = TA, and second the equation (18) gives the solute field
c(z) _ -..:1F(z)
-e , (25)
CL

which leads to the equilibrium partition coefficient

ke = -Cs = exp [AF( z -t -00)] = exp ( -

-Ll RT T- TB») .
VmLB(TA (26)
~ AB
This expression has the same form as in [6], where the equilibrium partition
coefficient has been obtained from a simplified set of equations in which the
effects of interface attachment kinetics are eliminated.
Using (4), from equation (24) at To = TA we obtain
LBTA
mo = m(O) = LA TB (TA - TB). (27)

The solute concentration at the interface z = 0, in accordance with (25), is

(28)

Thus equations (23), (27), and (28) give the temperature at the equilibrium
solid-liquid interface

(29)

which yields the following equation for equilibrium liquidus slope

= LB TA(T -T ) (_ Vm [WB - WA LB(TA- TB)]) (30)

me LATB A B exp RTA 64 + 2TB .

5 Large-Velocity Asymptotics
To investigate the behavior of the solute concentration under rapid solidifica-
tion conditions, we consider an asymptotic analysis in the limit V » 1. The
 Planar Solidification from Undercooled Melt 155

solution (20) can be written in the form

Cl(Z) = ~ 1 00
e-Vz'/D+.dF(z-z')-.dF(z)dz'. (31)

Applying the Laplace's theorem about asymptotic expansions [8] to this in-
tegral, we obtain the solute field with an accuracy to second order

Cl(Z) = c(z) = 1- D 8t:1F. (32)

COO V 8z
According to (32), a maximum value of solute is reached at z = O. Therefore
the segregation coefficient at large velocities is given by

k = Coo = 1 =1+ D 8t:1F /

clz=o 1 _ D 8t:1F / V 8z z=O
V 8z z=o
-1- ~ Vm (8 lB _ 8!A)/ (33)
- 41AV RT 8c/J 8c/J l/>=1/2·

For the sharp-interface model the nonequilibrium segregation coefIicient

has been derived by Aziz [9] and in the limit of dilute alloy is

k(a) = k e + VjVD (34)

1+ VjVD'
where diffusive speed VD is a characteristic trapping velo city. For VjVD »1
the segregation coefficient k(a) can be approximated by

k(a) = 1 - (1 _ k e ) VD . (35)
V
A comparison of the equations (33) and (35) gives

VD - D V m (8lB _ 8!A)/ (36)

- 41A(1 - ke) RT 8c/J 8c/J l/>=1/2,To

According to (36), the diffusive speed VD depends on the diffusion coefficient,

on the temperature, and on the difference of free energy densities of the pure
materials.

6 N umerical Calculations
To validate the obtained solution, we will compare it with numerical solutions
of the steady-state governing equations (7) and (8). The material parameters
used in the numerical calculations are given in Table 1. These parameters are
similar to parameters employed in [6] and correspond to the alloy Ni-Cu [2].
156 D. Danilov

Table 1. Thermophysical propenies used in calculations

Parameter Value Reference

LA 2350 Jjcm 2 [6]
LB 1725 Jjcm2 [6]
TA 1728 K [6]
TB 1358 K [6]
D 10- 5 cm 2 jfi. [6]
u 2.8 x 10- 5 Jjcm2 [6]
IA 6.48 x 10- 8 cm [6]
Vm 7.4 cm3 jmol [2]
p, 24 cmj{s·K) [10]

The value ofkinetic coefficient was set to J.' = 24 cmj(s·K) accordingto the
work [10] in which good agreements between the model for dendritic growth
and experimental data have been achieved. The far-field concentration was
set to Coo = 0.1. The parameter € is related to the surface energy u and to
the interface thickness lA by equation €2 = 6ulA [2,6]. The parameters WA
and WB were chosen to be equal and are given as WA = WB = W = 12ujlA
[6].
Employing (3) and (4), a finite-difference approximation of the equations
(7) and (8) on a uniform grid gives the following set of equations

c~+! - cf = ~ cf+! - cf-I

L1t M1 € 2L1x
D cf+! - 2c~ + cf-I M 2 q~+! - qtl
+ M1 €2 L1x2 + M1€2 2L1x ' (38)

~ = c.~(1- c.~) [(88t/>

q.
IB ) k _ (8fA) k] t/>~+! - t/>f-l
i 8t/> i 2L1x'
(39)

which determine the discrete functions of the solute and of the phase field
at a point z = iL1x on a (k + l)th step of iteration with a time step L1t. To
equalize the velocity V of the moving frame with the velo city of the interface,
we introduce the equation

(40)
 Planar Solidification from Undercooled Melt 157

where Va (T) is the shift velo city of the interface relative to the moving frame,
Tj is an accessory parameter, T is a time variable corresponding to the it-
eration process, and the interface coordinate Zj is given by the condition
cp(Zj) = 1/2. A finite-difference approximation of equation (40) gives

vk+ 1 _ Vk _ EM1 z~+l - zj

(41)
Llt Tj Llt

Results of numerical calculations of the equations (37)-(41) are shown in

the following figures. Fig. 1 shows a good agreement between the analytical
solution [equations (11), (20), (23), and (24)] and the numerical results in the
wide range of undercoolings.

500r--r--------.---------.---------~

400

..!!.!
E 300
0
;>
~
·0
~ 200
>

100
•

0
10 15 20 25
Undercooling dT, K

Fig. 1. Kinetic curve "interface velocity V - initial undercooling L1T". The solid
curve shows the approximated solution, data points denote the result of the numer-
ical calculations.

Fig. 2 shows the solute and the phase fields. At the interface, the solute
field is smooth as against sharp-interface models, which gives a jump in con-
centration. At low undercoolings, the analytical solution gives some excessive
value of solute concentration in the vicinities of the interface. The series of
numerical calculations given in Fig. 2 demonstrate the degeneration of the
solute to uniform field with increasing interface velocity and undercooling.
158 D. Danilov

Such behavior of the solute field is in accordance with the equation (32), in
which the deviation from the far-field concentration is proportional to I/V.

-5.10-7 0 5-10-7 7.5_10-7

0.1125 1

0.11 T=1687.5K
AT=9.9 K 0.75
0.1075 V=79.5cmls

0.105 0.5

0.1025
c(z) 0.25
0.1

0
0.1125 1

0.11 T=l680K

~ AT=17.4K
V=291.7 cmls
0.75
E
i
0.1075 0&-
B
C
Q)
Cl
r::::
0
0
0.105

0.1025
c(z)
0.5

0.25
"
Q)
gj
J::
c..
0.1

0
0.1125 1

0.11 T=1672.5K
AT=24.9K 0.75
0.1075 V=479.4 cmls

0.105 0.5

0.1025
0.25
c(z)
0.1

-5.10-7 -2.5.10-7 0 2.5-10-7 5-10-7 7.5-lg-7

Coordinate z, cm

Fig. 2. Solute profiles and the phase field for different values of the initial under-
cooling. The solid curve shows the approximated solution, data points denote the
result of numerical calculations.

7 Conclusions
Planar solidification from undercooled melt was investigated by numerical
and analytical methods in the limit of small impurity concentration. Com-
paring the obtained results with those of [6] we note that while equilibrium
partition coefficient (26) has the same form as in the work [6], the expression
(30) for the liquidus slope differs from the one given in the work [6]. However
 Planar Solidification from Undercooled Melt 159

the evaluated equilibrium liquidus slope has very elose values, m e = 306.2 K
from equation (30) and m e = 306.9 K from equation (A13) in [6]. The dif-
fusive speed has a weak dependence on the temperature, Eq. (36), and can
be accepted to be constant as weH as the diffusive speed obtained in [6]. The
numerical calculations presented here show a good agreement with the ana-
lytical solution that indicate a sufficient accuracy ofthe expansion (10)-(12).
The obtained solutions simultaneously are valid for the bulk phases and
the interface. Therefore they can be suitable for the purpose of numerical sim-
ulations: for instance to construct initial splitting of adaptive finite-element
grid and as a first approximation for the numerical solution of transcendental
equations.

References
1. Kurz W., Fisher D. J.: Fundamentals of Solidification, 3rd ed. Aedermanns-
dorf: Trans Tech Publication, (1992) 305.
2. Wheeler A. A., Boettinger W. J., McFadden G. B.: Phase-field model for
isothermal phase transitions in binary alloys. Phys. Rev. A 45 (1992) 7424-
7438.
3. Bi Z., Sekerka R. F.: Phase-field model of solidification of a binary alloy. Phys-
ica A 261 (1998) 95-106.
4. Wheeler A. A., Boettinger W. J., McFadden G. B.: Phase-field model of solute
trapping during solidification. Phys. Rev. E 47 (1993) 1893-1909.
5. EIder K.R., Grant M., Provatas N., Kosterlitz J. M.: Sharp interface limits of
phase-field models. Phys. Rev. E 64 (2001) 021604.
6. Ahmad N. A., Wheeler A. A., Boettinger W. J., McFadden G. B.: Solute trap-
ping and solute drag in a phase-filed model of rapid solidification. Phys. Rev. E
58 (1998) 3436-3450.
7. Turnbull D.: Metastable structures in metallurgy. Metall. Trans. A. 12 (1981)
695-708.
8. Olver F. W. J.: Asymptotics and Special Functions. New York: Academic Press,
1974.584 p.
9. Aziz M. J.: Model for solute redistribution during rapid solidification. J. Appl.
Phys. 53 (1982) 1158-1168.
10. Galenko P. K., Danilov D. A.: Model for free dendritic alloy growth under
interfacial and bulk phase nonequilibrium conditions. J. Crystal Growth 197
(1999) 992-1002.
Initial Transients in the Symmetrie Model for
Direetional Solidifieation

Raul Benitez and Laureano Ramfrez-Piscina

Departament de Fisica Aplicada. Universitat PoliMcnica de Catalunya. Av. Dr.

Maraii.6n 44, E-08028 Barcelona. Spain.

Abstract. We study the initial transient during the directional solidification of a

dilute mixture in the symmetrical constant-gap approximation. We perform phase-
field simulations of the transient recoil stages and compare the results with pre-
dictions obtained from the sharp-interface model. In particular, we focus in the
evolution of the front position and of the transient dispersion relation, obtaining
quantitative agreement between theory and simulations. Results are applied to the
destabilization of the front by ßuctuations.

1 Introduction
The selection of a dendritic pattern during the directional solidification of
a dilute binary alloy is a complex problem which depends on initial condi-
tions[l], and in particular on the first wavelengths that appear in the desta-
bilization of the planar front induced by fluctuations[2,3]. Recent work has
focused in the importance of internal fluctuations in solidification patterns
[4,5]. Quantitative agreement with experiments has only been obtained for
the solidification of pure substances, whereas in solutal growth the origin it-
self ofthe fluctuations is still an open problem [6,7]. The aim ofthis work is to
use phase-field techniques to study the effects of transients and fluctuations
in the selection problem of directional solidification. For this objective we
have performed a quantitative comparison between phase-field simulations
and predictions for the sharp-interface model.

2 The Sharp Interface Model

In a directional solidification experiment, a thermal gradient Gi is moved in
the z direction along the sampie at constant pulling velo city Vi. Provided the
sampie is thin, we take the system as 2D and describe the interface position in
the moving frame ofthe gradient by z = t(p, i), where p = H:.+yy. The solid
phase is located in the region where z < t, and the liquid where z > { We
will consider the particular case of symmetrie directional solidification, which
assumes the same solute diffusivity D in both phases. We introduce diffusion
length I = D IV and time T = D IV 2 to scale variables as r = r 11 and t = tlT.
We also introduce a diffusive field in each phase ui(r, t) = G~gqo (i=l solid,
i=2 liquid), where Ci is the solute concentration, Coo the concentration far
H. Emmerich et al. (eds.), Interface and Transport Dynamics
© Springer-Verlag Berlin Heidelberg 2003
 Transients in the Model for Directional Solidification 161

awayfrom the solid-liquid interface, and ..::1Co = [C2 -C1 ]int the eoneentration
jump across the interface. In the moving frame and in reduced variables, the
fields Ui evolve aceording to the diffusion equation

8 8
8t 8z
2
(- - - - V' )u·(r t)
" = 0. (1)

The diffusion fields Ui must satisfy some moving boundary eonditions at

the interface position whieh ean be written as

[Ul - U2]int = -1 (2)

u21int = 1- l~~ +;1\: (3)
ii· [VUl - VU2]int = -n z (1 + e) (4)

Equations (1-4) define the so ealled sharp-interface deseription of the sym-

metrie directional solidifieation problem. Equation (2) relates solute eoneen-
trations at both sides of the interface. We have assumed the additional ap-
proximation of having a eonstant coneentration jump across the interface.
This particular assumption is equivalent to suppose that the mixture has
parallel solid and liquid branches in the T(C) eoexistence diagram, which is
valid only for liquid erystals and alloys with a partition eoefficient dose to 1.
Eq. (3) is the Gibbs-Thompson equation (loeal equilibrium at the interface).
In this relation IT = ImLI..::1Co/G is a thermallength imposed by the tem-
perature gradient, and do is the eapillary length. Eq. (4) deseribes the solute
eonservation across the interface, and ii represents anormal unitary vector
pointing to the liquid. Using Green's funetion techniques [8], it is possible
to derive a dosed integral expression for Ui at each side of the interface. In-
troducing the notation p = (p, z, t), Ps = (p,~, t), the eoneentration at the
interface verifies

~Ui(PS) = (_)i ( f~\~~::> dr' . Gi (Ps, r', to) . Ui (r' , to)

- ft: dt' f dp' . [1 + f,(p', t')] . Ui(PS) . G(ps, Ps) (5)
- ftto dt' f dS'ii' . [Gi(Ps,Ps)V'Ui(PS) - Ui(p's)V'Gi(ps,ps')])

where Gi(P,p') is the retarded Green's funetion of the diffusion problem

, _ O(t - t') {_ (p - p')2 + (z - z' + t - t')2 }

G(p,p) - [41l"(t _ t')]! exp 4(t - t') . (6)

Summing Eq. (5) over i = 1,2, and using the moving boundary eonditions
Eqs. (2-4), an integro-difIerential equation can be obtained for the interface
162 R. Benftez, Laureano Ramirez-Piscina

position during the transient

1 jecp,to)
-l e(p,to) =- dr'G(ps;r',tO)ul(r',to)-
T -00

(OO
Jecp,to)
dr'G(ps; r', to)u2(r', to) + t dt' fd P'(l + ~(p', t))G(ps,p~)
Jto
+!(1 + er!(e(p, t) - e(p, to) + t)) (7)
2 20
e
where = 0 stands for the steady position of an unperturbed planar interface.
Note that Eq. (7) includes transients from the initial condition at to.
The next step is to perform a linear stability analysis of the problem to
obtain a transient dispersion relation describing the time evolution of a si-
nusoidal modulation with wavevector klo Within an adiabatie approximation
[2,3] we derive the growth ratio of the mode as

w(kl, t) = lT1 e• - (1 + e). 2+ q2(1 + e). - (ql + q2) [lT

1 do 2] ,
+ T(kl) (8)

where

qa = C-~)" (1 + Ö + JH1 +~)2 + (kl)2 + w(kl, t). (9)

Therefore the prediction of the transient dispersion relation consists of two

steps. Firstly the numerical resolution of Eq. (7) for a planar interface is per-
formed by a Newton-Raphson method, yielding the transient front position
e(t). This function is then introduced into Eq. (8) to obtain w(kl, t) in this
adiabatie approximation.

3 The Phase-field Model

In the last few years, phase-field models have become a quantitative technique
to numerically simulate solidification phenomena [9]. This method introduce a
continue variable (the phase-field) ,per, t) which takes different constant values
in the solid and liquid phases and which localizes the solid-liquid interface in
a diffuse region of thickness W. This field is coupled with the solute diffusion
field, and the parameters are chosen in such a way that in the limit e ==
W /l --+ 0 the interface dynamies recovers the behavior described by the sharp-
interface model. We will use a phase-field model introduced by Karma et. al.
[10] for symmetrie directional alloy solidification with constant concentration
gap. For isotropie solidification with no kinetie dynamies, the equations in
reduced variables take the form

~~"e3ßt,p = e2'\j2,p + (1- ,p2)(,p - efo7l(l - ,p2)(U + f;f,)) (10)

ßtu = '\j2U + ~ßt,p. (11)
 Transients in the Model for Directional Solidification 163

In this model, al = 5f, a2 = 0.6267 are integral constants obtained when

performing a thin-interface limit [9]. Note that the equation for the phase-
field evolution contains three parameters: two main control parameters com-
ing from the sharp-interface model (lTll and doll), and the model-specific
parameter c.

4 Results and Discussion

We will consider the transient from the rest, Le. for an initial condition at
to = 0 consisting of an equilibrium solid-liquid planar interface located at
e(t = 0) = lTIl. In this case C1(r,O) = Coo - LlCo (ul(r,O) = -1), and
C2 (r,0) = Coo (u2(r,0) = O).This corresponds to taking u(r, 0) = -1 as

°
initial condition for the phase-field model. Trying to mimic a real experiment,
we consider here that at t = the pulling velocity suddenly takes the final
value V (1 in scaled variables).
We perform numerical integration of the phase-field Eqs. (10,11) with
an explicit finite-difference scheme with Llx = 0.8 and Llt = 0.08 . We first
study the 1D dynamical evolution during the transient, comparing the results
with the sharp-interface predictions. Fig. 1 presents the front position for two
different values of the control parameter lTll = 12.5 and 2.5. For each case,
we compare simulations for three different values of c (c = 0.5, e = 0.25
and c = 0.125) with the front position obtained with direct resolution of
the integral equation (7). Convergence to the sharp-interface limit can be
observed as c decreases, and good agreement is found for a value of e = 0.125.
We now estimate the transient dispersion relation w(kl, t) from phase-field

,,~ a) ,.> b)

1020300405060 20 30
t

Fig.1. Evolution of the interface position during the transient for (a) IT /1 = 12.5
and (b) IT/l = 2.5, and convergence to the sharp-interface as c -+ o.

simulations and compare it with the sharp interface prediction of Eqs. (8,9).
To this end we simulate for each desired value of t a (planar) ID interface
evolving from to = 0 to t. At that moment we introduce a sinusoidal interface
perturbation with wavevector kl, and continue the simulation in 2D. The
164 R. Benftez, Laureano Ramfrez-Piscina

spectral analysis of the front allows us to locate the regime where the mode
evolution is linear. This is represented in Fig. 2a, where three definite regions
can be observed. From the linear region the value of the transient growth rate
w(kl , t) is calculated. Fig. 2b shows the growth rate at different times obtained
for two different modes kl = 1.25 and kl = 2.5 in the case of lTll = 12.5 and
doll = 0.06923. Quantitative agreement is observed for all times. In the last

O.7 S , - - - - - -- - - - - - ,
OS •
o.~b) ~
.().2S~ .

1-- firnt
kl=1.25
harmonie (tI-2.5)
1

21 28 29 30 31 32 33 34
t

Fig. 2. a) Localization of the linear growth regime of the mode kl = 1.25. b) Time
evolution of two different modes kl = 1.25 and kl = 2.5 for lT /1 = 12.5.

part of this work we have introduced internal fluctuations in the diffusion

equation of the phase-field by following the formaIism introduced in Ref.
[11]. In simulations the growth of a range of wavelengths can be observed
until a selected mode dominates. The spectral analysis of the front reveals a
good agreement during the linear regime between the growth of each mode
and the transient dispersion relation predieted by the sharp interface model.
In Fig. 3 we present the amplitude of one ofthe modes (kl = 2.5, ITII = 12.5)
averaged for 15 different noise realizations. The agreement is quite good in
the regime with positive w (Le. where an increasing Ak(t) is predicted).

5 Concluding Remarks

In this paper we have studied the initial stages of a directional solidifica-

tion experiment in the context of the symmetrie model. We have presented
predietions for the transient recoil of the sharp interface model and for the
transient dispersion relation by using an adiabatic approximation. We have
performed phase field simulations of this initial regime. Simulations quanti-
tatively agree with theoretical calculations. In partieular the analysis of the
evolution of single modes and of fronts with fluctuations, both from theory
and from simulations, should be of relevance in the dendritie selection prob-
lem. Work in progress [12] will extend these results to the one-side model and
to the Langevin formaIism for the modes of Ref. [2].
 'Iransients in the Model for Directional Solidifieation 165

0.5

~ -1
- phase-field with noise
..s -1.S - sharp-interface
-2

-2.5

Fig. 3. Time evolution of the mode kl = 2_5 (lTI' = 12.5) with the phase-field with
noise, and eomparison with the sharp-interface prediction.

We acknowledge financial support from Direcci6n General de Investigaci6n

Cientffica y Tecnica (Spain) (Project BFM2000-0624-C03-02) and Comis'-
sionat per a Universitats i Recerca (Spain) (Project 2001SGR 00221).

References
1. J_ A. Warren, J. S. Langer, Stability of dendritie spacings. Phys. Rev. E 42
(1990) 3518-3525.
2. J. A. Warren, J. S. Langer, Prediction of dendritie spacings in a direetional-
solidifieation experiment. Phys. Rev. E 47 (1993) 2701-2712.
3. B. Caroli, C. Caroli, 1. Ramirez-Piscina, Initial front transients in direetional
solidifieation of thin sampies of dilute alloys. J. Cryst. Growth 132 (1993) 377.
4. E. Brener, D. Temkin, Noise-indueed sidebranching in the three-dimensional
nonaxissymmetrie dendritie growth. Phys. Rev. E 51 (1995) 351-359.
5. A. Pocheau, M. Georgelin, M.: Validation eriterion for noise-indueed mechanism
of sidebranching in direetional solidifieation. Eur. Phys. J. B. 21 (2001) 229-240.
6. W. Losert, B. Q. Shi, H. Z. Cummins, Proe. Natl. Aead. Sei. USA 95 (1998)
439-442.
7. R. Gonzälez-Cinea, L. Ramirez-Piseina, J. Casademunt, A. Hernändez-
Machado, Noise-indueed sidebranching in a phase-field model for solidifieation.
Phys. Rev. E 63 (2001) 051602.
8. B. Caroli, C. Caroli, R. Roulet, On the linear stability of needle crystals: evo-
lution of a Zel'dovich loealized front deformation. J. Phys. Franee 48 (1987)
1423-1437.
9. A. Karma, W. J. Rappel, Quantitative phase-field modeling of dendritie growth
in two and three dimensions. Phys. Rev. E 57 (1998) 4323-4349.
10. W. Losert, D. A. Stillman, H. Z. Cummins, P. Kopczynski, W. J. Rappel, A.
Karma, Selection of doublet eellular patterns in directional solidifieation through
spatially periodie perturbations. Phys. Rev. E 58 (1998) 7492-7506.
11. A. Karma, J. W. Rappel, Phase-field model of dendritie sidebranching with
thermal noise. Phys. Rev. E 60 (1999) 3614-3625.
12. R. Benitez, L. Ramirez-Piscina, In preparation.
Dynamics of a Faceted N ematic-Smectic B
Front in Thin-Sample Directional Solidification

Tamas Börzsönyi1 ,2, Silvere Akamatsu 1 , and Gabriel Faivre1

1 Groupe de Physique des Solides, CNRS UMR 75-88, Universites Paris VI et

VII, Tour 23, 2 place Jussieu, 75251 Paris Cedex 05, France
2 Research Institute for Solid State Physics and Optics, Hungarian Academy of
Sciences, H-1525 Budapest, P.O.B.49, Hungary

Abstract. We study the dynamics of directional-solidification fronts presenting

facets in one direction. We focus on the phenomena occurring immediately above
the Mullins-Sekerka instability threshold when the facet is tilted with respect to the
growth direction. The experimental system we use is a nematic - smectic B front in
the planar configuration. We observe 1.) drifting shallow cells, which we explain by
means of a linear stability analysis including diffusive and kinetic anisotropies; 2.)
a new type of oscillatory solitary wave, called IIfaceton ll • For a numerical approach,
we use aphase field model with sharp capillary and kinetic anisotropies.

1 Introduction
Extensive studies have been devoted to the directional solidification (a thin
sampIe is pulled with a constant velo city V in a fixed temperature gradient
G) patterns of fully non-faceted systems, in which the surface tension 'Y and
the kinetic coefficient ß (defined as the ratio of the kinetic undercooling to the
growth velo city) are non-singular functions of the orientation of the interface
with respect to the crystal lattice. In contrast, little is yet known ab out
the directional-solidification dynamics of faceted crystals. Here we study the
dynamics of solidification fronts that are rough in all but one direction, and
have no forbidden directions. In such a system, when the facet is tilted with
respect to the isotherms, the dynamics of the front must obviously be that of
a non-faceted crystal as long as the deformation of the front remains smalI,
that is, below the cellular threshold Vc and in a small range of V above Vc •
Facets only appear at higher V when the deformation of the interface is large
[1,2].
The nematic-smectic B front of the liquid crystal CCH4 (MERCK) fulfils
such conditions [3,4]. A long-range order exists in the smectic B phase per-
pendicularly to the molecular layers, so that this phase actually is a lamellar
crystal. We use 12-JLm thick sampIes in the planar configuration (director par-
allel to the plane of the sampIe), so that the front -including facets, if any-
remains perpendicular to the sampIe plane. The observations are recorded
with a polarizing microscope equipped with a CCD camera. The video signal
is analyzed with digital image processing. Further details about our setup are
given elsewhere [5,6].

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 Dynamics of a Faceted Nematic-Smectic Growth Front 167

2 ResuIts
2.1 Experiments
The sequence of morphologies ohserved as a function of V for all (except 0°
and 90°) values of the tilt angle Bo of the facet is illustrated in Fig. 1. Drifting
shallow cells are observed from Vc (~ 2 /-Lms- 1 ) to ahout 8 /-Lms- 1 • In the
same range of V, solitary waves called "facetons" consisting essentially of a
drifting single facet are also observed. At higher V, arrays of narrow tilted
faceted fingers showing various types of oscillatory behaviors are ohserved.
We focus here on the morphologies observed near the threshold (drifting cells
and facetons ) .

Fig. 1. Left: Spatiotemporal diagram showing the trajectories of three oscillatory

facetons (on the left) and of shallow cells (on the right). V = 6.5/-Lms- l . Bo = 55°.
Right : Growth morphologies observed at the indicated values of V for Bo = 25°.
Ca) Planar front; (b) drifting shallow cells; (c, d) facetons; (e,f) arrays of faceted
fingers. G ~ 50 Kem- I .

The ohserved drifting cells certainly are Mullins-Sekerka cells. The fact
that they drift indicates that the system has a strong diffusive or kinetic
anisotropy [1,7]. The value of their drift velo city Vd depends on Bo , as it
should (see Fig. 2 below), but not on the cell amplitude. Facetons appear
when the amplitude of the cells is large enough for their sides to be tilted to
an angle larger than Bo. This does not occur spontaneously in the considered
range of V, but only under the effect of strong external perturbations (see
ref. [6]). Once created, facetons are stable. They propagate through the whole
sampie erasing the shallow cells that exist in front of them, and leaving behind
a planar front on which shallow cells slowly reappear. The total width of these
localized objects is of about 200/-Lm. Their drift direction is globally opposite
to that of the shallow cells, but, most frequently, their drift velocity is not
constant. A elose examination of the spatiotemporal diagrams shows that
facetons oscillate between two different quasi-stationary states. The times of
168 T. Börzsönyi, S. Akamatsu, G. Faivre

residence in these two states shows a great variability (the reason of this is
not known), but the drift velocity ofthe facet in these states is well defined.
In the so-called "blocked" state, the drift velocity is such that the normal
growth rate of the facet is much lower than V, as revealed by the fact that its
trajectory merges into its image. On the contrary, in the "unblocked" state,
the dynamical behavior of the facet is essentially the same as that of the
near-by rough regions of the front. Their drift velo city is then of an opposite
sign, and of a much smaller absolute value, than in the blocked state.
Knowing that the growth rate of a facet is governed by the motion of
molecular steps Bowing out of specific sources [8,9], these observations strong-
ly suggest that the transition between blocked and unblocked states corre-
sponds to the making, or the break, of a contact with step sourees. Observa-
tions give no direct information about the nature of the involved step sourees.
A possibility is that the steps Bow from the rough region of the front contigu-
ous to the colder edge of the facet when this region is of a concave shape, and
stop to Bow when it is of a convex shape. The blocked-to-unblocked transi-
tion would then be of a purely morphological nature (the lattice defects of
the crystal would play no essential part in it). If this is true, it should be
possible to reproduce the observed transition numerically with aphase field
model.

2.2 Linear Stability Analysis

We have performed numerically a linear stability analysis of the planar steady
state including simultaneously the anisotropy of ß, and those of the diffusion
coefficients in the nematic and in the smectic phase. The calculation is stan-
dard in its principle [2,7,10] but lengthy, and cannot be reproduced here.
The aim is to fit the calculated dependence of Vd on 00 to the measured
one, and hence to determine (or, at least, fix bounds to) the values of the
anisotropy parameters in our system. Let us review the involved parameters.
In the nematic (smectic) phase, the diffusion coefficient D" (Dff) parallel to
the director is different from the diffusion coefficient D 1.. (DfJ normal to the
director . The anisotropy of the interfacial properties is characterized by the
functions ,(0 - 00) and ß(O - 00 ), where 0 is the angle between the normal
to the interface and the growth axis z. The stability analysis is performed
at fixed value of 00, It involves the values of D zz = cos2 0oD" + sin 2 00D1..,
D zz = sin 2 00Dil +cos2 00D1.., D zz = sinOocosOo(DII-D1..), and those of,+,"
(surface stiffness), ß and ß' , where ," = tP,/d02 and ß' = dß/dO, at 0 = O.
The angular dependence of, is known from previous experiments [4], butits
actual value is not. We therefore took ,(Ore!), where Ore! is an orientation of
reference to be defined later, as an adjustable parameter. We have advanced
previously qualitative arguments showing that diffusion anisotropy cannot be
the main cause of the drift of the cells in our system [6]. In a first approach,
we therefore ignored diffusion anisotropy. '!'wo adjustable parameters (ß and
ß') are then left for a given value of 00 , These parameters of course cease to
 Dynamics of a Faceted Nematic-Smectic Growth Front 169

be independent when one imposes that the nmction ß(B) is smooth. Given
the value of ß(Brel)' the best-fit value of ß(B) is then uniquely determined.
Calculations performed with various values of the diffusion anisotropy in the
two bulk phases delivered values of ß(B) that were not substantially different
from those obtained by setting this anisotropy to zero.
The measured values of Vd(B o ) and the function ß(B) calculated numeri-
cally for the indicated values of the parameters are shown in Fig. 2. The fact
that Vd goes to zero at B° = 0° and 90° is in agreement with the symmetry of
the system. Another zero of Bo = 0° that is not prescribed by any apriori ar-
gument, seems to exist at Bo ~ 12°. We have chosen Brei = 12°, and imposed
that ß'(B)rel = O. Note that the calculated function ß(B) is an increasing
function of B for B > 12°, and that the ratio of its maximum to its minimum
value is large (about 3). The meaning of this as concerns the structure of
the interface on a molecular scale is not obvious. The range 0° < B < 12°,
in which ß(B) is a (slightly) decreasing nmction is perhaps the vicinal range,
but this is not certain. For consistency, it must be assumed that ß has a
discontinuity in B = 0 (facet devoid of steps), with ß(O) » ß(O±).

0.2

0.15
~ 0.3
C'!l. 0.1
;> 0.2
0.1 0.05
0
00
0

Fig.2. (a) Symbols: Measured values of Vd/V for V = 6.5p,ms- 1 • Curve: calculated
values of Vd/Vc for the values of ß given in (b). (b) Calculated ß(8) curve for
ß(12°) = 0.05, do (12°) = 0.001, K = 0.25. do = aoG/.tJ.T; (a o : Gibbs-Thomson
capillary coefficientj .tJ.To: thermal gap of the alloy) is the reduced capillary length.
ao is proportional to "( + "(" .

2.3 Phase-field Simulations

In this Section, we discuss the possibility of modeling the growth of our
faceted system with the help of the phase field model proposed in [11,12].
The equations we use are:

r(B)(öt - Vöz)</J = (1- </J2) [</J + a(</J2 - 1) (u + z~zo)] (1)

+V'(W(B)2V'</J) + Öx (rV'</J1 2W(B) ö~~)) + Öz (1V'</J1 2W(B) Ö~~))
170 T. Börzsönyi, S. Akamatsu, G. Faivre

(2)

For the angular dependence of W(O), we take W(O) = Wo (1 + (Wa -

l)lsin(O)1 where W a denotes the anisotropy (Wmax/Wmin) of the surface
tension. This nlllction is elose to the one previously determined experimen-
tally [4]. As for the angular dependence of 7(0), we assumed that interfa-
cial kinetics has very sharp minima (slow kinetics) corresponding to the
directions of faceting, as explained above. This has been realized by using
7(0) = 70(1 + cI(l -lsin(O)nc 2 )
The W(O) and 1/7(0) functions are shown in Fig. 3 for the parameters
W a = 3, Cl = 5· 106 and C2 = 1000. Numerical simulations have been
carried out by using these functions in order to trace the dynamics of the
front in the presence of a purely capillary anisotropy, and in the presence of
a purely kinetic anisotropy. The initial state was a stationary planar front
(at V = 0.06) which we perturbed by adding a spatially and temporally
uncorrelated noise with the amplitude of 0.1 (zero mean value) to <p near the
front (at -0.9 < <p< 0.9). Due to the perturbation the front became unstable
and converged to the pattern of stationary faceted fingers. By analyzing the
near-threshold behavior using a diffusional anisotropy in both phases and
by tuning 7(0) the results of the linear stability analysis have been verified.
The final morphologies are shown in Fig. 3. They bear some similarity with
facetons in their "unblocked" states. A morphology with facets in a "blocked il
state has not yet been obtained.

:lL~~~
.qoo -100 0 100 200

~~tI' : I.~
-200 -100 0
8(deg)
100
11
200

Fig. 3. The functions W(9) and 1/r(9) and the corresponding solution of stationary
faceted fingers at V = 0.06, D = 2, dx = 0.8, dt = 0.03 and lT = 250.

As a conelusion, we wish to underline some general characteristics of the

modeling of systems with large anisotropies using phase field models.
1. Since a~Jo) has a jump at 0 = 7r /2 in the vicinity of the singularity a
special treatment is needed. We found that by replacing the singularity by a
 Dynamics of a Faceted Nematic-Smectic Growth Front 171

paraholic function in the case when 7r /2 - E < () < 7r /2 + E using E = 7r /50,

gave a smoother dynamics of the front.
2. By introducing a very slow (l/T ---+ 0) kinetics in a narrow range of ()
(see Fig. 3) , we found that the propagation of the facet was not stationary
hut small steps are running along the facet, as it is seen in Fig. 4, where
three snapshots of the front position with respect to a perfect (rectilinear)
facet (which would he a horizontalline in this graph) and the normal velo city
of the facet is plotted.

50 100
x (pjxels)

Fig.4. a.) The deviation of the "Hat part" of the front from a perfect (rectilinear)
facet, b.) The normal velocity of the "facet" as a function of time.

References
1. F . Melo and P. Oswald, Phys. Rev. Lett., 64, 1381 (1990) .
2. F. Melo and P. Oswald, J. Phys. 11, I , 353 (1991). Also see F . Melo, These de
Doctorat, Universite Lyon-I (1991) .
3. A. Buka, T. T6th-Katona and L. Kramer, Phys. Rev. E 51, 571 (1995).
4. T. T6th-Katona, T. Börzsönyi, Z. Varadi, J. Szabon, A. Buka, R. Gonzälez-
Cinca, L. Ramirez-Piscina, J. Casademunt, and A. Hernandez-Machado, Phys.
Rev. E 54, 1574 (1996).
5. S. Akamatsu, G. Faivre and T . Ihle, Phys. Rev. E 51, 4751 (1995) .
6. T . Börzsönyi, S. Akamatsu and G. Faivre, Phys. Rev. E 65 11702 (2001).
7. S. R. Coriell and R. F. Sekerka J. Cryst. Growth 34, 157 (1976).
8. W. K. Burton, N. Cabrera and F. C. Frank, Phil. Roy. Soc. London 243, 299
(1951) .
9. A. A. Chernov, Contemporary Physics, 30251 (1989) .
10. A. Classen, C. Misbah, H. Müller-Krumbhaar and Y. Saito Phys. Rev. A 43,
6920 (1991).
11. W. Losert, D.A. Stillman, H.Z. Cummins, P. Kopczynski, W .-J. Rappel and A.
Karma. Phys. Rev. E 58, 7492 (1998) .
12. W .-J. Rappel and A. Karma. Phys. Rev. E 53, R3017 (1996) .
Last Stage Solidification of Alloys: A
Theoretical Study of Dendrite Arm and Grain
Coalescence

M. Rappaz, A. Jacot and w. J. Boettinger

- Laboratoire de Metallurgie Physique, Ecole Polytechnique Federale de

Lausanne, MX-G, CH-1015 Lausanne, Switzerland (M. Rappaz, A. Jacot)
- Calcom SA, PSE-EPFL, CH-1015 Lausanne, Switzerland (A. Jacot)
- Metallurgy Division, National Institute of Standards and Technology,
Gaithersburg MD 20899 USA (W.J. Boettinger)

This extended abstract is based on a paper submitted to Metallurgical

and Materials Transactions (January 2002)[1]

1 Introduction and Theoretical Aspects

Solidification of metallic alloys has been extensively studied (dendrite tip ki-
netics, microsegregation, coarsening of dendrite arms, etc.) but surprisingly,
in the absence of eutectic reactions (i.e., low concentration alloys), little is
known about the last stage of solidification when the primary phase regions
impinge. Yet, the details of the final stages of solidification of multigrain
and/or dendritic materials have significant impact on casting defects such as
hot tearing. Indeed, models of hot tearing [2,3] require input of a value of
fraction solid or temperature, sometimes called the coherency point, that de-
fines the situation where the two-phase liquid plus solid material is sufficiently
strong to withstand thermal contraction stresses without rupture.
For pure substances, coalescence or bridging of two planar solid-liquid
interfaces can be modeled by considering that the excess free energy of the
planar solid-liquid interfaces deviates from 2,sl, as the thickness of liquid
between the two interfaces approaches atomic dimensions. Here, we define
the interface energies as those for isolated interfaces, i.e., for interfaces of two
semi-infinite media. Fig. 1a shows a schematic representation of the excess
surface energy, ,(h), as a function of the liquid width, h, for a pure mate~
rial for the two most common cases encountered in solidification : (i) for two
solids originated from the same nucleus, there is no grain boundary energy
when the liquid film disappears (i.e., ,gb = 0) ; (ü) most grain boundaries are
"repulsive", i.e., ,gb > 2,sl [4,5]. The first case is referred to as "intragranular
coalescence", whereas the second one corresponds to "intergranular coales-
cence". The thickness-dependent interfacial energy of these two interfaces can
be described, somewhat arbitrarily as :

,(h) = 2,sl + (Tgb - 2,sl)exp ( ~h) , (1)

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 Last Stage Solidification of Alloys 173

where 8 is a measure of the interface thickness. For temperatures T different

from the melting temperature Tm, the excess free energy, G(h), of the two
interface structure per unit area is given by :

where Gl and G s are the bulk free energy per unit volume of the liquid
and solid, respectively, L is the heat of fusion per unit volume, LJ.Sj is the
entropy of fusion, LJ.T = (Tm - T) is the undercooling. Fig. 2 shows G(h)
for the intragranular and intergranular cases of Fig. 1, when the temperature
is above and below the melting point, respectively. In the first case, a local
maximum exists and the problem is equivalent to nueleating a liquid layer in
between two superheated grains : a thin liquid layer will not remain unless
h
its thickness becomes larger than the critical value h*,
above which the
whole domain melts. During solidification (Le., undercooled liquid present),
two such elose interfaces within a single grain are "attractive" and the liquid
film morphology is unstable (Le., a decrease of the liquid layer thickness
further decreases the excess interfacial energy). For intergranular coalescence
with ,gb > 2,sl, a minimum exists, Le., a stable liquid film of thickness h*
can remain in between two repulsive grains even below the melting point. It
is straightforward to show that such a situation occurs for an undercooling
LJ.T* given by :

JlT* _
.L.l -
,gb - 2,sl1
LJ.s j 8 exp
(-h*)
8 . (3)

Therefore, the film will disappear (h* = 0) at the coalescence or bridging

undercooling, LJ.n, given by :

LJ.n 'th LJ.I1 _ ,gb - 2,sl (4)

-8-' Wl b- LJ.Sj ,

where LJ.Fb is somewhat like a differential Gibbs-Thomson coefficient associ-

ated with the elosing grain boundary.
In aHoys, this undercooling will be elearly influenced by solute rejection
into, and diffusion parallel to, the liquid film located in between the two im-
pinging arms, as weH as by diffusion of solute in the two solids perpendicularly
to the interfaces (back-diffusion).
Two numerical approaches have been undertaken in the present study. For
ID situations, a sharp-interface front-tracking method [6], which incorporates
the ,(h) expression (1), has been employed. This model allows to consider the
influence of back-diffusion on coalescence. In order to handle 2D situations,
a multi-phase field approach based upon a Lagrangian multiplier method [7]
has been used. In this case, the free energy functional describing the system
is given by :
174 M. Rappaz, A. Jacot, W. J. Boettinger

Ii h

Fig. 1. Thickness dependence of the interfacial energy for intra- and inter-granular
situations.

IntragranuIar, AT< 0 Intergranular, AT > 0

2Ysl

Fig.2. Evolution of the Gibbs free energy for a pure substance, as a nmction of
the gap between two interfaces for two situations: intra- and inter-granular.

(5)

-L 0.5E;k !PigradPk - PkgradPi!2 - A(PI + P2 + P3 - 1)] dV,

i,k<i

where the p/s are the phase fields corresponding to the two grains (i = 1,2)
and to the liquid (i = 3), g is the bulk free energy, including double-weH
potential barriers Wik between two phases i and k, Eik is the other interfacial
energy contribution between these two phases, which multiplies the square
of the gradients of these phase fields, A is a Lagrangian multiplier associated
with the constraint PI + P2 + P3 = 1. More specificaHy, one has :

g(Pl' P2, P3, T) = LlSfLlTp~ [10 - 15P3 + 6p~] + L WikP;P~, (6)

i,k<i

Taking the functional derivatives of Eq. (5) with respect to all the variables,
PI, P2 and P3, the phase evolution equations are then given by :

1 BPi 8R
(7)
Mep Bt - 8Pi'
 Last Stage Solidification of Alloys 175

where the phase field mobility, Mq/j as weIl as Wik and Eik were chosen to
match the sharp interface model parameters (Le., 'Yi and kinetic coefficient
J-t), viz. :

W ZJ.. -- 3'Yij , EiJ' = ~J:

'YiJ'U, (8)
8
For binary alloys, the solute diffusion equation was appended to the sys-
tem of phase field equations [8]:

OC {[ (1 - k)c ]}
ot=V D VC-P3+k(1_P3)VP3 =V{D[VC-(Cl-Cs )V P3]},
(9)

where D is a diffusion coefficient averaged over the solid and liquid phases.

2 Results and Discussion

The time independent 1-D solutions of the phase field equations for a pure ma~
terial are first shown in Fig. 3 for an intergranular case b9b = OAJm- 2 , 'Ysl =
0.lJm- 2 ) and an undercooling L1T = 4K. Using the present form of the free
energy, the identity of the phase at any position is the one with the largest
value of Pi. Since Pl = P2 at the center of the grain boundary, a value of
P3 > 1/3 indicates that a thin liquid layer is present between the two grains.
As can be seen in Fig. 4, P3 = 0.5 at the center of the grain boundary,
thus indicating that a thin layer of liquid still exists at 4 K undercooling
far this repulsive case. It has been checked that the undercooling for which
P3 = 1/3 at the grain boundary follows essentially Eq. (4). In particular, the
multi-phase approach predicts that coalescence at a neutral boundary (i.e.,
"1gb = 2'Ysl) precisely occurs at zero undercooling.
Turning now to the dynamic evolution of two planar interfaces of an alloy,
Fig. 4 shows the evolutions of the average concentrations of the liquid and
solid phases, as predicted with the phase-field and sharp interface models.
These evolutions are superimposed on the phase diagram. The maximum liq-
uid phase field value at the centerline of the boundary calculated with the
multi-phase field model is also represented as a function of temperature, with
a scale that can be read on the top horizontal axis. All the parameters were
the same in both models, except for the parameter 8 which was adjusted in
the sharp-interface model in order to match the coalescence point calculated
with the phase field method. As can be seen, the overall behavior of the curve
calculated with the multi-phase field approach is qualitatively similar to the
prediction of the sharp interface model. At the start of solidification, one can
notice in both models a small deviation from the liquidus line which is due
to the use of the average liquid concentration instead of the interfacial value
(Le., solutal undercooling). As solidification proceeds, the amount of liquid
176 M. Rappaz, A. Jacot, W. J. Boettinger

CenterHne
,-___. . . . . . . . . . J
grainl
0,8

0,6

0,4

0,2

o ~------~~~~--------~
o Distanee
IOO/)

Fig.3. I-dimensional profile of the phase variables at i1T = 4K, as calculated

with the multi-phase field simulation for a pure system. 19 b = OAJm- 2 ,lsl =
O.lJm- 2 ,i1s/ = l06Jm- 3 K-l,Ö = 2nm.

decreases and astate of complete mixing of solute in this phase is reached.

Therefore, one follows the liquidus line until the two interfaces start to inter-
act. In the multi-phase field approach, the start of coalescence can be clearly
identified by a very rapid decrease of the liquid phase field at the center-
line of the grain boundary. Surprisingly, during this initial rapid decrease,
the average concentration in the liquid slightly increases, even though the
two interfaces are repulsive bgb = 3's t). Then, the liquid phase field in the
centerline decreases at a much slower rate and the average composition of
the liquid follows the trend of the curve calculated with the sharp interface
model. The coalescence point (4)3 = 1/3 at the grain boundary centerline)
calculated with the multi-phase field approach was used to adjust the value 8
of the sharp interface approach and consequently the coalescence line (dashed
line parallel to and 200 K below the liquidus in Fig. 4).
Finally, Fig. 5 shows 2D results obtained with the multi-phase field model
for a square domain 10 x 1OJ,tm2 • As initial conditions, a seed of grain 1 was
placed in the upper right corner and the remaining part of the domain was
set as liquid. The initial fraction of grain 2 was set to zero everywhere. The
concentrations of the seed and of the liquid were set to the equilibrium values
given by the phase diagram for a very small undercooling (i.e, kco and Co,
respectively). The boundary condition for the concentration on the limits
of the domain was set to that of a closed system. For the liquid phase, a
similar condition was applied on all the boundaries. For the solid phase, the
following symmetrie conditions were applied : grad4>l . n = -grad4>2 . n.
Such a boundary condition, which is equivalent to that of a closed system for
 Last Stage Solidification of Alloys 177

Ps = (PI + P2), allows to simulate the growth of two other mirror dendrite
arms of grain 2 in a left and bottom symmetrie square domain. The domain
was cooled with a constant heat extraction rate corresponding to a relatively
low cooling rate of -lOK/8 (higher rates lead to dendritic morphologies).
The domain was enmeshed with fairly coarse volume elements, except along
the left and bottom boundaries where coalescence is expected (mesh size in
these regions on the order of 8).

$, (x-O)
0.4 0.5 0.6 0.7 0.8 0.9

960

920

880
g

I
840

800
E-<

760

720

680

0 0.1 0.2 0.3 0.4

coru:entration [-]

Fig.4. Solidification paths of a repulsive grain boundary (,gb = 3,sz) predicted

with the sharp interface and multi-phase field models for the same set of param-
eters, except for the value of Ö (sharp interface: Ö = 4.5 1O-9 m , phase field:
o = 2.0 10-9 m). The liquid phase field at the centerline of the grain boundary
calculated with the phase field model is also shown.

In Fig. 5 (left) , the shape of grain 1 is shown at 3 different times (or

temperatures) for an attractive case with "(gb = ,sl' After an initial circular
growth stage (not shown here), the solute layers of both grains interact on
the left and bottom boundaries and the grain becomes more square-shape,
with an arc of a circle at the quadruple junction (top left, T = 922.4K). The
bottom interface is still fully liquid (P3 = 1), but the first instabilities of the
liquid film, induced in the present case by the noise of the numerieal calcula-
tion, can be observed. Just 0.758 later (T = 877.55K), the first bridges can be
visualized directly on the grain pietures (middle left of Fig. 5) and also in Fig.
6, where the liquid phase field profile along the bottom grain boundary has
been plotted (continuous curves). In this attractive case, the grain boundary
quiekly closes after these first initial bridges and at t = 498, T = 877.55K
178 M. Rappaz, A. Jacot, W. J. Boettinger

(bottom left ofFig. 5), most ofthe boundary is solid. As can be seen in Fig. 6,
a sharp transition from P3 < 1/3 to P3 > 1/3 can be observed in the last pro-
file. The remaining liquid region has changed its morphology, from a concave
morphology in the two upper pictures, to a convex droplet at the bottom. In
this last case, one can notice the formation of a well-defined tripie point with
a dihedral angle of 120 deg. corresponding to "tgb = "tsl. The multi-phase field
method is definitely capable of reproducing the observed coalescence of den-
drite arms within a grain (change from concave to convex pockets of liquid),
as weIl as the equilibrium condition at tripie junctions.

Fig.5. Maps of the liquid phase field, P3 (white region), as calculated with the
multi-phase field model for two situations: attractive grains ('Ygb = 'YsI, left) and
repulsive grains (-ygb = 3'Ys!, right). The time and temperature at which these maps
and profiles are shown during eooling are also indieated.
 Last Stage Solidification of Alloys 179

More interesting is the case shown in Fig. 5 (right) for repulsive grain
's
boundaries with 19b = 3 l. Solidification starts exactly in the same way
and at the corresponding time t = 438, the temperature (T = 922A3K)
and the shape of the grain are essentially the same. However, one can notice
that the initial attractive behavior of the interfaces predicted by the multi-
phase field approach when ~3 > 004 - 0.5 is indeed reproduced here. The
instability is already visible in the top figure, but does not develop into bridges
(~3 < 1/3). Indeed, after this initial attractive stage is over, the liquid phase
field along the bottom boundary remains between 004 and 0.5 (central picture
in Fig. 5, right) , as can be checked in Fig. 6 (dashed curves). This means that
the grain boundary is still wet. At that time, the shape of the solid-liquid
interface has not a constant-sign curvature near the bottom left corner : the
remaining liquid droplet has the shape of a cross. The convex shape of the
liquid droplet occurs near the grain boundaries, in order to ensure continuity
of the concentration field, since Cl at the grain boundary is larger in the initial
attractive stage. At t = 498 (T = 877.75K), the remaining pocket of liquid
has about the same volume as in the attractive case of Fig. 5 (left), but has
the surprising shape of a diamond. At that time, Fig. 6 indicates that the
boundary has coalesced on a fairly long distance (~3 < 1/3 for nearly 80%
of the boundary), but there is no real triple junction in this case. Instead,
the slope of d~3/dx for ~3 = 1/3 along the boundary is greatly reduced
and about 7 - 8% of the boundary is characterized by 1/3 < ~3 < 1. In
other words, the remaining intergranular liquid has a diamond shape, with
very thin wedges of liquid (smaller than 8) extending from the corners of the
diamond along the boundaries.

3 Conclusion

The last stage solidification of a binary alloy, which does not exhibit eutectic
solidification, has been investigated from a fundamental point of view. Using
first a sharp interface model, it has been shown that dendrite arms belonging
to the same grain ("tgb = 0) are attracted as the thickness of the interden-
dritic liquid film is on the order of the interaction distance, 8. This locally
increases the liquid concentration above the liquidus line. Neutral liquid-
solid interfaces, characterized by "tgb = 2"tsl, end their solidification on the

,s
liquidus line. Finally, for most grain boundaries, "tgb is thought to be larger
than 2 1 (repulsive case) and accordingly a coalescence line with undercool-
ing Lln = Lln8- 1 , where Lln = (,gb - 2,sl )Ll8f\ can be drawn below the
liquidus. Solidification ends when the combination of temperature and liquid
concentration of the thin liquid film between the grains reach this line. The
trajectory of the concentration-temperature path on the phase diagram is
strongly influenced by back-diffusion [1]. A multi-phase field approach to the
problem with a solution method based on a Lagrange multiplier technique has
also been developed. For pure substances, the equilibrium condition where a
180 M. Rappaz, A. Jacot, W. J. Boettinger

1 r
~
I
I
I
I
~ I
....!... 0,8 I
T=918.9

j
I
I
I
I
I
0,6 I
I
~
1\
tfI'-_ . . .I \,-----
..
I

1 0,4 \
\
' r--- -- -'
T= 914.8
......
t:r
--- " ....... ._--. -- --------._-_._---
~ 0,33 ---~----
.~-------------------
0,2 T= 877.75

T= 877.55

°°
I I I I

2 4 6 8 10
Distance [pm]

Fig. 6. Liquid phase field profile along the bottom grain boundary of Fig. 5 for the
attractive case (-ygb = 'Ys/, continuous lines) and repulsive case (-ygb = 3'Ys/, dashed
lines). The two curves for each case with the indicated temperatures correspond to
the middle and bottom pictures of Fig. 5.

thin liquid layer remains for repulsive interfaces below the melting point is
retrieved. For the dynamic simulation of alloys, the overall features of attrac-
tive and repulsive interfaces are reproduced, but some details require further
examination. This work is a first step in understanding the coalescence be-
havior of dendritic alloys during the last stage solidification and ultimately
modeling hot-cracking tendency.

4 Acknowledgements
The authors would like to thank the Office Federal de l'Education et de la
Science, Bern, Switzerland, for the partial financial support of this research
within the framework of the Brite-Euram project VIRCAST (OFES contract
# 99-00644-2)

References
1. M. Rappaz, A. Jacot, W. J. Boettinger, Last Stage Solidification of Alloys : a
Theoretical Study of Dendrite Arm and Grain Coalescence, submitted to Met.
Mater. Trans. (2002).
 Last Stage Solidification of Alloys 181

2. T. W. Clyne and G. J. Davies, Br. Foundryman, 68 (1975) 238.

3. M. Rappaz, J. -M. Drezet and M. Gremaud, Metall. Matls. Trans. A, 30 (1999)
449-455.
4. N. Eustathopoulos, Int. Met. Rev, 28 (1983) 189-210.
5. D. A. Porter and K. E. Easterling, Phase Transformations in Metals and Alloys,
1981 Van Nostrand Reinhold, UK, p. 122.
6. H. Combeau, A. Mo, J.-M. Drezet and M. Rappaz, Met. Mater. Trans., 27A
(1996) 2314-27.
7. B. Nestler and A. A. Wheeler, Physica D, 138 (2000) 114.
8. J. Tiaden, B. Nestler, H. J. Diepers, I. Steinbach, Physica D, 115 (1998) 73-86.
Phase-field Modeling of Eutectic Solidification:
From Oscillations to Invasion

Roger Folch* and Mathis Plapp

Laboratoire de Physique de la Matiere Condensee, CNRSjEcole Polytechnique,

91128 Palaiseau, France

Abstract. We develop a phase-field model of eutectic growth that uses three phase
fields, admits strictly binary interfaces as stable solutions, and has a smooth free
energy functional. We use this model to simulate oscillatory limit cyc1es in two-
dimensional lamellar growth, and find a continuous evolution from low-amplitude
oscillations to successive invasions of one solid phase by the other when the lamellar
spacing is varied.

1 Introduction
Solidification is both a fascinating example of pattern formation out of equi-
Iibrium and a phenomenon of practical importance in metallurgy. The most
common solidification microstructures found in industrial alloys are dendrites
and eutectic composite structures (rods or lamellae). Eutectic growth occurs
when two solid phases of different composition can solidify from the same
melt. The interplay between the redistribution of chemical components in
front of the moving phase boundary and the effects of capiIlarity along the
curved interfaces gives rise to a wealth of different patterns and nonlinear phe-
nomena such as bifurcations, limit cycles, solitary waves, and even chaotic
states [1].
Phase-field modeling has become the method of choice for simulating so-
lidification fronts. Its main advantage is that it avoids explicit tracking of
the solidification front by introducing ''phase fields" that vary continuously
between constant values corresponding to the bulk phases, thus replacing
the sharp fronts by diffuse interfaces with a finite thickness. The connection
to the traditional free-boundary formulation of soIidification is established
by the technique of matched asymptotic expansions around the equilibrium
front in the limit where the thickness of the diffuse interfaces is much smaller
than all other relevant length scales. Recently, the computational efficiency
of phase-field models for single-phase solidification of pure substances and
binary alloys has been drastically enhanced by pushing this perturbation
analysis to second order [2]. This, combined with a random walk algorithm
for an efficient simulation of the diffusion equation, has made it possible to
simulate quantitatively the dendritic growth of a pure substance in three
dimensions for experimentally relevant parameters [3].
* [email protected]
H. Emmerich et al. (eds.), Interface and Transport Dynamics
© Springer-Verlag Berlin Heidelberg 2003
 Phase-field Modeling of Eutectic Solidification 183

Therefore, it would be highly desirable to extend this second-order asymp-

totic analysis to phase-field models for multi-phase solidification; we will
specifically address the case of eutectic growth, where two solid phases de-
noted by a and ß grow from the liquid. The first phase-field models for
eutectic growth used the standard phase field to distinguish between solid
and liquid, and a concentration field [4,5] or a second phase field [6] to dis~
tinguish between the two solids. However, a solid-liquid interface then in-
volves variations of the two variables and the asymptotic analysis becomes
extremely cumbersome. The more recent multi-phase-field approach [7] as-
signs one phase field to each phase and interprets these fields as local volume
fractions. Across an interface, in principle only the volume fractions of the
two bulk phases limiting the interface need to vary, and one of them can
be eliminated in terms of the other. Hence, this method allows for equilib-
rium interface solutions depending on a single variable that are equivalent to
the usual ''binary'' interfaces of the standard phase-field model. However, to
achieve this the model has to be carefully designed, and so far strictly binary
interfaces have only been obtained by using a singular free energy (double
obstacle potential) [8], which might complicate the asymptotic analysis.
We present a phase-field model for eutectic growth that uses three phase
fields, yields exactly binary interfaces at equilibrium, and has a smooth free
energy functional. Furthermore, the free energy is designed to keep the inter-
faces binary to first order in the usual non-equilibrium asymptotic expansion,
which makes it a promising starting point for a second-order analysis.
To illustrate the capabilities of the model in its present form, we simulate
lamellar eutectic growth of a simple model alloy with a symmetrie phase dia-
gram. We show that small-amplitude oscillations of the lamellar width, "giant
oscillations", and successive invasions of one solid phase by the other, each
of which is seen in experiments, all belong to the same branch of oscillatory
limit cycles that is parametrized by the lamellar spacing.

2 Model
We use three phase fields Pi E [0,1] representing the local volume fractions
of each phase (a, ß or liquid), and thus L:iPi = 1. Their dynamics is derived
from a free energy functional F,
8Pi 18F
(1)
8t --:;:'8Pi'
where T is a relaxation time and the variational derivative has to take into
account the constraint Ei Pi = 1. Furthermore, the concentration field c
obeys the conserved dynamies
(2)
where J.t == 8F/ 8c is the chemical potential and M is a phase-dependent
mobility.
184 R. Folch, M. Plapp

The free energy functional :F = Iv IdV is the volume integral of a free

energy density I that is conveniently decomposed into I = I grad + hw +
le, where hw is a tripie well potential with minima on each pure phase
- the equivalent of the double weH potential of the standard phase-field
model, Igrad contains the gradients of the phase fields and ensures a finite
interface thickness and surface tension, and le couples the phase fields and
the concentration and drives the system out of equilibrium.
Due to the constraint Li Pi = 1, the free energy density and the interface
solutions can be conveniently represented on a Gibbs simplex, that is an
equUateral triangle where the distance to each side of the triangle from a
given point represents the value of one of the phase fields. Thus, each vertex
corresponds to a pure phase, and each side to a purely binary interface. Our
goal is to obtain strictly binary interfaces, i.e., the equilibrium solutions of
(1) for an interface connecting two phases should exactly run along the edges
of the simplex. One way to achieve this is to choose

(3)

and to design the function fTw in such a way that the three minima are con-
nected by "saddles" that run along the edges and that have vanishing deriva-
tives in the direction normal to the edges. Such functions can be constructed
using the geometry of the Gibbs simplex, as will be detailed elsewhere. The
simplest choice,

(4)

is plotted in Fig. la and reduces to the standard fourth-degree double well

potential of the binary phase-field model on the edges. With these choices,
the free energy is completely symmetrie with respect to the interchange of
any two phases, and as a consequence all three surface tensions are equal.
However, the general case of unequal surface tensions can also be treated by
adding other terms in the free energy density that modify the height of the
saddle points.
Out of equilibrium, the tripie well is ''tilted'' by the term le in order to
favor one phase over the others. Since we want to keep the interfaces binary,
this term needs to be carefully constructed. The elementary building blocks
are tUt functions gi that satisfy gi = 1 for Pi = 1, gi = 0 for Pi = 1, j =I- i,
and that have vanishing derivatives in the direction normal to the edges of
the simplex. One possible choice is

gi = 1- [(1 + Pi - pi)2(2 + Pi - Pi)2 + (1 + Pie - pi)2(2 + Pi - PIe)2

+ (Pie - Pi)2 (3 - Iple - Pi 1)2] /12 . (5)
 Phase-field Modeling of Eutectic Solidification 185

a\:71
I
I I I
a, 1 I I

ß (a) ~)
Fig.1. Triple-well potential /Tw (a) and elementary tilt function -gi (b) drawn
as "landscapes" over the Gibbs simplex.

In Fig. Ib, we plot -gi that lowers the weIl for the liquid with respect to the
two others. With the help of these functions,

(6)

interpolates between the concentration-dependent free energies fi (c -

A i )2/2+Bi of each phase i. Here, c = (C -CE )/L1C is a scaled concentration
variable, where L1C = Cß - Ca is the width of the eutectic plateau, and Ca,
Cß, and CE are the equilibrium compositions of the two solids and the liquid
at the eutectic temperature. An arbitrary phase diagram can be constructed
by performing the weIl-lrnown double tangent construction, which yields c!j =
Ai + (B j - Bi)/(Aj - Ai) for the equilibrium concentration c!j of phase i
coexisting with phase j, and by choosing appropriate Ai (T), Bi (T), where
T is the temperature. Note that by the same procedure, a peritectic phase
diagram can be constructed [9]. .
FinaIly, we need to specify the mobility function in (2). Since [J2 fda~ ==
1, the choice M(pa'Pß,Pl) = DPl yields a constant diffusivity D in the liquid,
and no diffusion in the solid (one-sided model).

3 Simulations
In thin-sample directional solidification of eutectic alloys with global com-
position in the eutectic range, the basic pattern is a periodic array of al-
186 R. Folch, M. Plapp

ternating lamellae of each solid phase, growing perpendicular to the large-

scale soIidification front and parallel to the imposed temperature gradient.
This basic state exhibits various instabilities that have been carefully stud-
ied both experimentally in the transparent organic alloy CBr4-C2CI6 [1] and
by numerical simulations using the boundary integral method [10]. For fixed
experimental parameters, the stability of the basic state is controlled by the
width of one lamella pair, or lamellar spacing. For sampie compositions elose
to the eutectic point, the first instability that is encountered for increasing
spacing is a period-preserving oscillatory instability, that is, the lamellae start
to oscillate, with all the lamellae of the same phase oscillating in phase, and
the global spacing left unchanged. Beyond the onset of this instability, stable
limit cycles are found, with an oscillation amplitude that increases with spac-
ing. In the boundary integral simulations, this branch of solutions terminates
when the amplitude becomes too large because the thinner lamellae pinch off.
In contrast, in experiments "giant oscillations" with very large amplitudes can
be obtained [11]. Furthermore, when the experiments are started from a sin-
gle solid phase growing into the liquid with a few widely spaced nuclei of
the other solid phase on the solidification front, the nuclei grow and spread
along the front forming ''invasion tongues" [12]. When two tongues growing
in opposite directions collide, they do not coalesce; instead, a narrow channel
of the other phase remains and the whole process starts over with the role
of the two phases reversed. The interface dynamies is hence still oscillatory,
even if the patterns look completely different. In this regime, the boundary
integral method becomes inapplicable because it uses the quasistationary ap-
proximation of the diffusion equation that is valid only if the interface motion
is slow compared to the diffusive solute redistribution. Here, we investigate
oscillatory limit cycles with our phase-field model, that does not have this
restriction.
The sampie is solidified with pulling speed V in a constant temperature
gradient G directed along the z direction. For simplicity, we use asymmetrie
phase diagram that has constant concentration jumps across the interfaces
(parallel liquidus and solidus lines). This is implemented by choosing A a =
-0.5, Aß = 0.5, Al = 0, Ba = Bß = 0 and BI = -(z - Vt)/lT, where IT =
(mLlC) / G is the thermallength, with m being the liquidus slope (equal for
both phases). The other relevant physicallength scales are the diffusion length
ID = D/V and the capillary length da = r/(mLlC), where r is the Gibbs-
Thomson constant. In terms of the model parameters, do = (2.J2/3)W. All
simulations discussed here were performed with lT/do = 530 and ID/da = 212
and at exactly eutectic composition (that is, c = 0 far ahead of the front).
The equations are integrated using a standard explicit finite-difference
scheme on a regular grid of spacing 0.8 W for which we have checked that
the discretization has converged. Since the diffusivity vanishes in the solid,
the diffusion equation only needs to be solved in interfacial and liquid regions.
Moreover, far ahead of the solidification front, it is solved using a variant of
the random walker algorithm of Ref. [3] that will be detailed elsewhere. Two
 Phase-field Modeling of Eutectic Solidification 187

types of initial conditions are used. To obtain lamellar steady states, simu-
lations are started from two flat lamellae in contact with the liquid. If the
lamellar state is unstable, the numerical noise generated by the random walk-
ers triggers the instability. Alternatively, simulations are started from a flat
single solid phase at equilibrium with the liquid. Upon pulling, the interface
recoils towards colder temperatures as a diffusion layer of the rejected com-
ponent is gradually formed. When the undercooling reaches a predetermined
value, nucleation is mimicked by placing a large enough semicircular nucleus
of the other phase on top of the interface.
We restrict our attention here to period-preserving oscillations without
lateral drift. Therefore, it is sufficient to simulate two adjacent half lamellae
or half a nucleus of one solid phase on top of the other with reflecting (no-
flux) boundary conditions on the sides of the simulation box that are parallel
to the lamellae. The lamellar spacing A is varied by changing the lateral size
of the simulation box. In the figures below, for clarity we have reconstructed
a whole lamella pair by reflecting the system with respect to one side of the
box.
We start by constructing the branch of stable lamellar steady-state so-
lutions and find the minimal undercooling spacing at Amin ~ 85do, in good
agreement with the prediction of the Jackson-Hunt theory [13], Amin ~ 79do.
Next, we start from a single nucleus and monitor the dynamies for increasing
spacings. We find a bifurcation toward oscillatory limit cycles for A close to
2Amin. As shown in Fig. 2a, for A = 2Amin the oscillation already has an
amplitude of about a quarter of the lamellar spacing, but its shape is still
close to a sinusoidal wave. At A = 4A min (Fig.2 2b), the oscillation is highly
asymmetrie, and has an amplitude close to halfthe lamellar spacing, Le., only
a thin channel of the other solid phase is left. Finally, at 8.48Amin (Fig. 2c),
three different regimes become apparent: (i) initially, the nucleus grows slowly
and spreads along the interface to form an invading finger with a well-defined
shape, (ii) the finger speeds up with an approximately constant acceleration,
leaving a straight border with the other solid phase behind, and (iii) when it
approaches a finger growing in the opposite direction (in our simulations, its
mirror image generated by the reflecting boundary conditions), it ''feeis'' the
diffusion field generated by the other and slows down very rapidly, allowing
for the other phase to emerge through a narrow channelleft between the two
fingers and thus restart the process with the role of the two solids reversed.
Remarkably, our largest simulations reproduce many characteristie fea-
tures of the "successive invasions" observed in the experiments [12], in par-
tieular the approximately constant acceleration of invasion tongues and the
leftover of narrow channels between "colliding" fingers. From our simulations,
it hence appears that this regime can be understood as period-preserving
oscillations with very large amplitudes. Indeed, the morphology at 4Amin al~
ready presents stages (i) and (iii), but not (ii). Apparently, to observe this
intermediate, constant acceleration regime, it is sufficient to have enough ''free
space" left in front of the propagating finger (Le., large lamellar spacings). It
188 R. Folch, M. Plapp

Fig. 2. Snapshot pictures of typical oscillatory structures. Growth is upwards, the

two solids are black and white, and the greyscale in the liquid is proportional to
the solute concentration. (a) low-amplitude oscillations ()../)..min = 2), (b) giant
oscillations ()../)..min = 4), (c) successive invasions ()../)..min = 8.48). The scale is the
same on all three figures.

is also interesting to note that in peritectie alloys, simulations of spreading in

the same geometry produce stages (i) and (ü), but then colliding fingers co-
alesce, and to obtain an oscillatory dynamies explicit nucleation events have
to be introduced [9,14].
The fact that we can switch from gentle oscillations to long invading
fingers just by changing the lamellar spacing demonstrates that both belong
to the same branch of oscillatory limit cycles. This branch of solutions is
indeed an attractor, since simulations at fixed spacing and started from very
different initial conditions always converge to the same cycle.

4 Conclusion
We have developed a phase-field model of eutectie solidification that has a
smooth free energy landscape and yields exactly binary interfaces away from
the trijunction points. These properties make it a promising starting point
for a second-order asymptotie analysis. Furthermore, we have used the model
to simulate oscillatory limit cycles in a model alloy that has asymmetrie
phase diagram, and we have found that low-amplitude oscillations, "giant
oscillations", and successive invasions, all observed in experiments, lie on a
single branch of solutions that is parametrized by the lamellar spacing and
that bifurcates from the steady-state branch at about twiee the minimum
undercooling spacing.
Here, we have focused on period-preserving oscillatory modes; many other
instability modes exist, for example the tilt instability or period-doubling
oscillations. These instabilities can be studied using less restrictive bound-
ary conditions. Preliminary simulations show that for the case studied here,
 Phase-field Modeling of Eutectic Solidification 189

namely an alloy with symmetrie phase diagram at its euteetic eomposition,

period-preserving oseillations remain the only stable limit eycles. In eontrast,
if we use off-euteetic sampie eompositions or an asymmetrie phase diagram,
and henee break the eomplete symmetry between the two solid phases, other
modes beeome aetive. Therefore, our model ean be used to obtain a more
eomplete bifureation diagram, which is the subjeet of ongoing work.
We thank Silvere Akamatsu and Gabriel Faivre for many stimulating
diseussions and Jean-Fran<.;ois Gouyet for his help with Fig. 1. R. F. ae-
knowledges financial support from projeet FMRX-CT96-0085 of the Euro-
pean Commission and from Centre National d 'Etudes Spatiales (Franee).

References
1. Ginibre, M., Akamatsu, S., Faivre, G.: Experimental determination of the sta-
bility diagram of a lamellar eutectic growth front. Phys. Rev. E 56 (1997)
780-796
2. Karma, A., Rappel, W.-J.: Quantitative phase-field modeling of dendritic
growth in two and three dimensions. Phys. Rev. E 57 (1998) 4323-4349; Karma,
A.: Phase-Field Formulation for Quantitative Modeling of Alloy Solidification.
Phys. Rev. Lett. 87 (2001) 115701
3. Plapp, M., Karma, A.: Multiscale random-walk algorithm for simulating in-
terfacial pattern formation. Phys. Rev. Lett. 84 (2000) 1740-1743; Plapp, M.,
Karma, A.: Multiscale Finite-Difference-Diffusion-Monte-Carlo method for sim-
ulating dendritic solidification. J. Comp. Phys. 165 (2000) 592-619
4. Karma, A.: Phase-field model of eutectic growth. Phys. Rev. E 49 (1994) 2245-
2250
5. EIder, K. R., Drolet, F., Kosterlitz, J. M., Grant, M.: Stochastic eutectic
growth. Phys. Rev. Lett. 72 (1994) 677-680
6. Wheeler, A. A., McFadden, G. B., Boettinger, W. J.: Phase-field model for
solidification of a eutectic alloy. Proc. R. Soc. Lond. A 452 (1996) 495-525
7. Steinbach, 1., Pezzola, F., Nestler, B., Seeßelberg, M., Prieler, R., Schmitz, G.
J.: Aphase field concept for multiphase systems. Physica D 94 (1996) 135-147
8. Garcke, H., Nestler, B., Stoth, B.: A multi phase field concept: numerical sim-
ulations of moving phase boundaries and multiple junctions. SIAM J. Appl.
Math 60 (1999) 295-315
9. Lo, T. S., Karma, A., Plapp, M.: Phase-field modeling of microstructural pat-
tern formation during directional solidification of peritectic alloys without mor-
phological instability. Phys. Rev. E 63 (2001) 031504
10. Karma, A., Sarkissian, A.: Morphological instabilities of lamellar eutectics.
Metall. Mater. Trans 27A (1996) 635-656
11. Akamatsu, S., Faivre, G.: Private communication.
12. Akamatsu, S., Moulinet, S., Faivre, G.: The Formation of Lamellar-Eutectic
Grains in Thin Sampies. Metall. Mater. Trans. 32A (2001) 2039-2047
13. Jackson, K. A., Hunt, J. D.: Lamellar and rod eutectic growth. Trans. Metall.
Soc. AlME 236 (1966) 843-852
14. Nestler, B., Wheeler, A. A.: A multi-phase-field model of eutectic and peritectic
alloys: numerical simulation of growth structures. Physica D 138 (2000) 114-
133
Phase-field Theory of Nucleation and Growth
in Binary Alloys

Laszl6 Granasy, Tamas Börzsönyi* and Tamas Pusztai

Research Institute for Solid State Physics and Optics, POB 49, H-1525 Budapest,
HungarYj *Universites Paris VI at VII, Tour 23, 2 place Jussieu, 75251 Paris
Cedex 05, France

Abstract. We present aphase field theory for binary crystal nucleation. Using
the physical interface thickness, we achieve quantitative agreement with computer
simulations and experiments for unary and binary substances. Large-scale numer-
ical simulations are performed for multi-particle freezing in alloys. We deduce the
Kolmogorov exponents for dendritic solidification and for the "soft-impingement"
of crystallites interacting via diffusion fields.

1 Introduction: Diffuse Interface and Nucleation

The solid-liquid interface is known to extend to several molecular layers, as
indicated by experiment [1], computer simulations [2], and statistical me-
chanical treatments based on the density functional theory [3]. This plays an
essential role in modeling nucleation, as the typical size of critical fluctuations
is comparable to the physical thickness of the interface. Modern nucleation
theories, which consider the molecular scale diffuseness of the interface, re-
moved the many orders of magnitude difference seen between nucleation rates
from the classical sharp interface approach and experiment [4].
The phase field theory (PFT) predicts a diffuse interface. In conventional
applications for solidification at small undercoolings, the diffuse interface is
both an advantage and a disadvantage: An advantage, since all mesh points
can be handled uniformly (no need for front tracking), and a disadvantage,
as the physical interface thickness is by orders of magnitude smaller than the
size of the modeled particles, thus due to computationallimitations, the true
interface thickness cannot be realized. Therefore, one is compelled to use an
artificially broad interface, which is accompanied with artifacts. Imaginative
work by Karma and coworkers (see [5] and references therein) established
operations, such as the introduction of interface currents and specific choice
of the phase field parameters, which remove these artifacts and ensure the
convergence of the diffuse interface results to the sharp interface solution.
In modeling nucleation, we face a different situation. Here, we are able
to choose the model parameters so that the true interface thickness is recov-
ered. We demonstrate that under such conditions the PFT is quantitative
for nucleation in unary and binary systems. Note that convergence to the
sharp interface model is not a requirement, as the sharp interface model is
inaccurate for nucleation [4].

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 Phase-field Theory of Nucleation and Growth 191

2 Phase-field Theory of Nucleation

We adopt the standard binary phase field theory (for review see [6,7]). Our
starting point is the free energy functional

F = J
drf':T (\7<p) 2 + f(<P'C)} ' (1)

where <p and c are the phase and concentration fields, J(<p, c) = WTg(<p)+[l-
P(<p)]fs + P(<p)/L is the local free energy density, W = (1- C)WA + CWB the
free energy scale, the quartic nmction g(<p) = <p2(1-<p)2/4 ensures the double-
weH form of J, while the function petP) = <p 3 (10 - 15<p + 6<p2) switches on
and off the solid and liquid contributions fS,L, taken from the ideal solution
model. (A and B indicate the constituents.)
For binaries the model contains three parameters €, WA and WB that
reduce to two (€ and W) in the one-component limit. They can be fixed if the
respective interface free energy ,,(, melting point TJ, and interface thickness
8 are known. Such information is available for the Lennard-Jones, ice-water,
and Cu-Ni systems [8-10].

2.1 Nucleation in 3D

The critical fluctuation (nucleus) is a non-trivial time-independent solution

of the governing equations:
o<p 8F
ot=-M<I> 8<p' (2)

oc = \7M \78F (3)

ot c 8c'
where (8F/8x) stands for the functional derivatives (x = <p,c), Mx are mo-
bilities. For spherical symmetry (a reasonable assumption), the phase field
equation reduces to \7 2<p = .1p(<p, c)/(€2T). Here .1p(<p, c) = WTyI(<p) + [(1-
c).1JA + c.1JB]P'(<P) is the local chemical potential difference relative to the
initial liquid, prime stands for differentiation with respect to the argument,
the local concentration is related to the phase-field as c(<p) = c",oe- Y /(1 -
Coo +cooe- Y ), where y = V(WB - WA)g(<p)/ R+V(.1JB -.1fA)[P(<p) -1]/ RT,
while .1Ji are the volumetrie free energy differences between the pure liquid
and solid phases. Solving these equations numerically under boundary condi-
tions that prescribe bulk liquid properties far from the fluctuations (<p -t 1,
and c -t Coo for r -t (0), and zero field-gradients at the center, one obtains
the free energy of critical fluctuation as W* = F - Fo. Here F is obtained
by numericaHy evaluating (1) after having the time-independent solutions
inserted, while Fo is the free energy of the initial liquid. This is compared
with W* = (167r /3hll.1j2 from the sharp interface "droplet" model of the
classical nucleation theory, where "(j = "(TJ).
192 L. Granäsy, T. Börzsönyi, T. Pusztai

The homogeneous nucleation rate is calculated as J = Joexp {- W* / kT},

where the nucleation prefactor J o of the classical kinetic approach is used,
which proved consistent with experiments [11].

2.2 Multi Particle Solidification in 2D

To treat nucleation and growth of particles on equal footing, we extend the
standard binary phase field theory [6,7] via incorporatinga non-conservative
orientational field (), which assumes a constant value between 0 and 1 in the
crystal that determines crystal orientation in the laboratory frame as done
recently by Kobayashi, Warren, and Carter [12]. We assume, however, that
() is defined in the liquid as well, where it fluctuates. Via this, we capture
the feature that the short-range order in the solid and liquid are usually
similar, with the obvious difference that the building units have a uniform
orientation in the crystal, while their orientation fluctuates in the liquid.
Following [12], we assume that the grain boundary energy acts in the solid
and is proportional to IVOI. We realize this by adding lori = MIVOI to Is,
where M is taken composition independent. The governing equation is
ao 8F
at = -Mo 88· (4)
For <p < 1, equation (4) switches in orientational ordering, and chooses the
value of 8 that survives as the orientation of the particle relative to which
the anisotropy of'Y is measured. Unlike previous work, in our approach the
orientation field 8 is coupled to the phase field, and extends to the liquid phase
where crystallographic orientation develops from orientational fluctuations.
While our model incorporates grain boundary dynamics, Mo is set so in the
solid that grain rotation is negligible on the time scale of freezing.
Since PFT is a mean field type approach, nucleation is introduced as
follows. Method 1.: By including noise into (2)-(4). Method H.: The siro'-
ulation area is divided into domains according to the Iocal composition.
The time-independent solution is found for these compositions. Critical fluc-
tuations of statistically correct numbers following Poisson distribution are
placed into these areas in every time step. The added small-amplitude
noise makes these fluctuations either grow or shrink. In nucleation-growth
processes the transformed fraction often follows the Kolmogorov scaling,
X(t) = 1 - exp{-(t/to)P}, where the "Kolmogorov exponent" p is repre-
sentative of the mechanism of the phase transformation, and is evaluated
from the siope of the plot ln[-ln(l - X)] vs. lnt. We define the crystalline
fraction via <p < 0.5. Simultaneous modeling of nucleation and growth raises
the question of interface thickness: When using the nanometer size physical
interface thickness, growth of large objects cannot be treated. If, in turn,
we use a broad interface for large particles, relying on methods for quantita-
tive description of dendritic solidification [5], we face with difliculties when
addressing nucleation and particle impingement. Therefore, following previ-
ous work [10], we perform the calculations for a hypothetical system of a
 Phase-field Theory of Nucleation and Growth 193

broader interface (8 = 41.6 nm), a reduced interfacial free energy b/6), and
an increased diffusion coefficient ('" 100 x D L). The consistent treatment of
nucleation and growth in our model, in which both processes emerge from
the same governing equations, allows us to describe the life of particles from
birth to impingement on other particles, a feature essential for understand-
ing the variation of Kolmogorov exponent with transformed fraction during
"soft-impingement" .
Equations (2)-(4) were solved numerically on 2000 x 2000 and 7000 x
7000 grids, using parallel processing (MPI protocol) on a computer cluster
consisting of 36 PCs of 1.0 - 1.6 GHz Athlon processors.

3 Results and Discussion

First we compare the phase-field predictions for the nucleation rate with
computer simulations and experimental results for unary systems. In the
Lennard-Jones system all relevant properties are known from molecular dy-
namics simulations [8] . The phase field profiles indicate that for typical under-
coolings the critical fluctuations are ramified, and do not show bulk crystal
properties. While the phase-field predictions for the nucleation rate agree
with results from computer simulations [13], those from the classical sharp
interface theory are too low by eight to ten orders of magnitude (Fig. la) .
Similar relationships are observed between predictions and experiment [14]
for the ice-water system (Fig. Ib). While the input data are less reliable for
the Cu-Ni system, the critical undercoolings computed for the realistic range
ofnucleation rates (J = 10-4 to 1 drop-1s-l for electromagnetically levitated
droplets of 6 mm diameter) are also in a reasonable agreement with experi-
ment [15] (Fig. lc). Note that without adjustable parameters, a quantitative

40 r-c:-::->-~---, r:---r-~--'---" 40 1800

mlJ Ar lc~-water (b) 30
Ul 35 1600
"'s 20
10 &
~ 30
t>D
I-
0
..8 25 1200
-10 Cu-Ni
20 5 '-0---'--e.J...O-
T-> ('-K-) --'--22""0---"'--2..1.1
4 0 -20 10000 02 0.4 0.6
C

Fig. 1. Comparison of the predicted 3D nucleation rates with computer simulations

and experiment (squares) on unary and binary systems. (PFT - phase filed theorYi
CNT - classical sharp interface theorYi dotted lines -limits due to the error of 1/.)

agreement has been achieved with simulations and experiment, indicating

that the phase-field theory is a powerful tool to attack nucleation problems,
provided that the physical interface thickness is used.
194 L. Gränäsy, T. Börzsönyi, T. Pusztai

Multi-particle freezing has been studied in 2D using the thermodynarnic

properties of Cu-Ni, and a large anisotropy for the interface free energy
(-ro {I + socos[4-a]} with So = 0.25). With appropriate choices of Me and M
single orientation nucleates per particle. The freezing morphologies and the
variation of exponent p with crystalline fraction are shown as a function of
liquid composition in Fig. 2. Here the particles were nucleated by Method
11. (Method I yields less regular shapes.) At small dimensionless nucleation
rates j = 0.49 dendritic particles evolve, yielding p ~ 3, that obeys the
relationship p = 1 +d (constant nucleation and growth rates in d-dimensions),
which follows from the steady-state traveling tip solution of the diffusion
equation. At large j pillow-shaped particles form. Thevariation of p with
transformed fraction depends on the liquid composition. Similar behavior
has been observed during the formation of nanocrystalline alloys.
At reduced Me, different orientations nucleate at the perimeter, of which
those with their fastest growth direction perpendicular to the interface sur-
vive, yielding spherulitic growth morphology (Fig. 3).

..
....
. .... ...
.. .. .
6.---..--.,...-...,....--.---,
x.o.2 x.o.8
4
0.1'------1
2 .................•...•...• 2 ••.................•....

(e) (f) (g) (h)

0.2 0.8 0.2
Tl 0.8 o 0.2
Tl 0.8 0.2 0.8
Tl
Fig. 2. Transformation kinetics in 2D VS. composition in the Cu-Ni system at
1574 K (x = (C - CI)/(C. - CI), where CI and C. are the solidus and liquidus
compositions). (a)-(d) Snapshots of concentration distributions (black - solidusj
light grey - liquidus)j (e)-(f) Kolmogorov exponent vs. normalized transformed
fraction 1J = X/ X max , where X max is the maximum crystalline fraction.

This work has been supported by the ESA Prodex Contract No. 14613/00
/NL/SFe, by the Hungarian Academy of Sciences under contract Nos. OTKA-
T-025139 and T-037323, by the EU grant HPMF-CT-1999-00132, and forms
part of the ESA MAP Project No. AO-99-101.
 Phase-field Theory of Nucleation and Growth 195

Fig.3. Spherulitic growth at x = 0.23: (a) Composition and (b) orientation maps.
Note the self-organized radial grain structure.

References
1. Huisman, W. J., Peters, J. F., Zwanenburg, M. J., de Vries, S. A., Derry, T. E.,
Albernathy, D., van der Veen, J. F.: Layering of a liquid metal in contact with
a hard wall. Nature 390 (1997) 379-381
2. Davidchack, R. L., Laird, B. B.: Simulation of the hard-sphere crystal-melt in-
terface. J. Chem. Phys. 108 (1998) 9452-9462
3. Ohnesorge, R., Löwen, H., Wagner, H.: Density functional theory of crystal-fluid
interfaces and surface melting. Phys. Rev. E 50 (1994) 4801-4809
4. Granasy, L., Igl6i, F.: Comparison of experiments and modern theories of crystal
nucleation. J. Chem. Phys. 101 (1997) 3634-3644
5. Karma, A.: Phase-field formulation for quantitative modeling of alloy solidifica-
tion. Phys. Rev. Lett. 81 (2001) 115701-1-115701-4
6. Warren, J. A., Boettinger, W. J .: Prediction of dendritic growth and microseg-
regation patterns in a binary alloy using the phase-field method. Acta Metall.
Mater. 43 (1995) 689-703
7. Boettinger, W. J ., Warren, J. A.: The phase-field method: Simulation of alloy
dendritic solidification during recalescence. Metall. Mater. Trans. A 27 (1996)
657-669
8. Broughton, J. Q., Gilmer G. H.: Molecular dynamics investigation of the crystal':'"
fluid interface. I-VI. J. Chem. Phys. 19 (1983) 5090-5127; 84 (1986) 5741-5768
9. Granasy, L.: Cahn-Hilliard-type density functional calculations for homogeneous
ice nucleation in undercooled water. J. Mol. Struct. 485--486 (1999) 523-536
10. Conti, M.: Growth of a needle crystal from an undercooled alloy melt. Phys.
Rev. E 56 (1997) 3197-3202
11. Kelton, K. F .: Crystal nucleation in liquids and glasses. Solid St. Phys. 45
(1991) 75-177
12. Kobayashi, R., Warren, J. A., Carter, W. C.: Vector-valued phase field model
for crystallization and grain boundary formation. Physica D 119 (1998) 415-423
13. Baez, L. A., Clancy, P.: The kinetics of crystal growth and dissolution from the
melt in Lennard-Jones systems. J. Chem. Phys. 102 (1995) 8138-8148
14. Taborek, P: Nucleation in emulsified supercooled water. Phys. Rev. B 32 (1985)
5902-5906
15. Willnecker, R., Herlach, D. M., Feuerbacher, B.: Nucleation in bulk undercooled
nickel-base alloys. Mater. Sei. Eng. 98 (1988) 85-88
Modelling of Phase Transformations in
Titanium Alloys with a Phase-field Model

Benoit Appolaire and Elisabeth Gautier

LSG2M, Eeole des Mines, Pare de Saurupt F-54042 Naney Cedex, Franee

Abstract. We have used aphase field model devoted initially to solidifieation for
studying some phase transformations in metastable ß titanium alloys. First eal-
eulations have been earried out to determine if we ean rely on the model to get
quantitative information on transformation kineties. For a simple ID problem, the
phase field model, a sharp interface model and an analytieal solution have been
eompared. The agreement was good even for quite eoarse meshes. Then, 2D eal-
eulations have been done starting from micrographs of real microstruetures. For a
moderate departure from equilibrium, we obtained kineties of the order of mag-
nitude of some experimental measurements. For a larger departure and with the
same parameters as before, the ealeulations failed beeause of the propagation of a
spurious intermediate state, which eould be avoided in decreasing drastically the
mesh size.

1 Introduction
Titanium alloys have more and more applications due to their desirable com-
bination of properties. First, titanium alloys have low densities which provide
very attractive strength to weight ratios allowing lighter structures. Secondly,
they show superior corrosion and erosion resistance in many environments.
Finally, they have a high temperature capability. Although titanium aJIoys
are still considered as expensive materials, their cost can be justified on the
basis of their versatile properties for many applications. Their good mechani-
cal properties are achieved thanks to proper thermo- mechanical treatments,
optimised to design the complex microstructures hierarchy.
From a metaJIurgical point of view, the microstructure formation obtained
by solid/solid phase transformations is easy to study by microscopy when
considering ß metastable alloys. Indeed, for these alloys it is possible to freeze
the microstructures during a transformation without subsequent occurence
of any displacive transformation at low temperatures.
During their thermo-mechanical treatments, most of the titanium alloys
are deformed in the high temperature range where the most stable phase
is ß (bcc) and further cooled to room temperature where the material is
a mixture of ß and 0: (hcp) phases. Different morphologies of the 0: phase
are observed depending on the cooling rate and the previous deformation
conditions:
- for smaJI departures from equllibrium conditions, 0: allotriomorph grains
at the ß grain boundaries are first observed (Fig. la).

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 Phase-field Modelling in Titanium Alloys 197

- For greater departures from equilibrium conditions, a plates of

widmanstätten have grown from or near the allotriomorph
grains (Fig. 1b).
- And finally, for the greatest departures from equilibrium conditions, aci-
cular a grains nucleated on defects are observed inside the ß grains
(Fig. lc).
Furthermore, deforming the material prior to the cooling prornotes the ap-
parition of the allotriomorphs together with the widmanstätten plates.

Fig. 1. Scanning electron microscopy (SEM) images of microstructures of a ß tita-

nium alloy (a) isothermally treated at 790°C during 5min.j (b) isothermally treated
at 790°C during 10min.j (c) isothermally treated at 700°C during 6min..

To explain these different morphologies, some mechanisms have been pro-

posed ([1] and references therein) mainly from post-mortem observations with
microscopy. However some features are unclear which could be explained
thanks to 2D /3D numerical modelling.

2 Which Model to Choose?

Among the different modelling techniques which can tackle this kind of prob-
lem, the phase field models have proved to be the best suited for the moment.
This is in great extent because the interfaces of complex morphologies have
not to be tracked explicitely. Another advantage of this kind of models, which
is of great importance for solid state transformations, is the possibility to in-
corporate easily phenomena which can be cast into the thermodynamical
formalism, such as elasticity [2].
Two methodologies distinguish between phase field models: on the one
hand those devoted to solid state transformations with fuH coherency at the
interfaces [2]; and on the other hand those devoted to solidification. We have
chosen the second one because of the nature of the a/ ß interfaces that obset'-
vations with Thansmission Electronic Microscope have shown to be at least
semi-coherent [1].
We have used the model of Kim et al. [3] developed initially for the solid-
ification of binary alloys because of its following features:
198 B. Appolaire, E. Gautier

- first, it is thermodynamically consistent, Le. the evolution equations are

gradient flows for the free energy functional of the system. Hence, in
principle, it should be quite easy to incorporate elasticity.
- Secondly, it is not linked to a particular phase diagram. The homogeneous
free energy is defined with a kind of rule of mixture between the energies
of both phases, augmented with the 'double wen potential' accounting
for the energy penalty of the interface.
- Thirdly, it is based on a special definition of the interface as a mixture
of the two phases with different compositions but with equal chemical
potentials, which prevents a chemical contribution to be added to the
interfacial energy.
- Finally, a thin interface asymptotic analysis has been performed, in a
simplified form compared to the original one of Karma et al. [4], which
may allow the use of coarser grids than the classical sharp interface ana-
lysis, so as to enhance computational efficiency.
We have adapted this model to the case of our multicomponent titanium
alloy with the composition reported in Tab. 1. The 0: phase can be assumed
stoechiometric as suggested by experimental measurements. The ß phase is
described as an ideal solid solution for simplicity at a first step, because the
agreement between our experimental measurements and a more sophisticated
description based on the CALPHAD method was not satisfactory.
With the previous hypotheses, the evolution equations are straightforward to
obtain and are not reported here for brevity.
Lastly, the necessary thermo-chemical data necessary for the calculations
(i.e. equilihrium data and diffusion coefficients) have been obtained by suit-
ahle and careful experiments.

Table 1. Composition of the ß CEZ alloy.

Molar % Ti Al Mo Cr Fe Zr Sn
Nominal balance 8.83 2.07 2.01 0.78 2.10 0.84
0: phase balance 10.90 0.19 0.27 0.04 2.10 0.84

3 Can we Rely on the Model to Get Quantitative

Predictions?
In order to check if the model of Kim et al. provides quantitative informa-
tion (Le. kinetics) together with the morphological evolutions, we have first
performed comparisons for a simple ID planar growth problem between the
 Phase-field Modelling in Titanium Alloys 199

phase field model, a numerieal sharp interface model and the analytical so,..
lution for a system of semi-infinite extent.
A nucleus of the growing aphase has been put at one side of the system,
small as eompared to the size of the system in order to avoid boundary effeets
on the growth rate, at least at the begining of the transformation. Moreover,
zero fiux boundary eonditions were applied on both sides of the system. The
initial eomposition in the ß matrix has been set at the nominal one (Tab. 1).
Then the system was quenched down to 830 oe (eorresponding to an under-
eooling of 60 Oe) and held at this temperature.
The main ealeulation parameters relative to this test ease are reported in
Tab. 2. Beeause the interface has been assumed to be at loeal equilibrium,
the phase field mobility was ealeulated in such a way that the interfacial
kinetics vanishes, thanks to a thin interface asymptotic analysis [3].
Even for a large interface width (Le. 3 J.Lm), a very good agreement bet-
ween the different methods has been achieved on the growth velo city as shown
in Fig. 2a. The agreement is also good for the time evolutions of the interfacial
eoneentrations of all the ehemieal species exeept for Molybdenum (Fig. 2b).
The moderate diserepaney for Molybdenum which is the slowest diffusing
species (DMo = 5.8.10- 15 m 2J8 at 830 Oe) is due to solute trapping. Indeed,
this effeet is minimized by choosing smaller mesh sizes beeause the inter-
facial Peclet number for Molybdenum is not small enough for large mesh
sizes (it is 0 (10- 1 )). Nonetheless, at this stage, we eould hope to perform
ealeulations on real systems with a moderate eomputational power (Le. a
single proeessor workstation). Thus, aseries of ealeulations has been earried

Table 2. Parameters of the calculations.

System size mesh size interface interfacial interf. kinetic

(J.'m or J.'m 2 ) (J.'m or J.'m 2 ) width (J.'m) energy (Jjm 2 ) coeff. (Wjm 4 )
10 100 0.5 3 1 0
2D 20x20 0.15xO.15 0.6 0.3 0

out in 2D, starting from real SEM microstruetures, with the main ealeula,-
tion parameters reported in Tab. 2. As a eompromise between aceuraey and
moderate eomputational power, we picked up square zones of 20x20 J.Lm2 in
the SEM images, that have been reeasted with a pixel resolution of 128x128
eorresponding to the eomputational grid.
The first ealeulation was performed starting from equilibrium microstrue-
tures at 870 oe. The initial eomposition was set at the equilibrium one in the
ß phase for this temperature. Then, the system was quenched down to 830 oe
and held at this temperature for 1000 s. Figs. 3a, b and e show snapshots of
the mierostruetures respeetively at 0 s, 100 s and 1000 s.
200 B. Appolaire, E. Gautier

1&-08
0.038
PhAse fteld melhod - - -
_eothaw moOlod
I
:r[=-"=:-:-1
Analylioal_ - ~ 0.036
Ö

j
i;'

~ 1&-09
0'<)3'

j
~

I:s 0 ,002
AnaIy1Ical aoIulIon -

(a) (b)
'e-10 0.03
0 '000 2000 3000 4000 5000 0 1000 2000 3000 4000 5000
ThIe(o) TIme (0)

Fig. 2. Comparison of the phase field model with a pseudo-enthalpy model and an
analytical solution for the growth of a planar interface: (a) growth rate; (b) molar
fraction of Mo at the interface in the ß matrix.

Fig.3. Snapshots of the 2D microstructures (0: precipitates are in black and the
grey levels visualize the molar fraction field of Al) when quenched from 870 oe to
830 0 e and held at this temperature for (a) Os; (b) 100s; (c) 1000s.

The 0: phase amount and mean size increase as expected. However, be-
cause we did not consider any anisotropy (interfacial energy or interfacial
kinetics), the 0: precipitates become rounded as the transformation proceeds,
contrary to what is observed experimentally. Indeed, in experiments, growth
is slightly favoured along ß grain boundaries. Nonetheless, a qualitative com-
parison between SEM images of equilibrium microstructures at 830°C and
the snapshot at 1000 s reveals a quite good agreement for the mean size of the
0: precipitates. Concerning the kinetics, the transformation is nearly complete
after 1000 s in agreement with resistivity measurements of transformations in
similar conditions.
We repeat the same kind of calculation but quenching microstructures
obtained at 820°C down to 750°C (so involving a driving force for the trans-
formation greater than in the preceding case). In Fig. 4 snapshots of the phase
field at successive times show the appearance of a spurious intermediate state
 Phase-field Modelling in Titanium Alloys 201

(in grey) which is destabilized as the transformation progresses. This spuri-

ous state gives rise to alternating bands of Cl: and ß phases which finally split
into strings of small Cl: spherical precipitates. Following Glasner et al. [5],
this phenomenon can be attributed to the existence of an internal minimum
of the free energy density with respect to the phase field when a parameter
proportional to the driving force and to the interface width is greater than
a critical value. This may be a serious limitation of the model when applied
to solid state transformations because the driving forces involved are usually
large, and should be offset against a very small interface width to get rid of
the spurious state.

Fig. 4. Snapshots of the phase field (black and white correspond to the Cl: and ß
phases, and grey represents the spurious intermediate state) when quenched from
820°C to 750°C and held at this temperature for (a) Os; (b) 70s; (c) 500s.

References
1. Aaronson, H.I.: Atomic mechanisms of diffusional nucleation and growth and
comparisons with their counterparts in shear transformations. Met Trans. A
24 (1993) 241-276
2. Jin, Y.M., Artemev A., Khatchaturyan, A.G.: Three-dimensional phase field
model of low-symmetry martensitic transformation in polycristal: simulation
of (; martensite in AuCd alloys. Acta Mater. 49 (2001) 2309-2320
3. Kim, S.G., Kim, W.T., Suzuki, T.: Phase-field model for binary alloys. Phys.
Rev. E 60 (1999) 7186-7197
4. Karma, A., Rappel, W.-J.: Quantitative phase-field modeling of dendritic
growth in two and three dimensions. Phys. Rev. E 57 (1998) .4323-4349
5. Glasner, K., Almgren, R.: Dual fronts in aphase field model. Physica D 146
(2000) 328-340
Spreading of Liquid Monolayers: From Kinetic
Monte Carlo Simulations to Continuum Limit

M. N. Popeseu and S. Dietrich

Max-Planck-Institut für Metallforschung, Heisenbergstraße 1, D-70569 Stuttgart

and Institut für Theoretische und Angewandte Physik, Universität Stuttgart,
Pfaffenwaldring 57, D-70569 Stuttgart, Germany

Abstract. We study with kinetic Monte Carlo simulations the two-dimensional

spreading of a fluid monolayer that is extracted from a reservoir. The asymptotic
spreading behavior is analyzed within a continuum limit and compared with pre-
vious predictions. The influence of a chemically patterned substrate (longitudinal
stripes) on the spreading behavior of such fluid monolayers is briefly discussed.

1 Introduction

Reeent experiments on dynamics of wetting [1] performed with nano-liter

volumes of liquid spreading on atomically smooth surfaces have shown one
or few layers of a preeursor film with molecular-thickness and macroscopic
extent advancing in front of the maeroseopie wedge of the droplet. For both
spreading of nano-droplets and eapillary rise geometries the linear extent X (t)
of the extracted film grows in time aeeording to X(t) = At1 / 2 as long as the
reservoir (droplet) is not emptied. Several models have been proposed [2-4]
to explain these observations. For the ease of a single monolayer preeursor,
a mieroseopie approach based on the dynamies of a lattice gas of interacting
particles has been developed by Burlatsky et al. [2] and reeently generalized
by Oshanin et al. [3]. Several mean-field assumptions allowed them to derive
analytieally the t 1 / 2 law for the spreading and the prefactor A.
In this paper we present results of kinetic Monte Carlo (KMC) simulations
of a model closely related to the one in Refs. [2] and [3]. The asymptotie
behavior of the transversally averaged density profile is analyzed within a
eontinuum limit and eompared with the previous mean-field predietions [2,3].
Finally, we briefly diseuss the influenee of a chemieally patterned substrate
with longitudinal stripes on the spreading behavior of such liquid monolayers.

2 Model and KMC Simulations

The following rules deseribe the model.

(i) The spreading geometry is reet angular (x-y plane). The half-plane x< 0
is oeeupied by a reservoir which maintains a fixed average density Co on the
line x = O. Co is defined as the average of the Ioeal oeeupation number

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 Spreading of Liquid Monolayers 203

.,,(rjt) E {O,l} on the lattice sites (x = O,y) (see (ii) below). At time t = 0
the half-plane x > 0 is emptYj thus, the liquid spreads in the x-direction.
(ii) The substrate is homogeneous, and the substrate-fluid interaction is mo-
deled as a periodic potential (square lattice) in which the particle motion
proceeds by activated jumps to nearest-neighbor weHs. Evaporation from the
substrate is not allowed. The activation barrier UAdefines the hopping rate
[} = Vo exp[-UA/kBT) , where Vo is an attempt frequency which defines the
time unit l/vo, kB is the Boltzmann constant, and T is the temperature.
This hopping rate is absorbed into the one-particle diffusion coefficient D o =
[}a 2 / z on the bare substrate, where z is the coordination number of the
underlaying lattice, and a is the lattice constant.
(iii) The pair interaction between fluid particles is assumed to be strongly
repulsive at short range, preventing double occupancy of the weHs, i.e., hard-
core repulsion, with a long-ranged attractive interaction -Uo/r6 for r/a ~ 1,
mimicking dispersion forces.
(iv) The prabability of a jump from lattice site r to r' is biased by the
fluid-fluid energy landscape and is given by

p(r -+ r' j t) = Z-l (rj t) exp{ß[U(rj t) - U(r' j t)]/2}, (1)

where l/ß = kBT, Z(r, t) = L exp{ß[U(rj t) -U(r'j t)]/2} is the normaliza-

r'
-
tion factor, U(rj t) = -Uo L .,,(r/l· t)
Ir _;/11 6 ' and the summation has been
Ir-r"I~3a
restricted to three lattice units for computational convenience. This is similar
to the cut-off used in molecular dynamics simulations for the attractive part
of Lennard-Jones interactions.
(v) The advancing edge rt of the monolayer at time t is defined as the union
of the most advanced particles from each line along the spreading direction.
In order to contral the rate of two-dimensional "evaporation", an additional
constraint is imposed by rejecting the moves from r t toward sites r' ahead
of rt for which IU(r'jt) + Uol < Uc , where Uc ~ 0 is a fixed threshold value.
We have used Uc = UO/3 6 , a value for which the evaporation is negligible.
The simulations have been carried using periodic boundary conditions
along the transversal, Le., y-direction (appropriate for simulating an infinitely
wide substrate), a hopping barrier ßUA = 3.5, and a time increment Llt which
is a random variable distributed according to P(Llt) = N[}exp( -N[}Llt) , N
being the number of particles in the film at time t [5]. The density p(rj t) is de-
fined as p(rj t) = (.,,(rj t)}, where (-) means average over different KMC runs.
The transversally averaged density C(x, t) and the average position of the
L
advancing edge of the monolayer are defined as C(x, t) = (L -1 L .,,(x, Yj t»
y=l
and X(t) = (L- 1 LX), respectively, where L is the width of the lattice.
rErt
204 M. N. Popescu, S. Dietrich

3 Results and Discussion

A. Time dependence of spreading

Figure l(a) shows the scaled time dependence ofthe front line, X(t)/../i, for
several values of Wo = ßUo and Co. The results confirm that the spreading
follows X (t) = At 1/ 2 , and the time independent dimensionless prefactor
A == A/VDo depends on both Wo and Co as shown in Fig. l(b). For

0.15
(a) 1.4 (b) -_ ...... ------
... - .. -
......................................
...................................
.
...~
0.10 ..... 1.2
S... ::."
..... e""
DOcDcOOCXIaccacacaocDOCc:x::ICDCacDcoaaoc
~ 1.D
~ 0.05 '<I:

C._I.:]
1° Co.O.S •
open: W. _ 1.0
D 0.8 •
ßII9d: W. _ 0.8 (> Co · O.6

00 0.6 0.7 0.8 0.9

2><10' 4xl0'
C.

Fig.1. (a) Position of the front line as a function of time for Wo = 0.8 (filled sym-
bols) and Wo = 1.0 (open symbols), respectively. (b) Dependence of the prefactor
A on the density of the reservoir, Go, for Wo = 1.0 (circles) and Wo = 0.8 (squares),
respectively. Symbols: KMC results, lines: mean-field prediction from Ref.(2).

comparison, we also show (lines) in Fig. l(b) the theoretical predictions 1

from Ref. [2]. The results of the simulations are smaller, and the decrease
with increasing Wo seems to be more pronounced than predicted. These
discrepancies are most probably due to the fact that the simulations capture,
even far away from the advancing edge, the existence of a very small
asymmetry between the rates of jumping toward the reservoir and the
advancing edge, respectively, which has been neglected in the mean-field
description [2,3].

B. Continuum limit
Assuming < 'TJ(r;t)'TJ(r';t') >=< 'TJ(r;t) >< 'TJ(r';t') >, Le., neglecting
correlations, the following master equation for the local density p(r; t) is
obtained:
dp(r' t)
d(nt) = ~{-p(r;t)p(r -t r')[l-p(r';t)]+[l-p(r;t)]p(r' -t r)p(r';t)}.
r

1 In order to use Eq. (8) in Ref. (2), the value J.t for the ratio of the probabilitiesfor
jumping ahead and back from T t is needed. We have determined this ratio from
our KMC data and obtained J.t ~ 0.95 for Wo = 0.8 and J.t ~ 0.93 for Wo = 1.0.
 Spreading of Liquid Monolayers 205

In the continuum space and time limit (a -* 0, n- 1 -* 0, D o finite), by taking

Taylor expansions for p(r -* r') and p(r'jt) around rand keeping terms up
to second order in the spatial derivative of p(rj t), one obtains [3]

atp = D o{V'2 p + ßV' [p(l- p)V'U]}. (3)

Using 2 U(rj t) !::::: -gUop(rj t), where 9 = 2: Ir /al- 6 is a geometrical

a~lrl~rc
factor depending on the lattice type and on the cut-off range r c , assuming
that the monolayer is homogeneous along y, which is supported by the sim-
ulation results, and changing variables to >. = x/JDot, one obtains that
asymptotically (t » 1) the density C(>') should satisfy [7]

rFC >. dC d [ dP]

d>.2 +"2 d>' - gWod>' C(1 - C) d>' = o. (4)

An important consequence of Eq. (4) is the existence of a critical value Wo,c

[6] given by Wo,c = 4/g. For the present case of a square lattice with a cut-off
at r c = 3 one has 9 !::::: 4.64, and thus Wo,c !::::: 0.86. Here we will discuss only
results for values Wo < Wo,c not too elose to the critical value, in particular
Wo = 0.6. A detailed analysis for Wo ~ Wo,c will be presented elsewhere [7].
In Fig. 2(a) we show the density profiles as a function ofthe scaled variable
>. (and as a function of x in the inset) for Co = 1.0. The data collapse for

~
KMC
-1'-'1
- - nwrHlold

~ 0.5
U

• 1_ 5.,,{
D 1_2.,,/
• 1-3.5x1rJ'
o 1_5XfrJ' •
o ~~~~--~~~·~ 0.5 1 1.5
o 1 2
A.
A. = x/(D. tJ'f2

Fig. 2. (a) Density proilles C(A) as a function of the scaling variable A (respectively
x in the inset). (b) Density proilles C(A) for Co = 1.0 and Co = 0.6 obtained
from Eq. (4) (fulllines), KMC simulations (symbols), and the mean-field approach
(dashed lines) from Ref. [2]. The horizontal dotted line shows Cl = 0.12.

times t > 1 X 105 is good and thus validates the scaling assumption leading
to Eq. (4). We have obtained similar results for all the values Wo and Co
which have been investigated. In order to numerically integrate Eq. (4) for a

2 The first order term, ...., V p, is zero due to symmetry and summation over all r'.
206 M. N. Popescu, S. Dietrich

quantitative analysis of the predicted density profiles, it must be augmented

by corresponding initial or boundary conditions. One boundary condition,
Le., C('x = 0) = Co is obvious, and a second boundary condition can be
obtained [7] at ,X = A based on the constraint (v) in Sec. 2 as Cl == C('x =
A) ::: 0.12 (for Uc = UO/3 6 ). In Fig. 2(b) we show the predicted density
profiles (fuUlines) C('x) for Co = 1.0 and Co = 0.6, respectively, along with
results of KMC simulations (symbols) at time t = 5 X 105 (which can be
regarded as being elose to the asymptotic limit) and with predictions of the
mean-field approach (dashed lines) from Ref. [2]. At low densities of the
reservoir the mean-field results and the predictions of Eq. (4) are almost
identical because the gas is dilute and thus the correlations are less effective.
At high densities ofthe reservoir, however, there are significant differences in
the qualitative behavior of the two theoretical predictions, and the simulation
results elearly favor the behavior predicted by Eq. (4).

4 Patterned Substrates
As a simple example of spreading on heterogeneous substrates we are in-
vestigating the case of chemical patterns in the form of longitudinal (along
the x-direction) wettable stripes on a non-wettable substrate. We choose the
same lattice constant a for both. The difference between wettable and non-
wettable regions is modeled, as shown in Fig. 3(a), through an additional
energy barrier l1U at the border between the two types of substrates, which
acts like a spatially varying chemical potential.

p ~ 1

Fig.3. (a) Potentiallandscape at the border between the two types of substrates
(left: non-wetting, right: wetting). (b) Density profile p(r;t) at time t = 105 für
spreading (along the vertical direction) on a periüdic distribution of longitudinal
wettable strip es (width = 16a) separated by non-wettable stripes (width = 6a) für
Co = 0.9 and Wo = 0.4. The color co ding (right) is a linear function of density.

This potential difference can be changed in real time by manipulating the

substrate properties, e.g., by using photo responsive coatings as in Ref. [8].
Adjusting 11U to a large enough value should facilitate a confining of the
 Spreading of Liquid Monolayers 207

spreading. Such an example of control is shown in Fig. 3(b). For the parame-
ters used, ßUA = 0.8 (wettable), ßUA = 1.0 (non-wettable), and ß.6.U = 6.5,
respectively, the long-time confined-spreading is achieved even with very nar-
row (6a) separating, non-wettable stripes.

5 Summary and Conclusions

We have studied with KMC simulations a model for the spreading of a liquid
monolayer being extracted from a reservoir, and we have briefly discussed
the influence of a chemically patterned substrate on the spreading behavior
of such liquid monolayers. Our main results are: .
=
1) The model captures the dependence X(t) At1 / 2 for the spreading edge,
but the previous mean-field approach [2] overestimates the rate of spreading.
2) Mean-field theory correctly predicts the density profiles C(A) for small
values Co and Wo, but breaks down at large values; an explicit inclusion of
the particle attraction as in Eq. (4) leads to correct predictions for all Co.
3) Patterning of the substrate facilitates a confining of the liquid.
Work in progress [7] addresses further questions like the effect of impurities
and of external fields on the spreading behavior.

Acknowledgments
Research done within the priority program ''Wetting & Structure Formation
at Interfaces", Grant DI 315/7-3, of the Deutsche Forschungsgemeinschaft.

References
1. F. Heslot, A. M. Cazbat, and N. Fraysse: Diffusion-controlled Wetting Films. J.
Phys.: Cond. Matt. 1 (1989) 5793-5798; F. Heslot, A. M. Cazbat, P. Levinson,
and N. Fraysse: Dynamies of Wetting of Tiny Drops: Ellipsometric Study of the
Late Stages of Spreading. Phys. Rev. Lett. 62 (1989) 1286-1289; N. Fraysse, M.
P. Valignat, F. Heslot, A. M. Cazbat, and P. Levinson: The Spreading of Layered
Microdroplets. J. Coll. Int. Sei. 158 (1993) 27-32.
2. S. F. Burlatsky, G. Oshanin, A. M. Cazbat, and M. Moreau: Microscopic Model
of Upward Creep of an Ultrathin Wetting Film. Phys. Rev. Lett. 76 (1996) 86-89.
3. G. Oshanin, J. De Coninck, A. M. Cazabat, and M. Moreau: Microscopic Model
for Spreading of a Two-dimensional Monolayer. J. Mol. Liquids 76 (1998) 195-219.
4. P. G. de Gennes and A. M. Cazbat: Films Diffusifs de Mouillage. C. R. Acad. Sei.
310 11 (1990) 107-111; D. B. Abraham, R. Cuerno, and E. Moro: A Microscopic
Model for Thin Film Spreading. preprint, cond-matj0201218.
5. E. Adam, L. Billard, and F. Lanc;on: Class of Monte Carlo Algorithms for Dy-
namic Problems Leads to an Adaptive Method. Phys. Rev. E 59 (1999) 1212-1216.
6. G. Giacomin: Van der Waals Limit and Phase Separation in a Particle Model
with Kawasaki Dynamies. J. Stat. Phys. 65 (1991) 217-234.
7. M. N. Popescu and S. Dietrich: unpublished.
8. S. Abbott, J. Ralston, G. Reynolds, and R. Hayes: Reversible Wettability of
Photoresponsive Pyrimidine-Coated Surfaces. Langmuir 15 (1999) 8923-8928.
A Multi-mesh Finite Element Method for
Phase-field Simulations

Alfred Schmidt

Zentrum für Technomathematik, Universität Bremen, Germany

[email protected]

Abstract. In phase field models, different eomponents of the solution (tempera-

ture and phase variable) exhibit a strongly different loeal behaviour. An optimal
diseretization should use a mesh for each eomponent. We present a general frame-
work for the adaptive solution of eoupled systems and its applieation to phase field
simulations.

Free boundary problems give rise to coupled systems of partial differ-

ential equations, whieh are prominent examples of systems where a careful
numerical discretization is needed in order to resolve the solution's behaviour.
Usually, a high resolution is essential near the free boundary, while coarser
meshes are sufficient in the bulk. A reasonable numerical method should use
locally adapted meshes to fulfil both the needs of accuracy and efficiency. For
coupled systems like the phase field system, where components of the solution
show strongly different local behaviour, an optimal discretization should use
specially adapted meshes for each component.

1 Models for Solidification

An undercooling or oversaturation of a liquid leads to a rapid solidification

of the material. Models for this behaviour include energy andj or mass trans-
port by diffusion andj or convection as weH as energy jmass conservation con-
ditions across the interface, like the Stefan condition. On a certain (meso-)
scale, additional surface effects at the phase boundary play an important role,
depending on curvature C or velo city V of the moving interface r,
usuaHy
described by a Gibbs-Thomson relation like

ccCr +cvVr +() =0 onr.

Anisotropie surface effects lead to dendritie growth.

Mathematical models for solidification with surface effects differ mainly
in the treatment of the free boundary, defining it as a sharp interface, a level
set, or a diffuse interface (phase field). Numerical methods for solidification
simulations are developed foHowing the same lines.

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 Multi-mesh FEM for Phase-field Simulations 209

1.1 Sharp Interface Models

In a sharp interface model, the phase boundary is a (smooth) hyper-surface
of the underlying space (a surface in 3D or a curve in 2D). The advantage
of the treatment of the interface as a smooth surface is its lower dimension
and that curvature, which appears in the Gibbs-Thomson law, is weH de-
fined. Drawbacks are that the standard model allows no changes in interface
topology, no nucleation. The notion of varifolds allows for some possibilities
in this direction.
The free boundary problem leads to adegenerate parabolic equation for
the interface motion, like an anisotropie mean curvature fiow equation, cou-
pled with heat or mass diffusion in solid and liquid phases. Additionally,
convection Can be added by coupling to Navier-Stokes' equations in the time
dependent liquid phase.

Numerical Methods for Sharp Interface Simulations. The sharp in-

terface model leads to numerieal methods with separate discretizations of
bulk (3D) and interface (2D) (or 2D & 1D).
In the bulk, a fine mesh is essential near the interface to sufficiently resolve
the behaviour of solutions (temperature, concentration). Adaptively refined
meshes in finite element methods are crucial to get enough resolution, espe-
cially in 3D. A numerical sharp interface finite element method with adaptive
meshing in 2D and 3D was presented in [7,8], with additional convection in
2D in [2].

1.2 Diffuse Interface Models

Phase field models can be formulated either with a double obstacle potential
or a smooth double wen potential. We want to restriet ourself here to an
obstacle formulation, as introduced by Blowey and Elliott [3].
The model describes the evolution of temperature () and phase variable X
in a domain [} and is a system of two (degenerate) parabolic equations:
8t (8 + AX) - K-,.18 =f
c 8t X - c div( a(V' X)) + A(X) - ~ ßX 3 ,8
in [} x (0, T) plus boundary and initial conditions. The function a may in-
clude anisotropie solidification parameters, and A is a (set valued) maximal
monotone graph, the sub differential of the double obstacle potential,

A(s) ={ °
(-00,0] if s =-1
if sE (-1,1)
[0, +00) if s = +1

with the effect that values of X are in the interval [-1, +1]. fis a given heat
source density, A, K-, and ß are non-negative, material-dependent coefficients,
210 A. Schmidt

and c: > 0 a small parameter. The phase variable X is equal to -1 (solid) or

+1 (liquid) everywhere but in a narrow transition region of width O(c:). For
c: -+ 0, the solution eonverges against a solution of the sharp interface model
with a eorresponding Gibbs-Thomson relation. Figure 1 shows the evolution
of the phase boundary during the solidifieation of an undereooled liquid.

Fig.1. Solid-liquid interface at different times

Both eomponents of the solution, temperature and phase variable, show

a strongly different loeal behaviour, see Fig. 2.

Fig. 2. Graphs of phase variable and temperature on 1/8 domain

- The phase variable is constant outside of a moving narrow strip of width

O(c:), where the phase transition oeeurs. Here, l\lxl = O(C 1 ).
- The temperature satisfies the heat equation outside this strip, thus it
is smooth. Inside the moving strip, the gradient of temperature changes
 Multi-mesh FEM for Phase-field Simulations 211

rapidly. In the limit c ---+ 0, the Stefan condition holds at the sharp
interface,

In order to resolve this behaviour with a finite element approximation, dif-

ferent requirements to the discretization hold for both components.

- The mesh for discretization of X must have a local mesh width of hs < Ce.
Outside this strip, where X is constant, the mesh might be arbitrarily
coarse. In order to be able to track nucleations, the phase field system
should be solved in the whole domain, not only near the current interface.
- To resolve the temperature behaviour, a much coarser mesh is sufficient
in the strip than is needed for the phase variable. On the other hand, the
mesh must have a sufficient fineness also in the rest of the domain.
Aseparate discretization for both temperature and phase variable is needed
in order to meet all requirements in an efficient numerical method.

2 Adaptive Finite Element Methods for Coupled

Systems of PDE

Besides phase transitions, many physical problems lead to coupled systems of

partial differential equations, too. Especially in case of nonlinear phenomena,
the components Ui, i = 1, ... , n may show a strongly different behaviour
(smoothness of solutions etc.) in the common underlying domain n c IRd •
The usual adaptive discretization for systems of PDEs uses the same mesh
for all components, locally refined based on aposteriori error indicators for
the sum of error contributions on each mesh element.
An optimal adaptive discretization should use different locally refined
meshes for different components of the solution, controled by separate er-
ror indicators for each component.
In [6] we present a general concept for adaptive finite element methods
for stationary or time dependent coupled problems.
Solution components Ui are discretized in finite element spaces Xi, which
are based on different locally refined simplicial grids Si, where all meshes Si
are refinements of the same macro triangulation So of n, see Fig. 3 for a
simple example.
The advantages of this approach are

- separately adapted mesh for each component of the solution;

- altogether less degrees of freedom, thus more efficient;
- by the common macro mesh, a direct local hierarchy of the triangulations
Si is given. This enables an exact evaluation of each other component Uj
on the elements of mesh Si.
212 A. Schmidt

Fig. 3. Two different refinements of a macro triangulation So

A somewhat similar idea for phase field simulations was proposed by EI-
liott and Gardiner [5], who use two separate meshes for temperature and
phase variable. A fine mesh with mesh size h « c for the phase variable is
used only near the transition region (selected by a 'mask'), a coarse mesh
with mesh size H = 4h for the temperature. Use of the 'mask' reduces com-
putations for the phase variable signifieantly, but does not allow to track
nucleations, e. g. Both meshes are aligned, but no error estimators or loeal
mesh refinements are used.

2.1 Aspeets of Implementation

The adaptive multi-mesh method is implemented in the finite element toolbox

ALBERT, a joint development with K. G. Siebert [9,10]. The toolbox uses
simplicial meshes in 2D (triangles) and 3D (tetrahedra) and loeal refinement
by biseetion of elements, which induees a hierarchical strueture of meshes
and finite element spaces.
The multi-mesh coneept intro duces the traversal of a eommon, virtually
refined mesh SM, see Fig. 4, for ealeulation of eoupling terms, which involve
several components or functions from finite element spaces defined on differ-
ent meshes, like In U/Ji with Ui E Xi, PE Xj.

Fig. 4. Virtual mesh SM with loeally maximal refinement from SI, S2

2.2 Numerieal Analysis and Adaptive Methods

We reeall the standard aposteriori error estimates and adaptive finite element
methods for a sealar elliptie problem. The error lIu-UIl between the solution u
 Multi-mesh FEM for Phase-field Simulations 213

and a diserete approximation U is estimated by a sum of loeal error indieators,

e.g.

Ilu - UII 2 ::; L 'fJ~,

SES
where the indicators 'fJs must be eomputable from the diserete solution U
and given data of the problem (a posteriori).
On an quasi-optimal mesh for a given error toleranee tol, the loeal error
indicators are equally distributed over all mesh elements,
tol 2
'fJ~ ~ #S for all SES,

eompare [1], e. g. A similar approach is used in ease of eoupled systems. The

goal is an estimate of the errors IIUi - Uill by a sum of loeal error indieators
on separate meshes, like
n
L IIUi - Ui1l 2 ::; L 'fJ~,s + ... + L 'fJ;,s·
i=l SESl SESn
In order to deseribe a quasi-optimal set of meshes, one possibility is to split
the total toleranee tol to the separate meshes
tol~ + ... + tol; = tol 2
and optimize every single mesh. This leads to deeoupled equidistribution
eonditions
2 tol;
'fJi,S ~ #Si for all S E Si,i = 1, ... ,no

3 Adaptive Method for Phase-field Models

In ajoint paper with z. ehen and R.H. Nochetto [4], we derive error estimates
and adaptive methods for the double obstacle phase field system. Denoting
by U = O+AX the energy density, and again diserete (finite element) functions
by upperease letters, the estimate looks like
lIu - UIILOO(O,T;H-l(,O» + vellx - XIILOO(O,T;L2('o»

+ (foT 118 - 8111'(0) + ae IIVx - vXIIl..(o) dt) 1/2

::; eomputable terms, loealizable to mesh elements.
In [4], we use a common mesh for the discretization of temperature and
phase variable. The loeal error indieators 'fJs eontain eontributions from both
eomponents,

Estimated error ::; 'fJo + m~ L 'fJ~,

SESm
214 A. Schmidt

where m is the time step index. A quasi-optimal common mesh fulfils the
equidistribution condition for the local indicators
tol 2
1]~ ~ #sm for all S E sm.
Here, the total mesh element count #sm is very large because of the fine
resolution in the strip, thus the local temperature error must be very small
also in elements far from the interface.

Adaptive FE Method with Separate Meshes

Aseparation of contributions from temperature and phase variable in the
error indicator is possible:

Estimated error ~ 1]0 + m~ ( L 1]~,s + L 1]~,S).

SES9' SES~

A split of the tolerance to both meshes, as described above, leads to a smaller

element count #So, thus larger local error tolerances for the temperature
error. It follows that the temperature mesh contains much less elements than a
common mesh. Also the mesh for the phase variable contains less elements, as
a fine resolution is needed only in the transition region. Fig. 5 shows element
counts over time. A time-dependent tolerance was used for the error in each
time step, which reflects the fact that the interface length grows during the
simulation.

NTal03

....""
.."" /
r
..""""" .11 /
20""
/ n,.,-I'"
/
u""
I."" I !IV
S""
."" -
O..QO 2..00 ...00 6.00 11.00 10.00
,

Fig. 5. Element counts for phase variable and temperature meshes over time

Figures 6 and 7 show the temperature and phase variable meshes with
zooms to the interface region. Finally, we present in Figs. 8-10 a comparison of
temperature and phase variable meshes from simulations with three different
error tolerances tal .
 Multi-mesh FEM for Phase-field Simulations 215

Fig.6. Temperature mesh (2608 elements) and zoom to interface

Fig.1. Phase variable mesh (23015 elements) and zoom

Conclusion
We described an efficient adaptive finite element method for phase field calcu-
lations, together with some 2D simulation results. A more detailed description
of the multi-mesh method and results from 3D simulations will be presented
in forthcoming artides.

Acknowledgements
This artide reports partly joint work with Z. Chen (Beijing), R. H. Nochetto
(College Park) , and K. G. Siebert (Freiburg) .
216 A. Schmidt

Fig. 8. tol=10: Meshes with 1122 resp. 7440 elements

Fig. 9. tol=7: Meshes with 1568 resp. 11813 elements

Fig. 10. tol=5: Meshes with 2245 resp. 20631 elements

 Multi-mesh FEM for Phase-field Simulations 217

References
1. BABUSKA, I. AND RHEINBOLDT, W., Error estimates for adaptive finite ele-
ment computations, SIAM J. Numer. Anal., 15 (1978), pp. 736-754.
2. BÄNSCH, E. AND SCHMIDT, A., Simulation of dendritic crystal growth with
thermal convection, Interfaces and Free Boundaries, 2 (2000), pp. 95-115.
3. BLOWEY, J. AND ELLIOTT, C., Curvature dependent phase boundary motion
and parabolic double obstacle problems., in Ni, Wei-Ming (ed.) et al., Degenerate
diffusions. IMA Vol. Math. Appl. 47, 19-60 , 1993.
4. CHEN, Z., NOCHETTO, R. H. AND SCHMIDT, A., Adaptive finite elementmeth-
ods for diffuse interface models. In preparation.
5. ELLIOTT, C. AND GARDINER, A., Double obstacle phase field computations of
dendritic growth. Report 96/19, University of Sussex.
6. SCHMIDT, A., Adaptive methods for coupled systems. In preparation.
7. - - , Computation ofthree dimensional dendrites withfinite elements, J. Com-
put. Phys., 125 (1996), pp. 293-312.
8. - - , Approximation of crystalline dendrite growth in two space dimensions,
Acta Math. Univ. Comenianae, 67 (1998), pp. 57-68.
9. SCHMIDT, A. AND SIEBERT, K. G., ALBERT: An adaptive hierarchical finite
element toolbox. Preprint 06/2000 Freiburg, 2000. Documentation.
10. - - , ALBERT - Software for scientific computations and applications, Acta
Math. Univ. Comenianae., 70 (2001), pp. 105-122.
 Transport of Point Defects in Growing
Si Crystals

Axel Voigt,Christian Weichmann

research center caesar, Crystal Growth, Friedensplatz 16, 53111 Bonn, Germany

Abstract. Silicon crystal growth is one of the key processes that determine the
yield and the profitability in semiconductor device manufacturing. The art of grow-
ing silicon crystals is today highly developed and the quality, purity and size of
today's crystals have reached an outstanding level. However, the so-ca1led grown-in
defects, which are related to the pulling of crystals, can not be perfectly controlled.
Such defeets can deteriorate devices, and the ongoing increase of integration density
requires wafer with even lower defect concentration. Hence the control and reduo-
tion of grown-in defeets still remains an important challenge for crystal growers.
The formation of grown-in defects is predominantly affected by intrinsic point de-
feets, which diffuse, recombine and aggregate in the silicon lattice during growth and
cooling of the ingot. We present a detailed defect model for silicon crystals which in-
corporates a1l relevant phenomena. Conservation equations for self-interstitials and
vacancies are introduced and a finite element method is provided for predicting the
concentration of point defects in growing Czochralski silicon crystals.

1 Introduction

Intrinsie point defects, i.e. self-interstitials and vacancies, impurity atoms

and defect cluster are present in the crystallattiee at finite temperature due
to energetie and entropie reasons. Vacancies and self-interstitials are formed
during solidification at the solid/liquid interface. It is generally assumed that
they are incorporated at their thermal equilibrium concentration. Once the
silicon atoms have crystallized, additional mechanisms for the creation are
known.

- The Schottky mechanism, where a bulk lattice atom jumps to an inter-

stitial site and diffuses to the surface, where it is added.
- The Frenkel mechanism, where a silicon atom leaves the lattice site by
simultaneously creating a vacancy and a self-interstitial.
- Recombination is the inverse process of the Frenkel mechanism. A va-
cancyand a self-interstitial diffuse towards each other and form a lattice
atom.
- The free energy of a larger defect is generally smaller than that of a larger
number of smaller defects. Therefore vacancies and self- interstitials can
also be consumed by preexisting agglomerates. But is is also possible that
they are released at the expense of the agglomerates.

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 Transport of point defects 219

Today it is generally assumed that vacancies and self-interstitials are simulta-

neously present in a silicon crystal under thermal equilibrium. The presence
of point defects in the lattice is due to thermal vibrations of the silicon atoms
at finite temperatures. For a detailed discussion of intrinsic point defect diffu-
sion in silicon, we refer to [5]. In a continuum approach we describe diffusion
Buxes of atomic species that are driven by the gradient of the chemical po-
tential, which is a function of temperature, pressure and the concentration
distribution of the diffusion species. Diffusion of point defect species due to
concentration and temperature gradients are governed by Fickian and ther-
mal diffusion laws. Pressure gradients occur in cooling crystals in the form
of thermal stresses.
Self-interstitials and vacancies are incorporated into the crystal in their
equilibrium concentrations and then diffuse and recombine rapidly in the high
temperature portion of the crystal. Depending on the growth conditions, one
of these defect species survives in excess and leads to the formation of the
corresponding microdefects. The source of most observable microdefects is
the condensation of supersaturated point defects as the crystal cools dur-
ing growth. In order to describe the formation of microdefects nucleation
and aggregation of defect clusters have to be considered. The model in [12]
can explain qualitatively the transition between self-interstitial and vacancy
related microdefects as a function of pull rate and thermal fields.

2 Point defect modeling

The conservation equations for intrinsic point defects include convective,

Fickian diffusion, thermal diffusion and recombination terms. The equations
can be derived from basic principles ofthermodynamic, see [2]. More detailed
models for the diffusion of point-defects in semiconductors can be found in
[4] and [7]. The governing equations for the concentrations ofinterstitials and
vacancies are

aC
at
i
+ U c . oe ° ( oe
v i - V ' Di v i - kT; V'Tc) -- k rec (ceqceq
QiDiCi i v - Ci v , e)
acv + U • V' c v -
7ft c V'. (e kT; V'Tc) -_ k rec (e
Dv V' v - QvDvCv
q
Ci C veq - CiCv ) ,

where Ci and C v are the concentrations of interstitials and vacancies, D i and

D v the diffusion coefficients, Qi and Qv the activation enthalpy for thermal
diffusion, c;q and czq the equilibrium concentrations, T temperature, U c
the velo city of the crystal, k the Boltzmann constant and k rec the reaction
coefficient for recombination.
Convection of self-interstitials and vacancies occurs as bulk transport in
the axial direction with a given pull rate Upull and in the tangential direction
with a given rotation rate Urat. Fickian diffusion is driven by concentration
gradients. The Bux of the diffusing species J = -DV'C is in the direction
220 A. Voigt, C. Weichmann

of lower concentrations. The diffusion coefficient D depends on temperature

and concentration. Due to the low concentrations of intrinsic point defects,
lo16 cm -3, the concentration dependence is neglected. Thermal diffusion is
driven by temperature gradients in the direction of higher temperature and
the flux is proportional to its activation energy Q, the diffusion coefficient
D and the temperature T. The effect of thermal diffusion was shown to be
small compared with the other transport mechanisms [3], and will therefore
be neglected. Recombination is identical for both species and describes the
elimination of self-interstitials and vacancies by the creation of perfect lat-
tice sites. Recombination is a three step process. First self-interstitials and
vacancies diffuse as delocalized point defects in the silicon lattiee by random
motion until they approach each other to a capture radius a r and become
localized. Then they have to overcome the recombination free energy barrier
and finally the self-interstitial will be trapped by the vacancy.

energy adjacent

separated

energy
difl'erenee
recombined
----------~--------------

configuration

Fig.l. Energy diagram for self-interstitial and vacancy recombination

The first part of the reaction coefficient describes the diffusion and the
second the activation dependence
41ra "Hlv-T"SIv
k rec = TT Cr (DI + Dv) e kT ,
vSi Si

with a r the capture radius, VSi the volume of a silicon unit cell, CS i the
atomic density, 6Hfv the enthalpy barrier for recombination and L::.SIv the
entropie contribution.
O.61+kT( -2.30+7.3S.1O- 3 T)
k rec =1.2'1O- 6 (DI+Dv)e- kT •

The thermophysical properties in the equations are experimentally because

of the high temperatures not directly measurable. One way to circumvent
 Transport of point defects 221

this problem is to calculate the microscopic properties with the help of sta-
tistical mechanics and molecular-dynamic simulations. The parameter set in
this study has been taken from [8]. The data were obtained with the use
of a simple Stillinger-Weber potential to approximate the microscopic pa-
rameters for point-defects in silicon crystals. To prevent finite size effects in
the molecular-dynamic simulations it is necessary to use millions of parti-
eIes to realize the needed length scales. Massive parallel algorithms are used
to overcome the complexity [1]. In the last years more complex potentials
have been developed to simulate the structural properties of silicon [9]. The
use of Tersoff-potentials for the purpose of predicting material properties is
in progress and will allow more precise point defect simulations. The esti-
mates for the diffusion coefficients and equilibrium concentrations are given
in Arrhenius form as
Hn>
D = DOe- kT ,
_SJ(T) _Hi(T)
Ceq = pe k e kT,

where pis the density of the lattice site, Hm the migration energy, Hf the
formation enthalpy and Sf the formation entropy. The following parameter
set is used [8].

D[ = 3.30.10- 1 . e-°i.,9P [C7 2 ]

D[ = 1.30.10- 3 . e-°i.,\1!6 [C7 2]
C[eq = 2.90· 1023 . e +1.4+3.S5.1Q-3
[~3]
T 3.46+3.0S·1Q-4 T
k • e- kT

= 4.97· 1022 . e [~3]

eq -3.7+3.53.1Q-3 T 2.4S+2.33·1Q- 4 T
CV k • e- ToT

By writing the transport equation in nondimensional form we define the

Peclet number Pe, which compares the convective with the Fickian diffusion
term, the Damköhler number Da, comparing recombination with Fickian
diffusion as well as a nondimensional number for thermal diffusion compared
with convection Td,
P _ ucL Da = krecceqL2, Td = ucLkT2
e- V ' D QDVT'
with a characteristic length L. Self-interstitials and vacancies are set to equi-
librium concentration at the solid/liquid interface. An essential boundary
condition for the point-defects at the crystal rim and cone can physically de-
scribed by the Schottky defect mechanism, which assurnes that point-defects
diffusing to the surface are added to the lattice and are therefore in equilib-
rium concentration,
Ci = cr,
Cv = c;/.
222 A. Voigt, C. Weichmann

Initial conditions are specified for time to assuming equilibrium concentration

in the crystal for self-interstitials and vacancies.

3 Finite element discretisation

If we multiply the reaction-diffusion equations by a test function <jJ E HJ ([}),
integrate over the region [}, use Green's theorem and apply the Dirichlet
conditions we get

In (~~ + (u e ' VC)) <jJd[} + In DVC· V<jJd[}-

In ~~~ VT· V<jJd[} = In krec (C?Cv - CICV) <jJd[}.

For existence of a unique weak solution see [6]. The discretisation combines
piecewise linear finite elements in space and a time discretisation including the
convection that is based on the method of characteristics. The characteristic
finite difference method is based on the approximation
t) t - T)
aC(x, t)
at + U c . 'MC(
v X,
~ C(x, t) - C(x -
'"
T
TUe,
,

with an appropriate time step size T. For each time step (t n , t n +1) let 1;+1 de-
note an adaptively constructed conforming mesh of the domain [} C [Rd. Let
W n +1 denote the finite element space of continuous piecewise linear functions
of 1;+1,
W n +1 = {W E CO (ii); WIT E Jg (T), TE m+1},
W~+l = W n +1 n HJ.
We approximate the concentrations as a linear combination of the basis ftmc-
tions such that
n
C n +1 = 2:: Cr+1<jJf+1,
i=l

where Cr+1 E IR and <jJf+1 E W~+l. The discrete problem reads as follows:
Given cn E wn find cn+1 E wn+1 such that for all <jJn+1 E w~+1

with en(x, tn) = cn(x - TnU c , tn). In each time step the equation leads to
the system of nonlinear algebraic equations
M Ge In+ 1 + A Gren+1
I + B Gr e In+1 + e Gv e In+1 -- RGr ,
n+1 + A Gv e V
MG e V n+1 + B Gv e Vn+1 + e Gr e Vn+1 -- R Gv ,
 Transport of point defects 223

with the diagonallumped mass matrix MG = (Mfl),

Mq
tt
= T~(-I.,:+1
'l't
-1.,:+1) n,
,'I't
n

the stiffness matrices A 01 = (ABI) and A Ov = (ABV),

ABI = (DIV'l/Ji+1, V'l/Jr 1)n,

A Gv _ (D VV'I'i
ij -
"-I.n+1 ) n,
"-I. n+1 ,V'I'j

the gradient matrices BOI = (BBI) and BOv = (BBV),

BC?I = (QIDI V'T-I.,:+l V'-I.1!-+1)
ZJ kT2 'l'z' 'l'J n,
BC?v = (QvDv V'T-I.n+1 V'-I.1!-+1)
ZJ kT2 'l'z' 'l'J n,
the recombination matrices C OI = (CBI) and COv = (CBV),

CBI = (kTeeCvl/Ji+1,l/Jj+1)n,
CBv = (kTee C1 I/Ji+1, I/Jj+1)n,

and the right hand sides RGI = (RBI) and RGv = (RBV),
R~h
Z
= ~(()!'z' 'l't
Tn
-I.,:+1)n+1
[}
+ (kTee ceqc eq -1.,:+1)
I V ,'I'z n
ROv
i
- ~(()n
- Tn
-I. n+1)n+1
v' 'l'i n + (kTee ceqc eq -I.n+1)
I V ,'I'i n·
Because of the nonlinearities we use a Newton method to solve the nonlinear
system. The generalized minimum residual method (GMRES) is used for
solving the linear system in each iteration.
The ingot length enlarges proportional to the pulling velo city during the
growth of the crystal, resulting in adeformation of the domain. Therefore
the numerical mesh has to adapt itself to the growing crystal. The shape of
the crystal and the temperature distribution in each timestep was calculated
in a thermal simulation taking account of conduction, melt convection, heat
radiation and a free solidification interface [10] and [13]. The meshes resulting
from this simulations are used for the point defect calculation.

4 Results

Quantitative results of the coupled heat transfer and point-defect system

were verified in [11], where an experimentally satisfied phenomenologicallaw
was used which qualitatively relates point defects to growth conditions [3].
Starting from this justification of the model, realistic Czochralski furnaces for
the growth of 200mm silicon crystals with different pulling velocities where
224 A. Voigt, C. Weichmann

'.

Fig.2. Defect concentration at various time steps for different pulling velocities

simulated, in order to relate the point-defect distribution in the crystal to

the pulling velo city. Due to the axisymmetry of the crystal the calculations
are performed in 2D.
In the same hot zone three different pull rates where used, Vpl = 0,348
mmjmin, V p 2 = O,454mmjmin and V p 3 = O,892mmjmin, to simulate the
growing process over a time interval of 250min. All simulations start with an
initial quasi steady state computation for a crystal length of lOem. Highest
self-interstitial concentrations occur in the crystal which is grown with the
lowest pull rate Vpl. This crystal is fully self-interstitial-rich after cooling
down. The crystal grown with pull rate V p 2 shows vacancy-rich as well as self-
interstitial-rich regions and the crystal grown with pull rate V p 3 is completely
vacancy-rich after cooling down. Fig. 2 shows normalized differences between
the concentration of self-interstitials and vacancies in the crystal at various
time steps.

5 Conclusions

Coupled transient heat transfer and transient point defect transport have
been successfully simulated with constant pull rates for growing silicon crys-
tals. The change in the phase boundary during the growth process influences
the defect distribution. In order to predict the concentration of intrinsic point
 Transport of point defects 225

defect in the crystal the thermal history of the crystal should therefore no
longer be neglected.

6 Acknowledgment
This work was partially supported by Wacker Siltronic Inc. and the DFG-
Sonderforschungsbereich 611 Bonn, TP C5.

References

1. Caglar, A.: (2000) On the numerical simulation of fullerene nanotubes:

G1Oo.ooo.ooo and beyond! Proc. of NIC Workshop on Molecular Dynamics on
Parallel Computers, World Scientific.
2. De Groot, S. R., Mazur, P.: (1984) Non-equilibrium Termodynamics. North-
Holland, Amsterdam.
3. Dornberger, E.: (1997) Prediction of OSF Ring Dynamics and Grown-in Voids
in Czochralski Silicon Crystals. PhD Thesis, Universite Catholique de Louvain.
4. Dunham, S. T.: (1992) A Quantitative Model for the Coupled Diffusion of
Phosphorus and Point Defects in Silicon. J. Electrochem. Soc. 1 2 - 35.
5. Fahey P., Griffin P., Plummer, J.: (1989) Point defects and dopant diffusion in
silicon. Rev. Modern Physics 61 291 - 306.
6. Lang, J.: (2001) Adaptive Multilevel Solution of Nonlinear Parabolic PDE Sys-
tems. Springer, Berlin.
7. Merz, M.: (1999) Analysis und Numerische Berechnung der Diffusion von
Fremdatomen in Homogenen Strukturen. HabilThesis, Technische Universität
München.
8. Sinno, T. R.: (1998) Defects in Crystalline Silicon: Intergated Atomistic and
Continuum Modeling. PhD Thesis, MIT, Cambridge.
9. Tersoff: (1986) New Empirical Model for the Structural Properties of Silicon.
Physical Review Letters 56(6) 632 - 635.
10. Voigt, A.: (2001) Numerical Simualtion of Industrial Crystal Growth. PhD
Thesis, Technische Universiät München.
11. Voigt, A., Nitschkowski, J., Weichmann, C., Hoffmann, K.-H. (2002) Control-
ling Point Defects in Single Silicon Crystals Grown by the Czochralski Method.
In: Breuer, M., Durst, F., Zenger, C. (Ed.) High Performance Scientific and En-
gineering Computing, Springer, Berlin, 229 - 236.
12. Voronkow, V.: (1982) The mechanism of swirl defects formation in silicon. J.
Cryst. Growth 59 625 - 643.
13. Weichmann, C., Nitschkowski, J., Hoffmann, K.-H., Voigt, A.: (2002) Time
dependent simulation of heat transfer in a Czochralski furnace. WIT Press. in
press.
Atomistic Simulation of Transport Phenomena
in Simple and Complex Fluids and Fluid
Mixt ures

Kurt Binder!, Jürgen Horbach!, Walter Kob2 , and Fathollah Varnik1

1 Institut für Physik, Johannes Gutenberg-Universität,
D-55099 Mainz, Staudinger Weg 7, Germany
2 Laboratoire des Verres, Universite Montpellier II,
Place E. Bataillon, cc69, 34095 Montpellier, France

Abstract. Computer simulations of fluids in thermal equilibrium can yield in-

formation on transport coeflicients such as self-diffusion and interdiffusion coeffi-
cients, viscosity, and thermal conductivity. While the estimation of self-diffusion
coefficients from the mean square displacements of the respective particles is rather
straightforward, the estimation of other transport coefficients is less straightfor-
ward, and can be based on either an analysis of time correlation functions of the
appropriate collective variables, or on nonequilibrium techniques where the linear
response to appropriate perturbations is measured.
To concretize ideas, we first discuss a very simple example, interdiffusion in a bi-
nary (AB) noninteracting lattice gas, based on Monte Carlo simulation of diffusion
in the framework of the vacancy mechanism. Measuring fluxes induced by chem-
ical potential gradients, Onsager coefficients are deduced, while the interdiffusion
coefficient is measured from an instationary Nonequilibrium Monte Carlo (NEMC)
method, studying the decay of concentration waves. Then Nonequilibrium Molec-
ular Dynamics (NEMD) methods are described, where fluid flow between plates is
used to extract both viscosity and thermal conductivity. As an alternative method,
the use of the Kubo formula for the viscosity and the analysis of time-dependent
specific heats (to estimate the thermal conductivity) are mentioned. Application
examples range from molten Si02 to polymer melts.

1 Introduction
Since the early days of computer simulation [1,2] it has been a very impor-
tant goal of such work to estimate transport coefficients and to elucidate
the relation between structural properties of the simulated model systems
and transport phenomena [3]. Both Molecular Dynamics methods [4-8], suit-
able for applications to simple and complex fluids and their mixtures, and
Monte Carlo methods [5,7-11], suitable for the study of diffusion phenom·
ena in solids [12,13] and viscous fluids such as polymer melts [14-17], are
used. Understanding the dynamics of fluctuations and associated transport
phenomena in equilibrium, one can proceed further to simulate the dynamics
of phase changes such as phase separation in mixtures (via "spinodal de-
composition" [18-21] or nucleation [22,23]), crystal growth from the vapor

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 Atomistic Simulation of Transport Phenomena in Fluids 227

or melt [24,25], growth of wetting layers at surfaces [26,27], dynamies of ad-

sorption (e.g. [28]) and related situations where interfaces are formed and
grow [29,30]. All these latter phenomena far from equilibrium, however, are
beyond the scope of the present article, which focuses on transport phenom-
ena in systems containing a single phase in equilibrium, so transport across
interfaces between coexisting phases will here not be considered. For the sake
of simplicity, all the examples treated are taken from the own research of
the authors, notwithstanding the fact that much excellent work from other
groups does exist on related topics: however, the purpose of this article is not
a comprehensive review, but rather to give a tutorial introduction to some
of the many methodologies that have been proposed in the literature, and to
describe the "state of the art" that one can reach with different simulational
strategies with representative examples.
One of the simplest cases is the simulation of diffusion in crystalline solids.
As a representative example, the simulation techniques to elucidate interdif-
fusion in binary alloys [31,32] will be described in Sec. 2. The related topic
of interdiffusion in polymer mixtures [33-36] has been the subject of a long-
standing controversy.
Sec. 3 then considers the estimation of selfdiffusion coefficients in fluids,
treating Si02 and Si0 2 -N8QO mixtures [37-41] and glassforming polymer
melts [42-45] as an example. Sec. 4 discusses the estimation of the viscosity,
both via Nonequilibrium Molecular Dynamics (NEMD) methods [46-60] and
via a Kubo formula [38], whiIe Sec. 5 discusses the estimation of the thermal
conductivity [59,61]. Sec. 6 finally summarizes a few conclusions.

2 Interdiffusion and Selfdiffusion in Binary Mixtures

(A,B)
First we recall the phenomenological theory of interdiffusion. One writes down
so-called constitutive equations for the current densities jA,jB of A- and B-
atoms in the mixture [62]. Note that this is just a postulate, relating the
current densities and the corresponding driving forces linearly. These driving
forces are the gradients of the chemical potential differences between A- or
B-atoms and vacancies, respectively: /-LA - /-Lv and /-LB - /-Lv. The coefficients
in these linear relations (AAA, ABB, ABA = AAB) are the Onsager coefficients,

(1)

(2)

Note that this (deterministic!) ansatz neglects statistical fluctuations, it is a

kind of mean field theory. By invoking the continuity equations, which simply
228 Kurt Binder et al.

o ® ® ® 0
o @

® ® ® ® ®
~
o ® ® ® 0
@--
r;.
® ® 0 @ ®
~
o
=0
L
.L:\
® ® ® 0
L I

® - ® 0 0
x
® ® ® ® ®
'-

(a) (b)

Fig. 1. (a) Schematic description of interdiffusion in a random mixed crystal, where

two different kinds of atoms (A,B) may occupy the sites of a regular lattice. Oc-
cupied sites are indicated by circles, an empty site by a dot. Diffusion proceeds by
the vacancy mechanism: A-atoms may jump to vacant sites with a jump rate FA,
B-atoms with a jump rate FB. For simplicity, it is assumed that these jump rates
do not depend on the occupation of neighboring lattice sites, implying that there
are no interactions depending on cf, the occupation variables of the lattice sites,
(= 1 if site i is taken by an A-atom, zero else) and cf. (b) Top: Schematic sketch
to indicate how a chemical potential difference dJL = JLA - JLv (or dJL = JLB - JLv)
over a linear dimension L in an L x L system with periodic boundary conditions
creates currents jA' jB' The lower part indicates a system prepared with a chemical
potential difference JLA - JLB = dJLsin(>'xIL), with LI>. integer, to create a concen-
tration wave dck, that is switched off at time t = 0, to monitor the decay of the
concentration wave with time t.

express the fact that the total concentrations are conserved, and hence are
exact,

- 0
at+ v 'JA= '
aCA n'
(3)
 Atomistic Simulation of 'fransport Phenomena in Fluids 229

,1
,
"",
(0)
I\AA
10- 2
....
""'.
111
c:
'''"
-
CII
'ü
;.:: " ~.768

"~ ~
CII
0
U

"" o.~, \ ..~

, .
L-
CII
cn
10-3 0.64
CI 0.1 /
111

0
c:
o.o/:
0.48 ~
0.02/
o.~• 0.01
10- 4
1 0.2 0.5 1.0
ratio of transition rates concentration

Fig.2. Onsager coefficient AAA plotted vs. the ratio of jump rates rA/rB (a) or
against the concentration CA of A-atoms (b), for a two-dimensional square L x L
lattice (L = 80) using a vacancy concentration Gy = 0.04, for the random binary
alloy ("ABV model") shown in Fig. 1. Parameter of the curves in part (a) is CA, in
part (b) is rA/rB. From Kehr et al. [31].

one obtains a closed set of equations (note CA + CB + CV = 1 for a lattice

model, Fig. 1a, where a lattice site is either taken by an A atom or a B
atom or empty). In the non-interacting case, the free energy simply contains
entropy of mixing terms,

(4)

and /La = 8Fj8ca . With some algebra, it is then possible to express the
interdiffusion coefficient in terms of the Onsager coefficients [31]:

(5)

However, there remain several problems: how are the Onsager coefficients
related to the jump rates rA, rB ofthe simple atomistic model (Fig. 1a)? Can
the nondiagonal term be neglected? After an, there are only 2 independent
jump rates (rA, rB) but three Onsager coefficients. In fact, analytical theories
in most cases assume AAB = 0 [33-36]. Assuming then that AAA is propor-
tional to rA and hence to the selfdiffusion coefficient DA of the A-atoms,
and likewise ABB <X D B one arrives at the so-called "slow mode theory"
[33,35]. It then is the slower diffusing species that controls interdiffusion. Ort
the other hand, assuming already in (1) and (2) that the vacancies always
are in equilibrium, V' /L = 0, then (5) does not hold, and one gets the ''fast
230 Kurt Binder et 31.

(al (bI
AABQndA BA
10-2
....
.1/1
c;

1~
", ",,
GI
'ü
.....
i;::
GI 0.2~
0
~ 10-3
GI
c:n
a
1/1.
',0.168
••
0.792'
~1~
c;
0
0.02~
o.01~
10-4
1 0.1 0.01 0.2 0.5 1.0
ratio of transition rates concentration

Fig.3. Same as Fig. 2 but for AAB = ABA. From Kehr et al. [31].

mode theory" [34]: now it is the faster diffusing species that controls interdif-
fusion! Now one could say, let the experiment decide which theory is correct
- however, different research groups have claimed experimental evidence for
both theories [36]! Since these theories drastically contradict each other, and
cannot both be correct in the same situation, there is clearly a problem in
the comparison between experiment and theory. In fact, it is not at all clear
whether the experiments really correspond to the simple model of Fig. la:
they are all done on fluid polymer mixtures [36] rather than mixed crystals
where single atoms diffuse. Even if one could do experiments studying inter-
diffusion for such mixed crystals, there would be interactions between the A
and B atoms and thus deviations from random mixing would occur, invali-
dating (4). In addition, elastic distortions (due to the fact that the atomic
radü of A and Bare not precisely equal) playa role, as weIl as lattice defects
(grain boundaries, dislocations etc. may greatly afIect diffusion processes in
crystals). So different experiments correspond to different physical situations
that often are incompletely characterized, and the model may be (perhaps?)
inadequate for a particular experiment. Thus the approximations made in
the mathematical treatment of the model cannot be tested conclusively by
a comparison with experiments! Here simulation has an advantage: precisely
the same model (Fig. la) can be simulated on which the theory is based,
all parameters that the theory uses are either given or can be independently
estimated (Fig. Ib). So no uncontrolled approximations occur.
Now answers to all above questions can be gotten from ''taylored'' simula-
tions, as indicated in Fig. Ib. In order to obtain the Onsager coefficients, one
simply foIlows the definitions, (1), (2): one imposes a small gradient 8J.L/ L
 Atomistic Simulation of Transport Phenomena in Fluids 231

O.03.---r-.,...-~----r--,--.,.---,r--,--.---,..,

GI
"'0
~
ct
E
CI
c:
o
~
!:
c:
GI
u
c:
o
u

•
O.oos'----'----'--.L.-----L-.........-...l.---''----'---'--.L.....J.
o 2 3 4 Sx103
Monte (orlo steps Ip.

Fig.4. Concentration amplitude <5Ck(t) plotted VS. time (in units of Monte Carlo
steps per particle), for the same model as in Figs. 1-3, using CA = CB = 0.48,
rA/ rB = 0.1, <5cA(t = 0) - <5CB(t = 0) = 0.02. Note that a single exponential
with Dint given by (5) results for cv ~ 0 only, while for cv i- 0 two exponentials
with different decay constants superimpose [31], resulting in separate curves for
A-particles (open dots) and B-particles (fuH dots). Three wavelengths A commen-
surate with L = 80 are shown, to test that the hydrodynamic limit (k ~ 0) actually
is realized. Curves are the result of an analytic calculation, based on (1)-(4) and
the "measured" Onsager coefficients. From Kehr et al. [31].

of the chemical potential in x-direction, acting on the A particles. One then

measures the resulting steady state currents of A-atoms and B-atoms. Be-
cause of the periodie boundary conditions, atoms that leave the simulation
box at the right boundary, re-enter it at the left boundary, and so a steady
state current is established after a transition period. This procedure yields
the ansager coefficients AAA, ABA. Then the same is done with a chemical
potential gradient acting on the B particles, to abtain ABA and A BB . So all
Onsager coefficients are gotten (Figs. 2, 3). In fact, it is not true that the
off-diagonal Onsager coefficient is always negligible in camparisan with the
diagonal ones: for small TAl TB one sees that AAB and AAA are of the same
order, and hence AAB must not be neg1ected!
For the Monte Car10 estimation of the interdiffusion constant, one f01-
10ws the recipe of linear response theory: one imposes a wav~like chemical
potential difference Öp = PA - PB (Fig. Ib, 10wer part) and thus prepares a
system with a superimposed "concentration wave" ÖCk (k = 21T I A) as an initial
state. At the time t = 0, the chemical potential difference is then switched
off, and one watches the decay of the amplitude of the concentration wave
with time, ÖCk(t) = öCk(0)exp(-Dintk2t). This decay should be exponential
232 Kurt Binder et al.

400

1:: 300
QJ
E
QJ

~
Ci.
VI
'ö 200
~
o
::J
cr
VI
I
C
o
QJ
E

Monte (arlo steps Ip.

Fig.5. Mean square displacements of tagged A (open circles) and B (solid circles)
particles as a function of Monte Carlo steps per particle, for a 503 lattice with
CA = CB = 0.45 and TAl TB = 0.1. From Kehr et al. [31].

(whieh needs to be checked in the simulation! [31]), and the decay constant
(for k -+ 0) yields the interdiffusion constant Dint. Fig. 4 shows the outcome
of such a linear response "computer experiment". Choosing a logarithmie or-
dinate scale and a linear scale for the time t, one verifies by observing straight
lines on this plot that the decay indeed is exponential with time, proportional
to exp( -nt). Using three different wavelengths (oX = L/2, L/4 and L/6), it is
verified that the decay constant r k scales with k as rk = D int k 2 . Actually if
the amplitude 8/-l is small but not infinitesimal, there are two different curves
for A and B atoms with slightly different amplitudes. The curves in Fig. 4
were calculated [31] from the above phenomenologieal theory, (1)-(4), using
the Onsager coefficients as estimated from the simulations, such as shown
in Figs. 2, 3. Thus, there are no adjustable parameters whatsoever, and the
agreement between the phenomenologieal theory and the simulations hence
is quite perfect! However, neither the slow mode theory [33,35] (in its tradi-
tional form, assuming AAB = 0, relating AAA, A BB to selfdiffusion constants
DA, DB) nor the fast mode theory [34] work for the present simple model.
One can obtain the selfdiffusion coefficients very simply from the Einstein
relations, considering the mean square displacements of ''tagged particles"
with time t
(6)
where d is the dimensionality of the system, and we average over all partieies
(with label i) of type a to improve the statisties, in order to estimate Da (a =
A or B). Fig. 5 shows that for our simple model (6) actually holds already
 Atomistic Simulation of Transport Phenomena in Fluids 233

for rather early times. Of course, Figs. 1-5 refer to a simplistic example:
the random hopping processes of Fig. la are straightforward to simulate by
Monte Carlo, and there are no problems of equilibration whatsoever. ''Real
life"-problems, to be discussed in the next sections, clearly are much harder
to handle, and thus, to our knowledge, only the present ''toy problem" has
received such a complete treatment, where selfdiffusion as weH as Onsager
coeflicients as weH as the interdiffusion constant have been obtained for the
same model. It remains achallenge for the future to perform a similar study
for an off-lattice model of a real system.

3 Estimation of Selfdiffusion Coefficients in Various

Models of Fluids
While in the case of the noninteracting mixture on a lattice considered in
the previous section the use of (6) to estimate selfdiffusion coeflicients is
straightforward, this task is much more demanding for chemically realistic
models of fluids: firstly, it is often diflicult to reach times that are sufliciently
large that the Einstein relation actually holds, while a different behavior is
detected for short and intermediate time scales.
We first exemplify these considerations with the case of molten silica
(Si0 2) and its mixtures with sodium oxide (Na20). These systems are of
great technical interest as prototypes of network glassformers, and also rel-
evant in the context of earth sciences and mineralogy. FUrthermore, a very
weH tested pair potential weH suited for Molecular Dynamics simulations has
been developed from quantum-chemical calculations [63,64]. This so-called
BKS potential [63] and its generalization [64] has the form (e is the charge
of a proton)

(7)

a, ß E [Si, Na, 0]. Here r is the distance between an ion of type a and an
ion of type ß. The values of parameters Aaß' Baß and Gaß can be found in
Refs. [63,64]. Note that here the qa are partial charges with values 2.4, -1.2
and 0.6 for Si, 0, and Na. FUrthermore the potential given by (7) is slightly
modified at short r if a or ß are Na. More details on this can be found in
Ref. [40]. It turns out that the potential gives a good agreement between
the simulated static structure factor [40] and corresponding experiments for
molten N~Si205 (NS2) [65].
The simulations have been done at constant volume, fixing the density at
{! = 2.37 gjcm 3 and using 8064 atoms for NS2 and 8016 atoms for N~Si307
(NS3) [40]. Since the long range Coulomb interactions, which require the use
of Ewald summation techniques, need very large computing times for these
relatively large system sizes, it would be convenient if one could work with
234 Kurt Binder et al.

smaller sizes. However, it turns out that finite size effects for smaller sizes
affect systematically the dynamical properties [66J (although no such size
effects were detected in static properties). For the integration of the equations
of motion, the velo city form of the Verlet algorithm [4-8] was used, with a
time step of 1.6 fs. The temperature of the system was controlled by coupling
it to a stochastic heat bath, Le. by substituting periodically the velocities
of the particles with the ones from a Maxwell-Boltzmann distribution with
the correct temperature. After the system was equilibrated at the target
temperature, the run was continued in the microcanonical ensemble (Le. the
heat bath is switched off).

101

~
A 10-1
"'....V

10-3

Fig. 6. Mean square displacements (r 2 (t» of (a) Si ions, (b) Na ions, and (c) 0 ions
for sodium trisilicate (NS3) on log-log plots versus time (in ps) at the temperatures
T = 4000 K, 3400 K, 3000 K, 2750 K, 2500 K, 2300 K, and 2100 K (from top to
bottom). Note that r(t = 0) is taken as the coordinate origin for each tagged
particle. From Horbach [67].

It should be noted that many oxide glass formers are already very viscous
at relatively high temperatures (e.g. for Si0 2 the experimental glass transition
temperature is around 1450 K). Therefore the lowest temperature at which
the Si02 system could be equilibrated (in a run over a physical time of 20 ns
[37,38]) was T = 2750 K, while for NS3 T = 2100 K and for NS2 T = 1900 K
could be reached [40].
Fig. 6 shows typical data for the mean square displacements of the
various ions obtainable in such simulations [67]. One recognizes three dif-
ferent regimes: for times up to about 10 fs, there is a ballistic regime,
(r 2 (t)) = 3kBTt2 /mo;, where mo; is the mass ofthe ion. For the corresponding
very small displacements, the variation of the potential due to the neighbor-
ing particles is not yet feit, and thus a free flight results. In the regime from
ab out 10 to ab out 100 fs, the curves bend over gradually to a slower motion,
or even a flat region (in particular for the lowest temperatures displayed in
 Atomistic Simulation of Transport Phenomena in Fluids 235

the figure). These ''plateaus'' in the (r 2 (t» vs. t curves can be interpreted
in terms of the "cage effect" [68]: the ions are confined in "cages" formed by
their nearest neighbors, and for low temperatures it takes a long time until
they can "escape from the cage". This "cage effect" leads to a slowing down
of all motions, and can (at least to some extent) be understood in terms of
the so-called mode coupling theory [68] for the glass transition. Only on the
time scale of the structural relaxation time T this cage breaks up, and the
slow crossover to diffusive motion, in accord with (6) sets in. At T = 2100 K,
a time of 1 ns in NS3 is hardly enough to reach this regime. Of course,
only when J(r 2 (t)) is of the order of many interatomic distances, one can
be sure that the asymptotic regime where (6) holds has been reached. Note
that there is an obvious problem of statistical accuracy, since the subsequent
system configurations generated in the simulation are correlated over a time
scale T needed to reach the diffusive regime which implies that one has to
simulate over a time which is significantly larger than T.

10-4

Cl
EA=5.18eV
exp:6eV
Brebec 1980

1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5

104rr [,,1]

Fig.7. Selfdiffusion constants for pure Si02, sodium disilicate (NS2) and sodium
trisilicate (NS3) plotted vs. inverse temperature. The bold straight !ines are fits
with Arrhenius laws, indicating the respective activation energies EA. The dashed
curves are guides to the eye. Experimental results for E;;i) are taken from Brebec
et al. [69] and for E~O) from Mikkelsen [70]. From Horbach et al. [40].

A further problem are systems in which the diffusion constants for the
various species are very different, such as in the discussed sodium-silicate sys-
tems. Fig. 7 shows that at low T the diffusion constant of Na is much larger
236 Kurt Binder et al.

than the one of Si or O. Hence one will find a mean square displacement which
increases quickly with time, see Fig. 6b, and one might be tempted to read off
a diffusion constant from runs that are relatively short. This might, however,
give rise to erroneous results, since the time scale might be insufficient to
aJIow the Si-O matrix to relax and to reconstruct. Hence it is necessary to
make runs that extend at least over the a-relaxation time r of the matrix
before the diffusion constant can be measured reliably. Note that with de-
creasing temperature this problem becomes more and more pronounced since
the activation energy of Na is significantly smaJIer than the one of Si or 0
(see Fig. 7).
Fig. 7 shows also that for pure Si0 2 the results for the activation ener-
gies agree reasonable weH with experiment [69,70]. Note, however, that the
experimental values are extracted from a regime of much smaller tempera-
tures (5.5 < 104 /T < 7.5) where no simulations are possible for pure Si02.
In the case of Si0 2 -Na2 0 mixtures, the Na ions cause a partial breakup of
the network of Si0 2 tetrahedra (each 0 sits at a corner of such a tetrahedron
and is shared between two tetrahedra). Therefore, the diffusion of Si and 0
in the mixtures is enhanced, in comparison with pure Si02. Unfortunately,
we are not aware of any experimental data on diffusion constants for either
NS2 or NS3; but there exist data for Na2S409 [72], e.g. for T = 2800 K and
p = 10 GPa DSi = 1.22 X 105 cm2/s and Do = 1.53 X 105 cm2/s. Simulating
this system at a density p = 2.9 g/ cm3 at T = 2800 K (then p = 10.13 GPa in
the simulation, using 7680 particles), one finds [40] DSi = 0.78 X 105 cm2/s,
Do = 1.17 X 105 cm2/s, and DNa = 2.35 X 105 cm2/s. Thus, the simulation
results underestimate the corresponding experimental data by less than 40%,
which is within the error bars of experiment. This agreement shows that the
model used in [40] gives a quite reasonable description of the diffusion in
sodium silicate melts.
In order to gain insight into the microscopic mechanism of the diffusion
of the different atomic species, it is useful to analyze the probability Paß(t)
that a bond between an atom of type a: and an atom of type ß is present
at time t, if it was present at time t = O. '!'wo atoms are defined as bonded
if their distance is less than the location of the first minimum rmin in the
corresponding pair correlation function 9aß(r). One finds [40] rmin = 3.6 A,
5.0 A, 2.35 A, 5.0 A, 3.1 A, and 3.15 A, for the Si-Si, Si-Na, Si-O, Na-
Na, Na-O, and 0-0 correlations, respectively. Hone plots Paß(t) versus
the logarithm of time one finds a plateau at intermediate time scales, which
corresponds to the plateau seen in Fig. 6, while at late times there is a decay
according to a stretched exponential function, Paß(t) cx exp[-(t/r)ßKWW],
with ßKWW ~ 0.54 [40] for Na, while for the other ions Paß(t) decays almost
like a simple exponential function (which would have ßKWW = 1, while the
best fit actually yields ßKWW ~ 0.9 [38]). Therefore, it is convenient to define
a lifetime raß of a bond from the condition Paß(t = raß) = e- 1 • As expected,
the selfdiffusion of oxygen is simply controlled by the lifetime of Si-O bonds
(Fig. 8): only if such a bond breaks, can the associated O-atom move. The
 Atomistic Simulation of Transport Phenomena in Fluids 237

4.8
4.3
.........
~ 3.8
c
:::J
-e 3.3
ctI
~ 2.8
oX
'§" 2.3
I:-> I

1.8
1. 3 +r-r-rr-rr-rTTT""rr-r-r-rr-r-r-.,--r-r-rr-r-r-,...,.."r-r-rr-r-r-rr-,.,--r-rr-r-r+

1800 2300 2800 3300 3800

T [K]
Fig.8. Products TaßDa for sodium disilicate plotted vs. temperature. From Hor-
bach et al. [40J.

O-diffusion can in fact be interpreted in terms of a kind of defect mechanism:

while normally each Si atom has 4 oxygen neigbors, defects can occur where
one Si "catches" a 5th O-atom, leaving a Si with only 3 O-atoms bound to it
in its neighborhood. These co ordination defects diffuse through the random
networks, much like vacancy-interstitial defects diffuse in a crystallattice.
However, Fig. 8 also shows that there is no simple relation between DSi
and the relaxation time TSi-O, and surprisingly D Na seems to be linked to
TNa-Na, and not to TNa-O, as one might naively expect. This result gives
evidence that the sodium diffusion is relatively independent of the matrix
and that the Na atoms hop independently through it, a picture that has been
confirmed by the investigation of the van Hove correlation function (see [3]
for adefinition) and other time dependent correlation functions [40,71].
The above remark that the regime where the Einstein relation, (6), is
reached when atoms have moved over a few interatomic distances holds only
for simple liquids, of course, but is not true for complex fluids such as polymer
melts. For a flexible macromolecule, one must distinguish between the motion
of a single monomeric unit and the motion of the chain as a whole, which
is formed from N such units (Fig. 9a) [42]. While both the single (effective)
monomer and the center of mass of the chain show a regime of ballistic mo-
tion at very short times in Molecular Dynamics simulations, as found for Si0 2
(Fig. 2), there occurs already at high temperatures (far above any glass tran-
sition) a regime of anomalous diffusion, before (6) holds. The coarse-grained
238 Kurt Binder et al.

111
ul
10'
10' (R')J(R.1
...
·1(/
,,.'1-'801
m 1/.11. 4 , 10·

i"
'
~ 10-t I J&A
CI
---~' --;~---------------
L. /11.11. fIl d;10-'
~

~1a-a 1.' ,," - T=O.60

CI )
--- T=O.30

/~mM
09,
10-' 10" -- T=O.23
'110 - MD,T=O.46
- Rouse
10"
I 10'"
10 10·' 10 10' 111 10 10 10 10- 10 10' 10'
t DtI(R.-)

(a) (b)

Fig.9. a) Log-log plot of the mean square displacement of an inner monomer,

91(t), and of the chains' center of mass, 9s(t), versus time for a temperature T = 1,
pressure p = 1, for a bead-spring model of a polymer chain with chain length
N = 10. Beads interact with a Lennard-Jones potential given by (8). The thick
solid line is based on the Rouse model, using the amplitudes of the Rouse modes and
the associated Rouse times observed in the same simulation. Characteristic lengths
indicated are Tsc, the size ofthe nearest neighbor cage, Rg, the gyration radii of the
chains, and Re, the end-to-end distance. Both the initial ballistic regime (3Tt 2 and
3Tt 2 IN, respectively), the regime of anomalous diffusion ofthe monomers (91 (t) <X
tO. 6S ) and the diffusion regime (91(t) = 9s(t) = 6Dt) are indicated by straight
lines. From Bennemann et al. [73]. b) Comparison of the mean square displacement
91(t), rescaled with the mean square radius of gyration, plotted vs. the scaled
time Dt/(R~}, for the off-Iattice model (curve marked MD) with the corresponding
Monte Carlo results for the bond fluctuation model at three temperatures. Straight
!ines indicate power laws of anomalous diffusion (tO. 65 ) and ordinary diffusion (tI).
From Binder et al. [74).

model used in Fig. 9a is a bead-spring model of a flexible polymer, where all

beads interact with a truncated and shifted Lennard Jones potential,

(8)

whiIe ULJ(r ~ rc) = 0, and Uo is chosen such that ULJ(r) is continuous

at r = r c . The parameters u = 1 and e = 1 define units of length scale and
temperature scale, respectively, and also kB == 1 and m 1 (time is measured =
in units of 70 = u(m/€)1/2). Bonded neighbors along achain interact through
a FENE potential

(9)
 Atomistic Simulation of Transport Phenomena in Fluids 239

with Ro = 1.5. Eqs. (8), (9) lead to an equilibrium bond length of lo = 0.96,
different from the distance ro = 21/ 6 (J" Rl 1.12 where ULJ has its minimum.
This competition between the two lengths preferred by bonded and non-
bonded interactions has the effect that this model of a polymer melt also
tends to glass-like freezing, and crystallization is avoided. For pressure p = 1,
however, the critical temperature of mode coupling theory is Tc = 0.45 [74],
and thus at the temperature T = 1 shown in Fig. 9a one is far away from the
temperatures where slowing down due to the glass transition occurs. In fact,
the anomalous diffusion seen in Fig. 9a can be weH understood as an effect
of chain connectivity, and can be accounted for by the Rouse model [75], at
least to some extent. In the regime of displacements exceeding the size of
the cages confining the monomers but smaller than the end-to-end distance,
r;c < 91(t) < R~ where91(t) == ([ri(t)-ri(O)]2), one observes 91(t) cx: tO. 63 (for
an ideal chain in a heat bath, the Rouse model would yield [75] 91 (t) cx: t 1/ 2 ).
Similarly, the center of mass diffusion 93(t) = ([rcM(t) - rCM(0)]2) behaves
!l(,
as 93(t) cx: tO. 8 as long as 93(t) < R~ Rl and (6) holds only for even larger
displacements.
When this model is studied at lower T closer to the glass transition, then
a plateau near 91 (t) Rl r;c develops, similar to the mean square displacement
shown in Fig. 6, see also Fig. 9b [74] and the data in [17,73,74]. The above
Rouse-like region then foHows at those times where the "cages" confining the
monomeric units have decayed.
Fig. 9b also illustrates the fact that the mesoscopic scale motions of poly-
mers can also be modeHed by even cruder lattice models of polymers, such
as the bond fiuctuation model [14,15]. In this latter model the elementary
dynamieal steps consist of attempted hopping moves of monomerie units
by one lattiee spacing in a randomly chosen direction. While such a model
clearly cannot capture the small-scale ballistic regime, and thus also the ini-
tial part of the dynamies in the cage differs from the corresponding off-lattiee
results, the Rouse-type regime already coincides with the latter, if appropri-
ately rescaled quantities are compared. Thus, not all lattice models behave
as simple as the model (Fig. 5) discussed in Sec. 2. At this point, we also
mention that by fine-tuning of effective interactions of a lattiee model one
even can approximately fit the behavior of atomistic off-lattice models for
real polymerie materials, such as polyethylene melts [16]. The accuracy of
this ''mapping'' has been tested by a comparison with Molecular Dynamies
simulations for a melt of ClOoH202 at T = 509 K [16].
Finally, we emphasize that another case, where the approach to the regime
where (6) holds is very slow, is the diffusion of small partieies in a frozen en-
vironment, e.g. gas moleeules diffusing in a polymerie matrix [76-80]. It is
found that (6) holds only if (r 2 (t)) is sufficiently large, such that in the ex-
plored region the locally random system is already homogeneous. This fact is
clear when one considers the related problem of diffusion on percolation clus-
ters ("ant in the labyrinth") near the percolation threshold [81]: The crossover
240 Kurt Binder et al.

5.2
Ux(Z)
p(z) F9 =O.05 0.08
4.2 p(Z=O)=O.99
- - T=1 0.06
3.2 ............... T=O.35

(a) 0.04
2.2

1.2 0.02

0.2
-10.0 - .0 0.0 5.0 10.0
0.00 W~~~~~~~U
- 10.0 - 5.0
0.0 5.0 10.0
z z
Fig. 10. Density profile of a polymer melt confined between two walls (at Zwall =
±10), for a force Fe = 0.05, density po = 0.99, and two temperatures as indicated
(left part). All quantities are measured in units of the Lennard-Jones potential, (8)
that acts between the monomers. Only the region -5 ::; Z ::; 5, where the density
oscillations ("layering") are weak, is used for the fit of (10) to the observed velocity
profile (right part). Solid lines are results of a fit to u", (z) = a - bz 2 and TJeff results
then from b as TJeff = poFe j(2b). From Varnik and Binder [59].

from anomalous to ordinary diffusion is observed only if v(r 2 (t)} exceeds the
correlation length ~p which diverges at the percolation threshold [81] .

4 Estimation of the Shear Viscosity

Both for molten oxides and for polymer melts the shear viscosity 'f/(T) is of
great interest, and a large variety of techniques to extract 'f/(T) from sim-
ulations has been proposed. The first NEMD method that we shall discuss
considers the flow properties of the fluid between parallel solid walls. The
flow is created by a constant external force pe that points in the x-direction
parallel to the walls on all fluid particles, and the resulting velo city profile
U x (z) in the z-direction perpendicular to the walls is recorded.
While for a study of a fluid at rest confined between walls it makes per,.
feet sense to consider the case of perfectly smooth, repulsive walls [82,83],
for a fluid flowing between walls a different treatment of the boundaries is
needed, in order to obtain a nontrivial profile ux(z) . In [59] structured walls
were assumed that are atomistically rough, choosing a triangular lattice (with
equilibrium sites r i,eq) for each wall, to which the wall atoms are harmon-
ically bound, Uharm(r) = !Kh(ri - ri,eq)2, with K h = 100. The beads of
the polymers interact with the wall atoms also with a Lennard-Jones poten-
tial, and in addition a repulsive wall potential (cx: z-9) is used to represent
the effect of the layers other than the topmost one of the solid forming the
 Atomistic Simulation of Transport Phenomena in Fluids 241

l • MD
500 - - VFT
\
400
\ .
~.

•'**,
\
\,
r;! 300 To=O.18±O.OO5

200

100

0
. .......
...... -
+-+-+-...
• •
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
T-To

Fig. 11. Viscosity of a polymer melt plotted VS. the temperature distance from
the Vogel-Fulcher temperature To = 0.194 ± 0.012, obtained from a least-square
fit (full curve) to the Molecular Dynamies results (circles with error bars). From
Varnik and Binder [59).

walls [84]. The parameters of the wall atom-monomer interaction are chosen
O'wm = 1j..;2, €wm = 2 [59]. In general, partial slip at the position Z = Zwall
ofthe wall is observed, oux(z)jozIZ=Zwall = ux(zwall)j8, 8 being a slip length.
Choosing the position Z = 0 in the center of the film, hydrodynamics then
implies the well-known Poisseuille type flow [52-60,84,85]
(10)
where Po = p(z = 0) is the fluid density in the center of the fluid film, and 'T/eff
is an effective viscosity, which reduces to the shear viscosity of the bulk fluid
for pe --+ 0 and if IZwa1l1 is large enough. In practice, this holds already if IZwa1l1
is a few atomic diameters, see Fig. 10. (Note that since the system is driven
one has to apply a thermostat in order to avoid that it heats up.) Similar to
the diffusion constant of the melt [17,44], the viscosity 'T/(T) can be described
by a Vogel-Fulcher relation 'T/(T) cx: exp(constj(T - To)) (Fig. 11). Thus
neither 'T/(T) nor D can be described by simple Arrhenius laws [17,42] which
could be expected, since polymers are ''fragile'' glass formers. On the other
hand, the simple Stokes-Einstein relation kBTj{'T/D) = const which holds for
the simple Rouse model [43,75] is not verified here at low temperatures [59].
An alternative method to estimate the viscosity does not use any NEMD
nor considers any confinement, but rather studies appropriate dynamical cor-
relation functions of a system in equilibrium, where periodic boundary con-
ditions are applied as usual. The shear viscosity 'T/ can be obtained as a time
242 Kurt Binder et al.

20
ksTh]D [Al
15

experiment

2.0

1.0 2.0 3.0

Fig. 12. Molecular Dynamies results for the viscosity of the BKS model (7) for SiÜ2
plotted vs. inverse temperature. The dashed straight line indicates an Arrhenius fit
with an activation energy EA = 5.19 eV. Experimental data [86) are compatible with
this value but would suggest a slightly different pre-exponential factor. Note that for
SiÜ2 the analysis of other correlation functions at very high temperature suggests
a critical temperature of mode coupling theory Tc ~ 3330 K, and 1J(Tc) ~ 8 Poise.
The insert shows the failure of Stokes-Einstein relations. From Horbach and Kob
[38].

integral over the autocorrelation function of the off-diagonal elements of the

pressure tensor [3] ("Kubo formula")

(11)

where V is the volume of the simulation box, and

N N
IIaß(t) = Lmiui(t)u~(t) + L LF/i(t)r~(t) (12)
i=l i=l j>i

In (12) the sum runs over all N particles i (which may have different masses
mi) in the system, ufis the a th cartesian component of the velocity Ui, rij
is the distance between particles i,j and F ij the force acting between them.
While (11) and (12) have not been used for the model of the polymer melt
described by (8) and (9), they were used for the model for Si0 2 described by
the BKS potential, (7). Fig. 12 shows a plot of 1](T) [on a logarithmic scale] VS.
 Atomistic Simulation of Transport Phenomena in Fluids 243

0.07
a.=O.OO1
0.06
0.05
0.04
-
..-...
N
>
x
0.03
0.02
0.01 T=3760K:1l=O.58P
ll Gl(=O.9P
0.00
0.0 6.0 12.0 18.0 24.0 30.0
z [A]

Fig. 13. Velo city profiles of Si0 2 melts obtained from NEMD simulations between
parallel walls. Three temperatures are included, as shown in the figure; solid curves
show fits to (10), and the corresponding viscosity estimates are quoted. Results
from the Green-Kubo formula, (11), are included for comparison. From Horbach
et al. [60].

1jT [38], to demonstrate the extent to which an Arrhenius law holds. (Note
that it is very remarkable that at sufficiently high T strong deviations from
the Arrhenius dependence are observed. This is discussed in more detail in
Ref. [38].) Also experimental data [86] are included, demonstrating reasonable
agreement. The inset shows that again no Stokes-Einstein relation holds.

If one compares 'f/(T) and corresponding data for the self diffusion con-
stants for the same model (Figs. 7 and 12) it is evident that the statistical
accuracy of ",(T) is much less. This must be expected, however, since for
Dsi(Do) every Si(O)-atom contributes an independent "measurement" in
(6), while '" as computed from (11) is a collective quantity of the whole sys-
tem. Thus the question arises whether ",(T) can be estimated more accurately
from the NEMD technique, described above. This question was followed up
in [60], and as Fig. 13 shows, the data obtained from both methods are com-
patible with each other, but there is no indication that the accuracy of the
NEMD approach is really better than that of the Kubo formula, (11). (Note
that a comparable amount of computer time had to be used to obtain '" via
the two methods.)
244 Kurt Binder et al.

T wan=1 p(Z=0)=0.795

2 • Fe =0.01
* F =0.05
e
• Fe=0.1
... Fe=0.2

1
- 10 -5 o 5 10
z
Fig.14. Temperature profile across a polymer film (chain length N = 10, density
p(z = 0) = 0.795, Zwall = ±10) at a wall temperature Twall = 1, for four different
choices of the external force pe driving the How. Curves are the result of a fit with
(13), with A being the single fitting parameter for each curve. From Varnik and
Binder [59].

5 Thermal Conductivity

Another transport phenomenon ofinterest in fluids is transport ofheat. Again

one can estimate the corresponding transport coefficient either by NEMD
methods or analyze suitable time correlation functions in thermal equilib-
rium.
In fact, the above simulation of a polymer melt flowing in x-direction
between walls can be used to estimate the thermal conductivity A, if one lets
the thermostat act only on the wall atoms, while the fluid flow in between
the walls is simulated in a strictly microcanonical ensemble. Due to the flow,
heat is generated in the fluid film, which is transported to the walls. This
gives rise to a temperature profile across the wall [87]
(ppeD 2)2
T(z) = Twall + 192A1Jeff [1- (Z/Zwall)4] , (13)

Twall being the temperature maintained (via the Nose-Hoover thermostat,

for instance [4-8,59]) at the wall. Fig. 14 shows results of a corresponding
simulation, together with a fit to (13) using data for the viscosity 1Jeff such
as shown in Fig. 10. For the range of temperatures reached in the film in the
case shown in Fig. 14, the thermal conductivity lies in the range 3 ~ A ~ 4.4.
 Atomistic Simulation of Transport Phenomena in Fluids 245

An alternative method to estimate A, which considers the autocorrelation

nmction of temperature fluctuations in the microcanonical ensemble, was
used in [61], again for the case of Si02 melts. One needs to measure the
autocorrelation function <pTT(q, t) oftemperature fluctuations 8Tq (t) at finite
wave vectors. The relaxation time r(q) extracted from this correlator can be
shown to behave as [61]

pCp 1
r ()
q = T q2 + const , (14)

where Cp is the heat capacity at constant pressure in thermal equilibrium.

For 3000 K ~ T ~ 4700 K, a nearly temperature independent value A ~
2.4 W j(Km) resulted [61], in good agreement with experimental estimates
(2 ~ A ~ 3 [W j(Km)] if T ~ 1000 K [88]).

6 Concluding Remarks

For simple lattice models it has been possible for a long time to estimate
reliably the transport coefficients and thus to describe interdiffusion in mixed
crystals (Sec. 2), or surface diffusion in adsorbed registered monolayers at
surfaces [89]. This is in contrast to MD methods for off-Iattice models of
practicaHy relevant fluids (fluid Si0 2, polymer melts, etc.) which only recently
have reached a level of accuracy that allows a meaningful comparison with
experimental data, and a significant test of theoretical concepts.
By examples drawn from the research group of the authors, it has been
shown that only the estimation of self-diffusion coefficients is rather straight-
forward, while the estimation of truly collective quantities (shear viscosity,
thermal conductivity, etc.) still is at a stage where ''feasability-type'' ex-
ploratory simulations are made. However, the results are fairly encouraging,
and it is hoped that in the near future much broader applications will ap-
pear, as weH as more refined techniques to extract all transport coefficients
of interest from suitably designed simulations. At the same time, there is
often a surprising scarcity of experimental data to compare with. Thus it is
hoped that the present article will stimulate both corresponding theoretical
and experimental investigations.
Acknowledgments: One of us (K. B.) is greatly indebted to the late K. W.
Kehr for stimulating his interest in the Monte Carlo study of diffusion phe-
nomena, and for a longstanding collaboration [13,31], including the develop-
ment of NEMC methods [31]. We are grateful to J. Baschnagel, C. Benne-
mann and W. Paul for their fruitful collaboration on the studies of diffusion
in glassforming polymer melts [15,17,42,44,73] and to A. Latz and P. Schei-
dler for their fruitful collaboration on the methods to estimate the thermal
conductivity of molten silica [61]. We acknowledge partial financial support
from the BMBF and generous grants of computer time from the NIC JÜlich.
246 Kurt Binder et al.

References
1. Alder, B. J., Wainwright, T. E.: J. Chem. Phys. 31, 459 (1959); Phys. Rev. A
1, 18 (1970); Rahman, A.: Phys. Rev. A 136, 405 (1964).
2. Erpenbeck, J. J., Wood, W. W.: In Statistical mechanics; Part B, Vol. 6 of
Modern theoretical ehemistry ed. by Berne, B. J., (Plenum Press, New York,
1977) 1-40.
3. Boon, J. P., Yip, S.: Molecular hydrodynamics. (Me Graw Hili, New York, 1980).
4. Ciceotti, G., Hoover, W. G., (Eds.): Molecular dynamics simulations of statis-
tical mechanical systems. (North-Holland, Amsterdam, 1986).
5. Allen, M. P., Tildesley, D. J.: Computer simulation of liquids. (Clarendon Press,
Oxford, 1987).
6. Rapaport, D. C.: The art o[ molecular dynamics simulation (Cambridge Univ.
Press, Cambridge, 1995).
7. Binder, K, Ciceotti, G. (Eds.): Monte Garlo and Molecular Dynamics of Con-
densed Matter Systems (Soeieta Italiana di Fisiea, Bologna, 1996).
8. Frenkel, D., Smit, B.: Understanding molecular simulation. Prom algorithms to
applications (Aeademic Press, San Diego, 1996).
9. Binder, K (Ed.): Monte Garlo methods in statistical physics (Springer, Berlin,
1979).
10. Binder, K (Ed.): Applications of the Monte Carlo method in statistical physics
(Springer, Berlin, 1984).
11. Landau, D. P., Binder, K: A guide to Monte Carlo simulations in statistical
physics (Cambridge University Press, Cambridge, 2000).
12. Mureh, G. E.: Atomic diffusion theory in highly defective solids (Trans. Teeh.
House, Adermannsdorf, 1980).
13. Kehr, K, W., Binder, K: In Ref. [10], 181-221.
14. Binder, K (Ed.): Monte Carlo and Molecular dynamics simulations in polymer
science (Oxford University Press, New York, 1995).
15. Binder, K, Paul, W.: J. Polymer Sei., Part B 35, 1 (1997).
16. Tries, V., Paul, W., Baschnagel, J., Binder, K: J. Chem. Phys. 106,738 (1997).
17. Binder, K, Basehnagel, J., Paul, W.: Progr. Polymer Sei. (in press).
18. Gunton, J. D., San Miguel, M., Sahni, P. S.: In Phase transitions and Griti-
cal phenomena, Vol. 8, ed. by C. Domb and J. L. Lebowitz (Aeademie Press,
London, 1983), p. 267.
19. Binder, K., Fratzl, P.: In Phase transformations in materials, ed. by G. Kostorz
(Wiley-VCH, Weinheim, 2001) 409-480.
20. Binder, K, Kalos, M. H.: In Ref. [9], 225-260.
21. Reister, E., Müller, M., Binder, K: Phys. Rev. E 64, 041804 (2001).
22. Binder, K, Stauffer, D.: Adv. Phys. 25, 343 (1976).
23. Novotny, M. A.: Phys. Rev. Lett. 74, 1 (1995); Kolesik, M., Novotny, M. A.,
Rikvold, P. A.: Phys. Rev. Lett. 80, 3384 (1998).
24. Müller-Krumbhaar, H.: In Ref. [9], 261-299.
25. Müller-Krumbhaar, H., Kurz, W., Brener, E.: In Phase transformations in
materials, ed. by Kostorz, G. (Wiley-VCH, Weinheim, 2001) 81-170.
26. Mon, K K, Binder, K, Landau, D. P.: Phys. Rev. B 35, 3683 (1987).
27. Binder, K: In Kinetics of ordering and Growth at surfaces, ed. by Lagally, M.
(Plenum Press, New York, 1990) 31-44.
28. Patrykiejew, A., Binder, K: Surface Sei. 273, 413 (1992).
 Atomistic Simulation of Transport Phenomena in Fluids 247

29. Barabasi, A. L., Stanley, H. E.: Fractal concepts in surface growth (Cambridge
Univ. Press, Cambridge, 1995).
30. Pal, S., Landau, D. P.: Physica A 267, 406 (1999).
31. Kehr, K W., Binder, K, Reulein, S.: Phys. Rev. B 39, 4891 (1989).
32. Murch, G. E.: In Phase transformations in materials, ed. by Kostorz, G.
(Wiley-VCH, Weinheim, 2001) 171-238.
33. Broehard, F., Jouffray, J., Levinson, P.: Macromoleeules 16, 2638 (1983).
34. Kramer, E. J., Green, P., Palmstrom, C. J.: Polymer 25, 437 (1984).
35. Binder, K: J. Chem. Phys. 79, 6387 (1983); Colloid & Polymer Sei. 265, 273
(1987).
36. For a review, see e.g. Binder, K, Silleseu, H.: In Encyclopedia of polymer sei.
& eng. Suppl. Vol., ed. by Mark, H. (Wiley, New York, 1989) 297-315.
37. Horbach, J., Kob, W., Binder, K: Phil. Mag. B 77, 297 (1998).
38. Horbaeh, J., Kob, W.: Phys. Rev. B 60, 3169 (1999).
39. Horbaeh, J., Kob, W.: Phil. Mag. B 79, 1981 (1999).
40. Horbach, J., Kob, W., Binder, K: Chemical Geology 174, 87 (2001).
41. Winkler, A.: Dissertation (Johannes Gutenberg Universität Mainz, 2002).
42. Paul, W., Baschnagel, J.: In Monte Carlo and Molecular dynamics simulations
in polymer science, ed. by Binder, K (Oxford Univ. Press, New York, 1995)
307-355.
43. Tries, V., Paul, W., Basehnagel, J., Binder, K: J.Chem. Phys. 106, 738 (1997).
44. Bennemann, C., Paul, W., Binder, K, Dünweg, B.: Phys. Rev. E 57,843 (1998).
45. Binder, K, Baschnagel, J., Paul, W.: Progr. Polymer Sei. (2002, in press).
46. Ashurst, T. W., Hoover, W. G.: Phys. Rev. Lett. 31, 206 (1972).
47. Lees, A. W., Edwards, S. F.: J. Phys. C 5, 1921 (1972).
48. Hoover, W. G.: Ann. Rev. Phys. Chem. 34, 103 (1983).
49. Evans, D. J.: J. Chem. Phys. 78, 3297 (1983).
50. Evans, D. J., Morris, G. P.: Comp. Phys. Rep. 1, 287 (1984).
51. Loose, W., Hess, S.: Phys. Rev. A 37, 2099 (1988); Physica A 162, 138 (1989).
52. Thompson, P. A., Robbins, M. D.: Phys. Rev. A 41, 6830 (1990).
53. Todd, B. D., D. Evans, D., Daivis, P.: Phys. Rev. E 52, 1627 (1995).
54. Travis, K. P., Todd, B. D., Evans, D. J.: Phys. Rev. E 55, 4288 (1997).
55. Koplik, J., Banavar, J. R.: Phys. Rev. Lett. 80, 5115 (1998).
56. Barrat, J. -L., Boequet, L.: Phys. Rev. Lett. 82, 4671 (1999).
57. Travis, K P., Gubbins, K E.: J. Chem. Phys. 112, 1984 (2000).
58. Sokhan, V. P., Nicholson, D., Quirke, N.: J. Chem. Phys. 115,3878 (2001).
59. Varnik, F., Binder, K: preprint.
60. Horbach, J., Kob, W., Binder, K: in preparation.
61. Seheidler, P., Kob, W., Latz, A., Horbach, J., Binder, K.: Phys. Rev. B 67,
104204 (2001).
62. De Groot, S. R., Mazur, P.: Non-Equilibrium thermodynamics (Amsterdam,
North-Holland, 1962).
63. van Beest, B. W. H., Kramer, G. J., van Santen, R. A.: Phys. Rev. Lett. 64,
1955 (1990).
64. Kramer, G. J., de Man, A. J. M., van Santen, R. A.: J. Amer. Chem. Soe. 64,
6435 (1991).
65. Misawa, M., Priee, D. L., Suzuki, K: J. Non-Cryst. Sol. 37, 85 (1980).
66. Horbach, J., Kob, W., Binder, K, Angell, C. A.: Phys. Rev. E 54, R5897 (1996).
67. Horbaeh, J.: Dissertation (Johannes Gutenberg Universität, Mainz, 1998).
248 Kurt Binder et al.

68. Götze, W.: In Liquids, Freezing, and the Glass transition, Ed. by Hansen, J.-P.,
Levesque, D., and Zinn-Justin, J. (North-Holland, Amsterdam, 1991) p. 287..
69. Brebec, G., Seguin, R., Sella, C., Bevenot, J., Martin, J. C.: Acta Metall. 28,
327 (1980).
70. J. C. Mikkelsen: Appl. Phys. Lett. 45, 1187 (1984).
71. Jund, P., Kob, W., Jullien, R.: Phys. Rev. B 64, 134303 (2001).
72. Poe, B. T., MeMillan, P. F., Rubie, D. C., Chakrabarty, S., Yarger, J., Diefen-
bacher, J.: Seience 267, 1245 (1997).
73. Bennemann, C., Baschnagel, J., Paul, W., Binder, K: Comp. Theor. Polymer
Sei. 9, 217 (1999).
74. Binder; K, Baschnagel, J., Bennemann, C., Paul, W.: J. Phys.: Condensed
Matter 11, A47 (1999).
75. Doi, M., Edwards, S. F.: The theory 01 polymer dynamics (Oxford, Clarendon
Press, 1986).
76. Müller-Plathe, F., Rogers, S. C., van Gunsteren, W. F. : Macromolecules 25,
6722 (1992); J. Chem. Phys. 98, 8995 (1993).
77. Gusev, A. A., Suter, U. W.: J. Chem. Phys. 99, 2221 (1993); J. Chem. Phys.
99, 2228 (1993).
78. Gusev, A. A., Müller-Plathe, F., van Gunsteren, W. F., Suter, U. W.: Adv.
Polymer Sei. 116, 207 (1994).
79. Weber, H., Paul, W.: Phys. Rev. E 54, 3999 (1996).
80. Paul, W., Weber, H., Binder, K: Ann. Phys. Leipzig 7, 554 (1998).
81. Stauffer, D., Aharony, A: Introduction to percolation theory (Taylor and Franeis,
London, 1992).
82. Varnik, F., Baschnagel, J., Binder, K: J. Chem. Phys. 113, 444 (2000).
83. Varnik, F., Baschnagel, J., Binder, K: Phys. Rev. E 65, 021507 (2002).
84. Jabbarzadeh, A., Atkinson, J., Tanner, R.: J. Chem. Phys. 110, 2612 (1999).
85. Bocquet, L., Banat, J.-L.: Phys. Rev. E 49, 3079 (1993).
86. Urbain, G. : Geochim. Cosmochim. Acta 46, 1061 (1982).
87. Todd, B., Evans, D.: Phys. Rev. E 55, 2800 (1997).
88. Mazurin, O. V., Streltsina, M. V., Shvaikovskaya, T. P.: Handbook 01 glass data,
Part A: Silica, Glass and Binary silicate glasses (Elsevier, Amsterdam, 1983).
89. Sadiq, A., Binder, K: Surface Sei. 128, 350 (1983).
 U nusual Viscosity Feature in Spinodal
Decomposition U nder Shear Flow

Jian Wangt,:I:, Wolfram Gronski+, Christian Friedrich+, Peter Galenkot

and Dieter Herlacht

t: Institute for Space Simulation, German Space Research Center, D-551147

Koeln, Germany
:j:: Freiburger Materialforschungszentrum, D-79104 Freiburg i. Er. Germany

1 Introduction

The process of phase separation of quenched binary systems, for example

quenched binary aHoys or polymer blends, is generally known as spinodal
decomposition [1]. It is weH known that domain growth in spinodal decom-
position foHows scaling power-Iaw. Hydrodynamic effect plays an important
role in spinodal decomposition. For example, without hydrodynamic effect,
domain grows foHowing a scaling power-Iaw with exponent 1/3, with hydro-
dynamics the exponent increases to 1 [2].
Under steady shear flow, the domain is usuaHy deformed along the shear
direction. The anisotropoic domain morphology leads to noval rheological
properties of the system. Both theoretical analysis and computer simulation
in 2d [3] showed that the extra viscosity due to the deformed domains un-
dergoes a maximum and then decreases. The maximum was supposed to be
the point when the stretched domains start to break up into small pieces.
Experimental results confirmed the signle peaked shear viscosity [4].
However, there were also experimental results that the viscosity developed
into two peaks [5]. Since early computer simulation by Ohta, Nozaki and Doi
[3] did not consider the hydrodynamic effect, in view of the strong effect of
the hydrodynamics on the speed of domain growth, there was great expecta-
tion that the hydrodynamic effect might also be responsible to the two peaks
in shear viscosity. Zhang, Zhang and Yang [2] recently made simulation com-
bining the phase seperation equation with the Navier-Stokes equation. They
observed two peaks in their simulations. However, it seems that their sim-
ulation stoped too soon to identify the second peak, and one might as weH
expect additional peak after the second. In theory[6], it has been shown that
even without the hydrodynamic equation, the phase seperation alone can re-
sult in multiple peaks in shear viscosity. The reason is that stretching and
and breakup can occur cyclicaHy. In experiemnt, wether a peak shows up also
depends on the resolution.

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
250 J. Wang et al.

2 Theory
Spinodal decomposition is weIl described by the Cahn-Hilliard equation, or
the time-dependent Ginzburg-Landau equation by choosing a phenomenolog-
ical free energy, for example [6,7]

(1)

The order parameter 'Ij; is defined as the volume fraction difference between
the two components of the binary system. The function f = - ~'Ij;2 + t'lj;4 is
a typical double-weIl potential, which implies that the two seperated phases
are more stable than the mixed one. The term !(V'Ij;)2 is the contribution to
the free energy due to the domain interface. The Langevin equation for the
evolvlution of order parameter is

(2)

where J.L(r, t) = -'Ij; + 'lj;3 - V 2 '1j; is the chemical potential, M is the mobility.
Considering convection, the above phase seperation equation may be mod-
ified as

(3)

Efficient algorithm, such as cell-dynamical system [8], has been developed to

solve the phase seperation equation.
For liquid, the velocity field may be derived from Navier-Stokes equation,
8v 2
P8i="'V v+'Ij;VJ.L-Vp (4)

where P is the mass density, ", is the liquid viscosity, and p is the pressure. In
the study of phase seperation of polymer blends, the Navier-Stokes equation
is usually approximated by neglecting the left hand side density term [2,7],

(5)
Applying the incompressible condition,

V·v=o (6)
one obtains a Poisson equation for the pressure

(7)

Once the pressure is solved, the velo city field can be obtained from the ap-
proximate Navier-Stokes equation above, which is also in the form of Poisson
equation for each component of the velocity.
 Unusual Viscosity Feature in Spinodal Decomposition Under Shear Flow 251

3 Simulation Detail
In case of steady shear experiment, there is an additional external velo city
field, V x = sy, where s is the shear rate and shear is applied in the x-direction.
Shear causes strain as time goes 'Y = st, where t is the time. To model the
strain effect in the simulation, Ohta, Nozaki and Doi [3] have proposed a
shear boundary condition,

(x, y, z) ---t (x + 'YL, y + L, z) (8)

The shear is applied in two (x,z) planes seperated by a distance Land in the
x direction. The strain will eventually increase to such a high level as time
passes, that the domians will be stretched into the shear direction. Periodic
boundary condition is applied in the x and z directions in simulation.
For periodic boundary system, one can solve the Poisson equation either
with Fourier transform method or real-space numerical method. For system
under shear boundary condition, the real-space numerical method is more
natural. In addition, quite efficient real-space algorithms, such as multigrid
methods [9], are available to solve the Poisson equation. In earlier simulations
[2], the phase seperation equation was solved under shear boundary condition,
while the Navier-Stokes equation was solved with Fourier transform method
which still implying using of periodic boundary condition. This might be in-
consistent. For periodic system of spinodal decomposition, we found the same
scaling growth power-Iaw of exponent 1 when the hydrodynamics equation
was solved with the real-space method, as Zhang, Zhang and Yang [2] ob-
served when the hydrodynamics equation was solved with Fourier transform
method.
To check how important the Navier-Stokes equation in explaining the
shear data, we made simulations with different shear rates and initial equi-
librium time steps. But we were still unable to convince ourselves an imme-
diate and apparent correlation between hydrodynamics and the two peaks in
shear viscosity. Often the second peak was too small to compare with that in
experiment.
The shear viscosity in computer simulation is defined as [3],

1 81/J 81/J
L1.'f} = -- < -, - > (9)
s 8x 8y
where s is the shear rate. L1.'f} is different from the liquid viscosity 'f}
introduced in the Navier-Stokes equation above. For uniform field, L1.'f} will
be zero, but 'f} may not be zero. L1.'f} is a quantity being able to characterize
the anisotropie morphology.

After switching on the shear, the domains start to strech in the shear
direction, which cause the morphology to change from uniform to anisotropy,
see Fig. 1. The black domain is 1/J > 0, and the white domain is 1/J < O. (or
252 J. Wang et al.

vice versa). As the shear viscosity reaches the maximum, the fully stretched
domains break up, which causes the shear viscosity to decrease. At this stage,
stretching and breakup coexist. It is likely that a second peak will occur, but
its magnitude can be quite different depending on the portions of stretching
and breakup domains. In experiment [5], the second peak appeared if the
shear was applied at a delayed time after quench. That inital equilibrium
allowed the domains to grow to such a narrow size distribution that it favours
most domains to stretch and break up synchronously. This synchronizing will
enhance the second peak to rise.However, in our computer simulations, simple
adjustment of the initial equilibrium time steps did not readily lead to an
apparent enhanced second peak.

Fig. 1. Typical anisotropical growth of the domain in spinodal decomposition under

shear.

In shear experiment, the distance between the shear planes (in y direc-
tion) is usually much smaller than the dimension in other direction (x and z
directions). Observing this geometrie setup of experiment, we designed sim-
ulations with a number of selected thicknesses. Fig. 2 displayes result with
 Unusual Viscosity Feature in Spinodal Decomposition Under Shear Flow 253

0.06 r----~------r---------_,

0.04

g-
~..,...
0
0.02

;>

-0.02 '------~------'-----~----~
o 10 20
Time t

Fig. 2. Shear viscosity for a system of lOOxlOOxlOO.

a grid of 100 x 100 x 100. The shear rate was 0.001 and initial equilibrium
time steps was 300. The results was obtained by solving the phase seperation
equation only without including the Navier-Stokes equation.
Fig. 3 displays simulation result on a grid IOOx50xIOO. The differenee of
Fig. 3 from Fig. 2 is the domain size. It is obvious that the boundary has a
signifieant effect on the peaks in shear viseosity. Reducing distanee between
the shear plane L has also resulted in reduetion of the periodicity in the
shear direction (x = X + 'Y L). That aeeeierated the seeond peak to appear in
experiment. The 3d simulation result is in agreement with other theoretical
prediction [6] that the phase seperation alone, without the hydrodynamic
equation, ean result in multiple peaks in shear viseosity. Experiments only
observed two peaks beeause the resulotion was not sensitive enough at the
beginning and the tail.

4 Conclusion
We simulate spinodal deeomposition under the shear fiow and explain the
unusual multiple peaks observed in viseosity measurement. Previously the
extra peaks have been attributed to the hydrodynamies effeet of the liquid.
Aeeording to the sealing exponent of 1, hydrodynamics tends to enhanee the
growth of domain during spinodal deeomposition. But this effeet is more or
less isotropie, so one may expect little eontribution to the peaks in shear vis-
254 J. Wang et al.

0.00 ,-------~----~--_r--------~--~--_,

0.02
.a
~
~
u
:;'"
0.01

10 20
Time t

Fig. 3. Shear viseosity for a systern of 100 x 50 x 100.

cosity. We found, on the other hand, the distance between the shearing planes
plays a significant role in the appearence of peaks observed in experiment.
This work was made under support from DFG-German Research Foun-
dation, research project He 1601/13.

References
1. K. Binder, in: Phase Transition in Materials, edited by R. W. Cahn, P. Haasen
and E. J. Kramer, Materials Scienee and Technology Vol. 5 (VCH Weinbeirn,
New York, 1990).
2. Z. Zhang, H. Zhang and Y. Yang, J. Chern. Phys. 113, 8348 (2000).
3. T. Ohta, H. Nozaki and M. Doi, , J. Chern. Phys. 93, 2664 (1990); Phys. Lett.
A145, 304 (1990).
4. J. Läuger, Chr. Laubner and W. Gronski, Phys. Rev. Lett. 75, 3576 (1995)
5. W. Gronski, J. Läuger, and Chr. Laubner, J. Mol. Struet. 383, 23 (1996) Chr.
Laubner, Diplornarbeit, Albert-Ludwigs University, Freiburg, 1996
6. F. Corberi, G. Gonnella and A. Lamura, Phys. Rev. Lett. 81, 3852 (1998)
7. H.Chen, and A. Chakrabarti, J. Chern. Phys. 108,6006 (1998).
8. A. Shinozaki and Y. Oono, Pbys. Rev. A45, R2161 (1992); Phys. Rev. E48,
2622(1993).
9. A. Brandt, Mathernaties in Cornputation 31, 333 (1977)
 Micro-macro Approach to Cluster Formation in
Granular Media

S. Luding

Institute for Computer Applications 1, University of Stuttgart,

Pfaffenwaldring 27, 70569 Stuttgart, Germany
Particle Technology, DelftChemTech, TU Delft,
Julianalaan 136, 2628 BL Delft, The Netherlands
e-mail: [email protected]@tnw.tudelft.nl
homepage: www.ical.uni-stuttgart.de/-lui

Abstract
Dissipation in granular media leads to interesting phenomena as there are
cluster formation and crystallization in non-equilibrium dynamical states.
The freely cooling system is examined concerning the energy decay and
the cluster evolution with time. Finally, an example is given for a possible
micro-macro transition with respect to the flow variables density, velocity,
and fluctuation kinetic energy.

Keywords:
inhomogeneous free cooling, clustering, cooperative phenomena, event-driven
molecular dynamics, micro-macro transition

1 Introduction
mteresting phenomena can be observed when granular media are studied
[1-4]. The subject of this paper is the pattern formation via clustering in
a dissipative, freely cooling system [5-8]. The special behavior of granular
media is connected to its ability to form a hybrid state between a fluid and a
solid: Energy input leads to a reduction of the density due to more collisions,
so that the material can flow, Le. it becomes 'fluid'. On the other hand, in
the absence of energy input, granular materials 'solidify' due to dissipation.
This makes granular media an interesting multi-particle system with a rich
phenomenology, however, theoretical approaches are non-classical and appear
often extreemly difficult, so that there is still active research directed towards
the understanding of granular media.
The basic ingredients of a granular model system are discussed briefly in
section 2. The inhomogeneous cooling and clustering are described in detail in
section 3. The basic idea is that in a freely cooling granular gas, fluctuations
in density and temperature cause a position dependent energy loss. Due to
strong local dissipation, pressure and energy drop rapidly and material moves

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
256 S. Luding

from 'hot' to 'cold' regions, leading to even stronger dissipation and thus
causing the density instability with ever growing clusters.
In order to describe and understand such inhomogeneous, non-equilibrium
systems, a particle simulation approach may be insuflicient due to the enor-
mous number of particles involved. However, a hydrodynamic theory of such
dissipative granular fiows is not available yet. Therefore, a first step towards
a micro-macro transition that it neccessary to understand the hydrodynamic
theory of granular media is presented. The density-, velocity-, and fiuctuation
kinetic energy-fields will be computed and discussed for the case of strong
clustering.

2 Models for Multi-particle Simulations

The constituents of granular media are mesoscopic particles. When those ob-
jects interact (collide) the attractive potentials of the individual grains can
be neglected. Two models for the repulsive particle-particle interactions are
discussed in the following. They account for the excluded volume of the par-
ticles via a repulsive potential, either 'hard' or 'soft' and also account for
dissipation in collisions via some coeflicient of restitution. The third ingre-
dient of a model granular material is friction which couples the rotational
degrees of freedom to the linear motion, but it is not discussed in this paper.
The difference between the two most frequently used discrete element
methods is the repulsive interaction potential. For the molecular dynamics
(MD) method, soft particles with a power-law interaction potential are as-
sumed, whereas for the event-driven (ED) method perfectly rigid particles are
used. The consequence is that the duration of the contact of two particles,
t e , is finite for MD, but vanishes for ED.

2.1 The Event-driven, Rigid Particle Method

(a) (b)
Fig.l. Typical velocities oftwo particles immediately before (a) andjust after (b)
collision.
 Micro-macro Approach to Cluster Formation in Granular Media 257

Consider two particles with diameter d l and eh. and masses ml and m2,
The normal unit vector for their contact is n, and Ti is the vector to the
position of the center of particle i (i = 1, 2). The relative velocity of the
contact points is V c = Vl - V2, with the velocity Vi of particle i. From
momentum conservation it follows

(1)

where V~ is the unknown velo city of particle i after collision. The change of
linear momentum of particle 1 is a function of the coefficient of restitution r:

(2)

with the reduced mass m12 = mlm2/(ml + m2).

For the simulation of rigid particles, an event-driven method is used
[9,10,7]. The particles undergo an undisturbed motion until an event occurs.
An event can be either the collision of two particles or the collision of a parti-
cle with a wall. From the velocities just before contact, the particle velocities
after a contact are computed following (1). Lubachevsky [9] introduced an
efficient scalar ED algorithm which updates only those particles involved in
the previous collision. The original algorithm is implemented and generalized
to take into account dissipation.

2.2 The Time Driven, Soft Particle Technique

Even without using the soft particle method [U-13] in this study, it is con-
venient to discuss briefiy the standard approach. Replacing LlP in (1) by
j(t)LltMD, with the molecular dynamics time step LltMD, allows the inte-
gration of the corresponding, discretized equations of motion with standard
numerical methods [12].
Since the modeling of realistic deformations of the particles would be
much too complicated, let us assume that the overlap of two particles is the
only quantity important for the interaction potential. The interaction of two
particles can be split into (at least) three independent forces, and is typically
short range, Le. the particles interact only when they are in contact. The first
force, an elastic repulsive force proportional to the overlap, accounts for the
excluded volume which each particle occupies. In the simplest case, a linear
spring can be used. The second force, a viscous damping force, models the
dissipation in the normal direction and is proportional to the relative velocity.
The simplest possible dashpot is again linear. This linear spring-dashpot
model can be solved analytically and leads to a constant contact duration t c
and a constant restitution coefficient r [14]. The third force, accounting for
friction, acts in the tangential direction, but will not be discussed here; for
more information see Refs. [1].
258 S. Luding

2.3 The Connection Between Hard- and Soft-sphere Models

In the ED method, the time during which two particles are in contact is
implicitly zero. The consequence is that exclusively pair contacts occur and
the instantaneous momentum change i1P in (1) suffices to describe the col-
lision. However, ED algorithms with constant r run into difficulties when
the time between events, tn, becomes too small - typically in systems with
strong dissipation - and the so-called ''inelastic collapse" occurs [6,15], i.e.
the collision rate diverges for a few particles in the system. Since this is an
artefact of the hard sphere model, it is unphysical and has to be avoided.
Because a diverging number of collisions is only possible if the contact du-
ration vanishes, the physical contact duration t e has to be reintroduced in
order to allow for realistic ED simulations. In MD simulations, on the other
hand, one has t e > 0, since every contact takes some finite time. Therefore,
only a limited amount of kinetic energy (i1E oe 1- r 2 ) can be dissipated per
collision. A finite contact duration implies a finite energy dissipation rate. In
contrast, the consequence of a diverging collision rate would be a diverging
energy dissipation rate.
In a dense system of real particles, energy dissipation becomes ineffective,
Le. the 'detachment effect' occurs [16,17]. This effect is not obtained with
hard particles and a constant coefficient of restitution r. Therefore, in the
framework of the so-called TC model, the restitution becomes elastic in na-
ture, r~) = 1, if collisions occur too frequently, Le. t~) :::; t e , for the collision n
of particle i. The time since the last collision is t~) and the cut-off parameter
t e for elastic contacts can be identified with the contact duration. Thus, an
additional material parameter is defined for the hard sphere model, that leads
to qualitative agreement between ED and MD simulations and, in addition,
avoids the inelastic collapse artefact. Recently, it has been shown that the
TC model does not affect physical observables of the system, like the energy,
as long as it is reasonably small [15].

3 Freely Cooling Granular Media

The simulations presented in the following involve N = 99856 = 3162 dis-

sipative particles with the restitution coefficients r = 0.9, 0.8, and 0.6, in a
periodic, quadratic system with volume fraction v = 0.25. The system size is
l = Ld with dimensionless size L = 560 and particle diameter d = 1 mm. In
order to reach an equilibrated initial condition, the system is first allowed to
equilibrate with r = 1 for several hundreds of collisions per particle, so that
a Maxwellian velo city distribution and a homogeneous density can be found.
Then, at t = 0 s, dissipation is activated and the quantities of interest are
examined. Snapshots of the simulation with r = 0.8 are presented in Fig. 2
at different, rescaled times T (see below).
 Micro-macro Approach to Cluster Formation in Granular Media 259

T = 0.047 T = 23.7

T = 380 T = 1519

T = 6076 T = 12150

Fig. 2. ED simulation with N = 99856 particles in a system of size L = 560, volume

fraction v = 0.25, restitution coefficient r = 0.8, and critical collision frequency
t;;1 = 10 5 S-I. The collision frequency is color-coded: red, green and blue correspond
to collision rates t;;1 ::::! 50 S-I, 10 S-1 and 2 S-I, respectively.
260 S. Luding

The first picture in Fig. 2 is taken in the initally homogeneous cooling

regime, whereas the next pictures show the different stages of the cluster
growth regime. The final snapshots are taken in the limiting state, where the
cluster has reached the system size. The parlicles are colored spots, where the
green/red areas in the cluster centers correspond to parlicles with collision
rate t;;l ~ 50 S-l. This is much smaller than the critical collision rate t;l =
105 s-l, so that only a very small number of parlicles will be affected by the
TC model.

3.1 Homogeneous and Inhomogeneous Cooling

In the homogeneous cooling state [18,6,19], one can expect that the kinetic
energy E = K (t) / K (0) of the system decays with time (the decay of energy
with time is also evidenced by the change of color from red over orange to
green and blue in Fig. 2) and follows the master-curve

1
E(7) = (1 + 7)2 ' (3)

with the dimensionless time 7 = (1 - r 2 )t/(4tE), the collision rate tEl =

8vg(v)v/(..fiid), the mean velocity v = VK(t)/Nm, and the increased con-
tact probability g(v) = (1 - 7v/16)/(I- v)2 due to excluded volume effects
at finite volume fractions v. Inserting the paranleters from the simulation,
1 - r 2 = 0.36, g(v) ~ 1.5833, and v = 0.02883m/s, one obtains an initial
collision rate t:E/(O) = 51.5s- 1 (corresponding to the red color in Fig. 2).
In Fig. 3, the normalized kinetic energy E and the normalized collision
rate t;;l /t E/, are presented, both as a function of the normalized time 7. At
the beginning of the simulation we observe a perfect agreement between the
theory for homogeneous cooling and the simulations. At 7 ~ 10 - 40 sub-
stantial deviations from the homogeneous cooling behavior become evident,
Le. the decay of energy is slowed down earlier for stronger dissipation. The
deviation from the analytical form increases until 7 ~ 103 where the clusters
reach system size and the behavior of E changes again to a slightly more
rapid decrease.
This change in behavior is evident from the collision rate t;;l. At first, in
the homogeneous cooling regime, the collision rate decays with t;;l cx: -JE.
Then the collision rate is almost constant (for r = 0.9 and r = 0.8) or even
increases (for r = 0.6), until at 7 ~ 103 it becomes very noisy, indicating
another change in the collective behavior. The long-time power law for the
decay of energy with time is -2 in the homogeneous cooling case. In the
cluster growth regime, however, we obtain slopes slightly smaller than -1
(the best fit leads to -0.920, -0.927, and -0.941 for r = 0.9, 0.8, and 0.6,
respectively, with errors ±0.002).
 Micro-macro Approach to Cluster Formation in Granular Media 261

10-2 10-1
.......
~ ......
....~
10-2
10-4
- - 1-=<>-6 - - r=0-6
------- r=O.8 -------- r=O.8
- - r=O.9 10-3 - - r=O.9
10-2 102 104 10-2 102 104
1: 1:

Fig.3. (Left) Kinetic energy E plotted against r for different values of r = 0.9,
0.8, and 0.6. The dotted line represents (3). (Right) Normalized collision rate tE/tn
plotted against r. The dashed line represents the collision rate ..fE in the homoge-
nously cooling case. .

3.2 Cluster Structure

In Fig. 4, the pair-correlation nmction during the clustering is plotted in order

to visualize the short-range structure in the clusters. In Fig. 5, zooms into
the bottom-right area of the system in Fig. 2 are presented in order to give a
more direct picture. In the initial state, the system is rather disordered and
homogeneous. In the cluster growth regime, some particles approach closer
and form loose clusters, however, the structure is still disordered, fluid-like.
Only in the very late stage of the simulation, where the clusters are very
large, one obtains crystalline, triangular lattice structures with a peculiar
distribution of collision rates as color-coded.
This more qualitative picture can also be verified by computing the parti-
cle correlation function 9 (r / rI), see Fig. 4. The data are displayed for different
r-values, showing that the structure in the system becomes more and more
pronounced with increasing time. Like the zoom-in in Fig. 5, also the cor-
relation function shows that the crystalline structure occurs only in the late
regime where the clusters are very large, with 'frozen' cores. Pronounced
peaks at 1, ..;3, 2, ... indicate a triangular order of the particles.

3.3 Cluster Growth

The cluster growth can be studied quantitatively in the spirit of Luding and
Herrmann [7]. All particle pairs with a distance smaller than some cut-off
distance 8 < (1 + S)d, with an arbitrary cut-off parameter S = 0.1, are
assumed to belong to the same cluster. After all particle pairs are examined,
one obtains a cluster-size distribution and its moments. The first moment, the
262 S. Luding

,=24300
1:=6076
T=15 19

I
JO ,=~!iO
1:=95
~
..::
'-
00
5
\ i\\ J\ 1\
I I h
\ ....J '\""" )
.J
--,' " "')\, ..,1.
I"- '\

1
0
0 2 3 4 5
rld
Fig. 4. Correlation function g( r / d) as obtained from the simulation in Fig. 2 with
T as given in the inset.

T = 23.7 T = 47.4

T = 760 T = 1519

.. ..~~
..... '. - '';:
T = 3038 T = 6076

Fig. 5. Zooms into the lower right part of the ED simulation from Fig. 2.

mean cluster size (M), and the size of the largest cluster M max are plotted
in Fig. 6 against the time T.
 Micro-macro Approach to Cluster Formation in Granular Media 263

6 loS
- - r=O.6
_....-_. r=O.6 . ..
---0---- r=O.g
'

1'=0_8
5 ,,-,, ~ ,,-, r=O.9 .....•.... r=O.9
104
4
11 Jo1
~3 ),~
102
2
10 1

O U-_..L-_...L.-_...L.-_-'-_-" 1 "----'----'-----'-----'------"
10-1 10 1 102 101 104 10. 1
t

Fig.6. (Left) Mean cluster size and (Right) Maximum cluster size as funtions of
time T for different coefficients of restitution r.

Both values are almost constant in the initial, homogeneous cooling

regime. In the cluster growth regime a rapid increase of both (M) and M max
is evidenced until, at larger T, the values reach a maximum and seemingly
saturate or even decrease in the final regime where the clusters have reached
system size. The cluster growth start es earlier for stronger dissipation, but the
largest cluster seems to grow more rapidly for weaker dissipation, however,
at a later time.

3.4 Micro-Illacro Transition

In this subsection, the macroscopic field quantities density, velo city, and fluc-
tuation kinetic energy are discussed. In order to obtain the fields from the
simulation data, the system is arranged on a 50 x 50 grid and the field quan-
tities are computed for every cello The density v(x, y) = n c7fa 2 /Vc and the
fluctuation kinetic energy e(x, y) = (m/2) I:iEJVi -v(x, y)F /n c , are plotted
in Fig. 7. The symbols are the number ofparticles per cell n c , the cell-volume
Vc , the particle radius a, mass m, and the mean velo city v( x, y) = I : iEc Vi / n c
in the cello
The iso-lines of the density are chosen such that the dilute area (v < 0.1
- pink) is visible as weH as the coexistence area between fluid and solid
regions (0.6 < v < 0.8 - between blue and green). The green islands are solid
clusters with triangular arrangement, whereas the pink islands are almost
empty regions. Between pink and blue, the fluid phase exists at almost all
densities.
The iso-lines for the fluctuation kinetic energy are chosen in arbitrary
units. However, it is evident, that the temperature and the density fields
are not correlated in a simple fashion. Dense regions are not neccessarily
cold neither are dilute regions neccessarily hot. The structure of the e-field
appears more detailed than the structure of the density field.
264 S. Luding

Fig.7. Density and ßuctuation kinetic energy iso-lines for the snapshot at time
T = 6076 in Fig. 2. The density is given as volume fraction, whereas the energy is
given in arbitrary units.

Fig. 8. Momentum ßux Held vv for the snapshot with T = 6076 in Fig. 2.

Finally, the velo city vector field is plotted as momentum fiux field in
Fig. 8 and shows the mass-flux in the system. A doser examination leads to
the condusion that the clusters are not stahle, hut have remaining internal
motion due to shear- or compressive modes. Note that also the velocity field
(energy due to fiux motion) is not directly correlated to the fiuctuation kinetic
energy.
 Micro-macro Approach to Cluster Formation in Granular Media 265

4 Summary and Conclusion

With a rather simple description of a granular material as an ensemble of
inelastic hard disks (spheres), we have investigated the interesting effect of
clustering in freely cooling systems. For short times the system is disordered
and gas-like, whereas the structures at larger times are dense, crystalline
clusters. The clusters grow until they reach the system size. Simulations at
very long times were possible with the TC model, which reduces dissipation
when contacts become too frequent.
As a first step towards a macroscopic, hydrodynamic description of clus-
tering in freely cooling granular media, the density, the velocity and the fluc-
tuation kinetic energy were computed from the simulation snapshots. The
density field is directly correlated with the snapshot, whereas neither veloc-
ity nor fluctuation kinetic energy are directly correlated with the density.
The latter fact implies that all three quantities are independent variables,
neccessary for a hydrodynamic description of the system.
Further investigations concerning the transport parameters pressure, vis-
cosity and heat conductivity in similar systems are in progress in two dimen-
sional model systems and are to be applied to a successful hydrodynamic
theory of granular flow.

Achnowledgements
We acknowledge the financial support of the Deutsche Forschungsgemein-
schaft (DFG) and of the Max-Planck PKS in Dresden.

References
1. Herrmann, H. J., Hovi, J. -Po and Luding, S.: editors. Physics oi dry granular
media - NATO ASI Series E 350, Dordrecht, 1998. Kluwer Academic Publish-
ers.
2. Pöschel, T. and Luding, S.: editors. Granular Gases (Berlin, 2001. Springer)
Lecture Notes in Physics 564.
3. Vermeer, P. A., Diebels, S., Ehlers, W., Herrmann, H. J., Luding, S. and Ramm,
E.: editors. Continuous and Discontinuous Modelling oi Cohesive Frictional
Materials (Berlin, 2001. Springer) Lecture Notes in Physics 568.
4. Kishino, Y.: editor. Powders (3 Grains 2001, Rotterdam, 2001. Balkema.
5. Goldhirsch, I. and Zanetti, G.: Clustering instability in dissipative gases. Phys.
Rev. Lett., 70(11):1619-1622, 1993.
6. McNamara, S. and Young, W. R.: Dynamics of a freely evolving, two-
dimensional granular medium. Phys. Rev. E, 53(5):5089-5100, 1996.
7. Luding, S. and Herrmann, H. J.: Cluster growth in freely cooling granular
media. Chaos, 9(3):673-681, 1999.
8. Cafiero, R., Luding, S. and Herrmann, H. J.: Two-dimensional granular gas of
inelastic spheres with multiplicative driving. Phys. Rev. Lett., 84:6014-6017,
2000.
266 S. Luding

9. B. D. Lubachevsky, B. D.: How to simulate billards and similar systems. J. 01

Comp. Phys., 94(2):255, 1991.
10. Luding, S.: Granular materials under vibration: Simulations of rotating spheres.
Phys. Rev. E, 52(4):4442, 1995.
11. Cundall, P. A. and Strack, O. D. L.: A discrete numerical model for granular
assemblies. Geotechnique, 29(1):47-65, 1979.
12. Allen, M. P. and Tildesley, D. J.: Computer Simulation 0/ Liquids. Oxford
University Press, Oxford, 1987.
13. Lätzel, M., Luding, S. and Herrmann, H. J.: Macroscopic material proper-
ties from quasi-static, microscopic simulations of a two-dimensional shear-cell.
Granular Matter, 2(3):123-135, 2000. cond-mat/0003180.
14. Luding, S.: Collisions & contacts between two particles. In H. J. Herrmann,
J.-P. Hovi, and S. Luding, editors, Physics 0/ dry granular media - NATO ASI
Series E950, page 285, Dordrecht, 1998. Kluwer Academic Publishers.
15. Luding, S. and McNamara, S.: How to handle the inelastic collapse of a dis-
sipative hard-sphere gas with the TC model. Granular Matter, 1(3):113-128,
1998. cond-mat/9810009.
16. Luding, S., Clement, E., Blumen, A., Rajchenbach, J. and Duran, J.: Anoma-
lous energy dissipation in molecular dynamies simulations of grains: The "de-
tachment effect". Phys. Rev. E, 50:4113, 1994.
17. Luding, S., Clement, E., Blumen, A., Rajchenbach, J. and Duran, J.: Interac-
tion laws and the detachment effect in granular media. In Fractal Aspects 01
Materials, volume 367, pages 495-500, Pittsburgh, Pennsylvania, 1995. Mate-
rials Research Society, Symposium Proceedings.
18. Haff, P. K.: Grain How as a Huid-mechanical phenomenon. J. Fluid Mech.,
134:401-430, 1983.
19. Luding, S., Huthmann, M., McNamara, S., and Zippelius, A.: Homogeneous
cooling of rough dissipative particles: Theory and simulations. Phys. Rev. E,
58:3416-3425, 1998.
 Part 111

TRANSPORT
 Urban Transport Phenomena in the Street
Canyon

Maciej M. Duras

Cracow University of Technology, Institute of Physics, ulica Podchorazych 1,

PL-30084 Cracow, Poland

Abstract. A field deterministic model of the vehicular dynamics in a generic urban

street canyon with two neighboring canyons is considered. The assumed hydrody-
namical model of vehicular movement is coupled to the gasdynamical model of the
air and emitted pollutants. The vehicles are assumed to move on distinct left lanes
and right ones. At the upstream and downstream ends there are coordinated trafIic
lights which introduce control parameters to the field model. The model of optimal
control of street canyon dynamics is based on the two optimal multi-criteria control
problems. The problems consist of minimization of the dimensionless functionals of
the cumulative total travel time, global emissions of pollutants, and global concen-
trations of pollutants, both in the studied street canyon, as weIl as together with
the two nearest neighbor substitute canyons, respectively.

1 The Field Models of Vehicles and Pollutants

In the present article we consider two coupled deterministic field models of
the urban street canyon. The vehicular fields are one-dimensional in spatial
variable x and they depend on time variable t. The considered fields are the
vehicular number density kt vt(x, t), vehicular velo city wt vt(x, t), emissivity
of exhaust gases e{,ct,vt(x, t), 'and vehicular heat emissivity ::ri,vt(x, t), where vt
is emission type's number, ct is number of the constituent of emitted exhaust
gases. It is assumed that there are nl = nL left lanes s = 1 and l is the left
lane's number, l = 1, ... , nL, whereas for n2 = nR right lanes s = 2 and l is the
right lane's number, l = 1, ... , nR, vt = 1, ... , VT, ct = 1, ... , CT. VT is total
number of types of vehicular emissions, and CT is total number of emitted ex-
haust gases. The above mentioned model consists of the equations of balances
of vehicular numbers [5] and vehicular equations of state (Greenshields' equi-
librium speed-density u-k model [6]) as weIl as exhaust gas emissivities and
heat emissivities. The vehicular flow in the canyon is multilane bidirectional
one-Ievel rectilinear, and it is considered with two systems of coordinated sig-
nalized junctions at the upstream and downstream ends oflanes [1],[2], [3], [4].
The considered vehicles belong to distinct vehicular classes: passenger cars,
and trucks. Emissions from the vehicles are based on technical measurements
of the following types of pollutants: carbon monoxide CO, hydrocarbons HC,
nitrogen oxides NO x • The field model ofpollutants is gasdynamical. The gases
are Newtonian viscous perfect and noninteracting. The considered fields are

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
270 M. M. Duras

three-dimensional in spatial variables (x, y, z) and one-dimensional in tempo-

ral variable t and they are mixture's density p(x, y, z, t), velo city v(x, y, z, t),
temperature T(x, y, z, t), pressure p(x, y, z, t), as weH as constituents' mass
concentrations Ci(X, y, z, t), and partial pressures pi(X, y, z, t), i = 1, ... , N,
where N is total number of mixture's elements. (N = NE - 1 + NA)' The
first NE - 1 = 3 gases are the exhaust gases emitted by vehicle engines
during combustion CO, CH, NO x . The remaining NA = 9 gases are the con-
stituents of air: O 2 , N2, Ar, CO 2, Ne, He, Kr, Xe, H2 • The field equations are
the balances of mixture's mass, linear momentum, energy, and equation of
state (Clapeyron's law) as weH as the balances of masses of constituents
and constituents' equations of states (Dalton's law). The studied sources are
mixture mass source Sex, y, z, t), mixture energy source a(x, y, z, t), and con-
stituents' mass sources Sf(x,y,z,t) (exhaust gases and consumed oxygen).
The boundary-initial problems are of mixed types (Dirichlet's and von Neu-
mann's types). The equations of dynamics are solved by the finite difference
scheme. The two separate multi-criteria optimization problems are posed by
defining the dimensionless functionals of cumulative total travel time, global
emissions of poHutants, and global concentrations of poHutants, either in the
studied street canyon or in the canyon and its two nearest neighbor sub-
stitute canyons. The vector of control is a five-tuple composed of two cycle
times, two green times, and one offset time between the trafik lights. The
optimal control problems consists of minimization of the two functionals over
the admissible control domain manifold.

2 The Governing Equations

Under the abovementioned specifications, the set of governing equations is

formulated as follows [7], [8]:

8k S
~
l,vt + d'IV (kl,vt
S wl,vt
S) = 0
ki vt(x, t)
wi,vt(x,t) = (wi,vt,f' (1- k~ ),0,0)
l,vtJam

~ + div(pv) = S

p(~: + (v 0 \7)v) + Sv = -\7p + 'T]LJ.V + (~+ ~)\7(diVV) + F

p(~: + v 0 \7E) = -( _~v2 + E)S + T : \7v + div( -q) + a
p. = ~.T
p mair
 Urban Transport Phenomena in the Street Canyon 271

OC'
p( ot~ +v 0 \7 Ci) = sf - CiS +
N-l
+ :E {(D im - DiN)' div[p\7(cm + k~m\7T)]}
m=l
mair
Pi = Ci . - - . p,
mi

where wi,vt,f is maximum free flow speed, ki,vtJam is jam vehicular density,
'T} e
is the first viscosity coefficient for the air, is the second viscosity coeffi-
cient, F = pg is the gravitational body force density, g is the gravitational
acceleration, D im = D mi is the mutual diffusivity coefficient from the i-th
constituent to m-th one, and D ii is the autodiffusivity coefficient of the i-th
constituent, and kT,m is the thermodiffusion ratio of the m-th constituent, E
is the mass density of intrinsic (internal) energy of the air mixture, T is the
stress tensor, symbol : denotes the contraction operation, q is the vector of
flux of heat. We assume that [2]:

. D ( kT,i nT )
Ji = -p ii Ci +T v ,

[pD ii ]
272 M. M. Duras

where Ei is the mass density of intrinsic (internal) energy of the ith con-
stituent of the air mixture, mi the molecular mass of the ith constituent,
k B is Boltzmann's constant, /1-i is the complete partial chemical potential of
the ith constituent of the air mixture (it is complete since it is composed
of chemical potential without external force field and of external potential),
mair = 28.966 [u] is the molecular mass of air, 8mk is Kronecker's delta, Cp,i
is the specific heat at constant pressure of the ith constituent of air mixture,
h is Planck's constant, ji is the vector of flux of mass of the ith constituent of
the air mixture, and '" is the coefficient of thermal conductivity of air. These
magnitudes were derived from Grand Canonical ensemble with external grav-
itational Newtonian field.

3 Optimal Control Problems

Let us consider the measurs of the total travel time (TTT): hTT [5], emis-
sions (E): JE , and concentrations (C): Je of exhaust gases in the street
canyon as weH as in the canyon and two its nearest neighbor substitute
canyons, JTTT,ext JE,ext JC,ext, respectively. Therefore the appropriate op-
timization problems may be formulated as foHows [2]. The measures depend
on traffic lights at the upstream and downstream ends of the lanes. The vector
of boundary control U reads:
U = (gI, C1 ,g2, C2 , F) E uadm ,
gm are green times, Cm are cycle times, F is offset time, and uadm is a set
of admissible control variables. The TTT, E, C functionals are given by:
VT l l
hTT(U) = 2:2: 2:
2 n. a Ts
kf,vt(x,t)dxdt.

ll
s=1 1=1 vt=1 0 0
Ts
= 2: 2: 2: 2:
2 n. CT VT a
JE(U) ei,ct,vt(X, t)dx dt.
s=1 1=1 ct=1 vt=1 0 0

Jc(u) = PSTP' N fi=1

1 ra rb r roTs Ci(X,y, z,t)dxdy dzdt.
Jo 10 Jo J
2 n. VT
JTTT,ext(U) = hTT(U) + a· L 2: 2: ki,vtJam . (C s - g8)'
s=1/=1 vt=1
2 n. CT VT
JE,ext(U) = JE(U) + a· L L L L el,ct,vtJam . (Cs - gs).
s=1 1=1 ct=1 vt=1
NE-1 2

JC,ext(U) = Jc(U) + PSTP . a· b· C· L Ci,STP' L(Cs - g8)'

i=1 8=1

The integrands kf vt> ef ct vt, Ci in the functionals depend on the control vector
U through the b~unda:ry conditions, through the equations of dynamics, as
 Urban Transport Phenomena in the Street Canyon 273

weIl as, through the sources. The value of the vector of control u directly
affects the boundary conditions for vehicular densities, velocities, and emis-
sivities. It also affects the sources. Next, it propagates to the equations of
dynamics and then it influences the values of functionals. PSTP is the density
of air at standard temperature and pressure STP, Ci,STP is concentration of
the ith constituent of air at standard temperature and pressure, a, b, c, are
dimensions of the street canyon, Ts is time of simulation.
We define two additional functionals. The first one is the composite to-
tal travel time, emissions and concentrations of the poIlutants in the single
canyon, whereas the second one in the canyon and two its neighbours:

H(a, u) = aTTTFTTT(U) + aEFE(u) + acFc(u).

Hext(a, u) = H(a, u) +
+aTTT,extJTTT,ext(U) + aE,extJE,ext(U) + ac,extJC,ext(u).

The scaling parameters aTTT, aE, ac, aTTT,ext, aE,ext, aC,ext make the two
functionals dimensionless, and

a = (aTTT, aE, ac) E Aadm,

aext = (aTTT, aE, ac, aTTT,ext, aE,ext, aC,ext) E A!~f,
where Aadm, A:~f are the sets of admissible scaling parameters. We formulate
two separate multi-criteria optimization problems consisting in minimization
of the functionals H, Hext with respect to control vector u over admissible do-
main and with respect to scaling parameters, while the equations of dynamics
are fulfilled:

H* = H(a*, u*) = min{a E A adm , u E U adm : H(a, u)}

H:xt = Hext (a:xt , u:xt ) = min{aext E A!~f, u E U adm : Hext(a, u)},
where H*, H:xt , are the minimal values of the two functionals, (a*, u*), and
(a:xt , u:xt ) , are scaling and control vectors at which the functionals reach the
minima, respectively.

4 Conclusions

The proecological urban trafik control idea and the model of the street canyon
have been developed. It is assumed that the proposed model represents the
main features of complex air pollution phenomena.

5 Acknowledgements

The author would like to gratefully thank the organizers of the Comphy02
Computational Physics of Transport and Interface Dynamics conference:
274 M. M. Duras

Heike Emmerich, Britta Nestler and Michael Schreckenberg, held in the Max-
Planck-Institut für Physik komplexer Systeme, Dresden, Germany, for the
invitation and creation of scientific atmosphere. The present work was partly
covered by the Polish State Commitee for Sientific Research KBN grant num-
ber F-l/63/BW/02.

References
1. M. M.Duras. Continuum Field Model of Street Canyon: Numerical Examples.
Engng. Trans. 48 (2000) 433-448.
2. M. M. Duras. Road Traffic Control in Street Canyons, Ph. D. Thesis, University
of Mining and Metallurgy, Cracow. (1998) 1-363.
3. A. Adamski, M. M. Duras. Environmental Traffic Control Issues in Street
Canyons. Polish J. Environ. Stud. 6 (1997) 13-20.
4. A. Adamski, M. M. Duras. Air Pollution Optimal Trafiic Control Issues in Inte-
grated Street. Polish J. Environ. Stud. 8 (1999) 7-17.
5. P. G. Michalopoulos, D. E. Beskos, J.-K. Lin. Analysis of Interrupted Trafiic
Flow by Finite Difference Methods. Transp. Res.: Part B: 18B (1984) 409-421.
6. B. D. Greenshields. A study of traffic capacity. Proceedings of Highway Research
Board 14 (1934) 448-477.
7. A. Cemal Eringen, G. A. Maugin. Electrodynamics of Continua. Springer-Verlag,
New York (1990) Vol. I, Chapters 3, 5; Vol. II, Chapters 9, 10.
8. L. D. Landau, E. M. Lifshitz. Theoretical Physics: Hydrodynamics. Nauka,
Moscow (1986)Chapters 2, 6.
9. M. M. Duras. Continuum Field Model of Street Canyon: Theoretical Description.
Engng. Trans. 48 (2000) 415-432.
10. M. Brzezanski. Evaluation of Transport Emission of the Amount on a Chosen
Section of the Road Network in Cracow. KONMOT'98 (1998) Table 2.
 Walker Behaviour Modelling by Differential
Games*

Serge P. Hoogendoorn 1

Delft University of Technology, Faculty of Civil Engineering and Geoscienees

Abstract. Gaining insights into pedestrian How operations and assessment tools
for pedestrian walking speeds and eomfort is important for e.g. planning and geo-
metrie design of infrastruetural facilities, as weil as for management of pedestrian
Hows under regular and safety-critical circumstances. Pedestrian How operations
are complex, and vehicular How simulation modeling approaches are generally not
applicable to pedestrian How modeling.
This paper focusses on pedestrian walking behavior. It is assumed that pedes-
trians are autonomous predictive controllers that minimize the subjective predicted
cost of walking. Pedestrians predict the behavior of other pedestrians based on their
observations of the current state as weIl as predictions of the future state, given
the assumed walking strategy of other pedestrians in their direct neighborhood. As
such, walking can be represented by a (non-cooperative or cooperative) differential
game, where pedestrians may or may not be aware of the walking strategy of the
other pedestrians.

1 Introduction

In optimizing the design of airports, public transit stations, stadiums, shop-

ping malls, etc., theory and models predieting spatiotemporal pedestrian fiow
patterns and individually experieneed walking eonditions, in terms of walking
times, comfort levels, etc., are important. This holds equally for timetable de-
sign, where walking times experieneed by publie transit riders when accessing
or transferring play an important role. Analysis tools can support infrastrue-
ture designers as weil as publie transport planners to optimize their design,
making it more eflicient, more comfortable, and safer. In this respect, we em-
phasize that pedestrian trafik and car-based trafik are very different. This
is why fiow theories and models of different modes generally are not inter-
changeable, justifying the development of a dedieated comprehensive dynamie
theory for pedestrians with related models.
In the theory underlying the model development deseribed in this eontri-
bution, key elements in predicting pedestrian behavior are activity scheduling
(choice to execute a particular activity, the order in which activities are per-
formed, the loeation where to perform an aetivity), route choice between ac-
tivities in the two-dimensional eontinuous space, and multi-directional walk-
* This research is funded by the Social Science Research Council (MaGW) of the
Netherlands Organization for Scientmc Research (NWO).

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
276 S. P. Hoogendoom

ing behavior. The main theoretical assumption is that of the ''pedestrian eco-
nomicus". I.e. pedestrians are subjective expected cost minimizers.
The focus of this contribution will be on pedestrian walking behavior,
assuming that pedestrian routes, described by continuous trajectories, are
given. The theory is operationalized in terms of behavioral models, deter-
mined by application of differential game theory. The walker model is con-
tinuous in time and space, and describes walking behavior considering the
pedestrian state (position and velocity) and control (acceleration). Under the
assumption of expected (predicted) cost minimization, pedestrians choose the
acceleration optimizing the predicted subjective walking cost, thereby con-
sidering the future actions (and reactions) of the other pedestrians (so-called
opponents). The subjective walking cost depends on the costs of drifting
from the planned route, of interacting with other pedestrians and obstacles,
and the cost of applying the control. Assuming that pedestrians do not take
into account the anticipated reactions of the other pedestrians, it turns out
that the resulting feedback models are similar to the social-forces model of
[7]. The modelIing framework thus provides a different interpretation of the
latter model, and its implied behavioral assumptions.
This contribution is outlined as folIows. Section 2 presents the overall
theory, distinguishing different levels of pedestrian behavior. The third section
focusses on the behavioral assumptions underlying walker model derivation.
Section 4 discusses the conceptual walker model. The fifth section presents the
mathematical model derivation using differential game theory. Furthermore,
the analogy with the social forces model is discussed. Section 6 presents the
approach to model calibration, while in section 7 some application results are
discussed briefly. The final seetion presents the general eonclusions as weIl as
future research directions.

2 Pedestrian Behaviour Framework

The main behavioral assumption is that allactions, let it be performing

an activity, walking along a eertain route, or,bypassing a slower pedestrian
performed within the infrastructure facility '>will provide utility, which the
pedestrian aims to optimize. Several empiricaI studies have shown the ap-
plicability of utility-based approaches to modelIing pedestrian behavior [3],
[9], [16], even though most decisions are made subconsciously. Therefore, we
argue that its use is justified.
The theory distinguishes the following mutually dependent behaviorallev-
eIs:

1. Departure time choice, and activity set choice (strategie level);

2. Activity scheduling and activity area choice; route-choice to activity areas
(tactieallevel);
3. Walking behavior (operationallevel).
 Walker Behaviour Modelling by Differential Games 277

Figure 1 depicts the chosen overall framework of pedestrian behavior.

Given the departure time to and the prior activity set E a pedestrian aims
to perform, the theory asserts that at the tactical level, pedestrians make a
simultaneous route-choice / activity schedule decision minimizing expected
subjective dis-utility. These choices at this level are influenced by both exter-
nal factors (e.g. presence of obstacles, stimulation of the environment), and
internal factors (e.g. time-pressure, attitudes of the pedestrian). Together
with the expected trafIic conditions, the result of the tacticallevel (Le. opti-
mal schedule S" and route r" (.)) serves as input for the pedestrian walking
process. In turn, pedestrian walking behavior will be a result of predicted
cost minimization at the operationallevel, where pedestrians incur costs due
to straying from their intended route, walking dose to other pedestrians and
obstacles, and frequent or severe accelerations and decelerations.

departure time t o
prior activity set l:

activity schedule S --- optimal schedule

and route rOO
s-

--- "kitTe,,]
external factors +
+ individual
walking behaviour
internal factors

t
route 1(.)
--- +
pede strian traffic
conditions

Fig. 1. Pedestrian behaviour framework

While staying in the facility, pedestrians either perform activities, or walk

from one activity area to the next. Each pedestrian entering the facility at to
has a prior activity set E. This set consists of the activities i the pedestrian
may perform. Some of these activities are discretionary, others are manda-
tory; some can be performed in random order; other activities can only be
performed once others are completed. In theory, optimal activity choice, ac-
tivity area choice, and route choice are mutually dependent. To describe this
dependency, we hypothesize that the total cost C can be defined by a nmction
ofthe route r(·), the schedule S = {i}, and the times {Ti} at which activities
i E S are performed. We hypothesize that pedestrians simultaneously choose
the optimal activity schedule S", the activity times Tt (and thus the activity
areas Aij) for each i E S", and the route r" (.) optimizing the total cost C.
The latter route costs will among other things depend on travel time, prox-
278 S. P. Hoogendoorn

imity of obstacles, number of sharp turns and rapid directional changes, and
stimulation of environment, and route attractiveness [4], [3], [7]. The impor-
tance of these factors in route-choice will vary substantially between different
pedestrian groups.
Contrary to network-based models where travellers can choose between a
finite number of alternatives, the theory allows pedestrians to choose from an
infinite set of routes (or paths) through the walking facility. This is achieved
by considering the so-called value function Wi (t, r), describing the minimal
cost to perform activity i as a function of time t and location r. The value
function depends on the prevailing fiow conditions, and as such defines a
mapping from these predicted conditions to the pedestrian route choice be-
havior. Hoogendoorn and Bovy [11] present an approach to determine optimal
continuous routes and activity scheduling in continuous space under uncer-
tainty. In the remainder of this contribution, we assume that the planned
path - described by the desired velocity v* (t, r) - has been determined.

3 Theory of Walking
This section concerns walking behavior. The theory and models aim to de-
scribe the essential microscopic processes in walking (e.g. speed choice and
acceleration behavior, mutual interaction between pedestrians, reaction to
obstacles, etc.). The model must be able to predict and explain the phe-
nomena in pedestrian fiow that are important from both a theoretical and
practical point of view, given the envisaged model applications. Examples
are the relation between density and speed, and the formation of lanes and
clusters, for a specific infrastructure design.
In the considered approach, pedestrians choose their way in time and
space according to subjective utility maximization by continuously observing
their surroundings and taking appropriate actions. This essentially means
that pedestrians are modeled as feedback-oriented dynamic optimal control
systems. Amongst the theoretical hypotheses underlying the walker model
are that pedestrians:

1. continuously reconsider their walking choices by using current observa-

tions and the resulting predictions into the subjective utility optimization
(rolling horizon).
2. are anisotropie particles that react mainly to stimuli in front of them;
3. anticipate on the behavior of other pedestrians by predicting their walking
behavior according to non-cooperative or cooperative strategies;
4. have limited predicting possibilities, refiected by discounting cost of ac-
tions over time and space, implying that they mainly consider pedestrians
in their direct environment;
5. will be more evasive when encountering a group of pedestrians than when
encountering a single pedestrian (described by assuming additivity of the
proximity costs);
 Walker Behaviour Modelling by Differential Games 279

6. a) minimize predicted discounted costs resulting from b) drifting from

the planned route, the proximity of other pedestrians and obstaeles, and
applying the control (acceleration)j

Let us briefly reflect upon these hypotheses by considering some empiri-

cal and experimental research findings from literature. Goffrnan [6] describes
how the environment of the pedestrian (infrastructure and pedestrians) is
continuously observed through a mostly subconscious process called scanning
in order to side-step small obstructions on the flooring and handling encoun-
ters with other pedestrians (hypothesis 1). The scanning area is an ellipse,
narrow to either side of the individual and longest in front of him (hypothesis
2). Moreover, the area of the ellipse changes constantly according to the pre-
vailing traflic density. Goffrnan [6] also describes how pedestrians react to one
another by bilateral, subconscious communication. This is how pedestrians
are generally aware of the future behavior of other pedestrians in their direct
environment (hypothesis 3). Other researchers also stressed the importance of
cooperation between pedestrians. In illustration, Sobel and Lillith [17] report
that pedestrians are reluctant to unilaterally withdraw from an encounter
until the last moment, possibly even resulting in physical contact between
the pedestrians. This brushing sends a signal to the 'offender' to cooperate.
Naturally, the predictions are limited with respect to the time-horizon as well
in a spatial sense (hypothesis 4). The latter is also due to the limited obser-
vation range. Not all encounters will result in a cooperative decision, as the
amount of space granted by a pedestrian depends on the cultural, social, and
demographie characteristies of the interacting pedestrians [2], [17].Willis et
al. [19] found similar results for one-to-many interactions: individuals tend to
move for groups (hypothesis 5).
HilI [9] convincingly argues that a pedestrian (or in fact, his brain) can be
seen as a "pilot 0/ a very special vehicle". In the remainder, we argue that the
task of guiding this vehiele (the body) is similar to the task of guiding a car,
and can hence be described in an analogous manner, e.g. see [15]. Experience
and knowledge have skilled the pedestrian in optimally performing the guid-
ance subtask (hypothesis 6a). It is obvious that pedestrians aim to walk along
the route that best meets their walking goals. However, pedestrians generally
dislike walking too elose to other pedestrians, yielding tension and irritation
[18]. This holds equally for frequent or severe accelerations, since these will
yield additional cost in terms of energy consumption. It is plausible that the
resulting walking behavior is a trade-off between these factors (hypothesis
6b).
Important behavioral parameters that will largely determine pedestrian
behavior are:

- the acceleration time T (reflecting the acceleration and direction changing

delay)
280 S. P. Hoogendoorn

- temporal discount factor 'fJ (rate at which costs are discounted over time),
and the spatial discount factor Ro (proximity cost reduction rate as a
function of the distance between two pedestrians);
- anisotropy factors ct and CO reflecting the difference in behavior when
reacting to stimuli in front or behind, relative to stimuli from the side;
- relative weighting factors Cl; for the respective walking cost components;
- physical dimensions of the pedestrians (radius pp), and;
- control restrictions (maximum speed Vo (t, r), given specific infrastructure
and prevailing traffic conditions).
These parameters are to be estimated from either microscopic data (sec-
tion 6) or macroscopic data (reproducing emerging flow characteristics, Le.
speed density curves, spatiotemporal patterns, etc.). The parameters are dis-
cussed in detail in the ensuing of this contribution.

4 Conceptual Pedestrian Walking Task Model

In the remainder of the contribution we describe how, given the planned ac-
tivity schedule and route, pedestrian behavior at the operational level can
be described. This section discusses a conceptual model to describe walking
behavior in terms of a closed feedback control system. Similar approaches
have been proposed to describe execution of other human tasks, such as car-
driving. Several authors have proposed using utility optimization for mod-
elling driving-task execution. We will show that utility optimization is also a
fruitful approach to describe walking task execution.

4.1 Walking Subtask Hierarchy

Pedestrians are operators performing walking tasks. Using the terminology
of [15], the walking task comprises both the control and the guidance of the
'pedestrian vehicle' [9]. The pedestrian vehicle control subtask includes all
activities pertaining to the split-second exchange of information between the
brain (the 'pilot'), the senses (eyes, ears) , and the actuators (legs, arms).
These actions are nearly always skill-based and are performed automatically
with no conscious effort. The pedestrian vehicle guidance subtask describes
the collection of decisions required to guide the pedestrian safely and com-
fortable over the available walking infrastructure and its elements, as weH as
the proper behavior when encountering other pedestrians.
The pedestrian is assumed to use an internal model for determining an
appropriate guidance contral decision, thereby aiming to optimize his task
performance. Experience and knowledge have skilled the pedestrian in per-
forming this walking task, and hence the assumption of optimal behavior is
justified. The optimization includes the limitations of the pedestrians in terms
of observation, information processing and internal state estimation, as weH
as processing times and reaction times. Moreover, the optimization includes
 Walker Behaviour Modelling by Differential Games 281

the pedestrian walking strategy. The final walking task level pertains to nav-
igation (route-choice). This level is beyond the scope of this contribution (see
[11], [13] for details) .

4.2 Walking as a Feedback-oriented Control System

A pedestrian is constantly monitoring his position and velo city relative to

the other pedestrians in the flow, as weH as relative to obstacles and walls
(hypothesis 1). This monitoring allows the pedestrian to perform corrective
actions, implying that walking is feedback-oriented. A continuous feedback
control system consists of a comparison between the input (positions, ve-
locities) and the controHed output (acceleration, deceleration and direction
changes).

u (q)

X(/)
Pedestrian -
~ Output malel
traffic system

u (P) (t + "tom +"t"" +"tcontrol )

,.J (t)

Interml state
Control Observation
estimation
response model model
model

Fig. 2. Continuous controlloop of walking task execution (for single pedestrian p)

Figure 2 depicts the continuous control loop for pedestrian p E Q, where

Q = {I, ... , n} denotes the set of all pedestrians in the walking facility. The
pedestrian trafiic system describes the actual pedestrian trafiic operations.
The state X(t) of the system encompasses all information needed to summa-
rize the system's history, typically including the positions rq(t) E n and the
velocities vq(t) E Va of all pedestrians q E Q.

X (t) := (r, v) where r := (rl, ... , rn ) and v := (VI, ... , V n ) (1)

The dynamics of the pedestrian trafiic system are described by kinematic

equals described the changes in the pedestrian system state over time. These
282 S. P. Hoogendoorn

are refiected by a eoupled system of ordinary differential equations

X(t) = F(t,X(t),u(t)) (2)
where u = (Ul' ... , u n ) denotes the control veetor, reflecting how the pedestri-
ans infiuenee the state X (e.g. by accelerating and changing directions). We
emphasize that the pedestrian traffie system state dynrunics (2) describe the
true changes in the system.
The output model maps the state X(t) onto the observable system output
vector Y(t). Since both the location and the velo city can be observed exter-
nally, in the remainder of the contribution, we will assume that state can be
observed directly, i.e. Y(t) = X(t).
The observation model of pedestrian p describes which of the elements of
the pedestrian traffic system's output can be observed by p, as a function of
the state X(t), the time TobB required to process the observations, and the
observation errors CObB' In mathematical terms:
y(p)(t + TObB) = a(p) (X(t)) Y(t) + CObB (3)
where a(P) is a matrix describing the observable elements of the system's out-
put Y(t). The output y(P)(t) typically describes the locations and velocities
of pedestrians that p can observe.
Pedestrian p will use his observation y(p) and experiences to estimate the
state of the pedestrian traffic system, which is described by the internal state
estimation model of figure 2. Without going into details, we assume that the
internal state estimate can be described by a function of the previous internal
state estimate, the observation, and the estimation delay Test
(4)

In the model described here, describes estimates of the locations and ve-
locities of pedestrians , where denotes the subset of Q describing which pedes-
trians q are included in the internal state x(p) and internal state estimate of
p. Note that these estimates generally contains more information than the
observations y(p) (t). E.g., p may be aware of the presence of a pedestrian q,
who is blocked from p's view (either by an obstacle or another pedestrian)
by previous observations and prediction q's walking behavior.
The control response model in figure 2 refiects the process of determining
the control actions of pedestrian p. The control actions describe the accel-
eration vector ap(t) applied by p. We hypothesize that the control decision
depends on the internal state estimate, the walking strategy, refiecting the
walking objectives c of the pedestrians, and the control delay Tcontrol
u(P)(t+Tcontrol) =v(P) (x(P)(t),c) (5)

where the control delay describes both the time needed to determine the
control decision, and the time needed to start the control movement. In de:-
termining the walking actions, p will predict and optimize the walking cost
 Walker Behaviour Modelling by Differential Games 283

J(p) incurred during some time interval [t, t + T). The prediction of the in-
ternal state x(p) will be described by the internal state dynamies

±(p) = f(p) (s, x(p) , u(P») subject to x(p) (t) = x(p) (t) (6)

This state dynamics include the reactions of the pedestrians q as weH. For
any control path applied during [t, t +T), p predicts the state dynamics using
eqn. (6), as weH as the walking cost

(7)

and chooses the optimal control path satisfying

*
u[t,HT) -
- arg mm
. J(p) «p) It (p) (t))
U[t,HT) ' x
A
(8)

The cost optimization process, its parameters, and its resulting control deci-
sions differ between the pedestrians, due to different objectives, preferences,
and physical abilities.
To derive mathematical models, the control framework discussed here will
be used. To this end, the foHowing section specifies the pedestrian kinematics
as weH as the walking cost J(p) as functions of the internal state x(p) and
control. In doing so, both observation errors and delays wiH be neglected.
Furthermore, the observation and estimation models are specified such that
pedestrian p is assumed to be able to estimate the locations and velocities of
all pedestrians q E Q, implying that p has perfect information regarding the
current state x(t) ofthe pedestrian traffic system. The model is specified such
that p will only react to pedestrians in his direct environment and directly
in front of him.

4.3 Pedestrian Kinematics (Interna! Model)

In order to apply the Pontryagin's Maximum Principle to establish the walker

model, the internal state dynamics model (6) as weH as the factors determin-
ing the walking disutility in terms of the cost function (7). To predict the
evolution of the dynamics of the state, p will use an internal dynamic model
reflecting the kinematics of the pedestrians q E Q in terms of ordinary dif-
ferential equations that describe the dynamies of the location r q and velo city
v q of pedestrians q E Q

(9)

subject to the initial conditions stemming from the internal estimation model
(4)

(10)
284 S. P. Hoogendoorn

where a q denotes the acceleration vector. In this formulation, ap denotes the

controllable acceleration vector of p, while a q for q #- p denotes the predicted
acceleration of pedestrians q.
The internal state of pedestrian p consists of the locations r q and the
velocities v q of all pedestrians q E Q (including p), i.e.

x := (r, v) where r := (rl, ... , r n ) and v := (Vi, ... , V n ) (11)

Pedestrian p can influence the internal state x directly via the control
vector u defined by the acceleration ap , i.e.

u(t) := ap(t) (12)

We can easily determine the internal state dynamics (6) using the state
definition (11), the kinematics (9), and the control definition (12)

x= (1', v) = (v,a) = f(x,u) (13)

We hypothesize that pedestrian p has knowledge pertaining to the walking

strategies of the other pedestrians (referred to as 'opponents'). Assuming
that the opponents q are also modelled by continuous feedback controllers, p
presumes that the acceleration a q of opponents q are a function of p's internal
state x, i.e.

(14)

4.4 Walking Discomfort (Resistance)

Pedestrians incur different types of discomfort (disutility or cost) while walk-

ing. These costs can be expressed as a function J(p) of the control applied
by p during the planning period [t, t + T), where t denotes the current time,
and T denotes the prediction horizon. The cost function is determined by
integrating the running cost Lover the planning period [t, t + T). We assume
that the pedestrian is aware of his limited predictive abilities. This is why
cost incurred in the near future will be of higher value than the costs that
may be incurred in the long run. To model this, the concept of discounted
costs is included in the modelling. Let 'TJ be the (temporal) discount factor.
For 'TJ > 0, and T sufficiently large, the following specification of the cost
function is applicable

(15)

where z denotes the integration variable. In the remainder of the contribution,

we assume that p will apply the contral function minimizing eqn. (15)
 Walker Behaviour Modelling by Differential Games 285

4.5 Cost Components

The running cost L reflects a variety of factors Lk. For the sake of simplic-
ity, we assume that the running cost is linear-in-parameters (although this
restrietion is not required upon application of the approach)

L = LCkLk (16)
k

where Ck denotes the relative weight of cost component Lk. In this section,
we will consider three walking cost components LI, L 2 , and La, respectively
reflecting discomfort due to straying from the optimal velocity (speed and
direction), walking too elose to another pedestrian (proximity costs), and
high acceleration or deceleration rates in the longitudinal as weIl as the lateral
direction. Dealing with pedestrian's reaction to obstaeles and walls is similar,
and is not discussed explicitly in this contribution.

4.6 Cost LI due to Deviation from Planned Route

We have hypothesized that pedestrians at the tactical planning level deter-
mine an optimal velo city nmction v* = v* (t, r p) minimizing the expected
cost of walking and performing activities (hypothesis 6b). We assume that
the disutility incurred due to straying from the velo city v* (e.g. due to inter-
actions) is described by a quadratic running cost nmction

(17)

4.7 Proximity Cost L 2

The cost caused by the presence of other pedestrians in the walking area will
be referred to as the proximity discomfort or proximity cost (hypothesis 6b).
The discomfort incurred due to walking too elose to pedestrian q is given by
a function of the (relative) location and (relative) velo city of p with respect
to q. will also depend on parameters specific for pedestrians p and q, such
as their size (required to determine the net distance between pedestrian p
and q). Moreover, the fact that pedestrians are anisotropie (react mostly to
stimuli in front of them) is ineluded in the proximity cost as weH. Since the
approach is applicable to any cost function L 2 that is not time dependent,
we will not further specify the proximity cost ftmction here. However, we
will assume that the total proximity cost L 2 is determined by the sum of
proximity costs caused by pedestrians q, Le.
L2 = LL~q) (18)
q#p

Adding the cost incurred due to different pedestrians q causes p to be more

evasive when a group is encountered, than when in case of a single pedestrian
(hypothesis 5).
286 S. P. Hoogendoorn

4.8 Acceleration and Deceleration Cost La

Finally, we assume that p incurs cost when accelerating (hypothesis 6b). We
distinguish between cost of longitudinal accelemtion (increasing / decreasing
speed) and of lateml accelemtion (side-stepping perpendicular to the current
direction). Let us define vectors 0: and ß

(19)

respectively denoting the unit vectors in the direction and perpendicular to

the current direction of p. The total cost of acceleration is found by the
following weighted sum

1 - () ( ß) 2
L3 ="2() (apO:
' )2
+ -2- ap
I
(20)

where 0 ~ () ~ 1 is a weighting factor reflecting the relative costs of longitu-

dinal acceleration compared to side-stepping.

5 Derivation of Walker Model

In this section, we derive a pedestrian walker model, by application of Pon-
tryagin's Maximum Principle to find the optimal acceleration u* = a;, satis-
fying eqn. (8). To this end, let us define the so-called Hamiltonian 1l by

1l(t, x, u, A) = e-'1t L(t, x, u) + A' J(t, x, u) (21)

= e-'1tC; Ilv; - v l + e-'1
p
2 t c2 L 2 (x) (22)

+ e-'1t c; [~ (a~o:)2 + 1; () (a~ß)2] + A'V + ,'a (23)

1l is an auxiliary function to determine the equations for the state and the
so-called co-state A, and required optimality conditions for the control u.
A := (A, ,) denotes the marginal cost (shadow cost, co-state vector, or adjoint
vector) ofthe internal state vector x: the vector A = (A}, ... , An) describes the

=
small change in the position vector r (r1, ... ,rn ); ,=
marginal change in the total cost (15) subject to eqn. (8) resulting from a
(,1, ... "n) describes
the marginal change in the total cost (15) due to small changes in the velo city
vector = v (v}, ... ,v
n ).

5.1 Application of the Maximum Principle

The required conditions for optimality (for unconstrained control), state that
the partial derivative of 1l to u equals zero
(24)
 Walker Behaviour Modelling by Differential Games 287

It can then be shown that the optimal control (acceleration) satisfies [10]

u*(t, x) = __1 e'T/t M- 1 I'p(t, x) (25)

C3

where the matrix M is defined by

M- 1 = ~ [ ai a1~2] + _1_ [ a~ -a1a2] (26)

() a1 a 2 a2 1 - () -a1 a 2 a1

Clearly, the optimal acceleration u * is a linear function of the marginal costs

I'p of the velocity vp.
Dynamic equations for the co-states A can be determined by considering
the partial derivatives of 1-l to the elements of the state x. Given that the
state-dynamics are time-invariant, as is the running cost L, we Can show
easily that the co-states A satisfy

(27)

Using eqn. (27), we can derive expressions for the marginal cost Ap and
I'p of the position r p and velocity vp of pedestrian p, as weH as the marginal
costs Aq and I'q of his opponents q. In turn, these expressions can be used
to express the marginal cost I'p in eqn. (25) in terms of the internal state x.
The resulting expression for the marginal cost I'p can then be substituted in
the optimal acceleration law (25), yielding the foHowing expression

u; = M- 1 [1- ry1 (Drpv;)' ] (v*p -T vP ) - A oM- 1 [D rp L 2 + 1]Dvp L 2 ]

- AoM- 1 (DrpaQ\p + 1]Dvp aQ\p)' 8;1 [DrQ\pL2 + 1]DvQ \pL2 ] (28)
where the acceleration time T and the interaction constant A o are defined by

(29)

We have employed the subscript notation Q/p to indicate the 2(n-1)-vectors

pertaining to allopponents q -=fi p, e.g.

(30)
and defined the 2(n - 1) x 2(n - 1) matrix 8 p as foHows

(31)

The optimal acceleration u; of pedestrian p is governed by three terms.

The first term

(u*)
P accel
= M- 1[I _ 1]~ (D r p
v*)']
p
(V; -T Vp ) (32)
288 S. P. Hoogendoorn

can be considered an acceleration term, and describes how p aims to keep

walking along the planned path described by the optimal velo city v;. From
eqn. (32) we can see that this is in part determined by the weights Cl and
C3 determining the acceleration time 'Tp , reflecting respectively the cost of
drifting from the planned velo city v; u
and of applying control p . The ratio
Cl / e3 reflects that when the cost of acceleration is relatively low (Le. C3 rela-
tively small) or when the cost of drifting from the optimal velocity is relatively
high (i.e. Cl large), the acceleration time 'T will be small while the acceleration
term will be large. In this case, the pedestrian will be more inclined to keep
walking at the optimal speed and in the optimal direction, even though this
may imply large accelerations from time to time. The matrix M-l describes
the preference for longitudinal acceleration and braking over side-stepping.
Neglecting the role of M- l , the term

(33)

shows that the pedestrian will generally aim to walk in the direction (v; - vp )

The matrix I - *
Le. into the direction where the drifting cost L l will decrease most rapidly.
(D rp v;) I corrects this direction in line with changes in
the optimal velo city that are caused by changes in the Iocation r p of the
pedestrian. In other words, it corrects for the fact that when the pedestrian's
location rp is changed, in general so will his optimal velocity v; = v;(rp ).
The second term

(34)

(35)

of eqn. (28) reflects what we will call first-order interactions. showing how
pedestrian p reacts to the other pedestrians q:l- p. Eqn. (34) shows that the
effects of multiple interactions is additive.
Considering the interaction of pedestrian p and q, expression (34) shows
that p will make the following trade-off in determining his walking direction
and speed. On the one hand, the term DrpL~q) reflects the direction in which
the proximity cost L~q) changes most rapidly. On the other hand, the term
'TJDvpL~q) reflect changes in the proximity cost L~q) due to changes in the
velocity vp of p. For instance, this can express the fact that the proximity cost
will be relatively large when two pedestrians approach each other very quickly,
even though the distances between two pedestrians is still considerable.
Finally, the third term
 Walker Behaviour Modelling by Differential Games 289

describes what we will can second-order interactions. More precisely, it re-

fleets the effects of changes in the locations r q and velocities v q of pedestrians
q :I P that are expected by p, given the anticipated acceleration strategy
a q = a q (x) due to changes in the velo city v p of P caused by acceleration. At
this point, note that when a q = a~ (p presumes constant acceleration on the
part of his opponents q), the second-order interactions are equal to zero.

5.2 Cooperative Walker Model

Empirical studies have shown the importance of cooperation between pedes-

trians in walking [6], [9]. So far, we have assumed that pedestrians only opti-
mize their own walking objectives, thereby taking into account the presumed
decisions of the opponents. As such, control decisions can be described in
terms of non co operative differential games.
To described cooperation, we assurne that pedestrians will not only con-
sider their own predicted walking cost, but also the cost incurred by their
opponent. We hypothesize that the interaction pedestrians will make a deci-
sion that optimizes the joint predicted cost. Let us consider the case of two
pedestrians PI and P2 interacting. Neglecting the influence of other, non-
cooperative pedestrians q, we assurne that the PI and P2 will make a simul-
taneous decision optimizing the following objective nmction

(37)

where J(Pl) and J(P2) are defined by eqn. (15). The objective of the pedestri-
ans is to find the joint control (Ul, U2) = (apl> a p1 ) optimizing the objective
nmction eqn. (37). Let us assurne that both the weighting factors Ck and
the discount factors 'fJ are equal for both pedestrians p and q. The necessary
conditions for optimality D u 1l = 0 now yield

(38)

which is equal to the expression for the non-cooperative interaction (25).

Expressions for the marginal costs "(Pi of the velocities vPi for i = 1,2 can be
determined using eqn. (27). If can be verified by the reader that the marginal
cost "(Pi depends on the derivative of the joint running cost L = L(Pl) + L(P2)
to v pi , and on the marginal cost APi of the location r pi ' In turn, the latter
depends on the derivative of the joint running cost L to the location r pi '
In sum, the following optimal acceleration law for cooperative encounters
between Pl and P2 results

ui* = M- I [I - ;j * ),]
1 (D T pi vpi (V;i -T V Pi ) - A0 M- l [D T pi L 2 + 'fJ D v Pi L]
2

(39)
290 S. P. Hoogendoorn

for i = 1,2, where L 2 = L~pt} + L~P2) is the joint proximity cost. The optimal
acceleration law eqn. (39) for co operative encounters is very similar to the
optimal acceleration law for non-cooperative encounters eqn. (28). The main
difference is that instead of considering the proximity cost of the individual
pedestrian, in the cooperative case the joint proximity cost is considered.
When the individual proximity cost are 'symmetrie', i.e. when they satisfy
D r pi L(pt}
2 = D r pi L(P2)
2 and D 'lIpi L<vt}
2 = D 'lIPi L(P2)
2 , the optimal acceleration
laws for the cooperative encounter and the non cooperative encounter are
equal. When anisotropy is considered in the proximity cost specification, the
assumption of symmetry of the proximity cost does not hold, and the non-
cooperative model and cooperative model are not equal.

5.3 Relation to the Social-forces Model

Let us assume that the proximity cost is described by the following simplified
expression

(40)

Furthermore, let us assume that the planned velocity v;

is independent of
the position T p • Finally, we assume () = ~, yielding M = I. Furthermore, let
us assume that p presumes no reaction on the side of his opponents q, i.e.
a q = O. The optimal acceleration law (32) then simplifies to

u
*= (V; - Vp ) An '" Tq - Tp e _Ih-..
--L RO
pll
(41)
P T Ro q:f:.p IIrq- rpll

which equals the basic social-forces model of [7] (i.e. not considering aniso-
tropy, bodily dimensions, and physical contact). This analogy provides a be-
havior interpretation of the social-forces model, namely that pedestrians are
dynamic feedback controllers who presume that the other pedestrians in their
vicinities will not react to their control decisions. Since the specification of
the running cost function eqn. (40) is symmetrie, the resulting optimal con-
trollaw is equal for both cooperative and non cooperative encounters. When
anisotropy is considered, the social-forces model implicitly presumes non co-
operative encounters. Furthermore, the proximity costs are additive.
Furthermore, since M = I, it appears that the social-forces model does
not differentiate between longitudinal and lateral acceleration. FurthermorE),
the social-forces model implicitly assumes that the planned direction of the
pedestrians does not change smoothly (D rp v;'
= 0).
5.4 Model Refinements
The preceding sections have discussed the basic model, which does not fulfill
several important pedestrian model requirements. Among the most important
 Walker Behaviour Modelling by Differential Games 291

requirements are the finite dimensions of the pedestrians and the effect of
physical contact, and the anisotropy requirement. The former describes the
fact that given the scale of the applications, pedestrian dimensions need to
be accounted for explicitly - contrarily to particles in fiuidic or gas fiows. To
this end, pedestrian p is described by a circle with radius pp. The inclusion of
the pedestrian dimensions in the model involves two steps. For one, the effect
of collisions between pedestrians is to be refiected in the pedestrian trafik
system dynamics eqn. (2). These effects may involve repellent forces as weIl
as frictional forces [8]. Secondly, pedestrians' behaviors will be infiuenced
by the finite space requirements as weIl as the effect of collisions between
pedestrians. This can be described by including a collision disutility into the
cost function eqn. (15). For details, we refer to [12].
Anisotropy can be included easily by proposing the correct proximity cost
term L 2 , yielding increased cost for pedestrians walking in front of p. At the
same time, the proximity cost function can be adapted to include pedestrian
grouping behavior [12].

6 Approach to Model Calibration

Model calibration and validation (determining the model parameters and

comparing model results with real-life observations) is an important issue for
any model application, and various approaches have been put forward in the
past. For microscopic pedestrian simulation models, model calibration is com-
plicated due to the absence of good microscopic pedestrian data. This is why
in most cases, model calibration is based on cross-comparing macroscopic
model predictions with real-life observations. In illustration, the simulatio~
model predicts the relation between density and speed as a function of the
model's parameters. This relation can then be compared to empirical re-
sults and the model parameters can be adapted accordingly. In illustration,
Hoogendoorn and Bovy [12] show how the speed density curve depends on the
anisotropy parameters and propose parameters ensuing realistic macroscopic
behavior. Similar analysis can be performed by comparing spatiotemporal
patterns in the trafik fiow, e.g. how congestion propagates upstream of a
bottleneck, etc.
Ideally however, microscopic data is used for model calibration. This is
why Delft University of Technology is performing experimental research with
the aim of collecting individual pedestrian data in a controlled environment.
During a pilot study, the behavior of twenty subjects was monitored using
video, from which pedestrian trajectories were derived automatically using
dedicated software tools. The subjects conducted a number of experiments,
mostly aimed at assessing the technical aspects of the study set-up, as weIl
as the experimental design. Among the experiments were walking in unidi-
rectional and bidirectional fiow, walking through a bottleneck, and walking
in random directions.
292 S. P. Hoogendoorn

Fig. 3. Example of pedestrians moving through a 8Dem bottleneck . The heads of

the pedestrians are recognised and tracked. The observation area is 10m x 4m.

To derive the pedestrian trajectories from the collected video data, a track-
ing tool was developed. This tracking algorithm performs the following steps:
1. Map the picture coordinates to the real-life coordinates.
2. Recognize the caps of the pedestrians, based on the colors of the caps.
3. Cluster the pixels identified as the cap and determine its centre, describ-
ing the candidate pedestrian locations.
4. Use the candidate pedestrian locations identified for the consecutive video
images to establish the most likely trajectory, based on Kalman filtering
techniques.
It turned out that the pedestrians' trajectories could be determined from
the data accurately. Moreover, it was concluded that the experiments should
be as simple as possible to ensure the desired pedestrian behavior and reduce
the number of faults to aminimum.
In the end, the microscopic data will be used to determine the model
parameters directly (e.g. the acceleration time T, the spatial discount factor
ct
Ro, the anisotropy parameters and cö). This can be done by comparing the
microscopic model predictions of the dynamics of a single pedestrian to the
measurements (i.e. the filtered accelerations) and determining the parameters
that minimize the difference.

7 Application Results
This contribution is aimed at showing the derivation of a walker model from
behavioral hypothesis pertaining to the way pedestrians perform their walk-
ing task. It turns out that for correct model parameters, the model can cor-
rectly replicate important macroscopic pedestrian flow characteristics, such
as the relation between speed and density [10], [12].
 Walker Behaviour Modelling by Differential Games 293

Having shown the equivalence with the social-forces model, it is deemed

obvious that the model behaves in a similar way and hence will be able
to simulate pedestrian flow features such as dynamic lane formation, the
formation of so-called bubbles, and oscillatory behavior in bidirectional flows
[7], [8]. Besides dynamic lanes, the model also shows the formation of stripes
in crossing pedestrian flows, in line with Japanese empirical studies [14]. It
turns out that the trade-off between longitudinal and lateral acceleration
cost (not featured in the social-forces model) substantially influence these
self-formed characteristics [10], [12]. The variability of the formed patterns
increases with descreasing cost of sidestepping due to the fact that pedestrians
are more inelided to look for free spaces in the pedestrian flow, rather than
to decelerate and wait their turn. In other words, the inclination to make an
evasive manoeuvre (side-stepping) rather than to decelerate upon interaction
has a negative influence of the lane formation, as weIl as on the formation of
strips in crossing pedestrian flows.
This holds equally for the heterogeneity of the pedestrian population (in
terms of the free speed distribution).

8 Conclusions and Future Research Directions

In this contribution, we have put forward a new theory and models for pedes-
trian walking behavior. The theory is based on the assumption that pedes:"
trians are cost minimizers. Important factors are the predicted cost due to
applying control (accelerating and braking, side-stepping), walking too elose
to other pedestrians and obstaeles, drifting from the planned path, and the
anticipated behavior of the other pedestrians.
In the approach, pedestrians are described by optimal predictive con-
trollers that minimize the predicted walking cost, given the presumed walk-
ing strategies of the other pedestrians. As such, different presumed strategies
have been considered, varying from 'no reaction of the other pedestrians' to
'cooperative optimization of the joint walking cost'. To operationalize the
theory, mathematical optimal control theory has been applied. The resulting
model equations bear strong resemblance to the social-forces model.

References

1. W. Daamen, P.H.L. Bovy, and S.P. Hoogendoorn. Modelling Pedestrians in

Transfer Stations. Proceedings of the Pedestrian and Evacuation Dynamics Con-
ference. Duisburg (2001).
2. J. M. Dabbs and N.A. Stokes. Beauty is Power: the Use ofSpace on the Sidewalk.
Sociometry. Vol. 38, no. 4 (1975) 551-557.
3. C. Y. Cheung and W.H.K. Lam. Pedestrian Route Choices between Escalators
and Stairways in Metro Stations. Journal of Transportation Engineering 124(3),
277-285.
294 S. P. Hoogendoorn

4. M. T. Chiolek. Spatial Behaviour in Pedestrian Areas. Ekistics 268 (1978) 120-

121.
5. W. H. Fleming and H.M. Soner. Controlled Markov Processes and Viscosity
Solutions. Applications of Mathematics 25. Springer-Verlag 1993.
6. E. Goffman. Relations in Public: Microstudies in the Public Order. New York.
Basic Books 1971.
7. D. Helbing. Verkehrs dynamik. Springer-Verlag 1997.
8. D. Helbing, 1. Farkas, and T. Vicsek. Simulating Dynamical Features of Escape
Panic. Nature 407 (2000) 487-490.
9. M. R. Hill. Spatial Structure and Decision-Making 01 Pedestrian Route Selection
Through an Urban Environment. Dissertation Thesis. University Microfilms In-
ternational 1982.
10. S. P. Hoogendoorn. Normative Pedestrian Flow Behaviour. Report VK2001.002,
Transportation and Trafik Engineering Section, Delft University Of Technology
2001.
11. S. P. Hoogendoorn and P.H.L. BovyPedestrian Route-Choice and Activity
Scheduling Theory and Models. Submitted to Transportation Research B 2002.
12. S. P. Hoogendoorn and P.H.L. Bovy. Normative Pedestrian Behaviour Theory
and Modelling. 15th International Symposium on Transportation and Traffic
Theory, Sydney 2002.
13. S. P. Hoogendoorn. Pedestrian travel behavior in walking areas by subjective
utility optimization. Transportation Research Annual Meeting. Paper nr. (2002)
02-2667.
14. R. Hughes. The Flow of Large Crowds of Pedestrians. Mathematics and Com-
puters in Simulation 53 (2001).
15. J. Reason. Human Error. New-York: Cambridge University Press 1990.
16. P. N. Senevarante and J.F. Moral!. Analysis Of Factors Affecting the Choice
of Route of Pedestrians. Transportation Planning And Technology, 10 (1986)
147-159.
17. Sobel, R.S., and N. Lillith. Determinant of Nonstationary Personal Space In-
vasion. Journal 01 Social Psychology. Vol. 97 (1975) 39-45.
18. LB. Stilitz. Pedestrian Congestions. In: D. Canter (ed.). Architectural Phsy-
chology. London Royal Institute of British Architects (1970) 61-72.
19. Willis et al. (1979). Stepping aside: correlates of Displacements in Pedestrians.
Journal 01 Communication. Vol. 29, no 4., 34-39.
 Investigations of Vibrations in the Complex
Dynamical Systems of Transmission Pipelines

Elena Mull and Vladimir Kravchenk02

1 Institute of Computer Information Technologies, Ternopil Academy of National

Economy, 3 Peremoga Sq., 46004 Ternopil, Ukraine
2 Institute of Cybernetics of National Academy of Sciences of the Ukraine, 40
Glushkov avenue, 03650 Kyiv, Ukraine

Abstract. Mathematical models for complex dynamical systems of transmission

pipelines with distributed and discrete parameters are formulated in this paper,
which consist of a partial differential equation and nonlinear boundary conditions.
The approximate analytical method of small parameter is developed for investi-
gation of vibrations in the considered systems. The conditions of vibrations self-
exeitation were obtained in analytical form as weIl as a formula for amplitudes of
possible vibrations. It is shown that it is possible to change purposefully frequen~
eies and amplitudes of vibrations and avoid an excitation of undesirable vibration
processes, including self-excited vibrations.

1 Introduction
Let us consider dynamics of transmission pipelines intended for lifting of
minerals from great depth. A pipe of significant length with big mass platform
on the end is a basic constructive element of such deep-water plants. Intensive
dynamical processes of different physical nature may be excited under the
influence of forced wave disturbances and nonlinear hydrodynamical forces
in the system. Water flow, which washes the pipe, can continuously transport
the energy into the system, exciting significant and dangerous vibrations, as
a result breaking of the plant is possible.
Hydrodynamical load is nonconservative, as it does not have potential
(force function) in contrast to gravitation. Changing of energy of system,
which interacts with fluid flow, is explained by the presence of energy source
in the flow and not by of dissipation of energy, which is always present in
real dynamical systems at the expense of internal friction. If load on the
pipe is nonconservative, i.e. applied forces can not be written as gradient
of some potential function, energy statical stability criterion has not applied
and the possibility of appearance of dynamical bifurcations and different self-
oscillating modes appears.
Whereas the platform of the mass M has high-drag form in contrast to
the pipe, the force of resistance appears under interaction of mass M with
environment and this force is nonlinear as a rule. It is necessary to solve the
considered and similar problems of instability during design and creation of
different constructions for work in the ocean.
H. Emmerich et al. (eds.), Interface and Transport Dynamics
© Springer-Verlag Berlin Heidelberg 2003
296 E. Mul, V. Kravchenko

2 Model
The mathematical model of vibrations excitation in the case, when the pipe
is washed by water fiow, is a dissipative wave equation of longitudinal vibra-
tions. Accounting energy dissipation in the material we will write the next
partial differential equation [1]:
02 U 02 ou
ot2 -a20X2(u+ßot) =0, (1)
where u(x, t) is longitudinal displacement of pipe points; a2 = lfl.p
; E is.
modulus of elasticity of the material; p is mass of material volume unit; ß is
a coefficient characterizing internal friction in the material.
For dynamical system of transmission pipeline with elastically attached
pipe, boundary conditions for equation (1) are the next [2]:
o ou(O, t)
x = 0, ES OX (u(O, t) + ß ot ) = ku(O, t), (2)

x= L, M02~~~,t) +ES :x (u(L, t) +ßou~~,t)) =

ou(L,t) (ou(L,t))3
(}:1 ot - (}:3 ot ' (3)
where M is a platform masSj S is pipe cross-section areaj k is longitudinal
rigidity of the elastic hanger at the point of pipe attaching; L is pipe lengthj
(}:1 8u~~,t) - (}:3 (8u~~,t»)3 is force of non-linear resistance, which appears as a
result of interaction between the body and environment; (}:1, (}:3 are constants
characterizing resistant environment.

3 Method
In this paper the approximate analytical method of small parameter is de-
veloped for problem investigation.
First of all for estimation of order of smallness of components from the
boundary problem (1)-(3), as weH as some simplification ofbulky calculations,
the boundary problem (1)-(3) was written in the dimensionless form [2,3]:

(4)

x=O, (5)

x= 1, (6)
where dimensionless variables and dimensionless parameters are determined
by the expressions:
_ x _ u
x=L' U= L' (7)
 Vibrations in Dynamical Systems of Transmission Pipelines 297

aß alL a3a pSL kL

Cl = L' c2 = Ma' c3 = ML' I' = M' r = ES· (8)
In most cases the dimensionless parameters Cl, c2, ca in dynamical systems
of transmission pipelines are significantly smaller than unit. This allows to
assume that these parameters are directly proportional to some common
small parameter C (c « 1):

(9)
And now let us begin from the problem of self-excitation of vibrations. It
is known [2-4], that in that case it is sufficient to account only the linear part
in the expression for the resistance force, i.e. to assume that C3 = O. Then we
will search for the solution of equation (4) in the form

A 2
u(x,r) = (AcosAx + BSin>..x)exp(-ccI"2 ±iA)r, (10)

where A is an eigenvalue of the nonconservative boundary problem; A, B are

arbitrary constants; e is the basis of naturallogarithm.
It is easy to show, that solution (10) satisfies partial differential equation
(4) accurate to c 2 • Satisfying boundary conditions (5)-(6) as weH accurate to
c 2 , we will have the next transcendental equation:

(11)

The solution of this equation at c = 0 is known and it determines eigen-

values W of the corresponding conservative boundary problem. Accounting
smallness of damping forces, we will approximately determine eigenvalues of
the nonconservative boundary problem by expression

(12)

where cAl is correction factor of system nonconservatism.

After substitution of expression (12) into transcendental equation (11),
we will obtain the next approximate formula for complex eigenvalues of the
nonconservative system:

(13)

Then such that system increment should be greater than its decrement,
we will obtain from the exponential function of solution (10) the condition of
excitation of self-oscillations in the considered dynamical system in the form:

clrw 2(w 2 + 1'2) +C21'(W2 + r 2) ~ 61;2 (w4 + w2r(r + 1) +

W21'(1' + 1) + J1.r(l'r + I' + r». (14)
298 E. Mul, V. Kravchenko

And now let us analyse not only the conditions of self-excitation of vibra-
tions, but their amplitudes as weH, that is much more complicated problem.
In the conservative system (at c = 0) periodie motions appear with any
amplitude dependent on initial conditions. But in the nonconservative system
(at c f:. 0) they have only fixed amplitudes, whieh conform to equality of the
energy present at the expense of negative resistance and energy dissipation
whieh is always present in existent dynamical systems [1]. So the solution of
the nonconservative boundary problem (4)-(6) we will search in the form

U(X,T) = UO(X,T) +Wl(X,T) + ... , (15)

where uo(x, T) is one from the solutions of the corresponding conservative

problem (at c = 0).
It is easy to obtain the next expression for such solution:
W
uo(x, T) = A( -r coswx + sinwx) COSWT, (16)

where A is sought-for amplitude, and it is necessary to select it close to

amplitude of boundary cycle of self-sustained oscillation system.
Taking into consideration relationship

(J-lr - w2) COSW = (r + J-l)wsinw, (17)

whieh determines eigenvalues W of conservative boundary problem (4)-(6) at

c = 0, we will have the next problem:
[)2 Ul a2Ul . W .
- [2) - [)-2 = AClW 3 (SIll WX + - coswx) SlllWT, (18)
T x r
x=O, -rul + 8U1 A 2·
[)x = clW SlllWT, (19)
[)2 Ul 8U1 . W
x= 1, - [2
T
) + J-l !l=
uX
= AW(SlllW + -r COSW) X

(clW 2 - c2) sinWT + A3c3w3(sinw + ~ COSW)3 sin3 WT. (20)

r
Now let us transform the last addend in condition (20) and neglect the
component of the force with harmonie 3w (as in the case of equation (17)
frequency spectrum is not equidistant, Le. 3w f:. W3, therefore we assurne,
that the component of the force with harmonie 3w does not cause considerable
amplitudes during vibrations). So we will write condition (20) in the form:
[)2 Ul 8U1 . W
x= 1, - [ 2)
T
+ J-l !l=
uX
= AW(SlllW + -r COSW) X

(clW 2 - c2 + 0.75A2c3w2(sinw + ~ COSW)2) sinWT. (21)

r
The boundary problem (18), (19), (21) may be considered as a mathemat-
ieal model of the conservative system with natural frequency w, on whieh the
 Vibrations in Dynamical Systems of Transmission Pipelines 299

external force of resonance frequency effects. It is known [1] that the station-
ary solution exists only on condition of orthogonality of external force and
normal mode of the system. This condition determines the amplitude, which
does not increase with time. Therefore for our problem let us choose the next
form of solution:

UI(X,T) = U(x)sinwT, (22)

and write the problem as fol1ows:

~~ + w2U = -Aelw3(sinwx +;. cos wx), (23)

dU
x=O, -rU + dX = Aw2 Cl, (24)

x= 1, 1': - w2u = Aw(sinw +;. COSW) x

(CIW 2 -c2 +0.75c3A2w2(sinw+ ~COSW)2. (25)
r
Searching for general solution of equation (23) in the form

U(x) = (BI + B2x) COSWX + (Cl + C2x) sinwx, (26)

where BI, B 2 , Cl, C2 are arbitrary constants, and satisfying boundary con-
ditions (24)-(25), we can write:
Clw - rBl = 0.5Aw2el, (27)
-0.5Ae l rw 2(1'2 + I' + w2) - A!t~2 (elJ.tW 2 + elw 2 - 2e2) - (1'2 + w2)rBl
+ AeUT + (u2
2,...,...
+ w2 )wC1 _ A-
~
2r
6
= 3-
es"
aAS
4rwr
2( 2+ 2~S • 2
w
8 (,.r-w 2 )
sm w. (28)
This allows to find quadrate of amplitude A of stationary vibrations as
follows:

(29)

It is easy to see from the expression (29) that condition of excitation of

vibrations will be obtained in the form (14) just as for linear problem. The
first approximation of stationary solution of the problem will be the next:

U = A( - coswx + sinwx) COSWT +

W

r
W
e-(C1 - 0.5Awcl + 0.5AwcITX) coswxsinwT +
r
Aelw3x . .
e(Cl - 2r ) smwXsmWT. (30)
300 E. Mul, V. Kravchenko

4 Discussion of Results and Conclusions

Analysis of condition (14) of vibrations excitation allows to draw the next
conclusion. Such boundary value Wb of frequency always exists in the dynam-
ical systems of transmission pipelines with discrete and continuous parame-
ters, that on the frequencies, which are greater than this boundary value Wb,
excitation of vibrations is impossible. The value of such boundary frequency
Wb is easy calculated from inequality (14).
So it becomes possible to obtain easy qualitative estimations for determi-
nation of the effect of different pipe parameters on frequency domain, inside
which excitation of vibrations is possible.
The numerous calculations showed that at small dimensionless parame-
ters of the system only the first eigenvalue Wl of the conservative boundary
problem was included in domain of vibrations excitation. Increase of the pa-
rameter of external friction C2 causes increasing of the corresponding value
of boundary frequency Wb and, as a result, vibrations can be excited now on
the first two frequencies in most cases of practical values of parameter J.L.
Thus, for each definite system of transmission pipeline it is possible to
define a character of possible vibration processes. It allows to avoid an exci-
tation of undesirable vibration processes, including self-excited vibrations.
The results of complex dynamical systems investigations, which were ob-
tained with the help of the approximate analytical method of small parame~
ter, were verified with the help of numerical method of normal fundamental
systems of solutions [5].
The developed approximate analytical method can be applied also to solu-
tion of different boundary value problems, which describe various distributed
systems, for example, dynamical systems of controlled machine units.

References
1. Philipov, A.P.: Oscillations of Mechanical Systems [in Russian]. Kiev, Naukova
Dumka (1965) 706.
2. Kuhta, K.Y., Kravchenko, V.P., V.A.Krasnoshapka, V.A.: Qualitative Theory
of Controlled Dynamical Systems with Distributed and Discrete Parameters [in
Russian). Kiev, Naukova Dumka (1986) 224.
3. Mul, E.V.: On Conditions of Excitation of Self-Oscillations in a Nonconserva-
tive Dynamic System with Distributed Parameters. Cybernetics and Comput-
ing Technology, Complex Control Systems, New York, Allerton Press, Inc., 111
(1998) 70-72.
4. Svetlitskiy, V.A.: Pipe-Lines and Hoses Mechanies [in Russian]. Moscow,
Mashinostroenie (1982) 279.
5. Kravchenko, V.P., Mul, E.V., Shut M.l.: Possibilities to Control Oscillations in
Systems of Machine Units with Distributed and Discrete Parameters [in Russian
and in English]. Moscow, Mekhanika Kompozitsionnykh Materialov i Konstruk-
tsii Vo1.5 4 (1999) 77-86.
 Information in Intelligent Transportation
System

J. Wahle1 and M. Schreckenberg2

1 TraffGo GmbH, Grabenstr. 132, 47057 Duisburg

2 Physics of Transport and Trafiic, Univ. of Duisburg, 47048 Duisburg, Germany

Abstract. Since Advanced Traveler Information Systems (ATIS) have been in-
troduced, their potential benefits as weil as their drawbacks have been discussed
controversially. This will continue as long as the drivers' reactions upon current or
even predictive information ab out the trafiic situation are not known. Thus, traffic
models that also consider this feedback-Ioop are necessary. In this paper, we ad-
dress a basic two-route scenario with different types of information and study the
impact of it using simulations. Different ways of generating current information are
tested. It is pointed out that the nature of the information very much infiuences
the potential benefits of the ATIS.

1 Introduction

In the recent decades, Intelligent Transportation Systems were developed

aiming at improving the travel efficiency and mobility, enhancing safety and
thus providing economic and environment al benefits. An integral part of these
are Advanced Traveler Information Systems (ATIS) which provide current or
even predictive information about the traffic flow to the road users [1-3].
Although such systems have reached a high technical standard and have
a high potential to alleviate traffic congestion, the impact of the systems on
traffic patterns is not known to a sufficient extend [5-12]. Furthermore, some
phenomena or behavior patterns are observed that may disprove the benefits
of the given information. In [5] the following three phenomena are listed:

Oversaturation ITS can offer a huge amount of information to the user. But
if people are confronted with too much information, they are oversaturated,
tend to ignore the information and develop simple heuristics to solve the
problem. Therefore, especially ATIS should offer personalized information [1].

Overreaction A more fundamental problem is the overreaction which occurs

if too many drivers respond to the information. Suddenly, a congestion may
be transferred from the original area to the alternative routes. People start
to anticipate the behavior of the other drivers'. This anticipation gets even
worse if predictive information is given. Thus, the behavior of the drivers has
to be incorporated in the forecast [9,11]. Otherwise, this behavior might cause

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
302 J. Wahle and M. Schreckenberg

oscillations of trafIic fiow among the alternatives, since fast routes attract
trafik and, in the process, become slower, whereas the previously slower ones
turn faster [5,6].

Concentration A set of drivers who go from one origin to one destination

tend to use different routes, since they have different preferences or perceive
the situation in different ways. This leads to a natural distribution among the
routes. ATIS can reduce these variations if all the drivers use the objective
information broadcasted. As a result a greater number of them may select
the best alternative and consequently drivers with similar preferences will
choose the same route, leading to congestion on this [13,14].
Although the effects of overreaction and concentration are similar, the
nature of both phenomena is different [5]. Concentration is intrinsic to any
system with information, whereas overreaction is a consequence of the fact
that drivers cannot forecast the behavior of the others. However, as intelligent
devices will provide even more information about link travel times, densities,
or route guidance to the road user, this ability will be crucial for the utility
of the ATIS. Therefore, understanding travelers' route choice behavior is an
important consideration for the development and effectiveness of such systems
(for an overview see [5,6,12]).
Many classical microscopic trafIic fiow models lack the description of this
kind of behavior. In this work, we try to model the impact of real-time in-
formation on the traffk patterns. Special attention is paid to study different
kinds of information and their specific impact on the trafIic patterns.
The remainder of this paper is structured as follows: In order to exam-
ine different types of information and their impacts a two route scenario is
proposed in the next section. The following sections discuss different kinds of
information and their individual impact on the trafik patterns.

2 Two-route scenario

For the simulation studies, a scenario with two routes is used. The drivers are
modeled according to cellular automaton model by Nagel and Schreckenberg
[15]. The setup is the following: Suppose that there are two routes A and
B with the same length, i.e., the number of cells L is identical (Fig. 1).
With a constant rate a road users enter the system. They have to choose
between the two alternative routes. Their decision is based on some kind of
information. In order to study the impact of real-time information on traffk
patterns, dynamic data has to be generated. Here, the data stem from the
simulation itself. The process of generating the information will be discussed
later. In the literature, such a scenario is a standard one used to study the
impact of ATIS, e.g., [6,7,16]. In general, two types ofroad users are identified:
statie and dynamie ones:
 Information in Intelligent Transportation System 303

A=867.
B =780.

Fig. 1. Trafik scenario with dynamic information. Agents enter the system and
have to select a route. Each route comprises L cells. Static drivers choose their
route with a probability PAS, whereas dynamic drivers decide with regard to the
information displayed on the board, e.g., travel time. The information is transmitted
to the board by Floating Cars when they leave the routes.

Static Agents Static agents can be either non-equipped with in-vehicle

devices or just ignore the dynamic information, e.g., they do not read infor-
mation boards or do not listen to the radio. Therefore, they decide on the
basis of static information, e.g., the length of the routes or road signs. In
the simulation experiment this is represented by the variable PAß, Le., the
probability that a static driver chooses the route A.

Dynamic Agents Dynamic drivers are equipped with ATIS or other de-
vices. Thus, they receive and process trafIic messages. In this scenario, the
messages are displayed on a board (Fig. 1). The board can represent a mes-
sage sign at the beginning of the routes or an in-vehicle device. Using this
information a dynamic driver always tries to optimize his trip, Le., select
the route with the minimal travel time. In that sense they behave ''rational''.
Note that this is a very particular kind of rationality: the agents trust the
information on the board and decide based on it. They do not over-react ,
i.e., there is no reasoning about the behavior of the others. If the informa-
tion for both routes is the same, a route is chosen randomly. The amount of
dynamic drivers is represented by the variable Sdyn. After the route choice,
both kinds of vehicles move according to the CA-rules [15].

Generation of Dynamic Data The decision of the dynamic agents is based

on real-time information, which is for instance provided by Floating-Cars
(FCs) . FCs are employed to investigate trafIic conditions on the routes, e.g.,
to measure speed or link travel times. In this paperwork, we study different
kinds of information. In the first scenario described, the FCs transmit their
travel times to a board when leaving the system (Fig. 1). Thus, they generate
dynamic trafIic information, which is the basis of the decision of newagents
entering the system. The amount of FCs is represented by the variable SFC.
The remaining vehicles do not transmit information. Additionally, the impact
of different kinds of information, like average speed or density on the route,
will also be studied (see Sect. 4).
304 J. Wahle and M. Schreckenberg

2.1 Simulation technique

Simulation run The simulations are performed in the following way: at the
beginning both routes are empty and the information on the board is set
to travel time on free routes tA = tB = L/(vrnax - p). Note that the mean
speed in the free-flow region is V rnax - p. Dynamic drivers choose their route
at random and static drivers with regard to PAB. With the probability SFC,
a static or dynamic driver is a Fe. After a certain time Fes finish their trip
and transmit their travel time to the board. Now, the dynamic drivers decide
on the basis of the information available. In general, every simulation lasts
for 50,000 time-steps. For the analyses the first 5,000 steps are ignored as
only then the system has reached a steady state. The results of a typical
experiment are depicted in Fig. 2.

Open boundaries Because of the model employed, special care has to be

taken when injecting new vehicles on the routes. In most investigations of
the dynamic properties of trafik flow models, periodic boundary conditions
are employed to study equilibrium dynamics (for an overview [17]).
Recently, systems with open boundary conditions have been discussed in
literature [18,19]. Open systems are characterized by an injection rate a and
an extinction rate ß, i.e., a vehicle enters the system with the prob ability a
or leaves it at the end with the rate ß. Various investigations of systems with
open boundary conditions yield that there are three different phases: the free-
flow, the congested traffic and the maximum current phase [19]. Note that
these phases are only stable if the rates are steady, i.e., fta(t) = ftß(t) = O.
For convenience, we set ßA(t) = ßB(t) = 1 and a(t) = const. For the injection
rates ofroute A and B, the following relation holds aA (t) + aB (t) = a, with
ftaA(t) =I 0 and !taB(t) =I O. The time dependence represents the dynamics
of the decisions.
The process of injection is carried out in the following way [18]: in front
of the system, an injection area with the size Vrnax + 1 is prepared. To add
a new vehicle, the position of the first car on the road is determined. Then
the additional vehicle is put into the injection area with a minimum gap of
V rnax to the detected car, and a speed of V rnax • Then the update is carried out
according to the mIes presented in [15]. If a vehicle remains in the injection
area after the update procedure it is deleted. This is necessary to avoid jams
close to that area, since otherwise the injection rate is replaced by the flow
out a = jjam and high-flow rates cannot be established.
There are two sets of parameters which stem from either the model used
or the scenario itself. The parameters are the following: Vrnax maximum speed
of vehicles, Pdec deceleration probability, a input rate, L length of the routes,
PAB decision of static drivers, SFC percentage of Fes, and Sdyn share of
dynamic/equipped drivers.
Obviously, Sdyn and SFC are interpreted as probabilities for creating a
dynamic road user or a Fe, respectively. With the probability PAB a sta-
 Information in Intelligent Transportation System 305

tie driver is put on route A. The system is symmetrie and thus only val-
ues of PAB ~ 0.5 are studied. However, in most simulations the default
value of PAB = 0.5 is used, sinee the route length L is the only informa-
tion statie drivers have. Note that for PAB = 0.5 a natural equilibrium is
reached, Le., an optimal solution of the problem. In all simulations the max::..
imum speed is set to V max = 3 eellsjtime-step. In this work, we are only
interested in the parameter Sdyn share of dynamiejequipped drivers and PAB
decision of statie drivers. The other parameters are studied in [20].
In order to reduee the stochastic noise in the quantities, a correlation
analysis is used. Such an analysis is especially useful to study oseillations,
which will oeeur for some quantities in the next section, it is valuable to em-
ploy eorrelation funetions. They detect and quantify eoupling effects between
variables. The auto-eorrelation for a variable 1j; is defined in the following
way:
1
aCt/I(r) = rT(1j;) [(1j;(t)1j;(t + r)} - (1j;(t)}(1j;(t + r))]. (1)

It provides information about the temporal evolution of 1j; (cr(1j;) is the vari-
anee of 1j;).

3 Floating-Car Data
In this section, Fes are used to provide dynamie data. They are inserted
into the system with the rate SFC. They flow with the traffie and transmit
their travel time to the information board, when they leave the system. The
dynamie drivers which enter the system reeeive the moving average (20 time-
steps) of the travel times eolleeted.
A typical experiment is depicted in Fig. 2. The parameters are set to
Pdec = 0.25, a = 1, L = 2,000, PAB = 0.5, Sdyn = 0.5, and SFC = 1, Le.,
every seeond driver proeesses dynamie information and every vehicle serves
as a Fe. The number of vehicles, travel times, and the moving average of
the speeds oseillate on both routes. Note that unlike other quantities the
fluetuations of the number of vehicles are not very strong sinee every time-
step only one vehicle ean enter and two vehicles ean leave the system (one
on route A and B).
The oseillations are due to the dynamie drivers who reeeive the informa-
tion on the board. Beeause they behave rational and trust the information,
they aim at minimizing their own travel time. Therefore, they use the route
with the shorter travel time. Sinee every dynamie driver behaves in the same
way the route with the smaller number of vehicles is haunted more (Fig. 2a,b).
At the same time, the speed on this route deereases and thus the travel time
inereases (Fig. 2e). Nevertheless, it takes some time for the Fes to transmit
this to the information board sinee they have to reach the end of the route.
During this time even more road users enter this route.
306 J. Wahle and M. Schreckenberg

350

I
i
\ j
\. :
\. , j
200
V - - ROUl.eA
----------- Route B
15OL-----~---- __ ~ ______ ~ _____ L_ _ _ _ _ _J
5000 6000 7000 8000 9000 10000
(a)

700L------L------~----~L_ ____ ~ ____ ~

5000 6000 7000 8000 9000 10000

(b)

2.7

2_6

1.'" 2.5

2.4

- - RouteA
----------- Route B
2.3
5000 7000 8000 9000 10000
Tm>eSlep
(c)

Fig.2. (a) Number of vehicles on each route; (b) travel time on the board; (c)
moving average (20 time-steps) of the mean speed vs. time. The parameters are
pdec = 0.25, a = 1, L = 2,000, PAß = 0.5, Sdyn = 0.5, and 8FC = 1. Every quantity
exhibits characteristic oscillations.
 Information in Intelligent Transportation System 307

Simultaneously, the situation on the other route gets better since it is

not used anymore by dynamic agents. The travel time on this route slowly
decreases and the situation turns around - this route gets more crowded
again. This oscillatory behavior is weH known in the literature, e.g., Bonsall
sees it as some kind of interaction:
fast links attract traffic but are, in the process, made slower [6].
In [5] this effect is named concentration:
Information tends to reduce the variations among drivers because it
increases uniformity of perceptions of network conditions around the true
value.
However, this effect seems to be intrinsic in most information systems. Nev:.
ertheless, the information influences the dynamics of the system in an un-
desirable way: it would be much more efficient to distribute the vehicles on
both routes randomly. This is not only due to the effect of concentration but
also due to the fact that the information is too old, the drivers tend to react
too late to the new situation. To quantify this behavior in more detail, the
influence of the parameter Sdyn will be investigated in the foHowing.

3.1 InHuence of the dynamic drivers

Since the oscillations are due to the behavior of the dynamic drivers, it is
interesting to study the influence of the parameter Sdyn, the rate of dynamic
drivers. Obviously, the amplitude of the oscillations depends on Sdyn. A dra-
matic situation occurs for Sdyn = 1, Le., every road user foHows the informa-
tion displayed on the board. In this scenario, it may happen that one route is
empty and there are no Fes on this route. Therefore, the travel time remains
constant, although the route is empty. In order to avoid such negative effects,
it is advisable to remove or slowly change information, which is too old. This
can be done by, e.g., setting default values after a certain time-out.
In Fig. 3 the auto-correlation for a number of cars on route A, aCNA(r),
is depicted. For small values of Sdyn, the correlation is not strong, whereas
systems with a higher number of dynamic drivers Sdyn are strongly correlated.
Similar results are found for the travel time and the speed.

Travel time distribution The effect of mixing the two types of agents can
be understood better by studying the travel time distribution of both species.
In order to provide the distribution, a histogram is used, Le., the number of
drivers in a certain interval of travel time Llt is aggregated. We have set
Llt = 5 time-steps. A distribution for different Sdyn is depicted in Fig. 4.
308 J . Wahle and M. Schreckenberg

- Sdyn= 0.1
/\ i., --- Sdyo= g.~

0.5
r
!/t\
,( i~.,.
r
,! \:
Sdyn=;"

11 \\ i 1;\

o
i! \\
----i-\--
i_I!, \\ ._,1,f/ \\
~"""'.F-·~::::-;;"';--...-\'I

-0,5
"I/
\\!i
\\/1\\j/
\" \,.fi ~:
\, /
~ :
~ : ~ : \j
'..:'
\/
-1 L -______L -______L __ _ _ _ ~L_ ____ ~L_ ____ ~

o 2000 4000 6000 8000 10000

Fig. 3. Auto-correlation aCNA (1) vs. T, for different values of Sdyn. The parameters
are L = 2,000, pdec = 0.25, a = 1, PAB = 0.5, Sdyn = 0.5, and SFC = 1. For small
values of Sdyn the strong correlations vanish.

In a society of static agents, Sdyn = 0, the agents visit both routes with
the same frequeney and the travel time distribution is more or less Gaussian,
with a small standard deviation (Fig. 4a). Note that using PAB = 0.5 an op-
timal solution is reaehed. For a mixt ure of the two species it is ehanging. The
dynamic drivers introduee fluctuations into the system. Their average travel
time inereases, whereas some statie drivers benefit from emptier routes. A
seeond peak at higher travel times is established due to the dynamic drivers
(Fig. 4b) . For higher values of Sdyn, this peak broadens and both peaks sepa-
rate (Fig. 4e). The majority ofthe static drivers profit from the dynamic ones.
For Sdyn = 1, every driver decides with regard to the eurrent information on
the board. A broad peak is found (Fig. 4d).
The behaviour of the dynamic drivers influenees existing trafiic patterns
in an undesirable way and leads to a sub-optimal situation, Le., the overall
flow of the system is redueed due to eoneentration of the equipped drivers
(see Fig. 9). The travel time distribution implies that the statie drivers profit
from this behaviour (see Fig. 4). Therefore, the question ean be raised:

1s it good to be ignorant to information?

 Information in Intelligent Transportation System 309

......... ........
(a) (b)

- 1bt • ..,..-o.7
- - - 0,.. ....00.1

........
(e) --
(d)

Fig. 4. Travel time distribution for statie and dynamie drivers for different values
of Sdyn. The parameters are L = 2,000, pdec = 0.25, Cl! = 1, PAB = 0.5, and
SFC = 1. (a) A homogeneous society of static drivers arrange in an optimal way for
PAB = 0.5. (b) A mixture between dynamie and static drivers is unfavourable for
the dynamic ones. They introduce a second, broader peak at higher travel times. (c)
This effect is amplified for higher Sdyn . The peak of the dynamic ones is broadened,
whereas the peak for the static ones gets sharper, Le., the minority of static drivers
profits even more. (d) For a homogeneous society of dynamic drivers Sdyn = 1 the
distribution is broadened further.

3.2 InHuence of the static drivers

So far, the static agents have chosen their route randornly, i.e., PAB = 0.5.
This means that they behave in an optimal way. In Fig. 5 the travel time
distribution for PAB < 0.5 is shown. For PAB = 0.2 the dynarnic agents
are not really able to profit from the information (cf. Fig. 4a), since the
static agents on route A have lower travel times. Only for PAB « 1, i.e.,
all static drivers use the same route, the dynarnic agents can profit from the
information and there are still agents, who use route B (see Fig. 5b). Besides,
such a behaviour is quite unrealistic.
310 J. Wahle and M. Schreckenberg

3000 3000
- dyo_A
-"_A
- dyoRool~B
2500 - -..Rou\cB 2500

2000 2000

d'!I 1500 d JSOO

~

~ ~
1000 1000

"'"
500

0 0
700 '50 800
..........
$SO 900 100 900

Fig. 5. Travel time distribution for different values of PAB. The parameters areset
to L = 2,000, pdec = 0.25, Sdyn = 0.5, and SFC = 1. (a) PAB = 0.2, i.e., 80 % of the
static drivers use route B. (b) PAB = 0, i.e., all static drivers use route B. Although,
the static agents are far away from the optimal coordination, the dynamic agents
do not profit very much and the advantage of using the information is small.

4 Different criteria
It becomes clear that the effect of concentration is intrinsic to information sys-
tems, which provide the same information to all users, since faster links at-
tract more traffic and in the process are made slower [6]. Additionally, Fe
travel time is not a good criterion to control and distribute .trafik in a net-
work, since it reßects the history of the network. In this section, different
kinds of information are tested: the gradient of the travel time, the global
density and the global speed.

4.1 Gradient of travel time

In control theory, it is weIl known that in order to control some systems it is
better to use the gradient of a quantity than the absolute values themselves
[21]. Here, the moving average of the difference between two consecutive travel
times of Fes Ll7i = (ti - ti-t) is used, with ti being the travel time of the
ith Fe on a route. Hence, the information displayed on the board is Ll7, the
moving average of the last twenty Ll7i, Le., the dynamic drivers do not follow
the absolute travel time but the trend of the measured values. Thus, changes
of the traffic states on both routes are reported much faster.

Sub-Optimal Static Agents Nevertheless, in some situations the trend

can be misleading. In Fig. 6 the travel time distribution for a sub-optimal
behaviour of the static drivers is depicted. The parameters are set to L =
2,000, Pdec = 0.25, a = 1, Sdyn = 0.5, SFC = 1, and PAB = 0, Le., all
 Information in Intelligent Transportation System 311

static drivers use route B. In this situation, the absolute travel time is the
better information. For PAB = 0, the bulk ofthe dynamic agents are on route
A. Thus, they improve the situation for the static drivers (Fig. 6a). If the
gradient of the travel time is used as source of information a lot of dynamic
agents use route B and increase the travel time further (Fig. 6b).

3000 2000
-dyDRou1oA - dyDRouIOA
- dyoRonteB - dynRouteB
-statRoulOB -statRoatcB
2SOO
1500
2000

8'S 1500
8'S 1000
:!! :!!
1000
500

- -
500

0
700 750 800 8SO 900 800 8SO 900

(a) (b)

Fig. 6. Travel Time Distribution for (a) absolute travel times and (b) the gradient
of the travel time. The parameters are L = 2,000, Pdec = 0.25, a = 1, Sdyn = 0.5,
SFC = 1, and PAB = 0, Le., all static drivers use route B. In this situation providing
the gradient of the travel time is misleading to the dynamic drivers. They introduce
a distinct peak at higher travel times.

This negative effect is also reflected by the average travel time for the
dynamic and static drivers. In the first scenario, tdyn = 760.15 time-steps
and t stat = 776.74 time-steps, whereas for the second scenario tdyn = 798.05
time-steps and t stat = 866.79 time-steps are necessary. The problem using th~
gradient is that the absolute value of the travel time is neglected. Of course a
criterion which combines the absolute value and its gradient, would provide
better results.

4.2 Global density and speed

Up to now, the information provided by FOs has been discussed. However,
one main problem is that this information is delayed. Better quantities to
characterise the current trafik state of a route are necessary, like the global
density and the mean speed. They are defined as follows: the density PA,B =
NA,B/ L, and the mean speed VA,B = Ef:~,B Vi;A,B. Note that in real systems
such precise quantities cannot be provided, since most measurements are
local.
Simulation experiments with global density and mean speed as informa-
tion displayed on the board are carried out. In the case of the density crite-
312 J. Wahle and M. Schreckenberg

rion, the dynamic drivers choose the route with the lowest mean density. H
the mean speed is given, the dynamic agents select the route with the highest
value.
However, in both cases the correlations vanish. In Fig. 7 the travel time
distribution is shown for Sdyn = 0.3 and PAB = 0.5. Using the global average
speed it is possible to coordinate the dynamic as weIl as static agents in a
more efficient way. Thus, giving a quantity which represents the current state
of a route as information, e.g., mean speed or density, helps to coordinate the
behaviour of the two different types of road users, so that all drivers are able
to profit.

2500
- - dyn Route A
--_.- stat Route A

2000

~ 1500
u
'S
~ 1000

500

0
725 750 775 800
Intervals

Fig.7. Travel time distribution for using the global speed as trafiic information.
The parameters are L = 2, 000, Pdec = 0.25, Cl! = 1, 8dyn = 0.3, 8FC = 1, and
PAB = 0.5. Using this information the traffic demand is distributed on both routes
in the same way.

Obviously, the difference between the two quantities is that if density is

the information, the speed fiuctuates, and vice-versa. In the case density is
provided, a higher number of vehicles is injected into the system. H speed is
given, the mean speed on both routes is quite high.
So far, it has been shown that the dynamic drivers can hardly profit from
the information provided. Only in very extreme cases, e.g., PAB = 0, the sit-
uation for some dynamic agents is better than for most static ones. However,
using a global quantity as information the dynamic road users outperform
the static ones. Such a situation is shown in Fig. 8, where the travel time
distribution for a sub-optimal behaviour of the static drivers (PAB = 0) is
given. In contrast to the previous experiments, all dynamic drivers profit
from the provided information since no one uses route A. This shows that
 Information in Intelligent Transportation System 313

information in principle can help to improve traffic patterns by coordinating

the behaviour among the agents. Note that in this experiment Sdyn = 0.25,
Le., there is a market penetration of 25 %. With a higher market penetration
this effeet is shortened.

4000
- - dyn Route A
_._.-._._.- stat Route B

3000

,
~ !i
1\
U
....0 /'\1 !

ci
2000
/' \
Z .",
\
I \

,) ."1, \\\
!OOO

0
f \
700 750 800 850 900 950 !OOO
Intervals

Fig. 8. Travel time distribution for using the global density as traffic information.
The parameters are L = 2,000, Pdec = 0.25, a = 1, Sdyn = 0.25, SFC = 1, and
PAB = 0, Le., all static drivers use route B. All dynamic drivers select route A.

4.3 Comparison
In the previous sections, four different kinds of dynamic information have
been discussed: FC travel time and its trend, the mean speed and density.
All types of information are evaluated from the road usens point of view.
From the traffic control point ofview, an important quantity to evaluate is the
capacity or the flow: j = NAVA +NBvB. The objeetive ofthe traffic controller
is to maximise the capacity of the network by providing information. In Fig. 9
the four different types of information are compared with each other for
different market penetrations. The parameters are L = 2,000, Pdec = 0.25,
0: = 1, SFC = 1, and PAB = 0.5, Le., the static drivers behave optimal.
It is clear that with a higher share of dynamic drivers Sdyn and the FC
travel time provided, the flow is reduced. This is due to the oscillations, which
destroy the trafik patterns (see Seet. 3). Providing the gradient of the travel
time is slightly better but also leads to a decrease in the flow with increasing
Sdyn' This is also true for the average speed as information, since it leads to a
high average speed but reduces the density. Thus, the flow is slightly reduced.
The only type of information, which enhances the efficiency of the system
is the average density. Therefore, this is the best type of information for traffic
314 J. Wahle and M. Schreckenberg

control purposes, since it helps to coordinate the agents in a sufficient way

so that even the natural equilibrium of the static drivers (PAB = 0.5) can
be optimised by reducing the fluctuations. This discussion makes clear that
the efficiency of an information system depends heavily on the nature of the
information provided.

1750 ".----~---...,....---.....---__._--__..

1500

i
~
'.:::1

l
1250

~
~
1000
FCD 1---+---1
Velocity "-1<--1
Density 1--&--<
Trend •._....-...
~O~--~---~---~--~--~
o 0.2 0.4 0.6 0.8

Fig. 9. Flow vs. Sdyn for different kinds of information. The parameters are L =
2,000, Pdec = 0.25, a = 1, 8FC = 1, and PAB = 0.5. For the Fe data and the trend
the ßow goes down rapidly. This even holds for the speed. From the trafik control
point of view, the density is the best criterion.

5 Summary and conclusion

In this paper, the impact of en-route information is studied by simulating

a simple two route scenario. For the analysis static and dynamic agents are
introduced - representing conservative or flexible route selection. Dynamic
travel times for both routes are generated by traflic flow simulations. The
data are transmitted by floating cars (FCs) to an information board, which
is the basis of the route choice of dynamic road users.
The en-route information leads to a concentration of drivers on the rec-
ommended routes, which is intrinsic to many information systems and leads
to a negative impact on the traflic pattern, like oscillations. These oscilla-
tions are studied using a correlation analysis and travel time distributions.
Different parameters are investigated systematically. It is found that the in-
formation transmitted by the FCs is often too old and reflects the previous
 Information in Intelligent Transportation System 315

states of the network. Thus, the dynamic agents do not profit from the in-
formation received.
Therefore, different types of message are tested: the gradient of the Fe
travel time, the global density and speed. The analysis yields that beside the
amount and the validity of the information an important question is the na-
ture of it. By its nature, travel time is not a good quantity since it intro duces
not only concentration but also oscillations into the system. For trafIic control
purposes, the best criterion is the global density since it optimises the fiow.
Nevertheless, the concentration of drivers on the recommended routes is in-
trinsic to many information systems, especially for a high market penetration
of information systems the benefits of providing information are questionable.
The analysis yields that one should be very careful to provide dynamic
information, since concentration is intrinsic to all information systems and
the overall performance of the system can get worse [22,7,16,10,14,23]. One
key question is: what kind of and how much information do road users need?
Another interesting issue is to study the effect of providing various types of
messages to different drivers. This may introduce a natural diversity which
avoids the effect of concentration.

References
1. J.L. Adler and V.J. BIue. Toward the design of intelligent traveler information
systems. Transp. Res. C,6 (1998) 157-172.
2. ITS International, editor. 6th World Congress on Intelligent Transport Systems,
CD-ROM, Toronto, 1999. ITS World Congress.
3. D. Helbing, H.J. Herrmann, M. Schreckenberg, and D.E. Wolf, editors. Traffic
and Granu.lar Flow '99, Heidelberg, 2000. Springer.
4. A. Ceder, editor. Proceedings 0/ the 14th International Symposiu.m on Trans-
portation and 'I'raffic Theory, Amsterdam, 1999. Pergamon.
5. M. Ben-Akiva, A. de Palma, and I. Kaysi. Dynamic network models and driver
information systems. Transp. Res. A, 25 (1991) 251-266.
6. P. BonsalI. The influence of route guidance advice on route choice in urban
networks. Transportation, 19 (1992) 1-23.
7. R.W. Hall. Route choice and advanced traveler information systems on a ca-
pacitated and dynamic network. Transp. Res. C, 4 (1996) 289-306.
8. H. Dia and H. Purchase. Modelling the impacts of advanced traveller informa:-
tion systems using intelligent agents. Road and Transport Research, 8 (1999)
68-73.
9. J. Bottom, M. Ben-Akiva, M. Bierlaire, I. Chabini, H. Koutsopoulos, and
Q. Yang. Investigation of route guidance generation issues by simulation with
DynaMIT. In Ceder [4], 577-600.
10. K. Lee, P.M. Hui, B.H. Wang, and N.F. Johnson. Effects of announcing global
information in a two-route trafiic flow model. J. Phys. Soc. Japan, 70 (2001)
3507-3510.
11. J. Wahle, A.L. Bazzan, F. Klügl, and M. Schreckenberg. Anticipatory trafik
forecast using multi-agent techniques. In Helbing et al. [3], 87-92.
316 J. Wahle and M. Schreckenberg

12. H.S. Mahmassani and R.-C. Jou. Transferring insights into commuter behavior
dynamics from laboratory experiments to field surveys. Transp. Res. A, 34
(2000) 243-260.
13. R. Arnott, A. de Palma, and R. Lindsey. Does providing information to drivers
reduce traflic congestion? Transp. Res. A, 25 (1991) 309.
14. J. Wahle, A.L. Bazzan, F. Klügl, and M. Schreckenberg. Decision dynamics in
a traflic scenario. Physica A, 287 (2000) 669-681.
15. K. Nagel and M. Schreckenberg. A cellular automaton model for freeway traflic.
J. Physique I, 2 (1992) 2221-2229.
16. K. Nagel and S. Rasmussen. Traflic at the edge of chaos. In R.A. Brooks
and P. Maes, editors, Artificial Life IV: Prac. of the 4th Int. Workshop on the
Synthesis and Sim. of Living Systems, pages 222-235, Cambridge, 1994. MIT
Press.
17. D. Chowdhury, L. Santen, and A. Schadschneider. Statistical physics of vehic-
ular traffic and some related systems. Physics Reports, 329 (2000) 199-329.
18. R. Barlovic, T. Huisinga, A. Schadschneider, and M. Schreckenberg. Effects of
open boundaries in a cellular automaton model. Physica A, 2001.
19. V. Popkov, L. Santen, A. Schadschneider, and G.M. Schütz. Boundary-induced
phase transitions in traflic fl.ow. J. Phys. A, 34 (2000) L45-L52.
20. J. Wahle. Information in Intelligent Transportation Systems. PhD thesis,
Gerhard-Mercator-University, 2002.
21. S. Sastry. Nonlinear Systems: Analysis, Stability, and Contral. Springer, New
York,1999.
22. W. Barfield and T.A. Dingus. Human Factors in Intelligent Transportation
Systems. Lawrence Erlbaum Associates !nc., Mahwah, New Jersey, 1998.
23. J. Wahle, A.L. Bazzan, F. Klügl, and M. Schreckenberg. The impact of real-
time information in a two-route scenario using agent-based simulation. Transp.
Res. C, 10, 2001.
 Experiments on Route Choice Behaviour

Reinhard Selten l , Michael Schreckenberg2 , Thomas Pitz l , Thorsten

Chmural , and Joachim Wahle2

1 Laboratory of Experimental Economics, Adenauerallee 24-42, 53113 Bonn,

Germany
2 Physics of Transport and Traffic, Gerhard Mercator University Duisburg,
Lotharstr. 1, 47048 Duisburg, Germany

Abstract. The paper reports laboratory experiments on a day-to-day route choice

game with two routes. Subjects had to choose between a main road M and a side
road S. The capacity was greater for the main road. 18 subjects participated in
each session. In equilibrium the number of subjects is 12 on M and 6 on S.
Two treatments with 6 sessions each were run at the computerized Laboratory
of Experimental Economics at Bonn University. Feedback was given in treatment I
only ab out own travel time and in treatment H on travel time for M and S. Money
payoffs increase with decreasing time. The main results are as follows: Though mean
numbers on M and S are very near to the equilibrium, the fluctuations persist until
the end of the sessions in both treatments. Fluctuations and the total number of
changes are significantly smaller under treatment H. Subjects' road changes and
payoffs are negatively correlated in all sessions.

1 Introduction

Understanding individual travel behaviour is essential für the design of Ad-

vanced Traveler Information Systems (ATIS), which provide real-time travel
information, like link travel times [2]. However, the response of road users to
information is still an open question [3], [4]. Moreover it is not clear whether
more information is beneficial [3]. Drivers confronted with too much infor-
mation may become oversaturated in the sense that information proeessing
becomes to difficult and users develop simple heuristies to solve the problem
[5].
Drivers mayaiso overreact to information and thereby cause additional
fluetuations. Thus, the behaviour of the drivers has to be ineorporated in the
forecast [3], [4], [8] ATIS can reduce fluctuations only if behavioural effects
are eorrectly taken into aeeount.
The Literature reports a number of experiments on route choice behaviour
[4]. Here we foeus on the route choice in a generic two route scenario, which al-
ready has been investigated in the literature [7]. However our aim is to present
experiments with a large number of periods and with sufficiently many inde-
pendent observations for meaningful applications of non-parametric signifi-
cance tests.

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
318 R. Selten et al.

Hone wants to investigate results of day to day route choice which can be
transferred to more realistic environments, it is necessary to explore individ-
ual behaviour in an interactive experimental set-up. Does behaviour converge
to equilibrium? Does more feedback reduce fluctuations? What is the struc-
ture of individual responses to recent experiences? Our experimental study
tries to throw light on these questions.

2 Experimental Setup

Subjects are told that in each of 200 periods they have to make a choice
between a main road M and a side road S for travelling from A to B. They
were told that M is faster if M and S are chosen by the same number of
people. No further information was given to the subjects. The number of
subjects in each session was 18. The time and depends on the numbers ns
and nM of participants choosing M and S, respectively:

tM = 6+2nM
ts = 12+3ns

The period payoff was 40 - t with t = tM if M was chosen and t = ts

if S was chosen. The total payoff of a subject was the sum of all 200 period
payoffs converted to money payoffs in DM with a fixed exchange rate of .015
DM fot each experimental money unit. One session took roughly one and a
half hours. All pure equilibria of the game are characterized by nM = 12 and
ns = 6.
Two treatments have been investigated. In treatment I the subjects re-
ceived feedback only about the time on the road chosen after each period. In
treatment II feedback was provided ab out both times. Six sessions were run
with treatment I and six with treatment Ir.

3 Equilibrium Predictions and Observed Behaviour

For each time series of participants on the side road it can be seen that
there is no convergence to the theoretical equilibrium. There are substantial
fluctuations until the end of the session. The same is true for all sessions
of both treatments. The overall average of numbers of participants on S is
very near to the equilibrium prediction. The standard deviation of the session
average from 6 is never greater than .17 and it is only .07 on the average.
The fluctuations can be measured by the standard deviation of the number of
participants choosing S per period. This standard deviation is between 1.53
and 1.94. In view of these numbers one can speak of substantial fluctuations
in each of the 12 sessions.
 Route Choice Behaviour 319

The fluctuations are a little larger under treatment I than under treatment
11. The effect is significant. The null-hypothesis is rejected by a Wilcoxon-
Mann-Whitney-Test on the significance level of 5% (one sided).
The game underlying the experiment has many pure strategy equilibrium
points. In all of them the number of participants on the side road is 6, but
the set of players who choose S can be any set of 6 players. The multiplicity
of pure strategy equilibria poses a coordination problem which may be one of
the reasons for non-convergence and the persistence of fluctuations. Feedback
on both travel times vs. feedback on only own travel time has a beneficial
effect by the reduction of fluctuations, but this effect is relatively small.
The fluctuations are connected to the total number of road changes within
one session. The Spearman-rank-correlation between the total number of road
changes and the standard deviation of the number of participants per period
on S is .795. This is significant on the level of 1% (one sided). The total
number of road changes under treatment I is greater than under treatment
H. A Wilcoxon-Mann-Whitney-Test rejects the null-hypothesis only on a very
weak significance level of 8.98% (one sided).
Under treatment I subjects who mainly choose only one of the roads
feel the need to travel on the other road from time to time in order to get
information on both roads. Under treatment H there is no necessity for such
information gathering. This seems to be the reason for the greater number
of changes and maybe also for the stronger fluctuations under treatment 1.

4 Response Mode
A participant who had a bad payoff on the road chosen may change his road
in order to travel where it is less crowded. We call this the direct response
mode. A road change is the more probable the worse the payoff iso
The direct response mode is the prevailing one but there is also a con-
trarian response mode. Under the contrarian response mode a road change is
more likely the better the payoff iso The contrarian participant expects that
a high payoff will attract many others and that therefore the road chosen will
be crowded in the next period.
The equilibrium payoff is 10. Payoffs perceived as bad tend to be below 10
and payoffs perceived as good tend to be above 10. Accordingly we classified
the response of a subject as direct if the road is changed after a payoff smaller
than 10 or not changed after a payoff greater than 10. An opposite response
is classified as contrarian. Table 1 shows the numbers of times in which a
subject changes roads (c_ for a payoff below 10 and c+ for a payoff above
10), or stays at the same road (8- for payoff below 10 and s+ for a payoff
above 10).
For each subject such a 2x2 table has been determined and a Yule coeffi-
cient Q has been computed as follows.
Q = c_s+ - c+s_ (1)
c_s+ - c+s_
320 R. Selten et al.

Table 1. 2x2 table for the computation of Yule coefficients.

change stay
payoff< 10 c_ s_
payoff> 10 c+ s+

The Yule coefficient has a range from -1 to +1. In our case a high Yule
coefficient reflects a tendency towards direct responses and a low one a ten-
dency towards contrarian responses.
In each of four sessions, one of them in Treatment land three in treatment
II, there was one player for whom no Yule coefficient could be determined
since these four subjects never change roads. These subjects are not consid-
ered in the evaluation of Yule coefficients.
Evidence for the importance of both response modes can be found in
the distributions of Yule coefficients within a session. If the two response
modes were not present in behaviour one would expect distributions of Yule
coefficients concentrated around o. However the number of subjects with
extreme Yule coefficients below -.5 or above +.5 tends to be greater than
the number of subjects with Yule coefficients in the middle range between -.5
and +.5. A Wilcoxon one sampie test supports this alternative hypothesis by
rejecting the null-hypothesis that none of both numbers tends to be greater
than the other, on the significance level of 0.01 (two sided).
If one classifies subjects with Yule coefficients above +.5 as direct respon-
ders and subjects with Yule coefficients below -.5 as contrarian responders,
then one receives 44% direct responders, 14% contrarian responders and 42%
unclassified subjects.

5 Payoffs and Road Changes

In all sessions the number of road changes of a subject is negatively correlated
with the subject's payoff. Even if subjects change roads in order to get higher
payoffs, they do not succeed in doing this on the average. This suggests that it
is difficult to use the information provided by the feedback to one's advantage.

6 Conclusion
The study has shown that the mean numbers on both roads tend to be very
near to the equilibrium. Nevertheless, fluctuations persist until the end of
the sessions in both treatments. This is of particular interest in view of the
fact that the experiments run over 200 periods which is unusually long and
should be enough to show a tendency of convergence to equilibrium, if there
is one.
Feedback on both road times significantly reduces fluctuations in treat-
ment II compared to treatment 1. However the effect is small. There is a
 Route Choice Behaviour 321

significant rank correlation between the total number of road changes and
the size of fluctuations. In treatment I road changes may serve the purpose
of information gathering. This motivation has no basis in treatment H. How-
ever road changes may also be attempts to improve payoffs. The finding of a
negative correlation between a subject's payoff and number of road changes
suggests that on the average such attempts are not successful.
Two response modes can be found in the data, a direct one in which
road changes follow bad payoffs and a contrarian one in which road changes
follow good payoffs. One can understand these response mo des as due to
different views of the causal structure of the situation. If one expects that
the road which is crowded today is likely to be crowded tomorrow one will be
in the direct response mode but if one thinks that many people will change
to the other road because it was crowded today one has reason to be in
the contrarian response mode. We have presented statistical evidence for the
importance of the two response modes.

References
1. Abbink, K., Sadrieh A., RatImage - Research Assistance Toolbox for Computer-
Aided Human Behavior Experiments. SFB Discussion Paper B-325, University
of Bonn, (1995)
2. Adler, J., Blue, V., Toward the design of intelligent traveler information systems,
Transpn. Res. C 6, (1998), 157
3. Ben-Akiva, M., de Palma, A., Kaysi, 1., Dynamic network models and driver
information systems, Transpn. Res. A 25, (1991), 251
4. Bonsall, P., The influence of route guidance advice on route choice in urban
networks, Transportation 19, (1992), 1-23
5. Gigerenzer, G., Todd, P.M., and ABC Research Group (eds.), Simple heuristics
that make us smart, Oxford University Press, (1999)
6. Hall, R., Route choice and advanced traveller information systems on a capaci-
tated and dynamic network, Transpn. Res. C 4, (1996), 289-306
7. Iida Y., Akiyama T., Uchida T., Experimental analysis of dynamic route choice
behaviour, Trans. Res. B 26, (1992), 17-32
8. Wahle, J., Bazzan, A., Klügl, F., Schreckenberg, M., Decision dynamics in a
traffic scenario, Physica A 287, (2000), 669-681
Transport Out of a Gravitationally Stable
Layer with the Help of a Faster Diffusing
Substance: PDE Simulations and Scaling Laws

Karsten Koetter, Malte Schmick, and Mario Markus

Max-Planck-Institut fuer molekulare Physiologie, Postfach 500247, 44202

Dortmund, Germany

Abstract. Hydrodynamic equations of an incompressible fluid with two dissolved

substances are integrated to simulate previously reported observations of double
diffusive convection. These observations had been performed in a setup that is
extremelyeasy to implement and to observe: a layer consisting of a mixture of a
surfactant, glycerine and water is placed below waterj fingers arising at the interface
transport the surfactant upwards. The calculated mean distance between fingers, as
weIl as their emergence time, both as functions of the initial concentration of glyc-
erine, are described by power laws. These results are in agreement with analytical
approximations and with the experimental observations.

1 Introduction
In the present contribution we examine double-diffusive convection [1-4] at
the horizontal interface between water (upper layer) and a mixture (lower
layer) of a surfactant (8), glycerine (T) and water. (In a previous work [5],
we have reported on experiments with this system). 8 is the alcohol ethoxy-
late CH3-(CH2h-O-[CH2-CH2-0h2-H. The densities are Ps = 0.95 g/cm3
and PT = 1.26g/cm3. A lower layer containing only 8 would rise, while it
would stay at the bottom if it contained T alone. We choose the mixtures in
such a way that they are gravitationally stable in the lower layer. However,
an infinitesimal, finger-like protrusion emerging upwards from the lower layer
is rapidly depleted of T and retains 8 owing to the difference in diffusion co-
efficients (Ds ~ 3.10- 12 m 2/s, DT ~ 10- 10 m 2/s). Therefore, the protrusion
grows and leads to an upwards transport of 8.
An analogous process is considered to be the most important mechanism
for transport of salt in oceans (see [1]). In fact, in low- and midlatitude
regions, upper layers of water are heated by the sun and become richer in
salt due to evaporation. Since the diffusion of salt is slower than that of
heat, salt fingers grow downwards in a configuration that is upside down if
compared to the one described in the preceeding paragraph [6,7].
An example of ecological importance is that of sewage that is injected
by a pipe on the bottom of the sea and is assumed to be safely deposited
because it is heavier than water. However, being colder and (eventually) less
salty than the layers above, salt fingers grow downwards and push dangerous

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 Transport Out of a Gravitationally Stable Layer 323

pollutants towards the surface of the sea [2,8]. Other examples of phenomena
involving double-diffusive convection are: unwanted finger-like distributions
of components in alloy manufacture out of the melt [3,9], layering of magma
[2,10,11], "helium-fingers" and "magnetic-moment-fingers" in stars [3,12-14],
and polymer transport in biological tissues [15,16].
We have shown experimentally [5] that the setup involving the surfactant-
glycerine-water mixture mentioned above has several advantages: i) if there
is sufficient space for the lower layer to expand due to gravity, a bulk
flow induced in the water straightens and confines the fingers quasi-two-
dimensionally, so as to allow comfortable visualization; ü) the extremely low
diffusion coefficient Ds causes long-lasting optical definition of the borders
of the observed fingers; iii) the micelles formed by the surfactant are so large
(length '" 4000 Ä) that they strongly scatter or absorb visible light and are
thus clearly detectable by the bare eye or by a video cameraj thus, no dyes
(see e.g. [2,17,18]) or special optical eqipment, e.g. to measure the refractive
index [9,19-21], are needed; and iv) the medium is not distorted by invasive
techniques, such as withdrawing samples or inserting conductivity probes
or thermocouples [9,17,18,20,21]. On the other hand, a disadvantage of our
setup is that the layer of the mixture is so thin that only features involving
the horizontal direction are observable (such as the distance between fingers)
but not those in the vertical direction (such as transport velocity).
In this contribution we will report on PDE simulations and analytical
estimates of the finger emergence time Tem and the mean distance between
fingers (A) versus initial concentrations. The results will be compared with
measurements.

2 Analytical Estimates
Aperturbation egrows as
.; '" exp (fot A(t') dt') , (1)

where
_1
A '" Tl !(R p ), (2)
see [4]. R p = [(pT - poyrz]/[(ps - po)Szl (Po: density of water). Tz and Sz
are the vertically averaged gradients 8T/8z and 8S/8z, respectively. H we
assume times small enough to approximate T ~ To , S ~ So at z = 0
(To, So : initial concentrations in the lower layer), then Tz = To/h(t) and
Sz = So/h(t) (h(t) '" t 1 / 2 To is the height of the layer, growing vertically
by diffusion). Thus, A '" To C 1/ 2 f(R p ) and R p '" R~ = To/So. Each set of
measurements was performed holding R~ = const. Therefore

(3)
324 K. Koetter, M. Schmick, M. Markus

The measurements were performed at a time t = T em , such that has an e

approximately fixed value, namely that at which the minima of the absorbed
light are within a given percentage p of the intensity measured with a void
dish. Therefore, Eqs. (1) and (2) yield To!t i = const, or

(4)

with "(T = 2/3.

-_1
For the wavelength A it has been shown [22] that A Tz I"V 4. Thus A I"V

(Tor!)-i. Setting t = T em and considering (4) we obtain

(5)

with "(J... = 1/3.

3 Experiments

Measurements were performed as described in Ref. 5. We now varied the

initial concentration To of glycerin in the mixture, while holding the ratio
R~ = Tol So constant, so as to fulfill the conditions required by the estimates
in Sec 2. The mixture was stirred for 1-2 minutes and then 50 pI were injected
with a pipette over a time of 20 s at a constant rate at the center of the bottom
of a cylindrical dish (diameter: 2 cm) containing water (~3 mm high). For
this injection we used a Precidor HT-pump (Infors, Basel). The mixture in
the dish was illuminated with white light from above, while the formation of
fingers was monitored (by determining the light absorbtion of S) from below
by a light microscope connected to an image acquisition equipment. A was
determined by two-dimensional Fourier transforms of the images, averaging
(at fixed T o and R~) over 50 independent measurements. We determined A
at the time t = T em ("emergence time") when p=5%.
Fig. 1 shows T em and A VS. T o for different values of R~. For R~=l,we
obtain "(J... = 0.34±0.02 and "(T = 0.42±0.03. The fact that the plots in Fig. 1
are approximately parallel reveals that these values of "(J... and "(T are nearly
independent of R~. Substituting the surfactant-glycerine-water mixture by
a non-dairy coffee-whitener, which also contains a surfactant, as weH as a
component that takes up the role of glycerine (see [5]), we obtain Fig. 2.
This plot yields "(J... = 0.38 ± 0.02 and "(T = 0.51 ± 0.03. Unfortunately,
the whitener has the disadvantage that we do not know quantitatively its
chemical composition.
 Transport Out of a Gravitationally Stable Layer 325

4 PDE Simulations

We solved the equations for an incompressible fluid with two dissolved sub-
stances 8 and T:

8t v + (vV)v = -Pg - 1
-Vp+ v~v (6)
Po Po
Vv=O (7)
ßt8 + (v V)8 = Ds~8 (8)
8t T + (v V)T = DT~T (9)
P = Po + (pT - po)T + (pS - Po)8 (10)
For initialization of the instability, we set the velocity v at t=O equal
to noise vectors with randomly and homogeneously distributed directions
and magnitudes (the latter between 0 and 10-2 mm/s). We set the viscosity
v = 1.1.10-6 m 2 /s as a constant throughout this work, following our capil-
lary measurements. We assumed cyclic boundary conditions in the horizontal
directionsj at the bottom and the top, we set no-slip boundary conditions for
v and no-flux boundary conditions for 8 and T. We used a finite volume
method on a colocated grid with an implicit three time level scheme, central
difference scheme and the SIMPLE algorithm for pressure correction [23]. We
set the time step to ~t=O.1 s. When considering the injection process, we set
the grid size to ~x = 4.10- 2 mm and the domain size to 10 x 10 x 3 mm3 ;
else, we set ~x = 10- 2 and the domain to 1 x 1 x 1 mm3 •
We assumed injection to occur out of a cylinder (vertical axisj diameter:
0.8 mm ; base centered and 0.5 mm above the bottom). For t < 20 s, we
set all variables equal to zero in a thin region around the cylinder (cylinder
wall; thickness: 0.16 mm). Inside the cylinder for t < 20 s, we set 80 =const,
To =const and v=(0,0,-1.6 mm/s). At t=20 s, we restarted calculations,
setting 8 = T =0 in the cylinder.
Fig. 3 shows results, not by considering injection, but by starting with
a vertically constant distribution of T and 8 in the lower layer. For R~=l,
we obtain T>' = 0.3 ± 0.01 and TT = 0.67 ± 0.02. As in Fig. 1, the fact that
the plots in Fig. 3 are approximately parallel reveals that these values of T>'
and TT are nearly independent of R~. Simulations considering injection yield
T>' = 0.33 ± 0.02 and TT = 0.43 ± 0.02, also independently of R~.
We also integrated the PDEs starting with T ans 8 homogeneously dis-
tributed within with a cone (height: 1 mm, diameter: 4.8 mm) that expands
due to gravity in a larger domain. In this way, we simulated the experiments
[5] in which a drop is injected at the center of a dish. An example of such a
simulation is shown in Fig. 4a, as compared with a corresponding measure-
ment (Fig. 4b). The gravitational expansion of the mixture induces a torodial
bulk flow of the surrounding water. This flow confines the mixture near the
bottom of the medium, as explained in [5]. Thus, views from above (as in
Fig. 4) permit to clearly visualize single fingers.
326 K. Koetter, M. Schmick, M. Markus

Note that in Fig. 4 only subsets (containing the centers of the total do-
mains) of the simulated and the observed areas are shown. This is sufficient
because - disregarding small fiuctuations - the structures are rotationally
symmetrie (see [5]). The distance A between fingers is visible here around the
centers of these pictures.

1()3
a) b)
~
E
1-0>
E
102 E
R'p=1 ~
R'p=1

101 0.1
0.1 . 1 0.1 10 T o [Vol%]
10 T o [Vol%]

Fig. 1. Finger emergence time Tem and mean distance between fingers A VS. the
initial glycerine concentration. R!"
= To/So.

103

~
E
1-0>

1()2
,al E
E
~
b)

101 L-~ __ ~ ____ ~~ ______ ~ 0.1 L-~_~ __ ~_~ _ _ _ _.....J

0.1 10 Co [Mass%] 0.1 10 Co [Mass%]

Fig.2. T em and A VS. the initial concentration of the whitener.

5 Conclusion
Varying To within factors of 20-30 in experiments (Fig. 1 and Fig. 2) and
simulations (Fig. 3) yields power laws comparable to analytical estimations
(Eqs. (4) and (5)). Moreover, the exponents are independent of R~, as pre-
dicted analytically and as indicated by the parallel displacements of the plots
 Transport Out of a Gravitationally Stable Layer 327

0.5 r-------..-------~----,
a)
~
E
1-0> E
E
;:t

0.1

10' L -_ _ ~ __ ~ ____ ~~ ____ ~

0.1 10 T o [Vol%) 0.1 10 To [Vol%)

Fig. 3. As Fig. 1, but from PDE simulations. (Initialization was done by a step
profile of the concentrations).

a)

b)

Fig. 4 . a) PDE simulations of surfactant fingers straightend and confined quasi-two-

dimensionally by the ßow of water that is induced by the radial expansion of the
lower layer (view from above) . For comparison with the experimental observations,
the darkest grey level corresponds to the maximum of S . b) Experimental observa-
tion corresponding to the simulation shown above. a,b) Simulation and experiments
were performed in larger domains; here only slices of these domains passing through
the center (12 x 2 mm 2 for simulations 6 x 1 mm 2 for experiments) are shown.

in Figs. 1 and 3. For the exponent "/>. (describing the mean distance between
fingers), the agreement between experiments, PDE simulations and the an-
alytical estimate of 1/3 is excellent. For the exponent "/T (corresponding to
the time of emergence of fingers) the analytical estimate of 2/3 agrees weIl
with the result obtained from PDE integration initialized with a step profile.
A lower value is obtained in experiments initialized by injection bT ~ 0.42),
and this value agrees with that resulting from PDE simulations with injection
(~ 0.43).
All in aIl, we obtain consistency between experiments, PDE simulations
and analytical estimates. The fact that rough analytical approximations agree
with PDE simulations, with weIl controlled experiments (Fig. 1) and with
328 K. Koetter, M. Schmick, M. Markus

experiments using quantitavely unknown mixtures (Fig. 2) points to a general

validity of the power laws reported here.

Acknowledgement
We thank the Deutsche Forschungsgemeinschaft (Grant Ma 629/5) for finan-
dal support.

References
1. Chen, C. F., Johnson, D. H.: J. Fluid Mech. 138 (1984) 405-416
2. Huppert, H. E., 'furner, J. S.: J. Fluid Mech. 106 (1981) 299-329
3. Turner, J. S.: Ann. Rev. Fluid Mech. 6 (1974) 37-56
4. Schmitt, R. W.: Phys. Fluids 26 (1983) 2373-2377
5. Koetter, K., Markus, M.: Europhys. Lett. 55 (2001) 807-813
6. Williams, A. J.: Science 185 (1974) 941-943
7. Tait, R. 1., Howe, M. R.: Nature 231 (1971) 178-179
8. Fisher, H.: Water Res. 5 (1971) 909-915
9. Coriell, S. R., Cordes, M. R., Boettinger, W.J., Sekerka, R.F.: J. Crystal Growth
49 (1980) 13-28
10. McBirney, A. R., Noyes, R. M.: J. Pet. 20 (1979) 487-554
11. McBirney, A. R.: J. Volcanol. Geotherm. Res. 7 (1980) 357-371
12. Schmitt, J., Rossner, R.: Astrophys. J. 265 (1983) 901-924
13. Goldreich, P., Schubert, G.: Astrophys. J. 150 (1967) 571-587
14. Spiegel, E. A.: Ann. Rev. Astron. Astrophys. 10 (1972) 261-304
15. Laurent, T. C., Preston, B. N., Sundelof, L.O.: Nature 279 (1979) 60-62
16. Harper, G. S., Comper, W. D., Preston, B. N.: J. Biol. Chem. 259 (1984)
582-589
17. Linden, P. F.: Deep-Sea Res. 20 (1973) 325-340
18. 'furner, J. S.: Deep-Sea Res. 14 (1973) 599-611
19. Huppert, H. E., Manins, P. C.: Deep-Sea Res. 20 (1973) 315-:323
20. Shirtcliffe, T. G. L.: J. Fluid. Mech. 57 (1973) 27-43
21. Taylor, J. R., Veronis, G.: Fluid Mech. 321 (1996) 315-:333
22. Stern, M. E.: Tellus 12 (1960) 172-175
23. Ferziger, J. H., Peric, M.: Computational methods tor fluid dynamics (Springer,
New York, 1997)
 Mieroseopic Parameters and Maeroseopic
Features of Traffie Flow

Peter Berg1 and Eddie Wilson 2

Department of Mathematics, 8imon Fraser University, 8888 University Drive,

Burnaby V5A 186, Canada, www.peterberg.net

Abstract. A major task of trafik modelling is to investigate how trafik parameters

on the microscopic scale, such as reaction time and inertia, infIuence the macro-
scopic fIow phenomena. In this article, we present a transformation that relates
microscopic car-following models to their macroscopic continuum counterpart. For
a specific type, the optimal-velocity model, it turns out that the related model is
analogous in the sense that it fuHills the same linear stability criterion. Moreover, it
predicts similar fIow patterns for similar trafik situations as they occur at on-ramps
or between platoons of vehicles of different fIuxes. However, the analysis suggests
that reaction time and multi-species fIows may have a profound effect on the fIow
pattern. To date, this has rarely been taken into account in continuum models. This
paper points out some of the problems and leaves some open questions regarding
macroscopic trafik modelling.

1 Microseopie Parameters of Traffle Flow

From personal experience we know that car drivers react mainly to traf-
fic events ahead. On fairly congested motorways, drivers are forced into a
certain reaction pattern rather than making their own decisions. Yet on a
macroscopic scale, the coupling of many individual vehicles can lead to cer-
tain traffic states (phases) like stop-and-go or synchronised flow [1]. It is,
therefore, a very interesting task to investigate how microscopic reaction al-
gorithms and parameters influence the overall macroscopic flow situation.
Figure 1 shows a general sketch of a one-lane highway. In what follows,
vehic1es will carry integer numbers n, which indicate their position in the
convoy. Generally speaking, the system is characterised by six parameters,
some of which vary from vehic1e to vehic1e:
1. Reaction time: Tn
2. Inertia of cars: In = I/an
3. Distance to the cars in front also referred to as headway: h n , (h n +hn - 1 ),
(h n + h n - 1 + h n - 2 ), ...
4. Own current speed, V n , and those of others, Vn-l, V n -2, ...
5. Top speed as hn --+ 00: V.:,nax
6. Minimum distance to the car in front in a trafik jam: h~in
It is at the heart of trafik flow modelling that one finds reasonable models
based on these parameters to reproduce real features of the system. One can

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
330 P. Berg, E. Wilson

approach this problem from both the macroscopic and microscopic side. We
will start with the latter and try to link it to the former.

road

Fig.1. An idealised sketch of road traflic. Vehicles are separated by headway h n

and carry speed v n .

2 Stochastic and Deterministic Microscopic

Description
At first it seems very questionable whether a complex system containing so
many variables and different parameters, like car trafik, can be described by
simple flow models. Often the simplifications lie in the following assumption:
identical vehicles (single-species traffic), one lane only and the neglect of
reaction time, varying road conditions as weIl as ramps.
Another fundamental question that arises from the modelling aspect is
whether trafik can be treated as a deterministic system, despite variations
in driver behaviour temporally and amongst other drivers. We will come
back to this issue in more detail in Chap. 6. However, deterministic models
match some of the trafik features quite weIl [2] and that is why we will focus
on them in this publication. For reasons of completeness though, a popular
representative of the class of stochastic models shall be briefly presented
before we proceed.

2.1 A Cellular Automata Model

A radical approach representing the road and the vehicles, would be to divide
it into cells whose lengths equal roughly the stretch that a vehicle occupies
in a traffic jam. The assignment of either of these two numbers to the cells
indicates whether it is occupied ('1') by a car or empty ('0'), as shown in
Fig. 2. Moreover, each occupied cell holds an integer number between zero
and vrnax that stands for the current speed of that car.
Nagel and Schreckenberg [3] proposed such a cellular automata (CA)
model based on a small number of update rules, but allows for stochasticity
 Microscopic Parameters and Macroscopic Features of TrafIic Flow 331

via some sort of "dawdling". The updates transform the state of the system
into a new state at a time interval, t1T, later via cell hopping of the cars.
This ansatz is very interesting for its discretisation of both space and time,
but will not be pursued here any further. It might also be mentioned that
some refer to CA models as car-following models, since the update rules also
inelude the distance between the vehieles.

o 1 I 0 I 1 1

road

Fig. 2. In a cellular automata model, the road is cut into blocks which may ('1')
or may not ('0') be occupied by the vehicles.

2.2 Car-following Models

Car-following models are generally based on ordinary differential delay equa-
tions (ODDE) of the form

(1)
It means that the reaction of a driver depends somehow on his speed, and
his distance to the cars in front at some point in the past due to the reaction
time. This discrete ansatz used to be considered as a different description
than the macroscopic continuum models.
So far, many car-following models have been proposed. Some inelude reac-
tion time and some neglect this effect which leads to a set of coupled ordinary
differential equations (ODE). Some are based on macroscopic trafik data and
some on speed-headway measurements [4].
However, linear stability analyses and numerical simulations reveal how
elose each suggestion comes to reality. Currently, this leaves the scientifk
community with the optimal-velo city model, also referred to as the Bando
model, which we will discuss next.

3 Optimal-velo city Car-following Models

In this paper, we concentrate on the most promising car-following model to
date with respect to the right instability criterion, and stop-and-go simula-
tions, the optimal-velocity (OV) model [5]
(2)
332 P. Berg, E. Wilson

It clearly neglects delay which might be a major criticism. Cars tend to adjust
to an optimal speed

VB(h n ) = tanh(h n - 2) + tanh(2), (3)

depending on the current headway hn as shown in Fig. 3. Here, we choose

the dimensionless form based on driver data [5].
The relaxation term resembles those of earlier proposed continuum mod-
els like Payne's [6], Kühne's [7] and Kerner-Konhäuser's [1]. This makes us
wonder whether they are somehow related, especially since the numerical
simulations show similar features to the latter model [2].
Figure 4 represents the fundamental diagram, the plot of the flow q ver-
sus the density p, where p = 1/h to first order. Two regions can be clearly
distinguished: the linear stable regime where the flow is given by q = pV(p),
and the linear unstable regime in which small perturbations of a station-
ary, homogeneous and uniform flow lead eventually to congested stop-and-go
trame.
Bando et al. found that the system becomes linear unstable when

2V~(ho) > 1 (4)

a

where ho is the initial headway of the platoon [5]. This defines the boundary
of the congested and the homogeneous flow in Fig. 4. Depending on the OV
function chosen, the model is generally stable in the free flow regime and
unstable in the denser region. Moreover, it shows meta-stabilities [4] and
other trafIic features in almost fuH accordance with the Kerner-Konhäuser
model [2]. This further pushes the question of whether there is a link between
these models, or whether this has another explanation.

4 Derivation of Macroscopic Counterparts

In this section, we give a brief summary of how to relate car-following models

to continuum analogues. For a more detailed discussion, see Berg et al. [9],
Lee et al. [10] and Helbing [11]. It is striking that the headway h is the variable
representing distances in the car-following model, whereas it is the density
p in the continuum models. Therefore, we seek for a relation between these
two variables. One way to do so is to change formally to continuum variables

Vn(t) --+ v(x, t), (5)

hn(t) --+ hex, t), (6)
and require

lo
h (X,t)
p(x+y,t)dy=.1 (7)
 Microscopic Parameters and Macroscopic Features of Trafik Flow 333

1.5

0.5

O~~~----L---~ __ ~ ____ ~ __ ~ __- - J

o 2 3 4 5 6 7
h

Fig. 3. The optimal-velocity function of the Bando model. Assumption of point-like

vehic1es leads to a minimum headway h min = o.

~~­
oo~....-
q

Fig.4. A qualitative sketch of the fundamental diagram of the OV model (2) con-
tains regions of linear stable and linear unstable flow (8)

Here, p defines h uniquely, but not vice versa. However, this is what we are
looking for.
334 P. Berg, E. Wilson

A Taylor expansion of the integral yields

1 2 1 3
hp + 2! h pz + 3! h pzz + ... = 1 (8)

Solving up to seeond order gives

h '" ! _ .!!=...- _ pzz + ~ (9)

- P 2p3 6p4 2p5 ·

Here, the series is assumed to eonverge sufficiently fast and henee, that higher
order derivatives playa minor role.
We assume that the expansion (9) ean now be used to transform any
ear-following model into a eontinuum version. However, we will foeus on the
OV model for astart, and eomment on general transformations later on. The
substitution of (9) into (2) and a Taylor expansion of VB(h(x,t)) delivers

-
Vt + VV z ~ a [V(p) - v] + aV-, (p) [pz pzz
2p + 6p2 -
p; ]
2p3 (10)

Here we set

V(p) = VB(l/ p) (11)

Moreover, the system is closed by the eonservation of ears

Pt + (vp)z = 0, (12)
which holds via
d
dt
l 0
h (Z,t)
p(x + y, t) dy =0 . (13)

In order to speak of an analogous model, we have to establish that it shows

similar dynamical behaviour in its numerieal simulations (next seetion), and
that it fulfills the same stability eriterion. In fact, the latter ean be proven
explicitly [9], and (4) reads now (V' < 0)

_ 2p~V'(pO) >1 . (14)

a
Strietly speaking, the stability eriterion is only eonserved for long wave-
lengths, but the aceuraey inereases with the number of terms in the ap-
proximation of hin (9).

5 Similarity of Flow Patterns

One suitable eomparison of the dynamieal behaviour of diserete ear-following
and eontinuum models, are travelling waves. In the ear-following model, this
 Microscopic Parameters and Macroscopic Features of Traffic Flow 335

is achieved by running a simulation for a sufficiently long time until the struc~
ture changes only minisculely. In contrast, the real advantage of continuum
models is shown by their reduction to a single ODE [9].
The boundary conditions of both simulations are linked by hex -+ -(0) =
L = l/p_ = l/p(x -+ -(0), and similarly hex -+ (0) = h+ = l/p+ =
1/ p(x -+ (0), where the speed equals the optimal velo city.
Both numerical results are then drawn in the same diagram. One example
is Fig. 5, which reveals how the wave structure of the car-following model
approaches the travelling wave solution of the continuum model. Note that
the solutions of the former have been shifted along the x-axis for reasons
of comparison. After t = 1000, the points lie almost perfectly on the curve
predicted by the continuum model.
When we increase the change in headway across the wave structure we
find, eventuaIly, oscillatory behaviour [12] as illustrated in Fig. 6. We now
observe slight deviations between the two solutions. This comes at no surprise
since larger gradients are now involved and would require higher order terms
in the asymptotic series (9) in order to be matched more accurately. Another
example arises from the simulation of an on-ramp in the discrete OV model
(2) [13]. Here, an initial homogeneous and stationary flow ofuniform headway
h o is perturbed by a constant ramp flux that sets in at t = O.
Figure 7 shows a ramp flux, qramp = 0.01 vehicles per unit time, that is
placed onto a lane of maximum flux at ho = 2.78. The structure that develops
reminds us of cnoidal water waves, a solitary wave train.
An analysis in the corresponding continuum picture reveals that this pro-
file can be explained to first order by a sech2 -solution of a Kortweg-de Vries
(KdV) equation via the neglect of dissipation [13]. In Fig. 8, the comparison
between this approximation, a sech 2 -like fit and the numerical data reveals
two things: firstly, the analytical solution of the KdV equation does not quite
reach the right height of the density hump and secondly, the numerical data
of the car-following model is not exactly symmetric unlike the sech2 -shape.
Both effects can be explained by the steep gradients and the dissipation in-
volved, which can only be matched by further derivatives of p in the analysis.
Strictly speaking, the comparison of the two model types (2) and (10) should
be extended to non-stationary solutions. Then the dependence of the linear
stability criterion in the continuum model on the wavelength might have a
significant impact and might show the limits of this transformation.

6 More Realistic Models

We now turn to the main part of this publication. We have seen that car-
following models can be related to a continuum counterpart via the neglect
of certain aspects: reaction time, multi-Iane flow, different driver and vehicle
types, as weIl as the influence of vehicles further ahead.
In this section, we address these issues to reveal where the problems of a
more realistic description of traffic flow lie.
336 P. Berg, E. Wilson

3.05
continuum - -
car-following : t=250
3 carfollowing: t=5OO
car-following: t=l000

2.95
>-
01
~
l)
01
2.9
Ql
.c
2.85

2.8

2.75
50 60 70 80 90 100 110
x

Fig.5. A comparison of the asymptotic behaviour of the car-following OV model

(2) and the travelling wave solution of the continuum counterpart (10). The car-
following model evolves from a Heaviside-like jump in headway at t = 0 with
boundary conditions as given in the text (4).

32
continuum - -
3 caF following

2.8

2.6
>-
<0
~
2.4
l)
cu
Ql
.r::
2.2

1.8

1.6

1.4
-680 -675 -670 -665 -660 ·655 ·650 -645
x

Fig.6. For certain changes in headway, the wave profile assumes oscillatory be-
haviour. This is still matched by the continuum model (10) although with less
accuracy.
 Microscopic Parameters and Macroscopic Features of Traffic Flow 337

3.5 ,..--""T""'--..---.----.----:-I>=2-=.7S.,----
_==""
1>=2.75 _ ...._..
<hmp = 0.01 1>=2.88 -
3
=·lJi-T9. 'Fi·-r=.. .r='i . n·~-·_. ·_. ·_ . .
2.5

1.5

-400 ·300 ·200 ·100 o 100

Fig.1. An on-ramp simulation of the OV model (2) shows solitary wave structures
in certain parameter regimes.

0.55 , - - - - - - . - - - - . .---,---.......,.....,---.---,
nomelical data •
analytical SOIulion -
p fil ..........·

0.5
/\

!(\\\
/' \

J11~\
0.45

0.4
••.i. \ ..
t---o-+o-""-" ' ....... -'-'-~---+-l
0.35 L-_--'~_~ _ _~_ _--'-_ _--'---'
· 190 ·1 80 ·170 ·160 -150 ·140

Fig. 8. The wave solution of the discrete car-following model can be described
approximately by a KdV equation derived from the continuum counterpart.

6.1 Multi-species Traffle; Temporal and Spatial Variations of

Driver Behaviour

A first objection to simplified car-following models like the OV model (2)

might be that it does not incorporate the diversity of driver types on the road.
A straightforward way to take this into account is to allow the parameter a
and the OV function VB to depend on the vehicle number

a -+ a(n), (15)
VB -+ VB(n) (16)
338 P. Berg, E. Wilson

A macroscopic (continuum) counterpart for multi-species flow would

now read [4]

Vt + VV x = a [V(p) - v] + aV'(p) [~; + ~;~ - ;;3] , (17)

Pt + (pV)x = 0, (18)
d
dt a(x, t) = at + vax = 0, (19)
d - --
dt V(x,t) = Vt +vVx = 0 . (20)
Equations (19) and (20) mean that the vehicle's and driver's characteristics
do not change with time and flow with the vehicle. It is worth pointing
out that the system (17)-(20) reminds one of Helbing's model [14], whose
modelling approach is based on the introduction of higher order statistical
variables, like the speed variance. In accordance, we also find a set of three
coupled partial differential equations, if we assume that either V or a does
not depend on time or space.
The initial conditions might be
p(x, 0) = po(x) , v(x,O) = vo(x),
a(x, 0) = ao(x) , V(x, 0) = Vo(x)
In the car-following model, we observed phase transitions caused by a small
amount of vehicles whose characteristics differ from the main platoon [4].
Hence, it remains an interesting question as to whether the continuum coun-
terpart above would show similar results. However, in real trafIic flow we do
not only find different driver and vehicle types. Generally, those character-
istics also change with time and space due to road layout, speed limits
and weather conditions, for example. It can be expected that these variations
might cause phase transitions of the flow, also.
The modelling aspect reads similarly to multi-species flow. In addition, we
have to add two functions, A(x, t) and B(x, t) say, to the right hand sides of
(19) and (20). The crucial problem is to determine their form, which remains
an unresolved issue when the vehicle characteristics change with time. It is
achalienging task to link these functions to the car-following counterparts,
where this effect is taken into account straightforwardly via the dependencies
an = a(n, t) and VB,n = VB(n, t).
In fact, it is much easier to prescribe spatial variations both for the car-
following model
a -+ a(x), (21)
VB -+ VB(X), (22)
and equivalently for the macroscopic counterpart
a -+ a(x), (23)
V -+ V(x) . (24)
 Microseopic Parameters and Macroseopie Features of Traffie Flow 339

In this case, the kinematic equations (19) and (20) are superfluous, and the
system consists again of only two partial differential equations (PDEs).
Nevertheless, it remains achalienging task to investigate what trafiic sim-
ulations predict for temporal and spatial variations.

6.2 Reaction Time

An advantage of car-following models over continuum models clearly shows
in the simulation of reaction time, which can easily be incorporated into
car-following models via differential delay equations (DDE).
As an example, the OV model (2) could be modified to give an ordinary
differential delay equation (ODDE) of the form

(25)
Hooper [15] analysed the linear stability of this model, and observed os-
cillations and even backward moving vehicles in his numerical simulations,
depending on the parameters chosen.
Generally, the reaction depends on the history of a past time interval,
which leads to an integro-differential equation. However, we focus on a dis-
crete delay here for reasons of simplicity.
The macroscopic counterpart now turns into a partial differential delay
equation (PDDE)

! v(x + .1x, t + T) = a [V(p) - v] + aV'(p) [~; + :;~ - ;;3] , (26)

where Xi + .1Xi at ti + T is obtained by integrating

x(t) = v(x, t), X(ti) = Xi (27)

from ti -4 ti + T. Suffice to say that the integration of this PDDE closed
by the conservation of cars (12), is not only a hard task but also very time
consuming in terms of the numerical requirements.
However, time delay might have a relevant impact on the flow and should
be taken into account. Unfortunately, this is only practical on the microscopic
level.

6.3 Multiple Look-ahead Models

A major problem of the dimensional OV model

(28)

are the extreme values of acceleration and deceleration. Here, Vdata is a func-
tion derived from real traffic data [5]. Hooper [15] suggested various ways
to fix this problem: limiting the maximum acceleration and deceleration by
340 P. Berg, E. Wilson

cut-offs, introducing time delay and multiple look-ahead models, which are
based on reactions towards numerous drivers ahead. Only the last approach
appeared as being successful. Multiple look-ahead in the dimensionless OV
model (2) can be most easily simulated in two different ways:

Model A: vn = a [tßjVB(bn-j) - vn] , (29)

3=0

Mo
deI B: ·
Vn = a [
"'ß
1::0
k Vi ("Lf=o
j B
. bn - l
j +1 ) - Vn
]
, (30)

with ßj > 0 and "L~=o ßj = 1. Both approaches are reduced to the standard
OV model (2) in case of reactions to the vehicle ahead only. The unresolved
issue here is how to obtain a macroscopie counterpart. Could downstream
information be incorporated by higher order spatial derivatives of P and v or
is a non-Iocal modellike Nagatani's [17]
(pv)t = a [Po VN(p(X + 1)) - pv], (31)
with VN(p) = VB(I/p), necessary? Here, non-Iocality refers to the spatial
coordinate rat her than time. Similarly to reaction time, we may deal with
an important effect on the fiow that cannot be simulated on the macroscopie
level.

6.4 Multi-Iane Flow

Like time delay and multiple look-ahead, the simulation of multi-Iane fiow
poses another unresolved problem to macroscopie fiow simulations. This is
because the discrete nature of trafiic fiow becomes most apparent upon lane-
changing, for two reasons. Firstly, headways and hence densities do not
change continuously. Secondly, lane-changing is a process that depends on
many conditions, e.g. speed and headway differences to cars in front and
behind, on the same lane and on the other lane.
How could these discrete pro ces ses be incorporated into a macroscopie
counterpart? So far, models have only taken density criteria into account,
but not speed criteria.
An example is the asymmetrie lane-coupling [18] described by the follow-
ing coupling term on the right hand side of the equation for the conservation
of cars
I-P2 I-PI]
s(x, t) =a [PI
+ W P24 - P2 Cl + W PI4 .
C2
(32)

Macroscopie fiow models will never be able to match the random and discrete
process of lane-changing, regardless of the inclusion of speed in the coupling
term [4]. Unfortunately, the interaction of lanes might also be of such impact
on the fiow that it should enter the modelling process, which can again only
be done realisticly in mieroscopic models.
 Microscopic Parameters and Macroscopic Features of Traffic Flow 341

7 Synchronised Flow

In the unstable regime of real motorway trafiic (p> 50 veh.jkm) , there is no

definite allocation of the density to the flow anymore. This leads to scattered
data points in the fundamental diagram obtained from flow measurements.
This area contains two phases: stop-and-go trafiic and the synchronised flow
regime [191. In the latter case, the lanes are strongly coupled, which leads
to a homogeneous movement of vehicles on alllanes, with an overall low
velocity but high throughput. Overtaking is impossible, and lane-changing
rare compared to free flow or traffic jams. It typically occurs at on-ramps
where a localised perturbation causes a linearly stable free flow of high speed
and flux to break down and move more slowly. Motorway data shows that
synchronised flow consists of a great number of steady state flows along the
road. Each state is characterised by its density and its flux. It has been
claimed of various models that they explain synchronised flow. On the other
hand, it is very challenging to distinguish between real synchronised flow and
a similar numerical phenomena.
One fundamental question is, therefore, whether we need more complex
models in which, for example, lane-coupling and ramps play a role. Perhaps
we have to go even further and describe synchronised flow as a phenomena
of a non-symmetrical1 delay system consisting of many particles whose mi-
croscopic features are different from each other and vary randomly with time
and space (e.g. road conditions and layout)? This is indeed how traffic flow
could be characterised most generally.
Considering this highly complicated system, it is surprising how successful
some simple flow models are.

8 Outlook

To date, it remains a great problem to work out whether traffic flow can
be described by simple models, neglecting stochasticity, reaction time, and
different vehicles and drivers on the road. This article shows and addresses
how microscopic features could enter the macroscopic description. Here, re-
action time and the simulation of multi-Iane flow pose serious hurdles since
the corresponding continuum models are neither practical nor exist as such.
In particular, the much discussed and unresolved issue of synchronised flow
might lead to a more comprehensive ansatz of modelling trafiic flow [141.

Acknowledgement Peter Berg would like to thank Professor Woods,

University of Cambridge, for his support throughout the author's PhD.
Moreover, we thank Steven Hooper who delivered vital results in his MSc
thesis.
1 The term "non-symmetrical" refers to the violation of actio=reactio between the
vehicles due to the reaction towards downstream events only.
342 P. Berg, E. Wilson

References
1. Kerner, B.S., Konhäuser, P.: Structure and parameters of clusters in traffic
fiow. Phys. Rev. E 50 (1994) 54-83
2. Herrmann, M., Kerner, B.S.: Local cluster effect in different traffic fiow models.
Physica A 255 (1998) 163-188
3. Nagel, K., Schreckenberg, M.: A cellular automaton model for freeway traffic.
J. Phys. I. France 2 (1992) 2221-2229
4. Berg, P.: Optimal-velocity models of motorway traffic. PhD Thesis, University
of Bristol (2001)
5. Bando, M., Hasebe, K., Nakayama, A., Shibata, A., Sugiyama, Y.: Dynamical
model of traffic congestion and numerical simulation. Phys. Rev. E 51 (1995)
1035-1042
6. Payne, H.J.: FREFLO: a macroscopic simulation model of freeway traffic.
Transp. Res. Record 722 (1979) 68-77
7. Kühne, R.: Macroscopic freeway model for dense traffic - stop-start waves and
incident detection. 9th Int. Symp. on Transp. and Traffic Theory, VNU Science
Press (1984) 21-42
8. Sugiyama, Y.: Dynamical model for congestion of freeway traffic and its struc-
tural stability. In: Bachem, A., Schreckenberg, M., Wolf, D.E., Traffic and gran-
ular fiow, World Scientific (1996) 137-150
9. Berg, P., Mason, A., Woods, A.W.: Continuum approach to car-following mod-
els. Phys. Rev. E 61 (2000) 1056-1066
10. Lee, H.K., Lee, H.W., Kim, D.: Macroscopic traffic models from microscopic
car-following models. Phys. Rev. E 64 (2001) 056126
11. Helbing, D.: From microscopic to macroscopic traffic models. In: Parisi, J.,
Müller, S.C., Zimmermann, W., A perspective look at nonlinear media, Springer
(1998) 122-139
12. Berg, P., Woods, A.W.: Travelling waves of an optimal-velocity model of free-
way traffic. Phys. Rev. E 63 (2001) 036107
13. Berg, P., Woods, A.W.: On-ramp simulations and solitary waves of a car-
following model. Phys. Rev. E 64 (2001) 035602
14. Helbing, D.: Traffic modelling by means of physical concepts. In: Bachem,
A., Schreckenberg, M., Wolf, D.E., Traffic and granular fiow, World Scientific
(1996) 87-104
15. Hooper, S.: Modification ofBando's car-following model ofhighway traffic. MSc
Thesis, University of Bristol (2000)
16. Bando, M., Hasebe, K., Nakanishi, K., Nakayama, A.: Analysis of optimal ve-
locity model with explicit delay. Phys. Rev. E 58 (1998) 5429-5435
17. Nagatani, T.: Density waves in traffic fiow. Phys. Rev. E 61 (2000) 3564-3570
18. Lee, H.Y., Kim, D., Choi, M.Y.: Continuum model for two-Iane traffic fiow. In:
Schreckenberg, M., Wolf, D.E., Traffic and granular fiow '97, Springer (1999)
433-438
19. Kerner, B.S., Rehborn, H.: Experimental properties of phase transitions in traf-
fic fiow. Phys. Rev. Lett. 79 (1997) 4030-4033
 An Adaptive Smoothing Method for Traflie
State Identifieation from Ineomplete
Information

Martin Treiber and Dirk Helbing

Institute for Economics and Traffic, Faculty of Traffic Sciences "Friedrich List",
Dresden University of Technology, D-01062 Dresden, Germany

Abstract. We present a new method to obtain spatio-temporal information from

aggregated data of stationary traffic detectors, the "adaptive smoothing method".
In essential, a nonlinear spatio-temporal lowpass filter is applied to the input de-
tector data. This filter exploits the fact that, in congested traffic, perturbations
travel upstream at a constant speed, while in free traffic, information propagates
downstream. As a result, one obtains velocity, How, or other traffic variables as
smooth functions of space and time. Applications include traffic-state visualization,
reconstruction of traffic situations from incomplete information, fast identification
of trafik breakdowns (e.g., in incident detection), and experimental verification of
traffic models.
We apply the adaptive smoothing method to observed congestion patterns on
several German freeways. It manages to make sense out of data where conventional
visualization techniques fai!. By ignoring up to 65% of the detectors and applying
the method to the reduced data set, we show that the results are robust. The
method works weIl if the distances between neighbouring detector cross sections do
not exceed 3 km.

1 Introduction

During the last decades, trafik dynamies and pedestrian flows have been
intensively studied. Regarding the observed phenomena and simulation ap-
proaehes we refer the reader to some reeent reviews [1-4]. Presently, scientists
are more and more getting interested in detailed empirical studies. Reasons
for this are the availability of better data and the need to verify and ealibrate
models. Both, theoretical [5-7] and empirical [8-17] studies indicate that the
phenomenology of eongested trafik is more eomplex than originally expeeted.
There seems to be a rieh speetrum of trafik states [5,6,12,13], hysteretie
or eontinuous temporal or spatial transitions among them [7,11-13,18,19],
and fluetuations or a erratieally appearing dynamics play a signifieant role
[13,16,20-22]. In order to make sense out of this, it is inereasingly important
to have suitable ways of data proeessing, in order to extract the information
relevant for scientifie investigations or specifie applieations. Here, we will pro-
pose a three-dimensional data-evaluation method allowing to visualize the
spatio-temporal dynamics of traflic patterns along freeways.

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
344 M. Treiber, D. Helbing

The developed "adaptive smoothing method" filters out small-seale fiue-

tuations and adaptively takes into aeeount the main propagation direetion
of the information fiow (Le. the dominating eharaeteristie line), which have
been determined by means of a spatio-temporal eorrelation analysis in other
studies [23]. The temporallength seale of the smoothing proeedure ean be as
small as the sampling interval, while the spatiallength seale is related to the
distanee between sueeessive deteetors, which ean be up to 3 kilometers long.
By ''filter'' we just mean a transformation of the data with specifie proper-
ties. Here, we use a spatio-temporallowpass filter, i.e., only (Fourier) eompo-
nents of low frequeney ean pass the filter, while high-frequeney eontributions
are eonsidered as fiuetuations and smoothed out. One particular feature of our
filter is that it is nonlinear and adaptive to the trafIic situation in distinguish-
ing free and eongested trafIic, as the propagation direetion of perturbations
differs. The results are three-dimensional visualizations of trafIic patterns,
which are quite robust with respeet to variations of the filter parameters and
very helpful in obtaining a c1ear picture of the systematic spatio-temporal
dynamics.
Therefore, this method is suitable for the reeonstruction of trafIic situa-
tions from ineomplete information, fast identifieation of trafiic breakdowns
(ineident deteetion), and experimental verifieation of trafiic models. First
results support the phase diagram of trafIic states oceuring at bottleneeks
[17,24], which, apart from free trafiic, predicts pinned or moving loealized
clusters, spatially extended patterns such as triggered stop-and-go waves,
and oseillating or homogeneous eongested trafIic, or a spatial eoexistenee of
some of these states [5,7,1]. The respeetively oeeuring spatio-temporal pat-
tern depends on the specifie freeway fiow and bottleneek strength, but also
on the level of fiuctuations, as these ean trigger transitions from, for example,
free trafIic to loealized cluster states [5,17,1]. We do not see sufficient support
for one "generalized pattern" [25] that would always be observed when traffie
fiow breaks down.

2 Description of the Method

The adaptive smoothing method is a data proeessing method for obtaining
traffie variables as smooth funetions of space and time out of stationary traffie
data. It has following heuristically motivated properties:

1. In ease offree traffie, perturbations (of, e.g., velocity or fiow) move essen-
tially into the direetion of traffie fiow [26]. More speeifieally, they prop-
agate with a eharacteristic velocity Cfree at about 80% of the desired
velocity Vo On empty roads [23]. Therefore, at loeations with free traffie,
perturbations with propagation veloeities near Cfree should pass the filter.
2. In ease of eongested traffie, perturbations propagate against the diree-
tion of traffie fiow with a eharacteristic and remarkably eonstant velo city
Ccong ~ -15 km [12]. With modern data analysis teehniques, it has been
 Adaptive Smoothing Method for Trafik State Identification 345

shown that such propagation patterns persist even in "synchronized" con-

gested traffic flow, where they are hardly visible in the time series due to
a wide scattering ofthe data in this state [26]. So, for high traffic densities
or low velocities, the filter should transmit spatio-temporal perturbations
propagating with velocities near Ccong more or less unchanged.
3. The filter should smooth out all high-frequency fluctuations in t on a
time scale smaller than T and spatial fluctuations in x on a length scale
smaller than 0'. The parameters T and 0' of the smoothing method can
be freely chosen in a wide range (cf. Tab. 1).
Let us assume that aggregated detector data z~j are available from n cross
sections i at positions Xi where jE {imin,'" ,jmax} denotes the index ofthe
aggregation intervals.
Usually, the aggregation interval
(1)
is fixed (between 20 s and 5 min, depending on the measurement device). A
typical value for German highways is Llt = 1 min.
The components of z~j represent the desired aggregated quantities as ob-
tained from cross section i during the time interval j. Typical examples in-
clude the average velo city Vi;, the vehicle flow Qij, the occupancy Oi;, or
some derived quantity such as the traffic density Pij. The input data z~j can
represent either averages over allianes or quantities on a given lane.
The adaptive smoothing method provides estimates z(x, t) for aillocations
x E [Xl,X n ] between the positions of the first and the last detector, and
for all times t E [tmin, t max ]. Extrapolations (in the sense of a short-term
traffic forecast ) will also be discussed. Without loss of generality, we assume
Xl < X2 < ... < Xn, and a traffic flow in positive x-direction.

The core of our "adaptive smoothing method" is a nonlinear filter trans-

forming the discrete input detector data z~j into the smooth spatio-temporal
functions z(x, t). To satisfy the first two requirements mentioned above, we
write the filter as
z(x, t) = W(zcong, Zfree)Zcong(x, t) + [1- w(zcong, Zfree)]Zfree(X, t). (2)
This is a superposition of two linear anisotropie lowpass filters Zcong(x, t)
and Zfree(X, t) with an adaptive weight factor 0:::; W :::; 1 which itself depends
nonlinearlyon the output of the linear filters as discussed later on.
The filter Zcong(x, t) for congested traffic is given by
1
L L
njmax •
Zcong(X, t) = Nocong (
X,
)
. l'J=3min
t t= .
<Pcong(Xi - X, tj - t)zij, (3)

with
tj = tmin + jLlt, (4)
346 M. Treiber, D. Helbing

and the normalization factor

n jmax

Ncong(x, t) =L L c/lcong(Xi - x, tj - t). (5)

i=l j=jmin

This normalization guarantees that a constant input vector z~j = Zo for all
i, j is transformed into a constant output function with the same components:
z(x, t) = zoo Notice that the normalization depends on both, location x and
time t.
The analogous expression for the lowpass filter Zfree(X, t) for free traflic is

1 n j",ax .
Zfree(X, t) = N,free ( x, t ) L L
. 1. .
c/lfree(Xi - x, tj - t)zij (6)
1.= J=Jmin

with
n jmax

Nfree(X, t) =L L c/lfree(Xi - x, tj - t). (7)

i=l j=j",in

The kerneis c/lcong(x, t) and c/lfree(X, t) of the linear homogeneous filters do

the required smoothing and are particularly transmissible for perturbations
propagating with the typieal velocities ccong and Cfree observed in congested
and free traflic, respectively (cf. Fig. 1).
H the propagation velo city of the perturbations was zero, the filter should
only perform the smoothing. We chose the (non-normalized) exponential
function

c/lo(x, t) = exp lxi - -:;:

( --;- Itl) . (8)

Instead, one could apply other localized functions such as a two-dimensional

Gaussian. However, it turned out that the exponential had more favourable
properties in our application.
The linear filter for nonzero propagation velocities can be mapped to
c/lo(x, t) by the coordinate transformations (cf. Fig. 1)
X=X' , t = t cong + -
I X
- = t free + --.
Ccong
X

Cfree
I
(9)

Thus, we obtain for the kerneis of the linear anisotropie filters

c/lcong(x, t) = c/lo(x' , t~ong) = c/lo (x, t _ _ X_) ,

ccong
(10)

c/lfree(X, t) = c/lo (x', t~ree) = c/lo (x, t - ~) . (11)

Cfree
 Adaptive Smoothing Method for Traffic State Identification 347

x
t'cong=t- +-
cong
= const. t'free =t- +-
free
= con t.

free
~ (x-x 0' t-t 0) > const'.

cong
cr ~ (x-x 0' t-t 0) > const'.

Fig. 1. Visualization of the effects of linear homogeneous filters with the kerneIs
tPfree(X, t) and tPcong(X, t), respectively. The shaded areas denote the regions con-
sidered in the calculation of a data point at (x, t). Triangles denote the mainly
contributing input data sampled in free traffic, squares the ones sampled in con-
gested traffic.

Figure 2 shows the action of the filters Zcong(x, t) and Zfree(X, t) for the
velo city fields Vcong(x, t) and Vrree(x, t).
Finally, we define the nonlinear adaptive weight function w(zcong, Zfree) E
[0,1]. Obviously, we must have W ::::J 1 for congested trafik, and W ::::J 0 for free
trafik, so we need some apriori estimate of the trafik situation at the point
(x, t).
Congested trafik is characterized by a high trafik density and low aver-
age velo city. Since, in contrast to the density, the velo city can be directly
measured with stationary detectors, we chose the velo city to determine the a
priori estimate. Different possibilities to estimate the velo city at point (x, t)
are:

- The measured velo city Vij of the detector cross section whose position Xi
is nearest to x in the time interval j containing the actual time t,
- the velo city Vcong(X, t) as calculated with the "congested-traffic" filter
Zcong according to Eq. (3) with kernel (10) (assuming apriori congested
traffic) ,
348 M. Treiber, D. Helbing

(a)

(b)

Fig. 2. Typical velocity fields Vfree(x, t) (top) and Vcong(X, t) (bottom) obtained by
application of the filters Zfree(X, t) and Zcong(X, t) to traflic data of a section of the
German freeway Ag South.

- the velocity Vfree(x, t) as ealeulated with the ''free-traffie'' filter Zfree ae-
eording to Eq. (6) with the kernel (11),
- or some eombination of the above estimates.

The first way to estimate the velo city is subjeet to errors, if the typicallength
seale A of oeeuring stop-and-go struetures is not larger than the distanee L1Xi
between two neighbouring deteetors. At this point, it is erucial that propa-
gating struetures in eongested traffie, especially stop-and-go waves, are very
persistent. It has been shown [27] that they ean propagate through freeway
interseetions or other inhomogeneities nearly unchanged, passing all pertur-
bations of free traffie on their way (see, e.g., Fig. 9). Therefore, whenever
at least one of the estimates indicates eongested traffie, the weight function
 Adaptive Smoothing Method for Trafik State Identification 349

should favour the filter for congested trafik. Specifically, we assume

W(Zcong, Zfree) = w(Vcong(X, t), Vfree(x, t)) = ~ [1 + tanh (Vc -:~(X, t))] ,
(12)

where

V*(X, t) = min (Vcong(x, t), Vfree(X, t)) . (13)

Vc and L\ V are parameters that can be varied in a wide range.

If not explicitely stated otherwise, for all simulations in the following
section we will use the parameters specified in Tab. 1. We will also show
that the four parameters ccong , Cfree, Vc and L\v can be varied in a wide
range without great differences in the output. In this way, we show that the
proposed adaptive smoothing method does not need to be calibrated to the
respective freeway. One can take the values from Tab. 1 as a global setting.
The smoothing parameters a and T have the same meaning and the same
effect as in standard smoothing methods, e.g., Eqs. (22) and (23) in Ref. [17].

Table 1. Parameters of the "adaptive smoothing method" defined by Eqs. (2)-(13),

their interpretation, and their typical values.

!parameter !TYPiCal Value !Meaning

(J' 0.6 km Range of spatial smoothing in x

T 1.1 min Range of temporal smoothing in t

Propagation velocity of
Cfree 80 kmjh
perturbations in free trafik

ccong Propagation velo city of

-15 kmjh
perturbations in congested traffic

Crossover from free to

Vc 60 kmjh
congested traffic

LlV 20 kmjh Width of the transition region

In summary, the proposed "adaptive smoothing method" is given by Eq.

(2) with the nonlinear weight function (12), the filter (3) with normalization
(5) and kernel (10) for congested trafiic, the filter (6) with normalization (7)
and kernel (11) for free trafiic, and the smoothing filter (8). Table 1 gives
350 M. Treiber, D. Helbing

an overview of the six parameters involved and typieal values for them. The
adaptive smoothing method includes the following special cases:

- Isotropie smoothing resulting in the limits Cfree ---+ 00 and ccong ---+ 00 (for
practieal purposes, one may chose Cfree = ccong = 106 km/h);
- only filtering for structures of congested trafik in the limit Vc ---+ 00
(for practieal purposes, one may set v;, » Vo with Vo being the desired
velocity);
- only filtering for structures of free trafik for Vc = LlV = 0;
- consideration of the data of the nearest detector only, if (T = 0;
- application of the actual sampling interval of the detectors, if T = 0,
Cfree ---+ 00 and ccong ---+ 00. (To avoid divisions by zero, zero values of (T,
T, and LlV are replaced by very small positive values in the software).

3 Application to German Freeways

We will now discuss data from the German freeways A8-East and A9-South
near Munieh, and of the freeway A5-North near Frankfurt. In all cases, the
trafik data were obtained from several sets of double-induction-loop detectors
recording, separately for each lane, the passage times and velocities of all
vehicles. Only aggregate information was stored with an aggregation interval
of Llt = 1 min. We will use the following input data z~j:

- The lane-averaged vehicle flow

(14)

ni
where j is the vehicle count at cross section i during time interval j on
lane l. The considered sections of the freeways have L = 3 lanes in most
cases.
- The lane-averaged mean velo city

L QI v:1
TT
Vij ="
L...Ji-j- -ij , (15)
1=1 Qij

where Vi~ is the average velocity at cross section i during time interval j
on lane l.
- The trafik density determined via the formula

Qij
Pij = V:-' (16)
~J

as occupancies were not available for all freeways.

 Adaptive Smoothing Method for Traffic State Identification 351

]
"!
(a)
S
..I<
""':
]
r-
]
00
]
C'l
---
S
uphill

..I<
S
..I<
0\
]
....
r-;
downhilI

S
..I<
00
.-
]
\0

'""" '" '"""

er. .,; -.6 00 N

t ~t === 1==Ff==f =1=_= =j ==~ t=

~
<'l er. M M r<l er.
""

Ir)
....-I
\D
_
r-
_
00 - 0\
- N0 N -Dl7N N Cf')
N -.::t
N
Cl Cl Cl Cl Cl Cl Cl Cl Cl Cl
Irschenberg

p (veh./kmllane)
50

(b)

x (km)

16.5 17
p (veh./km/lane)
50

(c)

x (km)

16.5

Fig.3. Complex congested traffic pattern on the German freeway A8-East from
Munich to Salzburg during the evening rush hour on November 2, 1998. (a) Sketch
of the freeway. (b) Plot of the spatio-temporal density p(x,t) using conventional
smoothing (resulting for the setting Cfree = Ccong = 106 km jh). (c) Plot of the same
data as calculated with the adaptive smoothing method.

Figure 3 shows an example of a complex trafik breakdown that occurred

on the freeway A8 East from Munich to Salzburg during the evening rush
352 M. Treiber, D. Helbing

hour on November 2, 1998. Two different kinds of bottlenecks were rele-

vant, (i) a relatively steep uphill gradient from x = 38 km to x = 40 km
(the ''Irschenberg''), and (ii) an incident leading to the closing of one of the
three lanes between the cross sections D23 and D24 between t =17:40 h and
t = 18:10 h. For further details, see Ref. [17]. Figure 3(c) shows the trafIic
density using the proposed data processing method with the parameters spec-
ified in Tab. 1. For comparison, (b) shows the result of conventional isotropic
smoothing ignoring the speed of information propagation, cf. Ref. [17]. While
the structures in the free-traffic regions are nearly the same in both plots, the
new smoothing method gives a better reconstruction of the congested traf-
fic patterns. In contrast to the conventional method, our proposed method
can resolve oscillations with wavelengths A comparable to or smaller than
the distance ..::lxi between two neighbouring cross sections as illustrated for
the three stop-and-go waves propagating through the whole displayed section
(A ~ 2 km). Notice that the conventional method generates artefacts for this
example.
The congestion patterns caused by the uphill gradient (the ''wings'' in
Fig. 3(b), (c)) have larger characteristic wavelengths and are resolved, at
least partially, also with the conventional method.
For comparison, we plot time series of some cross sections in Fig. 4. Let us
consider more closely the jammed traffic patterns caused by the temporary
bottleneck associated with the incident. While the fluctuations are irregular
at detector D22 (about 1.2 km upstream of the aforementioned incident),
they grow to three stop-and-go waves further upstream, cf. the arrows in
Fig. 5. However, due to measurement errors and other errors in determining
the density [1], the time series of the density is very noisy, especially that
of cross section D20. The adaptive smoothing method suppresses this noise
without suppressing the structure of the pattern.
To check, whether there are significant patterns in the data of detector
D20 at all, we plot the velocity data at D20 as weIl (Fig. 5). In contrast to the
density, the velo city is measured directlyresulting in less noise. One clearly
sees three dips corresponding to three stop-and-go waves which confirms the
results of the "adaptive smoothing method".
As a further check that the structures shown in Fig. 3(c) are not artefacts
of the data processing, we tested the method for a reduced data set. Figure 6
shows the result using only the even-numbered detectors as input, whose time
series are plotted in Fig. 4. The main structures remain nearly unchanged,
particularly, the three stop-and-go waves propagating through the whole sec-
tion. Notice that the whole plot is based on data of only four cross sections,
and that the data are extrapolated in the upstream direction over a distance
of nearly 2 km.
We now apply the processing method to data from a section of the freeway
A9-South near Munich. There are two major intersections 11 and 12 with other
freeways. Virtually every weekday, traffic broke down to oscillatory congested
traffic upstream of each of these intersections. (For details, see Ref. [17]).
 Adaptive Smoothing Method for Traffic State Identification 353

f,,"",80~
:
-520
D16(33Ak1ll)

~ 0 L -_ _~_ _ _ _ _ _ _ _L -_ _ _ _ _ _~_ _ _ _ _ _ _ _~

17 18 19 20

Q) 80
tä 60 D18 (36.8 km)
]40
i20
~ 0
17 18 19 20

i] 6080 D20 (39.0 km)

40
i20
~ 0
17 18 19 20

~80

!:~
-520
>
0::0
17 18 19 20
t (h)

Fig.4. Time series of the empirical density, approximated by means of Eq. (16),
for several cross sections.

Figure 7 (b) shows the inverse of the velocity, l/V(x, t) for a typical
congested situation. When plotting l/V(x, t), the structures of congested
traffic comes out more clearly than for the density. The method resolves
small density clusters in the region 508 km ~ x ~ 510 km between ab out
8:30 am and 9:30 am which disappear around x = 508 km. These structures
cannot be identified with the conventional smoothing method (see Fig. 7(c»,
which just shows a hilly pattern.
Figure 7(b) is an example for the spatial coexistence of homogeneous
congested traffic (a short stretch around x = 510 km), oscillating congested
354 M. Treiber, D. Helbing

150
~ ~D20(39
.0km)
~ 100 ••
> 50 •
o
17 18 19 20
t (h)

Fig. 5. Time series of the empirical velo city data determined via Eq. (15) for cross
section D20.

p (veh.fkmJ!ane)
50

x (km)

16.5 17

Fig. 6. Spatio-temporal density p(x, t) as in Fig. 3, but using only the data of cross
sections D16, D18, D20, and D22 as input (cf. Fig. 4).

trafik (around x = 509 km), and stop-and-go waves (around x = 507 km).
This may be a three-dimensional illustration of the so-called ''pinch effect"
[11,14], but we do not observe a merging of narrow clusters to form wide
moving jams. The narrow structures of short wavelengths rather disappear.
This may be an artefact of our smoothing method, so that video data are
required to get a more detailed picture. Simulations of this spatial coexistence,
however, indicate that narrow clusters rather disappear than merge, namely
when they do not exceed the critical amplitude in an area of metastable trafik
[7].
Notice that the bottleneck causing this congestion is located at ab out
x = 510 km implying that it is caused by weaving trafik and slowing down
 Adaptive Smoathing Method far Trafik State Identificatian 355

Junction (a) Il 12
Allershausen Neufahrn München-Nord
'"0( g N0 '"cO 9' .... '"
:i :i ~ ~'" ~;ri
N co
~ "l t-: '" '"~ ::i'" ~ OON ...c 00 N
~ vi
'" '" '"
~ '" S 8 0
'" '" '" '" '" '" '" '" '" '" "''''
0"'; N N
NM N N
11111'1111111
N

'" '"

I OOO/V (h/km)
100
(b) 0

7.5

IOOO/v (h/km)
100
(e)
o

7.5

Fig. 1. Two regions of stop-and-go waves on the German freeway A9-South near
Munich oecurring upstream of the freeway intersections 11 and 12. (a) Sketch of the
freeway. (h) Plot ofthe inverse l/V(x, t) of the spatio-temporal velocity V(x, t), us-
ing eonventional smoothing. (e) Plot ofthe same data using the adaptive smoothing
method.

at an off-ramp, not by an on-ramp as is most often the case on German

highways.
356 M. Treiber, D. Helbing

p (veh./klll/lane)
(a)

7~
P (veh./kmllane)
100
(b)
o

(e) 0

7 .)

Fig. 8. Same data and same processing method as in Fig. 7(c), but with (a) a prop-
agation velocity of ccong = -12 km/h instead of -15 km/h (b) Ccong = -20 km/h,
(c) with the crossover parameter Vc = 40 km/h instead of 60 km/ho

Now we demonstrate that the method is robust with respect to reasonable

parameter changes. Figure 8(a) shows the result for an assumed propagation
velo city of ccong = -12 km/h instead of -15 km/ho In plot (b), we assumed
ccong = -20 km/h instead. In both cases, the results are systematically better
 Adaptive Smoothing Method for 'Traffic State Identification 357

than with the isotropie proeedure. Sinee the propagation velo city of pertur-
bations in eongested trafik is always about Ccong = -15 km [12,23,28,29J, one
ean take this value as a global setting.
Due to the eomparatively high magnitude of the propagation velocity
Cfree, the related time shifts in the transformation (9) for free traffie are
smalI. Therefore, the method is insensitive to changes with respeet to this
parameter. It turned out that taking the isotropie limit Cfree -t 00 nearly
gives the same results.
The erossover parameters Vc and L1V ean be varied within reasonable
limits as weH. As an example, Fig. 8(e) shows the result after ehanging the
erossover velo city from Vc = 60 km/h to Vc = 40 km/ho
Finally, we applied the method to isolated moving loealized clusters and
pinned loealized clusters observed on the A5-North near Frankfurt. This free-
way is partieularly weH equipped with deteetors, so that the typical length
seales of the observed struetures are always larger than the distanee between
neighbouring deteetors. However, we obtained nearly the same results when
using only the even-numbered deteetors or only the odd-numbered deteetors
as input. Figure 9 shows that even a further reduetion of the information to
only 35% of the detectors yields good results.

4 Summaryand Outlook

We have proposed a new "adaptive smoothing method" for the three-

dimensional visualization of spatio-temporal traffie patterns, which takes into
aceount the characteristie propagation velocities observed in free and eon-
gested traffie. The method is robust with respeet to variations of its param-
eters, so that it ean be applied to new freeway seetions without ealibration.
In principle, it would be possible to determine the parameters (such as the
propagation velocity of perturbations) loeally (e.g., by means of a eorrelation
analysis). However, the results would look less smooth and regular, as the
small number of data to determine the loeal parameters would be associated
with eonsiderable statistical errors. Consequently, a large part of the varia-
tions in the loeal parameters would not refieet systematic variations of the
parameters. So, both the loeal and global parameter ealibration may produee
artefaets, but it is advantageous to use global parameter settings. Moreover,
there is empirieal support for surprisingly eonstant propagation velocities
Cfree and Ccong of perturbations in free and eongested traffie. The parameters
Vc and L1V are related to the transition from free to eongested traffie and,
therefore, ean also be weH determined. The spatial and temporal smoothing
parameters (1 and T ean be speeified aceording to the respeetive requirements.
Suitable parameters allow a good representation of traffie patterns even when
the distanees between sueeessive deteetors are about 3 kilometers.
We point out that the suggested "adaptive smoothing method" itself ean be
applied to ealibrate the characteristie propagation velocities Cfree and Ccong.
358 M. Treiber, D. Helbing

11 (a) 12 Junction
Nordwestkreuz Bad Homburger Friedberg
Frankfurt Kreuz

(b)
100

490

Ce)
100

490

Time (h)

Fig. 9. Freeway A5-North from Frankfurt to Kassel during the evening rush hour
on August 7,1998. (a) Sketch ofthe freeway. (b) Data reconstruction ofthe veloeity
using all detectors DIO-D26. (c) Reconstruction using a reduced data set of only 6
of all 17 detectors used in (b).

For this purpose, the ranges (j and T of spatial and temporal smoothing are
chosen small. The optimal propagation velocities minimize the offsets in the
propagation patterns.
 Adaptive Smoothing Method for Trafik State Identification 359

The aim of the "adaptive smoothing method" is to reconstruct the spatio-

temporal trafik data from incomplete information as good as possible to
allow a better understanding of the complex trafik dynamics. Potential ap-
plications are, for example, traffic state visualization, incident detection, or
the experimental verification of traffic models. Our method could be further
improved by taking into account information about the trafik dynamics such
as the continuity equation or a suitable equation for the average vehicle ve-
locity as a function of space and time. It could, then, be used to determine
short-term trafik forecasts.

Acknowledgments
The authors would like to thank the German Research Foundation (DFG) for
financial support through the grant He 2789/2-1. They are also grateful to
the Autobahndirektion Südbayern, and the Hessisches Landesamt für Straßen
und Verkehrswesen for providing the trafik data.

References
1. Helbing, D.: Traffic and related self-driven many-particle systems, Reviews of
Modern Physics 73, 1067-1141 (2001).
2. Chowdhury, D., Santen, L. and Schadschneider, A.: Statistical physics of vehic-
ular traffic and some related systems, Physics Reports 329, 199-329 (2000).
3. Helbing, D., Farkas, 1. J., Molnär, P. and Vicsek, T.: Simulation of pedestrian
crowds in normal and evacuation situations, in Pedestrian and Evacuation Dy-
namics, edited by M. Schreckenberg and S. D. Sharma (Springer, Berlin, 2002),
pp. 21-58.
4. Helbing, D., Molnär, P., Farkas, 1. and Bolay, K.: Self-organizing pedestrian
movement, Environment and Planning B 28, 361-383 (2001).
5. Helbing, D., Hennecke, A. and Treiber, M.: Phase diagram of traffic states in
the presence of inhomogeneities, Phys. Rev. Lett. 82, 4360-4363 (1999).
6. Lee, H. Y., Lee, H. -W. and Kim, D.: Dynamic states of a continuum traffic
equation with on-ramp, Phys. Rev. E 59, 5101-5111 (1999).
7. Treiber, M. and Helbing, D.: Explanation of observed features of self-
organization in traffic flow, e-print cond-mat/9901239 (1999).
8. Cassidy, M. J. and Bertini, R. L.: Some traffic features at freeway bottlenecks,
Transpn. Res. B 33, 25-42 (1999).
9. Daganzo, C. F., Cassidy, M. J. and Bertini, R. L.: Possible explanations of
phase transitions in highway traffic, Transpn. Res. A 33, 365-379 (1999).
10. Hall, F. L. and Agyemang-Duah, K.: Freeway capacity drop and the definition
of capacity, Transpn. Res. Rec. 1320, 91-108 (1991).
11. Kerner, B. S.: Experimental features of self-organization in traffic flow, Phys.
Rev. Lett. 81, 3797-3800 (1998).
12. Kerner, B. S. and Rehborn, H.: Experimental features and characteristics of
traffic jams, Phys. Rev. E 53, R1297-R1300 (1996).
13. Kerner, B. S. and Rehborn, H.: Experimental properties of complexity in traffic
flow, Phys. Rev. E 53, R4275-R4278 (1996).
360 M. Treiber, D. Helbing

14. Koshi, M., Iwasaki, M. and Ohkura, 1.: Some findings and an overview on vehic-
ular ftow characteristics, in Proceedings 0/ the 8th International Symposium on
Transportation and Traffic Flow Theory, pp. 403-426, V. F. Hurdle, E. Hauer,
and G. N. Stewart (Eds.) (University of Toronto, Toronto, Ontario, 1983).
15. Lee, H. Y., Lee, H. -Wo and Kim, D.: Phase diagram 0/ congested traffic ftow:
An empirical study, Phys. Rev. E 62, 4737--4741 (2000).
16. Neubert, L., Santen, L., Schadschneider, A. and Schreckenberg, M.: Single-
vehicle data 0/ highway traffic: A statistical analysis, Phys. Rev. E 60, 6480-
6490 (1999).
17. Treiber, M., Hennecke, A. and Helbing, D.: Congested traffic states in empirical
observations and microscopic simulations, Phys. Rev. E 62, 1805-1824 (2000).
18. Helbing, D. and Treiber, M.: Gas-kinetic-based traffic model explaining observed
hysteretic phase transition, Phys. Rev. Lett. 81,3042-3045 (1998).
19. Lee, H. Y., Lee, H. -Wo and Kim, D.: Origin 0/ synchronized traffic ftow on
highways and its dynamic phase transitions, Phys. Rev. Lett. 81, 1130-1133
(1998).
20. Banks, J. H.: An investigation 0/ some characteristics 0/ congested ftow,
Transpn. Res. Rec. 1618, 128-134 (1999).
21. Leutzbach, W.: Introduction to the Theory 0/ Traffic Flow (Springer, Berlin,
1988).
22. Treiber, M. and Helbing, D.: Macroscopic simulation 0/ widely scattered syn-
chronized traffic states, J. Phys. A: Math. Gen. 32, L17-L23 (1999).
23. Sollacher, R., Kerner, B. S., Konhäuser, P., Rehborn, H., Kühne, R., Schrecken-
berg, M. and Helbing, D.: SANDY - Nichtlineare Dynamik im Straßenverkehr,
in Technische Anwendungen von Erkenntnissen der Nichtlinearen Dynamik
(VDI - Technologiezentrum Physikalische Technologien, Düsseldorf, ISBN 3-
931384-25-X, 1999).
24. Helbing, D. and Treiber, M.: Critical discu8sion 0/ "sychronized ftow", submit-
ted to Cooper@tive Tr@nsport@tion Dyn@mics (2002).
25. Kerner, B. S. and Klenov, S. L.: A micropscopic model /or phase transitions in
traffic ftow, J. Phys. A: Math. Gen. 35, L31-L43 (2002).
26. Daganzo, C. F.: A behavioral theory 0/ multi-lane traffic ftow, Part I: Long
homogeneous freeway sections, (ITS Working Paper, UCB-ITS-RR-99-5, revised
June 20, 2000).
27. Kerner, B. S.: Theory 0/ breakdown phenomenon at highway bottlenecks,
Transpn. Res. Rec. 1110, 136-144 (2000).
28. Cassidy, M. J. and Mauch, M.: An observed traffic pattern in long freeway
queues, Transpn. Res. A 35, 143-156 (2001).
29. Mika, H. S., Kreer, J. B. and Yuan, L. S.: Dual mode behavior 0/ freeway traffic,
Highw. Res. Rec. 219, 1-13 (1969).
 Probabilistic Description of Nucleation in
Vapours and on Roads

Reinhard Mahnke

Universität Rostock, Fachbereich Physik, D-18051 Rostock, Germany

Abstract. The aggregation of particles out of an initially homogeneous situation

is weIl known in physics. Depending on the system under consideration and its con-
trol parameters the cluster formation in a supersaturated (metastable or unstable)
situation has been observed in nucleation physics as weIl as in other branches. We
investigate the weIl-known example of condensation (formation of liquid droplets)
in an undercooled vapour to conclude that the formation of bound states as a
phase transition is related to transportation science. We present a comparison of
nucleation in an isothermal-isochoric container with traffic congestion on a circular
one-lane freeway. The analysis is based, in both cases, on the probabilistic de-
scription by stochastic master equations. The construction of physically motivated
transition probabilities plays the central role in our analysis and comparison.

1 Introduction
The term cluster as a group 01 similar things is used in science in many differ-
ent fields. It is common on all scales from subatomic physics up to astronomy.
The existance of bound states of atoms or molecules as clusters with different
sizes in gases, liquids and solid matter is an intermediate step between the
elementary isolated particle and the macroscopic condensed state. The forma-
tion of bound states as an aggregation process is related to self--organization
phenomena [5,28]. In general the properties of clusters, especially the binding
energy, depend on the cluster size and shape [4,9]. The nucleation and growth
of atomic and molecular clusters is one of the basic topics in nucleation the-
ory [28,29]. An example of experimental verification is given in [27].
The formation of clusters takes place on different scales in physical (nu-
cleation of droplets, condensation of clouds) and nonphysical systems (insect
societies, socioconfigurations, transport communities). Nonphysical systems
are generaHy a reflection of the interaction of people or animals. We, however,
do not discuss here the social aspects of human interaction. We are particu-
larly interested in trafik flow theory where the car drivers (or, formally, the
vehicles) are considered as interacting particles. Physical systems like spin
glas ses are one of the most investigated complex systems to understand the
clustering of states. In chemically reacting systems the domain formation is
simulated by Monte Carlo experiments and by lattice-gas automata. Monte
Carlo methods are weH established in statistical physics [6,22].
The clustering behaviour can be considered as a phase transition be-
tween different states of matter. In dependence on the boundary conditions

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
362 R. Mahnke

a quenched system as for instanee an undereooled vapour or an overerowded

road undergoes a relaxation proeess to the new phase. The theoretical Ull-
derstanding of the physieal mechanism for the dynamies of the aggregation
phenomenon has important implieations for the more general task to real-
ize relaxation phenomena in equilibrium and nonequilibrium eomplex sys-
tems [28J.
The modelling of such supersaturated (metastable or Ullstable) systems is
dominated by a very eomplieated energy landscape with many local minima
and can be done by different means. Having in mind the homogeneous or het-
erogeneous (e. g. by electrieal eharges) nucleation and eondensation process
we divide the following theoretieal and eomputational methods [14J:
1. Moleeular dynamics
Based on the classical many-particle Hamiltonian with interaction, e. g.
of well-known Lennard Jones type, the integration of the equations of
motions shows the clustering proeess quite obvious.
2. Master equation approach
In this Monte Carlo simulation technique the master equation for the
chemical reactions of the attachment and the detachment of elementary
particles (monomers) on or out of clusters is solved with phenomenolog-
ically motivated transition probabilities. This method will be used later
on.
3. Monte Carlo simulation technique
The formation process of clusters on a two--dimensional lattice can be
modelled by a so-called nucleation automaton with a well-explained en-
ergy funetion. The relaxation to a one-big-cluster-state as the equilib-
rium situation is a searching proeess on the energy landseape to find a
minimum in a very high dimensional situation.
4. Cellular automata models
Chemically reacting systems with phase separations like the well-known
Schlögl model ean be simulated by probabilistic lattice-gas automaton
models where space, time and particle velocities are diserete. Each mole-
eule (oeeupied site) can undergo elastic and reactive collisions and de-
pending on the updating rule interesting spatio-temporal eonfigurations
of particles emerge.
5. Deterministic nonlinear dynamical systems
The Becker-Döring cluster equations and other kinetic equations of such
type provide a very famous dynamical model for the growth, shrink-
age, eoagulation and fragmentation of clusters (droplets) in a condens-
ing vapour. The trends to equilibrium can be studied in dependence on
the kinetic coefficients. Especially for the late Ostwald ripening regime
a number of dynamical systems as coupled nonlinear sets of equations
exists.
It should be mentioned that nonphysical systems, like traffie flow or eeo-
nomics, can also be described by methods of statistical thermodynamics. Such
 Probabilistic Description of Nucleation in Vapours and on Roads 363

an approach, based on the use of state function entropy and its maximization
principle, has been presented by Montroll, Reiss and co-workers [20,25,26].
Although our present investigation of traffic flow does not involve the en-
tropy maximization principle, it clearly supports the idea that nonphysical
systems like traffic flow and physical systems like supersaturated vapour can
be treated similarly.
The aim of the present paper is to give a comparison between aggregation
in supersaturated vapour and in traffic flow by providing unique description
of these two phenomena within the stochastic master equation approach. We
emphasize that phase transitions which are weH known and investigated by
statistical physics and thermodynamics take place in nonphysical systems too,
as already pointed out by MontroH [20]. We concentrate here on vehicular
traffic, but other fields like cancer growth and stock market development
behave similarly.

2 Stochastic Master Equation Approach

Stochastic processes enter into many physical descriptions of nature. Histori-

cally first the motion of a heavy particle in a fluid of light molecules has been
observed. The path of such Brownian particle consists of stochastic displace-
ments due to random collisions. Such motion was studied by the Scottish
botanist Robert Brown (1773 - 1858). In 1828 he discovered that the micro-
scopically smaH particles into which the pollen of plants decay in an aqueous
solution are in permanent irregular motion. Such stochastic process is called
Brownian motion and can be interpreted as discrete random walk or contin-
uous diffusion movement. The intuitive background to describe the irregular
motion completely as stochastic process is to measure values Xl, X2, •• • , X n , ...
at time moments tl, t2, ... , t n , ... of a well-defined time dependent random
variable x(t) and assume that a set of joint probability densities or distribu-
tions

(1)

exists.
This stochastic description in terms of macroscopic variables is called
mesoscopic. Typical systems encountered in the everyday life like gases, li-
quids, solids, biological organisms, human or technical objects consist of
about 1023 interacting atoms. The macroscopic properties of matter are usu-
ally the result of coHective behaviour of large number of atoms and molecules
acting under the laws of quantum mechanics. To understand and control these
coHective macroscopic phenomena the complete knowledge based upon the
known fundamentallaws of microscopic physics is useless because the problem
of interacting particles is much beyond the capabilities of the largest recent
and future computers. The understanding of complex macroscopic systems
364 R. Mahnke

consisting of many basic partic1es (in the order of atomic sizes: 10- 10 m) re-
quires the formulation of new concepts. One of the methods is the stochastic
description taking into account the statistical behaviour. Since the macro-
scopic features are averages over time of a large number of microscopic in-
teractions, the stochastic description links both approaches, the microscopic
and the macroscopic one, together to give probabilistic results [3,8,10].
Speaking about a stochastic process from the physical· point of view we
always refer to stochastic variables (random events) changing in time. A
realization of a stochastic process is a trajectory x(t) as function of time.
A stochastic process without any dynamies (like a coin throw or any
hazard game) is called a temporally uncorrelated process and the normalized
distribution Pl(Xl, tl) describes the process totally.
Introducing dynamies via correlations between two different time mo-
ments, this assumption enables us to define the Markov process by two quan-
tities totally, namely the first-order and the second-order probability den-
sity P2(Xl' tl; X2, t2), or equivalently by the joint probability Pl(Xl, td and
the conditional probability P2(X2, t2 I Xl, tl) to find the value X2 at time t2,
given that its value at previous time h (tl< t2) is Xl. In contradiction to
uncorellated processes, Markov processes are characterized by the following
temporal relationship

(2)
This Markov property enables us to calculate all higher-order joint probabil-
ities Pn for n > 2. To determine the fundamental equation of stochastic pro-
cesses of Markov type we start with the third-order distribution (tl < t2 < t3)

P3(Xl, tl; X2, t2; X3, t3) = P2(X3, t3 I X2, t2) P2(X2, t2 I Xl, td Pl (Xl, tl) (3)
and integrate this identity over X2 and divide both sides by Pl(Xl, tl)' We
get the following result for the conditional probabilities defining a Markov
process

P2(X3,t31 xl,td = ! P2(xs,tsl X2,t2)P2(X2,t21 xl,tl)dx2 , (4)

called Chapman-Kolmogorovequation. As already stated the Markov process

is uniquely determined through the distribution Pl (x, t) at time t and the
conditional probability P2(x', t' I x, t), also called transition probability from
X at t to x' at later t', to determine the whole hierarchy Pn (n ~ 3) by the
Markov property (2).
The memory in a Markov process is very short, only one time interval
from t to t' plays any role. H the trajectory has reached x at time t, the
past is forgotten, and it moves toward x' at t' with a probability depend-
ing on x, t and X', t' only. The entire information relevant for the future is
thus contained in the present. This gives sense to transform the Chapman-
Kolmogorovequation (4) in an equivalent differential equation in the short
 Probabilistic Description of Nucleation in Vapours and on Roads 365

time limit t' = t + 7 with small 7 tending to zero. The short time behaviour
of the transition probability P2 (. I .) should be written as series expansion
with respect to time interval 7 in the form

P2(x, t + 7 I x", t) = [1 - w(x, t)7] c5(x - x") + 7W(X, x", t) + 0 (7 2 ). (5)

The new quantity w(x, x", t) ~ 0 is the transition rate, the probability per
time unit, for a jump from x" to x =I x" at time t. This transition probability
multiplied by the time step 7 gives the second term in the series expansion
describing transitions from an other state x" to x. The first term (with the
delta function) is the probability that no transitions takes place during time
interval 7.
Inserting (5) in the right hand side of the Chapman-Kolmogorovequa-
tion (4) and integrating over x' (taking into account the normalization condi-
tions for probabilities) we obtain the differential formulation of the Chapman-
Kolmogorov equation

:tP1(x,t) = ! w(x,x',t)PI(x',t)dx' - ! w(x',x,t)PI(x,t)dx' (6)

called master equation in the (physical) literature [3,8,10]. The name 'master
equation' for the above probability balance equation is used in a sense that
this differential expression is a general, fundamental or basic equation. For a
homogeneous in time process the transition rates w(x, x', t) are independent
of time t. The short time transition rates w have to be known from the
physical context, often like a clever ansatz, or have to be formulated based
on a reasonable hypothesis or approximation. With known transition rates
w the master equation (6) gives the resulting evolution of the probability PI
over a long time period.
The master equation can be written in different ways. Since we are partic-
ularly interested in the case where the stochastic variable has discrete values
(natural numbers) within a finite range 0 ~ n ~ N, the master equation for
the time evolution of the probabilities PI (x, t), now P(n, t), then is written
as
dP(n,t)
--,:--:........:....
dt
=L {w(n,n')P(n',t) -w(n',n)P(n,t)} , (7)
n':j:n

where w(n',n) ~ 0 are rate constants for transitions from n to other n' =I n.
Together with the initial probabilities P(n,O) (n = 0,1,2, ... , N) and the
boundary conditions at n = 0 and n = N this set of equations goveming
the time evolution of P(n, t) from the beginning at t = 0 to the long-time
limit t -+ 00 has to be solved. The meaning of both terms is clear. The first
(positive) term is the inflow probability current to state n due to transitions
from other states n', and the second (negative) term is the outflow current
due to opposite transitions from n to n' .
366 R. Mahnke

Defining stationarity, sometimes caJIed steady state, as a time independent

distribution pst(n) by the condition dP(n, t)jdtlp=p.' = 0, the stationary
master equation is given by

0= L {w(n,n,)pst(n') - w(n',n)pst(n)} (8)

n'#n

This equation states the obvious fact, that in the stationary or steady state
regime the sum of all transitions into any state n must be balanced by the
sum of all transitions from n into other states n'. Based on the properties of
the transition rates per unit time the probabilities P(n, t) tend in the long-
time limit to the uniquely defined stationary distribution pst(n). In open
systems a constant probability flow is possible. This fundamental property of
the master equation may be stated as

lim P(n, t)
t--+oo
= pst(n) . (9)

Let us shortly discuss the question of equilibrium in a system without

e:lcternal exchange. The condition of equilibrium in closed isolated systems
is much more stronger than the former condition of stationarity (8). Here
we demand a balance between each pair of states n and n' separately. This
so-called detailed balance relation is written for the equilibrium distribution
peq(n) as

0= w(n,n')peq(n') - w(n',n)peq(n) . (10)

Of course, each equilibrium state is by definition also stationary. If the ini-
tial probability vector P(n, 0) is strongly nonequilibrium, many probabilities
P(n, t) change rapidly as soon as the evolution starts (short-time regime),
and then relax more slowly towards equilibrium (long-time behaviour). The
final (thermodynamic) equilibrium is reached in the limit t --+ 00. As a par-
ticular example of such a behaviour, the relaxation process in atomic clusters
has been studied in [19].
We are speaking about a onEHlimensional stochastic process if the state
space is characterized by one variable only. As in our investigations, this dis..
crete variable is a particle number n ~ 0 describing the amount of molecules
or the size of an aggregate. In chemical physics such aggregation phenomena
like formation andj or decay of clusters are of great interest. Examples are
the formation of a crystal or glass upon cooling a liquid or the condensation
of a droplet out of a supersaturated vapour. To determine the relaxation dy-
namics of model clusters we take a particularly simple Markov process with
transitions between neighbouring states n and n' = n ± 1. This situation is
called a one-step process. In biophysics, if the variable n represents the num-
ber of living individuals of a particular species, the one-step process is often
called birth-and-death process to investigate problems in population dynam-
ics. The random walk with displacements to left and right by one step is wen
 Probabilistic Description of Nucleation in Vapours and on Roads 367

known in physics and often plays a role as an introductionary example and

has been recently revisited and applied to new fields like econophysics [23].
Setting the transition rates w(n, n-I) = w+(n-I), w(n,n+I) = w_(n+
1), and therefore also w(n + I,n) = w+(n), w(n -I,n) = w_(n), the master
equation (7) reads
dP(n,t)
dt = w+(n -1) P{n -1, t) + w_(n + 1) P{n + 1, t)
- [w+{n) + w_{n)] P{n, t) . (11)
In general the forward and backward transition rates w+(n),w_{n) are non-
linear nmctions of the random variable nj the physical dimension is one over
time (S-1). We pay attention to particles as aggregates in a closed box or
vehicular jams on a circular road. Therefore in finite systems the range of the
=
discrete variable n is bounded between 0 and N (n 0,1,2, ... , N).
The general one-step master equation (11) is valid for n = 1,2, ... ,N -1,
but meaningless at the boundaries n = 0 and n = N. Therefore we have to
add two boundary equations as closure conditions
dP(O, t)
dt = w_(I) P{I, t) - w+(O) P(O, t) , (I2)
dP(N,t)
dt = w+(N - 1) P(N - 1, t) - w_ (N) P(N, t) . (13)

In finite systems with n =

0, 1,2, ... ,N one finds a situation which cor-
responds to steady state with a detailed balance relationship similar to (1O).
Therefore the stationary distribution pst{n) fulfils the recurrence relation

(14)

By applying the iteration successively and taking into account the normal-
ization condition (the sum of all the probabilities must be equal to 1), the
stationary probability distribution pst(n) in finite systems is finally written
as

n = 1,2, . .. ,N
1+ LN 11k w+{m - 1)
w_(m)
k=1m=1 (15)
1 n=O.
1+ L 11 w+w_(m)
N k
(m - 1)
k=1 m=1

The obtained result is a unique solution for the stationary probability distri-
bution in finite systems with closed boundaries. For an isolated system the
368 R. Mahnke

stationary solution of the master equation pst is identical with the thermody-
namic equilibrium pe q , where the detailed balance holds, which for one-step
processes reads

(16)

The condition of detailed balance states a physical principle. H the distri-

bution peq is known from equilibrium statistical mechanics and one of the
transition rates is also known (e. g. by a reasonable ansatz w+), the equa-
tion (16) provides the opportunity to formulate the opposite transition rate
w_ in a consistent way. By this procedure the nonequilibrium behaviour is
adequately described by a sequence of (quasi-)equilibrium states. The relax-
ation from any initial nonequilibrium distribution tends always to the known
final equilibrium. In physical systems the equilibrium distribution usually is
represented in an exponential form

peq(n) cx exp [-!1(n)j(kBT)] (17)

where !1(n) is the thermodynamic potential depending on the stochastic vari-
able n, kB is the Boltzmann constant, and T is the temperature [21].

3 Nucleation in Supersaturated Vapours

H we consider a vapour at equilibrium then a certain change of the thermo-
dynamic parameters enables us to remove the system into a nonequilibrium
state. The vapour becomes supersaturated. The basic quantity describing the
situation is the cluster distribution N at time t

(18)
which gives the number of clusters N n of size n. The free particles (molecules)
are called monomers of size n = O. It is supposed that N l molecules are
excited. These molecules may be named as precluster of size n = 1. The
bound states are clusters of size n ~ 2. Investigating a finite system the overall
number of particles Ntotal as wen as the volume V and the temperature T
are fixed. The particle conservation law
N
Ntotal = No + N l + L nNn = const (19)
n=2

takes into account that particles are either free, excited, or bounded in clus-
ters. There is always some difficulty to describe the initial stage of formation
of a cluster. We have introduced the precluster as an intermediate state be-
tween free and bounded states to provide an easy and unified description of
the aggregation process which is valid both in supersaturated vapour and in
traffic flow discussed later on.
 Probabilistic Description of Nucleation in Vapours and on Roads 369

A simple exarnple of a system of many particles is a gas consisting of iden-

tical molecules called monomers. H the gas is dilute (the density as number
ofmolecules per unit volume is small), the average separation length between
the monomers is large and correspondingly their interaction is negligible. The
gas is said to be ideal, if the average separation length is much larger than the
de Broglie wave length. We treat the monomers as undistinguishable particles
moving in a closed volume and making reactive collisions to form aggregates
called molecular clusters. Furtheron, we consider a simplified case where only
one single cluster of size n (i. e., N n = 1) coexists with No = Ntotal - n
unbounded (free) particles, which means that the cluster distribution (18)
reduces to

N(t) = (No,O, ... ,O,Nn = 1,0, ... ,0) (20)

and the overall particle conservation (19) to

Ntotal= No + n . 1 = const , (21)

where the stochastic variable n = n(t) is the number of particles bounded in

the cluster at time t. The nucleation box (volume V) embedded in a heat bath
(temperature T) displaying the situation schematically is shown in Fig. 1.

•
~(n).

~~)

Fig.l. Isothermal-isochoric nucleation in supersaturated vapour. Free moleeules

(black dots) called monomers in the initial stage (left) of the aggregation pro-
cess form a cluster (spherical droplet) of certain size n, coexisting with gas of free
molecules, afterwards (right).

The starting point is the one-dimensional one-step master equation (11)

describing condensation and evaporation of a single particle on or from a
molecular cluster, where w+(n) and w_(n) are transition rates of condensa-
tion and evaporation, respectively, which have to be formulated now.
The attachment probability per time of a monomer to a (spherical) clus-
ter of size n is proportional to cluster surface A(n) and to density of free
370 R. Mahnke

monomers No/V, i. e.

w+(n) = aA(n) (Ntotal - n) /V 1 ~ n ~ Ntotal . (22)

The coeflicient a is up to now not specified (interpretation: velocity of stick-

ing) . The surface of spherical droplet is given by

(23)

with known incompressible particle density inside the cluster Cclust = const
(liquid density as given experimental value). A special case is the formation
of aprecluster n = 1 out of an elementary particle (n = 0). The precluster
can be understood as an excited monomer which is able to react with some
other monomer to form a dimer (n = 2). In the free particle state n = 0 any
of the Ntotal monomers can become excited, so that we can write

(24)

where the dimensionless parameter pinthis case means the excitation prob-
ability per time multiplied by the time constant T.
By using the detailed balance relation (16), the evaporation rate w_(n) of
a monomer from a cluster of size n is calculated from the known attachment
rate (22), as shown in Appendix, i. e.,

w_(n)
- ln-I)
= aA(n) Ag1 exp (In kBT . (25)

Here In - In-l is the difference of binding energies between clusters of size

n and n - 1 and Ao (T) is de Broglie wave length of a monomer (particle of
mass m) given by

(26)
This is the wavelength of a quantum-mechanical free particle with energy
E = p2/2m = li2k 2/2m, where k is the wave number related to the wave-
length An.
Clusters as bound states of elementary particles (monomers) have negative
potential energy, which is the so-called binding energy. The potential function
In(T) is weH known from atomic and nuclear theory and also the Bethe-
Weizsäcker-formula [27-29] which in a simple nonlinear approximation reads

In(T) = JLoo(T) n + O"A(n) . (27)

The binding energy consists of a negative volume term (JLoo < 0) and a pos-
itive surface contribution. The quantity JLoo (T) is the chemical potential of
one monomer or, in other words, the energy necessary for taking away one
elementary particle (monomer) from a cluster with a fl.at surface. The param-
eter 0" can be understood as the surface tension of a fl.at surface. Ansatz (27)
 Probabilistic Description of Nucleation in Vapours and on Roads 371

is a good approximation for large enough sizes n and provides also correct
normalization condition 10 = 0 for a free particle (n = 0). Substituting (27)
into the detachment rate (25) we obtain the approximation
_ A() 1 {Poo(T)+O"[A(n)-A(n-l)]}
w_ (n ) - a n >.g exp kBT

1 (f.Loo(T») (20"k(n») (28)

~ aA(n) >.g exp kBT exp CclustkBT .

This result is valid for large enough clusters (starting with n ~ 10) and
contains the curvature k(n) of a size-n-droplet

k(n) = Ifr = (cclust 47r f3)1/3 n- 1 / 3 . (29)

The difference of surface areas A(n) - A(n - 1) in (28) can be evaluated by
using aseries expansion of A(n -1) around n and retaining the leading term
only, i. e.

n2/ 3 - (n _1)2/3 ~ n2 / 3 - n 2 / 3 (1- 3:) rv n- 1 / 3 . (30)

Taking into account the ideal gas model, the chemical potential f.Loo is re-
lated in a simple way to the equilibrium density (concentration) ceq(oo) of
monomers for Hat interface (r --)- 00) between liquid phase (droplet ) and
gaseous phase (free monomers). Thus, we have

(31)

However, in reality due to spherical droplets the interface is curved, therefore

the concentration of free monomers in equilibrium is larger than ceq (00).
For large enough clusters, the detachment probability

w_(n) = aA(n)ceq(oo) exp(lk(n» (32)

is obtained by inserting (31) into (28), where the length l = leT) defined as
(33)
is explained and visualised in Fig. 2.
Considering the cluster sizes n(t) as continuous variable whicl1 can be
measured experimentally, the equation of motion dnfdt with a given velo city
function v(n), showing the time evolution of the cluster size, is weIl known.
Putting forward phenomenological arguments like Fick's law the dynamical
equation of reaction limited aggregation reads
dn D
dt = v(n) with v(n) = f"A(n) (c/ree - ceq(n» , (34)
372 R. Mahnke

cc1ust t-----l.-! i I :

C C

coq(n) ------------ ------~--_____l coq(n) ---------------...L...------l

Fig. 2. Particle concentration c depending on the distance r from the center of

cluster. The left picture shows areal density profile, whereas the right one approxi-
mates the profile corresponding to a model of cluster with sharp border. The length
l shows the width of the interface between the dense phase (Cclust) and the dilute
surrounding (Ceq (n)) over a size-n-droplet.

where cfree = No/V = (Ntotal - n)/V = C - n/V is the density of free

particles, and C = Ntotaz/V is the total density of particles. The constant D
is called diffusion coefficientj the coefficient l called capillary length is a small
interface thickness defined by (33).
In the stochastic approach an equation of the same type as (34) can be
obtained for the average cluster size (n). Based on the master equation (11),
we get a deterministic equation for the mean value

(35)

which can be written in a certain approximation as follows

(36)

describing the time evolution of the average cluster size (n). According to the
definitions of the transition frequencies (22) and (32), the time evolution of
the mean cluster size (36) can be written in the same form as (34),

~(
dt n ) -_ a A«(n )) [NtotalV- (n) _ ceq (00) elk«n»] . (37)

By comparing these two equations we find the up to now unknown coefficient

a and the equilibrium concentration ceq(n), i. e.,

a=D/l (38)
and
ceq(n) = ceq(oo) elk(n) . (39)
 Probabilistic Description of Nucleation in Vapours and on Roads 373

The only difference between (34) and (37) is that in the latter case we always
have the average value of the cluster size (n) instead of n. Rewritting (37)
we get finally

d~~} = ~A((n}) (c- (~ -Ceq(oo)elk(n») . (40)

In the stationary state d(n}/dt = 0 holds. Equation (40) is valid at large

enough n only, whereas at (n) -t 0 it should be modified to ensure that
the transition rates and, therefore, the expression in the brackets does not
diverge. Then we obtain three stationary solutions. The homogeneous situ-
ation without any cluster corresponds to A((n}st) = 0, or to zero value of
the stationary cluster size (n) st = O. The other two solutions, describing the
heterogeneous situation, originate from the identity c/ree((n}st) = ceq((n}st)
or in extended version

C-
(n}st -_
V ceq (00) exp [0.[. (cc!ust 47r / 3)1/3 (n }-1/3]
st • (41)

From this equation we can find the stationary cluster size (n}st as function
of the total density c. This is a nonlinear equation which cannot be solved
analytically. However, it can be easily analysed (solved) graphically.1n Fig. 3
terms on the 1. h. s. (thick straight line) and on the r. h. s. (curved line)

,,,
,, r.h.s.

l---L---------
I,:, .i
o ner n~r
o~--~--------=_~----~~--~
n n stabfe

Fig. 3. Terms on the l. h. s. (straight lines) and on the r. h. s. (curved line) of (41)
depending on the cluster size n. The two crossing points at n = nstable (stable clus-
ter size) and n = n CT (unstable or critical cluster size) correspond to two different
solutions of (41). The thin straight line related to the critical density Cl has a single
common point with the curved line at n = n;T. The horizontal dashed line shows
the value of equilibrium concentration Ceq(oo).
374 R. Mahnke

vs (n) are shown. The two crossing points (n) = n cr and (n) = nstable
correspond to two different solutions of (41). The quantity n cr is known
as the critical cluster size in nucleation theory, whereas nstable represents
the stable stationary cluster size. Their meaning will be clarified in further
discussion. The crossing points exist only if the total concentration C exceeds
some critical value C > Cl which corresponds to a bifurcation point where
both solutions merge into one, as it is shown by thin straight line which has
only one common point with the curved line at the marginal (largest possible)
value of the critical cluster size (n}st = n~r. At C> Cl three different regions
can be distinguished for the cluster size (n}st:
I. At (n) < n cr we have d(n) / dt < 0 which means that the cluster dissolves.
11. At n cr < (n) < nstable we have d(n}/dt > 0 which means that the cluster
growth until reaches the stable stationary size nstable.
111. At (n) > nstable we have d(n}/dt < 0 which means that the cluster
reduces its size (dissolves) to the stationary value nstable.
According to this, the solution (n) = nstable(C) corresponds to a stable cluster
size, whereas the solution (n) = ncr(c) to an unstable stationary cluster size.
In Fig. 4 we have shown both solutions of (41) (branches nl(c) and n2(c)),

<n>st

n.;, -----------

Fig.4. The stationary cluster size {n}st depending on the total density C ofparticles
in a supersaturated vapour. The stable cluster size (the horizontal line and the
branch nl(c» is shown by thick solid lines, whereas the unstable cluster size (the
branch n2(c» is shown by dot-dashed line. Arrows indicate the time evolution of
{n}.

providing the stationary cluster size in a supersaturated vapour as a function

of total density c. Branch nl ( c) depicted by thick solid line corresponds to
the stable cluster size, whereas branch n2(c) indicated by dot-dashed line
 Probabilistic Description of Nucleation in Vapours and on Roads 375

correspond to the unstable (critical) cluster size. Several trajectories showing

the time evolution of (n) to one of the stable stationary values (thick lines)
are indicated by arrows.
Expansion of the exponent in (39), retaining the linear term only, yields

ceq(n}) = ceq(oo)(l +lk(n})) . (42)

By using this linearization around the critical cluster size n cr , (40) can be
written in the well- known form [28]

d~~} = Dceq(oo)A(n}) [k(n cr ) - k(n})] (43)

From this equation the above discussed property that clusters with an over-
critical size (n > ncr, i. e. k(n) < k(n cr )) grow, whereas those with an under-
critical size (n < ncr, i. e. k(n) > k(n cr )) dissolve, is obvious. In the actual
bistable situation the growth from an undercritical to an overcritical cluster
size cannot be described by deterministic equations of motion like (43) . This
phenomenon of noice-induced transitions over the critical value of cluster size
can be treated in the stochastic approach only.

Fig. 5. Three different stochastic trajectories showing the time evolution of the
cluster size n vs the dimensionless time t*. The lower dashed line indicates the
critical cluster size n cr = 54; the upper dashed line represents the stable cluster
size nstable = 650.

We have simulated by Monte Carlo method and have presented in Fig. 5

three different stochastic trajectories showing the time evolution, i. e. the
376 R. Mahnke

t*=0.003
-50
0.04

P(n,t*) -150
F(n)/(keD
F(n)/(kBT)
0.02
-250

-350
00 1000
n

Fig.6. The probability distribution P(n, t*) at three different dimensionless time
moments t* = 0.003 (left), t* = 0.04 (middle), and t* = 0.3 (right) calculated by av-
eraging over 20 000 stochastic trajectories simulated by Monte Carlo method. The
equilibrium distribution is shown by smooth solid line. The maximum of the equi-
librium distribution corresponds to the minimum of free energy, i. e. F(n)/(kBT)
(smooth curve).

cluster size n vs the dimensionless time t* = (aA(I) IV) . t, of the system with
Ntotal = 1000 particles starting with n values around the critical cluster size
n cr Rj 54 (the lower dashed line). The parameters of the system are choosen
such that aA(I)/(kBT) = 10 and Vceq(oo) = 160. In one of the cases the
cluster dissolves, whereas in other two cases it grows over the critical size.
Besides, in distinction to the prediction ofthe deterministic equation (40), in
one of the cases the growth up to the stable cluster size n Rj 650 (the upper
dashed line) occurs starting with n = 45 < n cr . The probability distribution
at three different time moments, i. e. t* = 0.003, 0.04, and 0.3, have been
calculated by averaging over a large number of stochastic trajectories starting
with the n = n cr = 54. The results are shown in Fig. 6. The probability
maximum moves towards larger values of the cluster size n with increasing
time, and at t* = 0.3 the probability distribution agrees approximately with
the equilibrium distribution peq(n) cx: exp( -F(n)/(kBT» (smooth solid line)
given by (17). The maximum of the equilibrium distribution corresponds to
the minimum of free energy, whereas the critical cluster size to the local
maximum of free energy, as shown in Fig. 6.

4 Car Cluster Formation on Roads

The aggregation of particles out of an initially homogeneous situation is weH
known and it takes place in physical and in nonphysical systems, like trafik
 Probabilistic Description of Nucleation in Vapours and on Roads 377

fiow, and can be described similarly [12,15]. Trafiic fiow theory has a long
history [24] and is still alive [7]. Theoretical approaches based on deterministic
car following models [1] and particle hopping models [2] are weIl known. An
experimental study of phase transitions in trafIic fiow has been presented by
Kerner [13], the stochastic description is provided in [11,16-18].
In analogy to the droplet formation in supersaturated vapour, we consider
a model of trafIic fiow on a one-Iane road according to which N cars are
moving along a circle of length L, as illustrated in Fig. 7 where two different
regimes of trafIic fiow are shown, i. e., free fiow (left) and congested trafIic fiow
with two car clusters (right) . For simplicity, we have restricted our further

U'
d~ree
.. t \

.
\ r.~~
I
• ~
~~
\J
t

Fig.7. Free trafik ßow (left) and congested traflic ßow (right) on the one-lane
circular road. Each black dot represents a car and its direction of motion is indicated
byarrow.

analysis to the simplest model where only one car cluster (jam) can exist. If
a road is crowded by cars, each car requires some minimal space or length
which, obviously, is larger than the real length of a car. We call this the
effective length l of a car. The maximal velo city of each car is v max . Following
the idea of Bando et al. [1], the behaviour of individual drivers is described by
the optimal velocity model. The desired (optimal) velocity Vopt, depending
on the distance between two cars Llx, is given in dimensionless variables
Wopt = vopt/vmax and Lly = Llx/l by the formula [16-18]

(44)

where the parameter d = D / l is the interaction distance. D is the distance

between two cars corresponding to the velo city value v max /2.
378 R. Mahnke

Measurements on highways have been shown that the density Cclust of

cars in congested trafik is independent of the size of the dense phase
(jam). As a consequence the distance between jammed cars, the spacing
LlXclust = iLlYclust = const 2:: 0, is weH known and has to be treated as
a given measured quantity.
The length of the cluster (jam) depending on the number of congested cars
n is defined by Lclust = in + (n - l)Llxclust. According to this, the average
distance Llxfree = iLlYfree between two cars outside the jam (or free cars)
distributed over the free part of the road with length Lfree = L - Lclust is
given by

Ll () _ Lji - N - S(n)LlYclust
Yfree n - N _ S(n) , (45)

where N is the total number of cars on the road (circle of length L) and
S(n) = n - 1 + 8n ,Q is the number of fixed spacings LlXclust between cars in
jam.
The trafik flow is described as a stochastic process where adding a ve-
hicle to a car cluster of size n is characterized by a transition frequency
(attachment probability per time unit) w+ (n) and the opposite process by
a frequency w_(n). The number n of cars in the cluster is the stochastic
variable which may have values from 0 to N. The basic equation for the evo-
lution of the prob ability distribution P(n, t) to find a cluster of size n at time
t with probability P is known as master equation (11). The one-dimensional
stochastic equation reads
1 dP(n,T)
-:;. dT = w+(n -l)P(n -1, T) - [w+(n)
+w_(n)] P(n, T) + w_(n + l)P(n + 1,T) . (46)
where T = tjr is the dimensionless time. The time constant r will be specified
below.
The main task is to formulate expressions for both transition probabilities
w+ and w_. We have assumed that the detachment frequency w_(n) or the
average number of cars leaving the cluster per time unit is a constant inde-
pendent of cluster size n. The ansatz for w+(n) is now formulated allowing
for LlXclust to be nonzero. Our general assumption is that a vehicle changes
the velo city from Vopt(Llxfree) in free flow to Vopt(LlXclust} in jam and ap-
proaches the cluster as soon as the distance to the next car (the last car in
the cluster) reduces from Llx free to LlXclust. This< assumption allows one to
calculate the average number of cars joining the cluster per time unit or the
attachment frequency to an existing cluster w+(n). The case w+(O) have to
be considered separately. This is a probability per time of a stochastic event
that a driver randomly reduces the velo city ofhis car to Vopt(LlXclust) without
obvious reason, thus forming aprecluster or cluster of size n = 1. An ansatz
for w+(O) similar to that we have used in the case of supersaturated vapour
 Probabilistic Description of Nucleation in Vapours and on Roads 379

is appropriate. Thus, we have the ansatz which in dimensionless quantities

reads
W+(O) = !!. N (47)
7
w+(n) = ~ Wopt(L1Y/ree(n)) - Wopt(L1Yclust) (48)
7 L1Y/ree(n) - L1Yclust
w_(n) = 1/7 = const , (49)
where b = V max 7/ i denotes a dimensionsless parameter and p is the stochas-
ticity. The parameter 7 is a time constant, which can be understood as the
waiting time for the escape (detachment) of first car out of the jam into free
flow [16].
The master equation (46) has been solved in Refs. [16-18] with an aim to
extract from this solution information about the formation of traffie jams and
ab out the various possible regimes of traffie flow depending on the parameters
of the system, as wen as to calculate the flux-density relationship, called
fundamental diagramm of traffic flow. Our purpose is to study the dynamics
of the cluster (jam) growth and dissolution by finding stable and unstable
stationary states, and compare the results with aggregation in supersaturated
vapour.
Like in the ease of the cluster growth in supersaturated vapour, the deter-
ministic equation (35) for growth of a single ear cluster on the road is written
as (36). The stationary cluster size (n)st ean be calculated from the condition
d(n)/dt = 0 or

w+«n») = bWopt (L1Y/ree «n»)) - W opt(L1Yclust) = 1

(50)
w_«n») L1Y/ree«n») - L1Yclust
consistent with the ansatz for transition probabilities (48) and (49). By using
the definition (44) of the optimum velocity function wopt(Lly) we obtain the
equation

b [(L1Y/ree«n»))2 (LlYclust) 2 ] L1 «») Ll ()

d2 + (LlY/ree«n»))2 - d2 + (L1Yclust)2 = Y/ree n - Yclust 51

which ean be solved with respect to L1Y/ree' One solution of the third order
equation (51) is trivial L1Ytree = L1Yclust. The other two solutions, whieh have
certain physical meaning, read

Ll (1,2) _ d (52)
Y/ ree - 2[d2 + (LlYclust)2]
x {bd ± ..jb2d2 + 4bLlYclust[d2 + (L1Yclust)2]- 4[d2 + (L1Yc!usd2j2} .
According to (52), the headway L1Ytree between cars in a free flow coexisting
with a single cluster has constant value depending merely on the control
parameters of the model. Now, by means of (45), which states the relation
380 R. Mahnke

between L1y free and n, we are able to calculate the stationary cluster size
(n}st. As already pointed out we set S(n) = n - 1 + 8n ,Q ~ n, assuming that
the cluster (if it exists) contains a large number of cars. This leads to the
equation

(n}st C (1 + L1y (1,2»)

free - 1
L1Y free L1 ' (53)
L/i (1,2)
- Yclust

where c = iN/ L is the total density of cars and the term on the left hand
side of the equation, in fact, is the relative part of the road crowded by cars.
In this case L1y};;~ has a constant value given by (52). Result (53) makes
sense at large enough densities where it provides a positive value of (n}st.
The solution with the largest value L1Y?:ee (positive sign in (52)) gives the
average stationary size of a stable cluster depending on the total density C
within the region c > Cl, where
1
Cl = (1) (54)
1 + L1YJree

is the critical density having the same meaning as in the case of cluster growth
in supersaturated vapour. There is another critical density C2, given by
1
C2 = (2)' (55)
1 + L1Yfree
which defines the region C > C2 where an unstable car cluster corresponding
to the solution (53) with the smallest value L1Y}~ee of the headway can exist.
In a special case of vanishing bumper-to-bumper distance in jam L1Yclust = 0
our result (52) reduces to

(1,2)_
L1y free -
b
2± V~
4" - d,2 , (56)

and (53) to

(n}st c-1
(57)
L/i =C+ b ~2 .
-± --d,2
2 4
The value of (n}st i/ L with sign '+' in (57) corresponds to the stable, whereas
that with sign '-' - to the unstable stationary cluster size. In Fig. 8 we have
shown both branches n1 (c) (stable) and n2 (c) (unstable) as a function oftotal
density c.
This result, obtained in the simple one-cluster model discussed here, re-
mains true in the multi-cluster model, outlined in [12] and discussed in detail
 Probabilistic Description of Nucleation in Vapours and on Roads 381

I
/
<n>st I
/
UI /
/
/
/
/
/
/
l./ n2 (C)
I

/
/l
/
,/
o ~-*---------L---4~~----~
o CI c

Fig. 8. The stationary cluster size (n)"t normalized to L/l (i. e., the relative part of
the road crowded by cars) depending on the total density of cars c. The stable cluster
size (branch nl (c) and horizonta lines) is shown by thick solid lines, whereas the
unstable cluster size (branche n2(c)) and a horizontalline) is shown by dot-dashed
lines. Arrows indicate the time evolution of (n). Parameters: b = 8.5, d = 13/6, and
LlYcl"st = O.

l00.---------~----------.---------~

80

60

40

40000 60000
T

Fig.9. A stochastic trajectory showing the total number of congested cars n VB

dimensionless time T.

in [11], ifthe size ofthe system is large (N -+ 00) and stochasticity pis small.
Note only that n, in general, is defined as the total number of congested cars.
Like in the case of supersaturated vapour, the growth of the car cluster
starting with the undercritical (i. e., smaller than the unstable) cluster size
382 R. Mahnke

can be described within a stochastic approach only. Fig. 9 presents a typical

stochastic trajectory showing a noice-induced switching from free flow with
n values fluctuating below the unstable cluster size (in this example N = 92,
C = 0.7886, b = 8.5, d = 13/6, LlYclu8t = 1/6, which corresponds to the
critical cluster size about 53 and the stable cluster size about 91) to the state
whith large stable cluster including almost all cars. Note, however, that we
have two different kind of bifurcations at c = Cl in Figs. 4 and 8. In distinction
to the actual model of trafik flow (Fig. 8), in the case of the supersaturated
vapour (Fig. 4) there is a jump in the cluster size. Besides, in the traflic flow
we do not have the critical cluster size at intermediate densities Cl < C < l!2,
which means that even a very small car cluster at the initial time moment
has tendency to grow up to the stationary size. At large densities (c> C2)
the critical cluster size n2(c) appears as weH in our trafik flow model. The
above discussed distinguishing feature disapears in an advanced traffic flow
model discussed further on.

5 An Advanced Model of Car Cluster Formation

Here we propose an essential innovation in our traffic flow model: now the
detachment frequency depends on the cluster size. In analogy to physical
systems like droplets in supersaturated vapour, one can expect that smaller
clusters dissolve easier, i. e., w_(n) increases remarkably when the cluster size
becomes smaller than some characteristic value no. In traffic flow, this idea
is based on a concept of higher irregularity of the cluster structure when the
size is small. The increase in w_(n) might be related also to some multilane
effect, i. e., due to several possible overtaking maneuvers, clusters consisting
of few cars are particularly unstable. In some approximation our one-lane
model effectively describes a multilane freeway. In this case n and N are the
number of congested cars and the total number of cars per one lane. The
above mentioned multi-lane effect is simulated by choosing an appropriate
ansatz for w_(n). Our specific choice is to replace (49) with

w_(n) =!T [1+ ß (~)8]

n+no
n ~ 1, (58)

where ß and s are positive constants. Assuming no ~ 1, the parameter

ß ~ (w_(l) -w_(oo»/w_(oo) shows the relative increase in w_(n) when the
cluster size n is reduced from large values to 1. The parameter s is responsible
for the fastness of w_(n) convergence to its asymptotic value w_(oo) = I/T
at n -t 00. The specific form of w_(n) is choosen somewhat arbitrarily to get
a simple but still realistic description.
The deterministic equation (51) for the the stationary mean cluster size
in our advanced model becomes
bWopt (LlYfree «(n))) - Wopt(LlYclu8t} = 1 + ß (~) B
(59)
LlYfree«(n) - LlYclu8t n+no
 Probabilistic Description of Nucleation in Vapours and on Roads 383

consistent with the ansatz for transition probabilities (48) and (58). Equation
(59) has been solved together with (44) and (45) in the approximation Sen) ~
n,
..:1 () _ (L/f)(1- c) - n..:1Yclust
Yfree n - c(L/f) _ n ' (60)

at realistic values of the dimensionless control parameters b = 8.5, d = 13/6,

..:1Yclust = 1/6 estimated in [16]. Three different values of the parameter
ß, ß = 0, 0.8, and 1.6, have been considered. Other parameters have been
choosen L/f = 500, s = 1, and no = 10. We have calculated the stationary
mean cluster sizes (n}st corresponding, in general, to different stationary solu-
tions of these equations. Besides, the stable free flow state with (n) st = 0 has
been identified as astate with d(n}/dT < 0 at (n) -+ O. The corresponding
dimensionless fluxes

J. = JT = b [Wopt(..:1Yclust )f(n}st
--y;- + Wopt (..:1Yfree( (n })) ( c - -L-
f(n}st)] (61 )

have been calculated as weH, where J is the flux averaged over the whole
road or, which is the same, over an infinitely long time interval at a fixed
coordinate. In this case we have neglected the stochastic fluctuations taking
n == (n}st. The results are shown in Figs. 10 and 11.

<n>st 1.5
,.._.,
400 i, I \
!,"
I, :' \
I \
300
fl 1.0 i \
i ,,
1 ,
n,(c
i ,, I \\.\
,,
200
f! ,,

100
i ,,
,,
n2(c)/i ,,
0.5
\,
i ,,
i , c 0.0
0.5 C2 Cd"'" 1.0 0.0 c, 0.5

Fig. 10. Stationary cluster size (left) and fiux (right) vs concentration calculated for
finite road at dimensionless control parameters ß = 0, b = 8.5, d = 13/6, LlYclust =
1/6, L/f. = 500. The thick and the thin solid lines correspond to stable free and
congested states, respectively. Dashed lines represent an unstable free fiow, dot-
dashed line shows the unstable (critical) cluster size.

In Fig. 10 we have illustrated the behavior of the stationary cluster size

(n}st (left) and the corresponding flux (right) in a special case of parameter
ß = 0 where our advanced model reduces to the previous one discussed in
Sec. 4. As a result, the stationary cluster size behaves in a similar way as
384 R. Mahnke

<n> 1.5
400
,
j
300
j 1.0
n,(c) j
j
i
j
200
li, 0.5
i
100
n.(cil
-_ ....../ C
0.5 Cd""' 1.0 0.5
<n>at 1.5
400 j.
j'
i
j
300 j 1.0
i
n,(c
200 II i
0.5
100 n.(C)/l
,I
.-
-----_.-.,; C
0.5 Cd""' 1.0 0.5

Fig.ll. Stationary cluster size (left) and Hux (right) vs concentration calculated
for finite road at ß = 0.8 (top) and ß = 1.6 (bottom). Other dimensionless control
parameters are no = 10, s = 1, b = 8.5, d = 13/6, L1Yclust = 1/6, L/l = 500. The
meaning of solid, dashed, and dot-dashed lines is the same as in Fig. 10.

in Fig. 8, with the only essential difference that L:1Yclust > O. In this case
the free flow with (n}st = 0 is stahle up to some critical density C = Cl.
Then a car cluster appears which growth linearly in size with increasing of
C from Cl to the maximum value Cclust = 1/{1 + L:1Yclust) (the density in
cluster), as shown in Fig. 10 hy thin solid line. This corresponds to one ofthe
stationary solutions (n) = nl{c) of (59). At densities C2 < C< Ccl u8 t another
positive solution (n) = n2(c) exists (dot-dashed line) which corresponds to an
unstahle or critical cluster size. (dn/dt > 0) if n < nl(c), The most prohahle
time evolution of n, given hy the deterministic equation (36), depending on
the density C and initial conditions is indicated in Fig. 10 hy arrows. Like
in the supersaturated vapour, a growth of cluster starting from undercritical
size n < n2 (c) and finishing with the stahle cluster size n = nl (c) is possihle,
as weil, hut only due to stochastic fluctuations.
In Fig. 11 we have shown the same pictures at ß > O. A distinguishing
feature is that a nonzero critical cluster size appears already at C = Cl where
a phase transition takes place with a jump-like increase in the stahle cluster
size (n}st from 0 to n~rit. This hehavior is the same as in supersaturated
vapour. At small values of ß, however, a density region with vanishing critical
 Probabilistic Description of Nucleation in Vapours and on Roads 385

cluster size still exists (ß = 0.8, top in Fig. 11) which disapers at larger ß
values (ß = 1.6, bottom in Fig. 11). Note also that at ß = 0.8 the free flow
becomes unstable somewhat above Cl (see j(c) plots in Fig. 11), while both
free and congested states of the system are stable at ß = 1.6. At positive ß, a
jump-like decrease of the flux j is observed at the critical point C = Cl when
switching from the free flow (thick solid line) to the congested flow (thin solid
line). This is a finite-size effect which disapears in the limit N --+ 00, since
n~ritfN --+ O.

6 Conclusion

To conclude this paper we recall the key points of the developed approach.
It has been shown that both supersaturated vapour and trafik flow can be
described by stochastic master equations. Solutions of the master equation
in each of the cases have shown the existence of phase transition from ini-
tially homogeneous state to a heterogeneous state where some of the particles
(molecules or cars) are bounded in a cluster. The cluster growth dynamics
in supersaturated vapour and in traffic flow has been studied in view of
both deterministic and stochastic theory. Finally, the comparison of results
has been made showing that at certain conditions the nucleation in traffic
flow is similar to that in supersaturated vapour. In particular, in the model
where the detachment frequency increases remarkably at small cluster sizes
n ;S no (cf. Sec. 5) an undercritical bifurcation exists at some car density Cl'
lt means that, like in the supersaturated vapour, small, i. e. undercritical,
clusters tend to dissolve, whereas those of overcritical size tend to grow if the
density exceeds Cl.

7 Appendix: Derivation of Thermodynamic Potential

and Corresponding Transition Rate

Our system is described by the cluster distribution N defined by (18). The

total Hamiltonian H (N) reads

(62)

with the contribution H n for the N n clusters of size n at coordinates r~n) and
momentum p~n) written as kinetic energy and interaction potential

(63)
386 R. Mahnke

The mass m~n) of a cluster containing n monomers (n ;::::: 1) is given by

(64)
where m == mo is the mass of one monomer. The canonical partition nmc-
tion, i. e. the statistical integral, is an integral over all space and momentum
coordinates. In the semiclassical approximation it reads [21]

(65)

where ß = 1j(kBT). This partition nmction can be divided in two factors,

one of which represents an ideal part due to the kinetic energy

Zideal (T, V, N) = rrn=O NnVNnh3Nn ( V rr mnkB T ) 3N

N
! 2
n
(66)

and the second part Zbinding (T, V, N) is responsible for the energy stored in
clusters. In a certain approximation, both terms together read

Z(T V N) =
, ,
rrNVNn [(J2rrmnkBT)3
NI h exp
(_~)lNn
k T (67)
n=O n' B
where the binding energy f n (T) is the minimum value of the potential energy
searched over all spatial arrangements of the n bounded monomers

fn(T) = mjn L ug·,n) (I ri - rj I) . (68)

i<j

From the canonical partition function we can calculate the thermodynamic

quantities using the relation to state nmction free energy

F(T, V, N) = -kBT In Z(T, V, N) . (69)

According to (67) and (69), in the isothermal-isochoric situation we obtain
N N
F(T, V,N) = kBTL [Nnln (An(T)3jV) + In N n!] + LNnfn(T) (70)
n=O n=O
with de Broglie wave length An(T) n- 1 / 2 Ao(T) =
n- / (h j(2rrmkBT)) 1/2 (at n
1 2 2 > 1). Using Stirling formula
In N n ! ~ N n In N n - N n in (70) we obtain an approximation for large
cluster numbers N n [15]
N N
F(T, V,N) = kBTL N n [ln (A n (T)3NnjV) -1] + L Nnfn(T). (71)
n=O n=O
 Probabilistic Description of Nucleation in Vapours and on Roads 387

In the case of one cluster of size n only (see (20)) and at No = Ntotal - n -+ 00
(i. e., in the thermodynamic limit by expansion oflnNo!) the free energy (70)
reads
peT, V, n) = kBT {(Ntotal - n) [ln (Ao(T)3(Ntota ! - n)IV) - 1]
+ (1 - on,o) In (An(T)3 IV)} + in(T) . (72)
For the equilibrium cluster distribution, the most probable value of the
cluster size n corresponds to the minimum of the thermodynamic potential, in
this case free energy F, as it is evident from (17). The latter equation (with
n == F) combined with the detailed balance condition (16) allows to find
the relation between transition rates weN' I N) and weN IN') of opposite
stochastic events (transition from state N to state N' and vice versa), i. e.
weN' IN) _ (F(T, V,N) - F(T, V,N'))
weN IN') - exp kBT . (73)

In our special case, where only one cluster of size n is possible, (73) in the
thermodynamic limit reduces to
w_(n) =
----''--'-....,.. V (1- 1In)3/2 exp (in(T) - in-leT)) : n > 2
(74)
w+(n - 1) Ag(T) (Ntotal - n) kBT -
w_(l) 1 (h(T)) (75)
w+(O) = Ntotal exp kBT .
From (74) we obtain an approximation
w_(n) V (in(T) - in-leT») (76)
w+(n) = Ag(T) (Ntotal - n) exp kBT '
consistent with Eqs. (22) and (25), which is true for large enough n. In a
rough approximation, we have assumed that (76) can be extrapolated up to
n = 1. Based on (75) the stochasticity parameter p in (24) can be written as

-:;p = w_(l) (
exp - kBT
h(T») . (77)

By (22), (25), and (77) all the transition rates are weH defined in a way
consistent with the basic principles of statistical mechanics.

Acknowledgement: The author would like to thank Jevgenijs Kaupuzs (Biga),

Ihor Lubashevsky (Moscow) and Peter Wagner (Berlin) for useful comments,
criticism and help.

References
1. Bando, M., Hasebe, K., Nakayama, A., Shibata, A., Sugiyama, Y.: Structure
stability of congestion in traffic dynamics, Japan J. Indust. and Appl. Math.
11 (1994) 203
388 R. Mahnke

2. Chowdhury, D., Santen, L., Schadschneider, A.: Statistical physics ofvehicular

trafiic and some related systems, Phys. Rep. 329 (200) 199
3. Gardiner, C. W.: Handbook 0/ stochastic methods (Springer, Berlin, 1983). (1st
ed.), 1985, 1990
4. Haberland H. (Ed.): Clusters 0/ atoms and molecules (Springer, Berlin, 1994).
5. Haken, H.: Synergetics. An introduction (Springer, Berlin, 1978) (1st ed.),
1983
6. Heermann, D. W.: Computersimulation methods in theoretical physics
(Springer, Berlin, 1990).
7. Helbing, D.: Verkehrs dynamik. Neue physikalische Modellierungskonzepte
(Springer, Berlin, 1997).
8. Honerkamp, J.: Stochastische Dynamische Systeme, VCH Verlagsgesellschaft,
Weinheim, 1990; Stochastic Dynamical Systems, VCH, New York, 1994
9. Jellinek, J. (Ed.): Theory 0/ atomic and molecular clusters (Springer, Berlin:,
1999).
10. van Kampen, N. G.: Stochastic processes in physics and chemistry (North-
Holland, Amsterdam, 1981, 1992).
11. Kaupuzs, J., Mahnke, R.: A stochastic multi-cluster model of freeway traffic,
Europ. Phys. J. B 14 (2000) 793
12. Kaupuzs, J., Mahnke, R.: Nucleation on roads, In: Nucleation and Atmo-
spheric Aerosols 2000 (Eds.: B. Hale and M. Kulmala), AlP Conf. Proceed.,
Vol. 535, p. 221
13. Kerner, B. S.: Phase transitions in traffic fiow, In: '!'raffiG and Granular Flow
'99 (Eds.: Helbing, D., Hermann, H. J., Schreckenberg, M., Wolf, D. E.),
(Springer, Berlin, 2000) p. 253
14. Mahnke, R.: Aggregation phenomena to a single cluster regime under different
boundary conditions, Z. Phys. Chem. 204 (1998) 85
15. Mahnke, R.: Nucleation in physical and non-physical systems, In: Nucle-
ation and atmospheric Aerosols 2000 (Eds.: B. Hale and M. Kulmala), AlP
Conf. Proceed., Vol. 535, p. 229
16. Mahnke, R., Kaupuzs, J.: Stochastic theory of freeway traffic, Phys. Rev. E
59 (1999) 117
17. Mahnke, R., Kaupuzs, J.: Probabilistic description of traffic fiow, In: Special
issue on traffic flow theory (Ed.: Zhang, H. M.), Networks and Spatial Eco-
nomics, Vol. 1, No. 1/2, March 2001, p. 103
18. Mahnke, R., Pieret, N.: Stochastic master-equation approach to aggregation
in freeway traffic, Phys. Rev. E 56 (1997) 2666
19. Miller, M. A., Doye, J. P. K., Wales, D. J.: Structural relaxation in atomic
clusters: Master equation dynamics, Phys. Rev. E 60 (1999) 3701
20. Montroll, E. W., Badger, W. W.: Introduction to quantitative aspects of sodal
phenomena, Gordon and Breach Sci. Publ., New York, 1974
21. Münster, A.: Statistical thermodynamics Vol. I, (Springer, Berlin, 1969).
22. Newman, M. E. J., Barkema, G. T.: Monte Carlo methods in statistical physics
(Clarendon Press, Oxford, 1999).
23. Paul, W., Baschnagel, J.: Stochastic processes From Physics to Finance,
(Springer, Berlin, 1999).
24. Prigogine, 1., Herman, R.: Kinematic theory 0/ vehicular traffic (Elsevier, New
York, 1971).
25. Reiss, H., Huang, C.: Statistical thermodynamic formalism in the solution of
information theory problems, J. Stat. Phys. 3 (1971) 191
 Probabilistic Description of Nucleation in Vapours and on Roads 389

26. Reiss, H., Hammerich, A. D., Montroll, E. W.: Thermodynamic treatment

of nonphysical systems: Formalism and an example (single-lane trafiic),
J. Stat. Phys. 42 (1986) 647
27. Schmelzer, J., Jr., Lembke, U., Kranold, R.: Nucleation and growth of AgCl
clusters in a sodium borate glass: Numerical analysis and SAXS results, J.
Chem. Phys. 113 (2000) 1268
28. Schmelzer, J., Röpke, G., Mahnke, R.: Aggregation phenomena in complex
systems (Wiley-VCH, Weinheim, 1999).
29. Wu, D. T.: Nucleation theory, In: Solid State Physics, Vol. 50, (Eds.: Ehren-
reich, H., Spaepen, F.), (Academic Press, San Diego, 1997) p. 37
Cellular Automata Simulation of Collective
Phenomena in Pedestrian Dynamics

Andreas Schadschneider1 , Ansgar Kirchner 1 ,2, and Katsuhiro Nishinari 3 ,1

1 Institut für Theoretische Physik, Universität zu Köln, D-50937 Köln, Germany

2 Theoretische Physik, Fakultät 4, Gerhard-Mercator-Universität Duisburg,
D-47048 Duisburg, Germany
3 Department of Applied Mathematics and Informatics, Ryukoku University,
Shiga, Japan

Abstract. Pedestrian dynamics exhibits a variety offascinating and surprising col-

lective phenomena. A 2-dimensional cellular automaton model is presented which
is able to reproduce these effects in a simple way. Inspired by the principles of
chemotaxis the interactions between the pedestrians are mediated by a so-called
ftoor field. This field has a similar effect as the chemical trace created e.g. by ants
to guide other individuals to food places. In our case the Hoor field modifies the
transition rates to neighbouring cells such that a motion in the direction of higher
fields is preferred. Already the inclusion of only nearest-neighbour interactions al-
lows to reproduce many of the collective effects and self-organization phenomena
(lane formation, How oscillations at doors etc.) encountered in pedestrian dynamics.

1 Introduction
Considerable research has been done on the topic of trafiic flow using meth-
ods from physics during the last decade [1-4,6,5] where the focus has been
especially on highway trafiic. In contrast to the latter, pedestrian flow [7]
is truely 2-dimensional and effects due to counterflow become important.
In vehiclar trafIic one usually deals with a well-defined ordering of vehicles.
Interactions with nearest neighbours (i.e. the next car ahead) are most im-
portant. In pedestrian dynamics, however, the situation is more complex and
interactions with other individuals (which might cross the path of walking)
in a certain range have to be taken into account. This gives rise to several
self-organization phenomena not observed in vehicular trafiic.

1.1 Collective Effects

Many interesting collective effects and self-organization phenomena have been
observed in pedestrian dynamics. Here we only give abrief overview and refer
to [6-8] for a more comprehensive discussion and list of references.
J amming: At large densities various kinds of jamming phenomena occur,
e.g. when many people try to leave a large room at the same time where the
flow is limited by e.g. a door (see Fig. 1) or narrowing. Therefore this kind of
jamming phenomenon does not depend strongly on the microscopic dynamics

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 CA Simulation of Collective Phenomena in Pedestrian Dynamics 391

of the particles. This clogging effect is typical for a bottleneck situation.

It is important for practical applications, especially evacuation simulations.
Other types of jamming occur in the case of counterflow where two groups

... .
..*.-r:.. . * ...
. .. -.... ,
I e.. : *•
ü ••* _ • • ....
a S;.I__ • I
. ......t:r..
. . . . . ..
•
• I •••••
• • •I
... • : :- .*. •*'i.:... •1-· •

.. *. •••
•••
.*.
.'- *·*1- •
.:.
••
. r:·.*· .at· I··.. ·"1 •
:. .: c•.....
. .......
1 :- I I :..

I ... rO. °11 0".. ." .*

••

Fig. 1. Clogging near a bottleneck (door). Shown are typical stages of an evacuation
simulation: (a) initial state (t = 0); (h) middle stages; (e) end stage with only a
few partic1es left.

of pedestrians mutually block eaeh other. This happens typically at high

densities and when it is not possible to turn around and move back, e.g.
when the flow of people is large.
Lane formation: In counterflow, Le. two groups of people moving in
opposite direetions, a kind of spontaneous symmetry breaking oecurs (see
Fig. 2 and Sec. 4) . The motion of the pedestrians can self-organize in such a
way that (dynamically varying) lanes are formed where people move in just
one direction [9] . In this way, strong interactions with oncoming pedestrians
are reduced and a higher walking speed is possible.

Fig.2. Lane formation in counterflow in a narrow corridor.

Oscillations: In counterflow at bottlenecks, e.g. doors, one can observe

oscillatory changes of the direetion of motion. Onee a pedestrian is able to
pass the bottleneck it beeomes easier for others to follow him in the same
392 A. Schadschneider, A. Kirchner, K. Nishinari

direction until somebody is able to pass (e.g. through a fluctuation) the bot-
tleneck in the opposite direction (see Fig. 3).

Fig. 3. Flow oscillations at a door with counterflow.

Patterns at interseetions: At intersections various collective patterns

of motion can be formed. A typical example are short-lived roundabouts
which make the motion more efficient. Even if these are connected with small
detours the formation of these patterns can be favourable since they allow
for a "smoother" motion.
Panies: In panic situations, many counter-intuitive phenomena can occur.
In the faster-is-slower effect [10] a higher desired velocity leads to a slower
movement of a large crowd. In the freezing-by-heating effect [11] increasing
the fluctuations can lead to a more ordered state. For a thorough discussion
we refer to [8,10] and references therein.

1.2 Modelling Approaehes

The number of models for pedestrian dynamics is much smaller than for ve-
hicular traffic [4,6]. Nevertheless, several approaches [7] have been suggested:
Socia! foree models: So far continuum models have been most successful
in modelling pedestrian dynamics. An important example are the social force
models (see e.g. [6,9,13] and references therein). Here pedestrians are treated
as particles subject to long-ranged forces induced by the social behaviour of
the individuals. The typical structure of the force between the pedestrian is
described by [10]

f.3,
•
= f~~oc)
'3
+ r~~hIl8)
'3
. (1)

Here ~;oc) is a repulsive social force modelling the tendency to keep a

certain distance to other individuals. Typically it is long-ranged: fi~oc) cx:
exp( -rij / Ü' ~fhIlB) is a physical force and describes friction and compres-
sion when pedestrians make contact. This idea leads then to equations of
 CA Simulation of Collective Phenomena in Pedestrian Dynamics 393

motion similar to Newtonian mechanics. There are, however, important dif-

ferences since, e.g., in general the third law ("actio = reactio") is not fulfilled
by social forces. It should be mentioned that the social force model is very
successful in modeling the collective effects listed in Sec. 1.1.
Active-walker Illodels: So-called active-walker models [12,14,15] have
been used to describe the formation of human or animal trails etc. Here
the walker leaves a trace by modifying the underground on his path. This
modification is real in the sense that it could be measured in principle. For
trail formation, vegetation is destroyed by the walker and some animals (like
ants) leave a chemical trace which can be detected by other animals.
Cellular autoIllata: Most cellular automata (CA) models for pedestrian
dynamics proposed so far are rather simple [16-19] and can be considered
as generalizations of the Biham-Middleton-Levine model for city trafik [20].
However, these models are not able to reproduce all the collective effects
observed empirically. The same is true for more sophisticated discrete models,
e.g. the one suggested by Gipps and Marksjös [21] or a disretized version of
the social force model [22].
In [23-26] a new kind of CA model has been introduced which - despite
its simplicity - is able to reproduce the observed collective effects. It takes
its inspiration l from the process of chemotaxis (see [27] for a review) as used
by some insects. They create a chemical trace to guide other individuals
to food places. This is also the central idea of the active-walker models. In
the approach of [23] the pedestrians also create a trace. In contrast to trail
formation and chemotaxis, however, this trace is only virtual although one
could assume that it corresponds to some abstract representation of the path
in the mind of the pedestrians. Its main purpose is to transform effects of
long-ranged interactions (e.g. following people walking some distance ahead)
into a local interaction (with the ''trace''). This allows for a much more efficient
simulation on a computer.

2 Basic Principles of the Model

First we discuss some general principles applied in the development of the
model [23,24]. To allow an efficient implementation for large-scale computer
simulations a discrete model is preferable. Therefore a two-dimensional CA
is used with a stochastic dynamics taking into account the interactions be-
tween the pedestrians. As mentioned above, similar to chemotaxis, we want
to transform long-ranged interactions into local ones. This is achieved by
the introduction of so-called fioor fields. The transition probabilities for all
pedestrians depend on the strength of the floor fields in their neighbourhood
in such a way that transitions in the direction of larger fields are preferred.
Interactions between pedestrians are repulsive for short distances. One
likes to keep a minimal distance to others in order to avoid bumping into
1 Such "learning from nature" is the central idea of a field called Bionies.
394 A. Schadschneider , A. Kirchner, K. Nishinari

them. This is taken into account through hard-core repulsion which prevents
multiple occupation of the cells. For longer distances the interaction is often
attractive. E.g. when walking in a crowded area it is usually advantageous
to follow directly behind the predecessor. Large crowds may also be attrac-
tive due to curiosity and in panic situation often herding behaviour can be
observed [10].
The long-ranged part of the interaction is implemented through the Hoot
fields. We distinguish two kinds, a statie ftoor field and a dynamic ftoor field.
The latter models the dynamic interactions between the pedestrians, whereas
the static field represents the constant properties of the surroundings.
The dynamic Hoor field corresponds to a virtual trace which is created
by the motion of the pedestrians and in turn influences the motion of other
individuals. Furthermore it has its own dynamics, namely through diffusion
and decay, which leads to a dilution and finally the vanishing of the trace
after some time.
The static Hoor field does not change with time since it only takes into
account the effects of the surroundings. Therefore it exists even without any
pedestrians present. It allows to model e.g. preferred areas, walls and other
obstacles. A typical example can be found in Sec. 5 where the evacuation
from a room with a single door is examined. Here the strength of the static
field decreases with increasing distance from the door.
The introduction of the Hoor fields allows for a very efficient implementa-
tion on a computer since now all interactions are local. We have translated the
long-ranged spatial interaetion into a loeal interaetion with ''memory''. There-
fore the number of interaction terms grows only linearly with the number of
particles. Another advantage of local interactions can be seen in the case of
complex geometries. Due to the presence of walls not all particles within the
interaction range interact with each other. Therefore one needs an algorithm
to check whether two particles ''see'' each other or whether the interaction is
blocked by some obstacle. All this is not necessary here.
Furthermore we do not need to provide the pedestrians with some BOrt of
"intelligence". The Hoor fields are sufficient to achieve the formation of com-
plex structures and collective effects by means of self-organization. Even for
an evacuation scenario [26] (see Sec. 5) the knowledge about the Hoor fields
is sufficient to find the exit. Therefore the pedestrians behave like simple
'particles' in a field without explicit intelligence. Since we also not make de-
tailed assumptions about the human behaviour it allows us to keep the model
simple. Nevertheless it is able to reproduce many of the basic phenomena.
In contrast to vehicular traffic the time needed for acceleration and brak-
ing is negligible in pedestrian motion. The velo city distribution of pedestrians
is sharply peaked [28]. These facts naturally lead to a model where the pedes-
trians have a maximal velo city Vmax = 1, Le. only transitions to neighbour
cells are allowed. Furthermore, a larger Vmax , Le. pedestrians can move more
than just one cell per timestep, would be harder to implement in two dimen-
sions and reduce the computational efficiency.
 CA Simulation of Collective Phenomena in Pedestrian Dynamics 395

For some applications it is useful to introduce another field called matrix

of preference which encodes the preferred walking direction and speed of
each pedestrian. From this direction, a 3 x 3 matrix is constructed which
contains the probabilities for a move of the particle. The central element
describes the probability for the particle not to move at all , the remaining
eight correspond to a move to one of the neighbouring cells2 (see Fig. 4).
The probabilities can directly be related to observable quantities, namely the
velo city and the longitudinal and transversal standard deviations (see [23,29]
for details). In the simplest case the pedestrian is allowed to move in one
direction only without fiuctuations. In the corresponding matrix of preference
only one element is one and all others are zero. In the following it is assumed
that a matrix of preference is given at every timestep for each pedestrian. In
principle, it can differ from cell to cell depending on the geometry and aim
of the pedestrians. Therefore it can be combined with some model for route
selection which assigns certain routes to each pedestrian.

0 p 0
-1,0

P P P
0,-1 0,0 0,1

0 P 0
1,0

Fig. 4. A particle, its possible directions of motion and the corresponding transition
probabilities Pi;.

3 Definition of the Model

CA are discrete in space, time and state variables. The area available for
pedestrians is divided into cells of approximately 40 x 40 crn 2 which is the
typical space occupied by a pedestrian in a dense crowd [30]. Each cell can
either be empty or occupied by exactly one particle (pedestrian). For special
situations it might be desirable to use a finer discretization, e.g. such that each
pedestrian occupies four cells instead of one, but we restrict to the simplest
case which is sufficient for the applications studied here.
We now define the stochastic dynamics of the model by specifying the
transition probabilities Pij for a motion to a neighbouring cell in direction3
2 In most cases motion in diagonal directions is not considered so that the matrix
of preference has at most five nonzero elements.
3 i,j E {-l,O, I}, see Fig. 4.
396 A. Schadschneider, A. Kirchner, K. Nishinari

(i,j). As explained in Sec. 2 the long-ranged interactions with other pedes-

trians and the surrounding is encoded in two fields, the dynamic Hoor field D
and the static Hoor field S, respectively. Furthermore a matrix of preference
M can be used to specify the walking direction (see Sec. 2), speed and Huc-
tuations for each individual. The transition probability Pij in direction (i,j)
is then determined by all three contributions. Explicitly it is given by
Pij = NekDD'ieksS'i M ij ( l - nij)eij. (2)

D ij and Sij are the strengths of the dynamic and static Hoor field at the target
cell and Mij is the matrix element of the matrix of preference for a motion
in the direction (i,j). N is a normalization factor to ensure '2: Ci,j)pij = 1
where the sum is over the possible target cells. The factor 1 - nij, where nij
is the occupation number of the neighbour cell in direction (i,j), takes into
account that transitions to occupied cells are forbidden. eij is a geometry
factor (obstacle number) which is 0 for forbidden cells (e.g. walls) and 1 else.
Finally, we have introduced two coupling constants kD and ks so that we can
vary the coupling strengths to each field individually.
The actual values of the parameters kD and ks depend on the situation
(see Sec. 5). A large ks implies that the pedestrians choose their path mainly
due to the surrounding without being distracted too much by other people.
A large coupling kD to the dynamic field, on the other hand, corresponds
to a strong herding behaviour. Here the pedestrian tries to follow the lead
of others, e.g. in the case of panics or insufficient knowledge about the sur-
roundings. The inHuence of these couplings will be discussed in more detail
in Sec. 5.
As mentioned before, the dynamic Hoor field is created by the motion
of the pedestrians and corresponds to a virtual trace. At t = 0 it is zero
everywhere. Whenever a particle moves from site (x,y) to one of its neigh-
bours (x + i, y + j), the field D zy at the origin cell is increased by one
(D zy -+ D zy + 1). Thus D zy has only non-negative integer values which
can be interpreted as the number of 'bosons' located at site (x,y).
The dynamic Hoor field is not only changed by the motion of the pedestri-
ans, but it is also subject to diffusion and decay which first leads a spreading
and dilution of the trace and finally to its vanishing after some time. Diffusion
and decay are controlled by two parameters 0: E [0,1] and 8 E [0,1]. In each
time step of the simulation each boson of the dynamic field D decays with thE;l
probability 8 and diffuses with the probability 0: to one of the neighbouring
cells.
The update rules of the full model including the interaction with the Hoor
fields then have the following structure:
1. The dynamic Hoor field D is modified according to its diffusion and decay
rules.
2. Using equation (2), for each pedestrian the transition probabilities Pij
for a move to an unoccupied neighbour cell (i,j) are determined by the
matrix of preference and the local dynamic and static Hoor fields.
 CA Simulation of Collective Phenomena in Pedestrian Dynamics 397

3. Each pedestrian chooses a target cell based on the probabilities of the

transition matrix P = (Pij).
4. The conflicts arising by any two or more pedestrians attempting to move
to the same target cell are resolved (see below).
5. The pedestrians which are allowed to move rocecute their step.
6. The pedestrians alter the dynamic floor field D zy of the cell (x, y) they
occupied before the move.
These rules have to be applied to all pedestrians at the same time (parallel
dynamics). This intro duces a timescale into the dynamies which can roughly
be identified with the reaction time t reac [23]. The existence of a timescale
in the dynamics of the model is essential if one wants to make quantitative
predictions for real processes. In the deterministic limit, corresponding to
the maximal possible walking velocity in our model, a single pedestrian (not
interacting with others) moves with a velo city of one cell per timestep, i.e.
40 crn per timestep. Empirically the average velo city of a pedestrian is about
1.3 m/ s [30]. This gives an estimate for the real time corresponding to one
timestep in our model of approximately 0.3 sec which is of the order of the
reaction time t reac and thus consistent with our microscopic rules. It also
agrees nicely with the time needed to reach the normal walking speed which
is about 0.5 sec.
One detail is worth mentioning. If a particle has moved in the previous
timestep the boson created then is not taken into account in the determina-
tion of the transition probability. This prevents that pedestrians get confused
by their own trace. One can even go a step further and introduce 'inertia'
[23] which enhances the transition probability in the previous direction of
motion.
Finally we like to comment on step 4, the resolution of conflicts which
occur if m ~ 2 particles choose the same destination cell in step 3. This is
relevant for high density situations. In order to avoid multiple occupancies of
cells only one particle is allowed to move while the others keep their position.
There are different ways to choose this particle. E.g. the winning particle
can be picked at random with probability I/rn or according to the relative
probabilities Pij with which each particle chose their target [23]. For the
problems studied here the details of the conflict resolution turned out to play
no important role. The influence of conflicts will be discussed in more detail
in Sec. 6.

4 Collective Phenomena
In this section we show that the model is indeed able to reproduce the col-
lective effects mentioned in Sec. 1.1. As most prominent example we want to
discuss lane formation out of a randomly distributed group of pedestrians.
This corresponds to a spontaneous breaking of the symmetry of the particle
number distribution in space. Simulations show that an even as well as an
398 A. Schadschneider, A. Kirchner, K. Nishinari

odd number of lanes may be formed. The latter corresponds to a spontaneous

breaking of the left-right symmetry of the system.

Wall.erModel
Wldll1

.
• •••
•• -. I .-. • . =. .
== I . -. : -. • •
• ••
•••• -
• •
Helghl
•
Rallo-Flnal
•• •• •• • • • •
Ratlo- B8gl0 f--____~'7 • • • ••• • • •••• •
1..-....·c·
7 • --..
. •• . .- •
· ..- .
. .8.
Ratio-Rate 0.00000
VBloCtty 1.000
S-Long 0.000 ~ ••• I • • • •• •••• • I
s- Trans 0.000
0810""-Tlmo L4_ _J /
~

..l • I •• • ••
• • • • ••
•
•

J-::-::-;-;:--,' . 11 ....... 1 •••••• .·:.1

c ~ •

,~__ 7 I I · • .. I•• I I·.... •••• i ....... I ••

1 • • • •••
7 . I I • c·. r·· ... · .1 1•
Floor-Oecay 0.005

FI.or-lmp,clZ~"7 I = I t··.·.. I •••• = =I • •••••

1·.·
Floor-lmpacl1 0.150
I
Oel. 60 ••• : ••••••
y ,J • .............. ;1_
Apply - ••• •• • • •
•

Fig. 5. Snapshot of a simulation of counterflow along a corridor. The left part shows
the parameter contro!. The central window is the corridor and the light and dark
squares are right- and left-moving pedestrians, respectively. The right part shows
the floar fields for the two species.

Fig. 5 shows simulations of a rectangular corridor which is populated

by two species of pedestrians moving in opposite directions. Parallel to the
direction of motion the existence of walls is assumed. Orthogonal to the
direction of motion both periodic and open boundary conditions have been
investigated. With periodic boundary conditions, the density of pedestrians
is fixed for each run. The numbers of left-and right-movers are equal and
each species interacts with its own dynamic floor field. For open boundaries,
we fix the rate at which pedestrians enter the system at the boundaries. The
pedestrians leave the system as soon as they reach the opposite end of the
corridor. Fig. 5 shows the graphical frontend running a simulation of a small
periodic system. Lanes have already formed in the lower part of the corridor
and can be spotted easily, both in the main window showing the cell contents
and the small windows on the right showing the floor field intensity for the
two species.
The formation of the lanes can also clearly be seen in the velocity profile
(Fig. 6) which has been measured at a cross section perpendicular to the
direction of flow. In a certain density regime, the lanes are metastable. Spon-
taneous fluctuations can disrupt the flow in one lane causing the pedestrians
 CA Simulation of Collective Phenomena in Pedestrian Dynamies 399

1
S
I'-
0.5

-0.5 '-L
'-L
-1
o 5 10 15 20 25

Fig. 6. Velocity profile of aperiodie system with p = 0.12 exhibiting lane formation.

to spread and interfere with other lanes. Eventually the system can run into
a jam by this mechanism.
Apart from lane formation we have also observed oscillations of the di-
rection of How at doors and the formation of roundabout-like How patterns
at intersections [23,25,29]. Therefore the model captures the main phenom-
ena correctly which is important for practical applications, e.g. evacuation
simulations or the optimization of escape routes.

5 Infiuence of the Floor Fields

In order to elucidate the influence of the coupling parameters ks and kD

we investigated an evacuation process in a simple geometry, namely a large
room with one door (see Fig. 1). In the following we do not use the matrix of
preference. All information about the location of the exits is obtained from
the Hoor fields. Fig. 7(a) shows the form of the static Hoor field S for such a
situation. For each lattice site S is calculated using some distance metric [26]
so that the field values increase in the direction of the door. Fig. 7(b) shows
a 3-d plot of the dynamic field D in the final stages of an evacuation process.

The value of ks, the coupling to the static field, can be viewed as a measure
of the knowledge of the pedestrians about the location of the exit. A large ks
implies a motion to the exit on the shortest possible path. For vanishing ks,
on the other hand, the people will perform a random walk and just find the
exit by chance. So the case ks « 1 is relevant for processes in dark or smoke:-
filled rooms where people do not have full knowledge about the location 01:
the exit.
The parameter kD for the coupling to the dynamic field controls the ten-
dency to follow the lead of others. A large value of kD implies a strong herding
behaviour which has been observed in the case of panics [10].
400 A. Schadschneider, A. Kirchner, K. Nishinari

Fig.7. Evacuation from a large room with a single door: (a) static floor field S;
(b) dynamic floor field D in the final stages of the evacuation.

In the following we consider a grid of size 63 X 63 sites with an exit of one

cell in the middle of one wall. The particles are initially distributed randomly
and try to leave the room. The only information they get is through the
floor fields. Fig. 1 shows typical stages of the dynamics for an initial particle
density of p = 0.3, which means 1116 particles. In the middle picture of
Fig. 1 a half-circle jamming configuration in front of the door is easy to spot.
A typical feature of the dynamics is a radial motion of 'holes' created by
particles escaping through the door.
Fig. 8(a) shows the evacuation times for fixed sensitivity parameter kD
of the dynamic field and variable sensitivity parameter ks of the static field.
For ks --t 0 the pedestrians do not sense the strength of the field. Therefore
they do not have any guidance through the surroundings and perform a pure
random walk which leads to a maximal value of the evacuation time for
k s = O. For ks --t 00 they have full information ab out the shortest distance
to the door and the evacuation time converges towards a minimal value. The
movement of the particles becomes almost deterministic. Therefore ks can be
interpreted as some kind of inverse temperature for the degree of information
ab out the inanimate surrounding of the pedestrians.
In the same way the sensitivity parameter kD of the dynamic field works
as an inverse temperature for the information ab out the virtual trace. If
ks is turned on from zero to infinity, a non-zero value of k D only means
additional noise to the pedestrians, the evacuation times are increased for
higher coupling strength to kD (see Fig. 8(a)).
Much more interesting is the behavior for fixed ks . In this case the evac-
uation times for various sensitivity parameters kD are shown in Fig. 8(b).
They converge to maximal values for growing sensitivity parameter kD of
the dynamic field. The most interesting point is the occurence of minimal
evacuation times for non-vanishing small values of the sensitivity parameter
kD of the dynamic field. This effect also depends on the value of the diffusion
 CA Simulation of Collective Phenomena in Pedestrian Dynamics 401

~.---------~--------~

5000

4000

'0000 3000

OO~~~::::~2::::~~~ -------------------- -------------

k8 kd

Fig. 8. Averaged evacuation times for a large room with an initial particle density
of p = 0.3 and ä = 0.3, Cl! = 0.3 for (a) fixed kD, and (b) fixed ks.

parameter Cl! and beeomes most pronouneed in the limit Cl! -+ O. Therefore a
small interaction with the dynamie field, whieh is proportional to the velo city-
density of the particles, is of advantage. It represents some sort of minimal
intelligenee of the pedestrians. They are able to deteet regions of higher loeal
flow and minimize their waiting times.
H the eoupling to the dynamic field is further inereased, the evaeuation
times inerease again from the small regime of the minimal times to maxi-
mal values. The interaction with other pedestrians beeomes more and more
unfavourable, one ean eompare this to the arising of a panie situation. The
weaker the eoupling to the statie field S is, the higher are the evacuation
times: the particles than have less information of the inanimate surrounding
(for example they eannot find the way to the door beeause of smoke in a fire
situation).
Three main regimes for the behaviour of the partieles ean bedistinguished
[26]. For strong eoupling to ks and very small eoupling to kD we find an
ordered regime where particles only react to the static floor field and the
behaviour than is in some sense deterministic. The disordered regime charac-
terized by strong eoupling to kD and weak eoupling to ks leads to a maximal
value of the evaeuation time. The behaviour here is typical for panic situa-
tions. Between these two regimes an optimal regime exists where the eombi-
nation of interaction with the statie and the dynamie floor fields minimizes
the evacuation time.

6 Friction Effects

In [23,26] the eonfiiets between pedestrians were solved in the following way:
when m > 1 particles share the same target eell, one (I E {l, ... , m}) is
chosen to move while its rivals for the same target keep their position. There
are two main ways to determine the moving particle I:
402 A. Schadschneider, A. Kirchner, K. Nishinari

1. According to the relative probabilities with which each particle chooses

its target cell, i.e. the probability for the l-th particle to move is :i(I)

~s=l
j (sj.
Pij

2. With equal probability, i.e. each particle moves with prob ability ~.
Both methods yield very similar results in the simulations.
We now extend the basic model by a new friction parameter J1, E [0,1] (see
[31] for details), in order to describe clogging effects between the pedestrians.
Whenever two or more pedestrians try to attempt to move to the same target
cell, the movement of all involved particles is denied with the probability J1"
i.e. all pedestrians remain at their site (see Fig. 9). This means that with
prob ability 1- J1, one of the individuals moves to the desired cello Which par-
ticle actually moves is then determined by one of the rules for the resolution
of conflicts described above. With this definition of J1, and the extended up-

•
•••
t
Fig. 9. Refused movement due to the friction parameter f." (for m = 4).

date rule it is easy to see that J1, works as some kind of local pressure between
the pedestrians. If J1, is high, the pedestrians handicap each other trying to
reach their desired target sites. This local effect can have enormous influence
on macroscopic quantities like flow and evacuation time [31]. Note that the
kind of friction introduced here only influences interacting particles, not the
average velo city of a freely moving pedestrian.
As one can see the use of a parallel update is essential. Any other form of
random- or ordered-sequential update will disguise the real number of arising
conflicts between the pedestrians in the system.
Fig. lO(a) shows the influence of the friction parameter on the evacuation
time T for the scenario described in Sec. 5. As expected, T is monotonically
increasing with J1,. The strongest effect can be observed in the ordered regime.
Here the evacuation time T is mainly determined by the clogging at the door.
For large values of J1" T increases strongly due to arching effects similar to
those in granular materials.
For fixed J1, and varying coupling strength ks a surprising result can be
observed (Fig. 10(b)). For J1, = 0 the evacuation time is monotonically de-
creasing with increasing ks since for large coupling to the static field the
 CA Simulation of Collective Phenomena in Pedestrian Dynamics 403

8-"""...
----
0

.9

10000

,,----
5000 - ~_------------
' - _ . _ - _ . _ ._ _. _ - - - - - - j

0.0 0.2 0.4 0.6 1.0

ks

Fig.l0. Dependence of evacuation times on the friction parameter p, (a) in the

ordered and disordered regimes and (b) as a function of ks for p = 0.3.

pedestrians will use the shortest way to the exit. For large Jl, however, T(ks)
shows a minimum at an intermediate coupling strength ks ~ 1. This is sim-
ilar to the faster-is-slower effect described in Sec. 1.1: Although a larger ks
leads to a larger effective velocity in the direction of the exit, it does not nec-
essarily imply smaller evacuation times if it increases the number of conflicts
elose to the door.

7 Conclusions

We have introduced a stochastic cellular automaton to simulate pedestrian

behaviour4 . The general idea in our model is similar to chemotaxis. However,
the pedestrians leave a virtual trace rather than a chemical one. This virtual
trace has its own dynamics (diffusion and decay) which e.g. restricts the in-
teraction range (in time). It is realized through a dynamic floor field which
allows to give the pedestrians only minimal intelligence and to use local in-
teractions. Together with the static floor field it offers the possibility to take
different effects into account in a unified way, e.g. the social forces between
the pedestrians or the geometry of the building.
The floor fields translate spatial long-ranged interactions into non-Iocal
interactions in time. The latter can be implemented much more efficiently
on a computer. Another advantage is an easier treatment of complex geome-
tries. We have shown that the approach is able to reproduce the fascinating
collective phenomena observed in pedestrian dynamics. As an example the
formation of lanes in counterflow has been discussed in Sec. 4.
In Sec. 5 we have investigated the influence of the coupling strengths ks
and kD to the static and dynamic floor fields on the dynamics, especially on
the evacuation time in a simple scenario. Interesting nontrivial phenomena
4 Further information and Java applets for the scenarios studied here can be found
on the webpage http://www.thp.uni-koeln.de/ ... as/as.html.
404 A. Schadschneider, A. Kirchner, K. Nishinari

could be observed, e.g. the nonmonotonie dependence of the evacuation time

on the coupling to the dynamic Hoor field. Further surprising effects are
caused by friction (Sec. 6), related to the resolution of conflict situations
where several individuals want to occupy the same space.
It is surprising to note that the properties of the model are in many
respects very similar to the social force model [9] although the interactions
are very different. In our approach pedestrians interact with the velocity-
density through an attractive coupling whereas the social-force model uses a
repulsive density-density interaction. It would be interesting to get a deeper
understanding of these similarities.

References
1. Wolf, D.E., Schreckenberg, M., Bachem, A. (Eds.): Traffic and Granular Flow
(World Scientific, Singapore 1996)
2. Schreckenberg, M., Wolf, D.E. (Eds.): Traffic and Granular Flow '97 (Springer
1998)
3. Helbing, D., Herrmann, H. J., Schreckenberg, Wolf, D. E. (Eds.): Traffic and
Granular Flow '99: Social, Traffic, and Granular Dynamics (Springer 2000)
4. Chowdhury, D., Santen, L., Schadschneider, A.: Statistical physics of vehicular
traflic and some related systems. Phys. Rep. 329, 199 (2000)
5. Nagel, K., Esser, J., Rickert, M.: Large-scale traflic simulation for transporta-
tion planning. Annu. Rev. Comp. Phys. 7, p. 151, ed. D. Stauffer (World Sci-
entific, 2000)
6. Helbing, D.: Traffic and related self-driven many-particle systems. Rev. Mod.
Phys. 73, 1067 (2001)
7. Schreckenberg, M., Sharma, S.D. (Ed.): Pedestrian and Evacuation Dynamies,
Springer 2001
8. Helbing, D., Farkas, 1., Molnar, P., Vicsek, T.: Simulation ofpedestrian crowds
in normal and evacuation situations. in [7], p. 21
9. Helbing, D., Molnar, P.: Social force model for pedestrian dynamics. Phys. Rev.
E51, 4282 (1995)
10. Helbing, D., Farkas, 1., Vicsek, T.: Simulating dynamical features of escape
panic. Nature 407,487 (2000)
11. Helbing, D., Farkas, 1., Vicsek, T.: Freezing by heating in a driven mesoscopic
system. Phys. Rev. Lett. 84, 1240 (2000)
12. Helbing, D., Schweitzer, F., Keltsch, J., Molnar, P.: Active walker model for
the formation ofhuman and animal trail systems. Phys. Rev. E56, 2527 (1997)
13. Hoogendoorn, S.P.: Walker behaviour modelling by differential games. These
proceedings.
14. Helbing, D., Keltsch, J., Molnar, P.: Modelling the evolution of human trail
systems. Nature 388, 47 (1997)
15. Chowdhury, D., Guttal, V., Schadschneider, A.: Cellular-automata model of
ant-trail and vehicular traffic: similarities and differences. cond-mat/0201207
16. Fukui, M., Ishibashi, Y.: Self-organized phase transitions in cellular automaton
models for pedestrians. J. Phys. Soc. Jpn. 68, 2861 (1999)
17. Muramatsu, M., Irie, T., Nagatani, T.: Jamming transition in pedestrian
counter flow. Physica A267, 487 (1999)
 CA Simulation of Collective Phenomena in Pedestrian Dynamics 405

18. Muramatsu, M., Nagatani, T.: Jamming transition in two-dimensional pedes-

trian trafIic. Physica A275, 281 (2000)
19. Klüpfel, H., Meyer-König, T., Wahle, J., Schreckenberg, M.: Microscopic sim-
ulation of evacuation processes on passenger ships. in Theory and Practical
[ssues on Cellular Automata, S. Bandini, T. Worsch (Eds.), Springer (2000)
20. Biham, 0., Middleton, A.A., Levine, D.: Self-organization and a dynamical
transition in traffic-flow models. Phys. Rev. A 46, R6124 (1992)
21. Gipps, P.G., Marksjös, B.: A micro-simulation model for pedestrian flows.
Math. and Comp. in Simulation 27, 95 (1985)
22. Bolay, K.: Nichtlineare Phänomene in einem fluid-dynamischen Verkehrsmodell.
Diploma Thesis, Stuttgart University (1998)
23. Burstedde, C., Klauck, K., Schadschneider, A., Zittartz, J.: Simulation ofpedes-
trian dynamics using a 2-dimensional cellular automaton. Physica A295, 507
(2001)
24. Schadschneider, A.: Cellular automaton approach to pedestrian dynamics -
Theory. in [7], p. 75
25. Burstedde, C., Kirchner, A., Klauck, K., Schadschneider, A., Zittartz, J.: Cel-
lular automaton approach to pedestrian dynamics - Applications. in [7], p. 87
26. Kirchner, A., Schadschneider, A.: Simulation of evacuation processes using a
bionics-inspired cellular automaton model for pedestrian dynamics. Physica A
(in press)
27. Ben-Jacob, E.: From snowflake formation to growth of bacterial colonies. Part
H. Cooperative formation of complex colonial patterns. Conternp. Phys. 38,
205 (1997)
28. Henderson, L.F.: The statistics of crowd fluids. Nature 229, 381 (1971)
29. Burstedde, C.: Simulation von Fußgängerverhalten mittels zweidimensionaler
zellulärer Automaten. Diploma Thesis, Universität zu Köln (2001); available
for download at http://www.burstedde.de/carsten/diplom. html
30. Weidmann, U.: Transporttechnik der Fussgänger. Schriftenreihe des IVT 80,
ETH Zürich (1992)
31. Kirchner, A., Nishinari, K., Schadschneider, A.: Friction effects and clogging in
a cellular automaton model for pedestrian dynamics. in preparation.
Modeling, Simulation and Observations for
Freeway Traffk and Pedestrian

Yuki Sugiyama1 * and Akihiro Nakayama2

1 City College of Mie, Tsu, Mie 514-0112, Japan
2 Gifu Keizai University, Ohgaki, Gifu 503-8550, Japan

Abstract. We perform the simulations of Optimal Velo city Model with a bottle-
neck and investigate the new aspects of this model under such boundary conditions
for the purpose of providing the interesting possibility of mathematical understand~
ing of "synchronized flow", and compare our results with the observation. We also
give the short review of extending OV model in 2-dimensional space for applying
the several kinds of phenomena of collective behaviors, such as granular flow and
pedestrian dynamics.

1 Introduction
The trafIic flow provides many interesting aspects of physics for collective
phenomena of non-equilibrium dynamics [1] [2]. Recently, the phenomenon
of trafIic flow in the bottleneck, so called "synchronized flow" affects many
researchers [3]. It is challenging problem to provide the mathematical un-
derstanding on the view point of non-equilibrium physics. In this paper, we
perform the simulations of Optimal Velo city Model with a bottleneck and
investigate the new aspects of this model under such boundary conditions for
the purpose of providing the interesting possibility of mathematical under-
standing of "synchronized flow".

2 Brief Review of OV Model (Traffic in Circuit)

Optimal Velo city Model is one of the successful models for describing the
dynamics of trafIic flow [4][5]. The model is formulated as

(1)

Xn is the position of the nth car, and L1x n = x n +1 - Xn is the headway. a is a

sensitivity constant. V(L1x n ), so called OV-function, determines the optimal
velo city according to the headway, which should have the form such as

V(L1X) = Vo{tanh(L1x - c) + const.}. (2)

* The present address is School of Informatics and Sciences, Nagoya University,
Nagoya 464-8601, Japan.

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 Freeway Trafik and Pedestrian 407

The model has the homogeneous flow solution; cars are moving uniformly
distributed with the same velo city V(l/k) , where k = NIL is the average
vehicle density in the case of circuit (N is the number of cars and L is
the length of the circuit). The homogeneous flow is linearly unstable under
the condition d~ V(X)IX=l/k > a/2 In this case, the jam flow is formed as
shown in Fig. 1. After relaxation all jam clusters are moving with the same
velo city opposed to the direction of the vehicle. When we take a different

200

150

§
.~ 100
.9

50

200 400 600 800 1000

time

Fig. 1. The formation of trafik jam. The hold line is the trajectory of a sampie car.

initial condition, we obtain a different figure of jam formation. But we always

recognize the universal profile of jam flow solution by the trajectory of vehicles
in the phase space of headway and velo city (L1x n , xn ) in Fig. 2. In the jam
flow solution all cars are moving along the specific closed curve, so called
"hysteresis loop" [4], which is a kind of limit cycle. The size of the loop
depends on a, not on k. As a becomes larger and exceeds the super-critical
value ac , the loop shrinks and disappears.
We show the fundamental diagram of OV model in Fig. 3, We have two Q-
k (the flow rate - the car density) relations corresponding to the homogeneous
flow and the jam flow solutions. We can observe the transition from the
homogeneous flow state to the jam flow state as in Fig. 3. In the region-I the
homogeneous flow is absolutely stable. The jam flow is absolutely stable in
the region-lI! defined by d~ V(X)IX=l/k > a12. In the region-I! (the cross-over
state), both flow-solutions are stable. The region-I! is defined by the size of
the hysteresis loop L1xF(a) - L1xJ(a) in Fig. 2.
408 Y. Sugiyama, A. Nakayama

>

i~

headwayAx

Fig. 2. The profile of jam :fI.ow solution. All vehicles move along the hysteresis loop
in the direction of arrow.
300 ....
homogenepus flow -
". sÖn~';';-~--
.........
250
....•.. :
200
~
3!
:ll 150
!
CI

S 100 .......
'..... ''
50
:......
11
i .... I

60 80 100 120
density k (vehic1eslkm)

Fig. 3. q-k diagram of OV model is drawn by the homogeneous :fI.ow solution and
limit cycle (jam :fI.ow) solution. The results of simulations are plotted together.

3 0 bservations in the Bottleneck

We first summarize some aspects of"synchronized fiow". In the original mean-

ing, the "synchronized fiow" is the phenomena that the oscillation of velocities
in trafiic fiow are synchronized between two parallelIanes in the existence of
bottleneck [6]. The data observed by Koshi indicates such phenomena as a
typical example [6][7]. Figure 4 shows the temporal sequential data for the
oscillation of average velocities in two parallelIanes at the three different
points (O.56km, 2.3km and 4.7km) in congested fiow on the upstream of the
bottleneck. It seems that the amplitude of oscillation is quite small at the im-
mediate upstream of the bottleneck. From Fig. 5, the correlation of velocities
between two lanes become clear as the upstream distance larger (than 1.2km).
 Freeway Trafik and Pedestrian 409

This data indicates the synchronization appears in the larger distance from
the bottleneck. As the correlation between two lanes become dear, the ampli-
tude of velocities become larger. Actually, the synchronized flow is followed
by jam flow in much larger distance. The synchronized flow occurs in some
intermediate distance from the bottleneck, between the small-amplitude flow
and the jam flow.

- _ Oller lane
•_. .. •• •• Inner Iarw

s~ F'roße of IndiVidual Lanes

(1 min. av&. speed)

Fig. 4. The temporal sequence of the velocity-changing of two lanes at three points
upstream from the bottleneck.

1.0 1.2

outet 18....
3a

. -,..
.. 30'
...\,t\ ••-.
..

,
N-5 N-54
r=0.31 r-O.71
-30

Speed ConwIation Betwwen Lanes

Fig. 5. The correlation of two lanes at three points upstream from the bottleneck.
410 Y. Sugiyama, A. Nakayama

4 OV Model with a Bottleneck

4.1 Modelling and Simulation
In this section, we perform the simulations of OV model with a bottleneck
in one-Iane trafIic. In the original meaning, the simulation of two-Iane traffic
is needed for observation of synchronization. But, we can suppose that one
of the most important aspects of "synchronized flow" is the structure of flow
on the upstream of the bottleneck, which consists of three distinct regions of
flow mentioned in the previous section. Such structure can be observed even
in one-Iane simulation with bottleneck. The characteristic behaviors of q-k
diagram for "synchronized flow" can be also investigated.
In our simulation of OV model, we define the bottleneck using OV function
as
Vbottle(.4x) = Vb{tanh(L1x - c) + const.} (3)
in the region of bottleneck, where the values of Vb and Vo in (2) are chosen
such that
Vbottle(L1X) < V(L1x), (4)
as in Fig. 6. This definition means the optimal velocity in a bottleneck (Here,
we suppose a tunnel as a bottleneck.) is reduced comparing with that outside
a bottleneck. This condition is reasonable from the view point of psycholog-

V(Ax)
------
-------V
V max
max
bottle

-+--~~---------~Ax
o
Fig. 6. The OV function in the bottleneck.

ical effect for drivers as weH as traffic regulations. The suppression rate of
bottleneck is measured by the ratio of the maximum velocities as
s - Vbottle (L1x -t 00) - v;,~~fe 1
bottle = V(L1x -t 00) = Vmaa: < . (5)

Here, we simply define Vb = Sbottle Vo in the bottleneck.

The simulation starts in the condition that vehic1es are homogeneously
moving on the upstream of the bottleneck and no vehic1es are on the down-
stream. The length of the bottleneck is 100m and the length of lane is enough
 Freeway Traffic and Pedestrian 411

long and the number of vehicles are enough not to suffer an artifact to the
behavior of trafIic flow around the bottleneck.

2000

1800

1600

1400

1200
U
Q)
~
Q)
1000
E
:.=
800

600

400

200

o
-4000 -3500 -3000 -2500 -2000 -1500 -1000 -500 o 500
location (m)

Fig.1. The spatio-temporal plot on the upstream of the bottleneck for Sbottle =
0.65, a = 1.6. The bottleneck is located from Om to 100m. Vehicles move in the
positive direction of location.

The figure 7 shows the formation of stable structure of trafIic flow around
the bottleneck. After relaxation three distinct spatial temporal patterns of
flow are formed on the upstream of the bottleneck. In the immediate upstream
the vehicles are distributed almost uniformly. The fine stripes are observed
in the further upstream region, which means the vehicle density is oscilla-
tory changing with small period in this region. The "oscillatory wave flow" is
followed by the jam flow, where jam clusters appear clearly. The difference
among these three patterns can be recognized more c1early in Fig. 8.
412 Y. Sugiyama, A. Nakayama

2000

U
Cl>
.e 1900
Cl>
§
1800
-4000 -3500 -3000 -2500 -2000 -1500 -1 000 -500 o 500
loeation (m)

Fig.8. The three stable patterns on the upstream of the bottleneck for Sbottle =
0.65, a= 1.6 after optimization.

4.2 The Structure of Flow UpstreaIIl of Bottleneck

Figure 9 shows the snapshots of velocity jheadway distributions correspond-

ing to Fig. 8. We can see the homogeneous flow with the identical velo city
and headway in the region immediate upstream of the bottleneck. In the
intermediate region 'the oscillatory wave' is observed for both velo city and
headway. In the final region the kink-like cluster appears, which implies stop-
and-go-wave as jam flow. Jam clusters propagates upstream, but the region
of the oscillatory wave is located near but in some distance of the bottleneck,
as you can see in Fig. 7.
Figure 10 is also the snapshots for the bottleneck of Sbottle = 0.7. In this
case the size of each region of three patterns changes. The region of homoge-
neous flow shrinks and the region for oscillatory wave is spread. These sizes
are determined by the strength of bottleneck. Moreover, the amplitudes of
velo city jheadway for the oscillatory waves are different between the two bot-
tlenecks. While, the amplitudes for the homogeneous flow and the jam flow
are originally characterized in the basic properties of OY model in circuit.
Actually, the amplitude of jam flow is the same as that on the circuit with
no bottleneck. In the contrast, the amplitude of the oscillatory wave is deter-
mined by the strength of the bottleneck, which is an additional element in
the system.
The sequence of these three patterns on the upstream of the bottleneck
can be qualitatively identified as the aspects of the flow accompanied with
synchronized fiow as mentioned in the previous section. The observed "syn-
chronized fiow" occurs in some intermediate distance from the bottleneck,
between the regions of the small-amplitude fiow and of the jam fiow. In our
simulation, the homogeneous fiow is identified as the small-amplitude fiow
in the observation. And, the oscillatory wave appears in some intermediate
distance from the bottleneck, which is followed by the jam fiow. The wave
has the characteristic amplitude, which is smaller than that of the jam flow.
It can be understood the fiow of oscillatory wave as the "synchronized fiow".
 Freeway Traffic and Pedestrian 413

i~l~~~ , il
-3500 -3000 -2500 -2000 -1500
Iocation (m)
-1000 -500 0

i~~: -3500 -3000 -2500 -2000 -1500

location (m)
-1000 -500 0

Fig. 9. The snapshots of velo city jheadway distribution on the upstream of the
bottleneck (the region between the dashed lines) of Sbottle = 0_65.

i~M1n~
-3500 -3000 -2500 -2000 -1500
loeation (m)
-1000 -500 0

60~i
-50 j'
~ 40 ij
~ 30 :
~CI)
20 ii
: :
.s::. 10 ii
o :i
-3500 -3000 -2500 -2000 -1500 -1000 -500 0
location (m)

Fig. 10. The snapshots of velocity jheadway distribution on the upstream of the
bottleneck of Sbottle = 0.7.

Figure 11 is the headway-velocity plots observed at four different distances

from the bottleneck in Fig. 10. Each figure has the vehicle-points within 100m
of the observation point in 5 minutes_ Figure lO(a) shows the transit-state
from the homogeneous fiow to the oscillatory wave.
Figure 10(b) shows the small closed loop as a clear trajectory in compari-
son with Fig. lO(a). The small closed loop can be understood as the profile of
the oscillatory wave. Within the interval from -500m to -1400m the profile
preserve its shape, that indicates the oscillatory wave is stable in this situa-
tion. This state is followed by the jam fiow solution through the transit-state
414 Y. Sugiyama, A. Nakayama

shown in Fig. lO(c). The profile of jam flow in Fig. lO(d) is just the same
as that on the circuit with no bottleneck in Fig. 2. The small closed loop in
Fig. lO(b) seems to be a kind of limit cycle solution as the jam flow solution
iso This small limit cyc1e have not observed in the simulation on the circuit,
that implies this solution is originally unstable. But this unstable solution
is convectively stabilized by the bottleneck. In this sense the small loop so-
lution can be called 'quasi-stable' solution which can be stabilized in some
boundary conditions such as a bottleneck.

(a) -400 m (b) -1000 m

30 ................ 30 ...................
,.

,/
; .........
......
25
.../'

o
o 5
"..
10
,,/'/
...".....
/'
,..ll

15 20 25 30 35 40 45
5

o ,----,-",,"'L.
o 5 10
0
-
!P
......../ ..
- - - ' - - - - ' ' - - - ' - -. . . . . . . . .- ' - - - - ' - - - '

15 20 25 30 35 40 45
headway (m) headway (m)

,"
(c) -1500 m (d) -3500 m

,~7/
30 30
,/
25 25 / /
/
i ,
;'
'*
0-
I /
:§. 20 20 /
:;:- 15 :;:- 15 I /
'g 'g / I
/ ,I I
~ 10 ~ 10
I / I
/ /
5 5
,/,,//
O'--.........""'----'----'--'--.........---L----I.---' OL-........-=~---'--'---'--'----L---'
o 5 10 15 20 25 30 35 40 45 o 5 10 15 20 25 30 35 40 45
headway(m) headway (m)

Fig. 11. The plots of vehicles-points in the phase space (headway, velocity) at four
points on the upstream from the bottleneck of Sbottle = 0.7.
 Freeway Traffic and Pedestrian 415

Figure. 12 is the plot of vehicle-points in q-k space corresponding to

Fig. 11. The plots of 1 min. average data corresponding to Figs. 10 (b) and
(d) in Fig. 12 are shown in Fig. 13. We take 1 min. average data every 10 m
point from -1000m to -1200m, and from -3500m to -37000m. We compare
the data with q-k relations for the homogeneous and jam flow solutions in
the system on the circuit in Fig. 3, where the relations are universal for OV
model. The data for -1000m-point are concentrate on the point between the
curve of the homogeneous flow solution and the line of jam flow solution. The
data for -3500m are distributed along the line of jam flow solution.

(a) -400 m (b) -1000 m

0.8 0.8

I I /\~
\
IrQ
\.
Q) 0.6 Q) 0.6

I\
TI TI
:;: : \ :c
! 0.4
Q)
.2:. 0.4

~ 0.2 I \. '.""." ...........

~
;:: 0.2 ,
i ,

....................., .

/
o L.---'--_-'------'--_-'------'-_
......
' ......
_ .."""
.••
..........._........ .
0'---'--"---'--"----'---......::"""
o 0.020.040.060.080.10.120.14 o 0.02 0.04 0.06 0.08 0.1 0 .12 0.14
density (vehicleslm) density (vehicleslm)

(c) -1500 m (d) -3500 m

0 .8 0.8
g
~ 0.6
~
I
Q) 0.6 ,

II
.2 TI
s: :c
!
Q)
Z. 004 0.4
l=
0
;::

0.2
0.2

0'---'-- "----'---'------'---......::""" 0L-~~~~~~~

o 0.020.04 0.06 0.08 0.1 0.120.14 o 0.020.04 0.06 0.08 0.1 0.120.14
density (vehicleslm) density (vehicleslm)

Fig.12. The plots of vehicles-points in q - k diagram (ßow, density) at four points

on the upstream from the bottleneck of Sbottle = 0.7. The dotted curve represents
the q - k relation of the homogeneous ßow solution.
416 Y. Sugiyama, A. Nakayama

More clearly presenting the dynamical structure in the sequence of ftow::-

patterns, we take 5 min. average data observed at several successive points
upstream of the bottleneck, and plot the data on the q-k space together with
the q-k diagram of homogeneous ftow and jam ftow solutions.

-1000 m -3500 m

0.8 0.8

I
~
0.6 I~ 0.6
.c
..," .........., .c
! 0.4 ! 0.4
.~.

~
0
"" 0.2
.......
,
.......
~
"" 0.2
........" .." ....
........ ............
........ . .......
0
0 0.02 0.04 0.06 0.08 0.1 0.120.14 0.02 0.04 0.06 0.08 0.1 0.120.14
density (vehicleslm) density (vehicleslm)

Fig.13. The plots of 1 min. averaged data in q - k diagram (ßow, density) at

two points on the upstream from the bottleneck of Sbottle = 0.7. The dashed curve
represents the q - k relation of homogeneous ßow solution, and the dotted line
represents that of jam ßow solution.

Figure 14 shows the data from Om to -3500m for S/Jottle = 0.7 and from
Om to -5000m for S/Jottle = 0.65. The data are taken along the sequence
of three stable patterns of ftow in Fig. 8. On the immediate upstream of
the bottleneck there exits the homogeneous flow , as the data points are
on the curve for the homogeneous flow solution. Originally, it is unstable
in the region for the density (the region III in Fig. 3). The homogeneous
flow is stabilized by the existence of the bottleneck. As the data is taken
in upstream direction through the oscillatory wave, the data points transit
from the homogeneous flow solution to the jam flow solution in q-k space. The
jam flow is stable in this region. In contrast, the homogeneous ftow and the
oscillatory wave flow are originally unstable, but they are stabilized by the
existence of the bottleneck. The data points of the upstream are aligned at
the same ftow rate with different densities in q-k space. The average density is
increasing as the distance is larger through the three patterns upstream of the
bottleneck. And this structure is convectively stabilized by the bottleneck.
 Freeway Trafik and Pedestrian 417

1 ,,
.,,,
,,,
,,
0.8 ,,
,,
,
U
Q)
,

~Q) 0.6 ~ 0.7

C3
:c .,+--0.65
-
• !
-!
Q) "}
> 0.4
3:
0
;;:::
0.2

o
o 0.02 0.04 0.06 0.08 0.1 0.12 0.14
density (vehicles/m)

Fig. 14. The plots of 5 min.- averaged data for successive observed points on the
upstream of the bottleneck. The solid curve and dotted line are the q-k relations for
the homogeneous fiow and jam fiow solutions, respectively. In the region between
dashed lines (the region III in Fig. 3), the homogeneous fiow is unstable, and the
jam fiow is stable.

4.3 Summary and Conjecture

We have found the new solution of OV model, which appears as the flow
of oscillatory wave on the upstream of the bottleneck. The amplitude of the
wave is rather smaller than that of jam flow solution, and the velocity of
the small cluster is different from the jam cluster. The properties of oscilla-
tory wave are manifestly different form the jam flow and the homogeneous
flow solutions. The flow of oscillatory wave is 'quasi-stable' solution, which
is originally unstable but convectively stabilized by bottleneck. The similar
oscillatory wave solution as ours, have been found by Mitarai and Nakanishi
on the analysis of OV model in open boundary with the existence of localized
perturbation [8]. They suggest the similar sequence of patterns on the flow
upstream of the localized noise, which sustains the structure of their flow.
The flow of oscillatory wave has small amplitude comparing with jam flow
and it is localized at some distance beyond the homogeneous flow upstream
of the bottleneck. In flow-density plot, the average density varies with the
418 Y. Sugiyama, A. Nakayama

same flow-rate. These properties indicate the possibility to understand the

"synchronized flow" as the "oscillatory wave" solution of OV model with bot-
tleneck. Our investigation in this paper is for one-Iane trafIic, but the above
properties can be preserved for two-Iane trafIic. We hope the synchronization
phenomena such as velocity changing for parallel two lanes, may be repro-
duced by two-lane simulation of OV model with the same situation studied
in this paper.

5 Modeling Pedestrians in 2-dimensional OV Models

OV model can be extended to 2-dimensional version for applying the phe-

nomena of several kinds of collective motions of interacting particles, such
as granular flow in liquid, pedestrians, evacuation dynamics and collective
bio-motions, etc. In principle, the two characteristics in OV model are in-
dependently appear in 2-dimensionj the following behavior by the attractive
force, and the exclusion behavior by the repulsive force.

attractive
.. _....._-.•.
~ ~

ropulsivo

ropulsivo

• -----_.-8-+
~
Fig. 15. The illustration of two behaviors in 2-d OV model.

In this paper we formulate the model expressed with the exclusive effect
by the repulsive force and apply the pedestrian flow.

d2 { N(n) d }
dt2Xn(t) = a Vce", - ~ [v mG", - V(r mn (t))]P(8mn (t))emn(t) - dtXn(t) . (6)

Vcez presents the free-moving in x- direction of the desired velo city for each
particle, which we set the same constant value for simplicity. The interaction
between two particles (m, n) are dependent on the distance and direction
as V(r mn ) and P(Omn). For OV-function Ver) we use the usual hyperbolic-
tangent type function. The anisotropy of interaction is introduced by choosing
P(Omn(t)) = cosOmn(t), because a pedestrian usually look forward in the
direction of movement. The effect of each particle (mth) surrounding nth
particle are collected.
 Freeway Traffic and Pedestrian 419

For the simulation of pedestrian flow in passage, we introduce two kinds of

particle, which have the different OV-function. In this case we can observe the
fast particles overtake the slow particles in uni-direction flow . Fig. 16 shows
the typical snapshots for different densities of pedestrians. In the low density

N=80

N=35 0

· .. ..., , ..·.. . , . · ·· . ...,

••
: ·. ·. ,·.. . ...,
• "
.
.. . .·· .·, .•.:. . .Ji. .. . ..
.: 0

'. a : .
0
, . 0• 0 '

....·· . ... .. .'....., ...... ...,......... ..:.. ···... ,.. . .. .'.

0
,f . 0 • 111. :- .

' ;
~ o • 111

.. :.
·. ..· ·.·· .'-,._.1."1_-... .·· .. . ... .·. . "··.··.· .. . ..'..
o' - o •

· '. . . .. , .
. . ... 11
' '
. -. Will .. .

0 • 0 0

I ...
I
;
e- ' f - - . aa-.A _ _ _. -.

N=600

I-.,. . .• ....- . •
~.:· III ·.·
....•..•.. a • • • \.
11
. . .. .... . .......
• • ••

,
..... ': :. · .··· 1 ...
. a •••- • • •. ...
•

-I
• .. •

11
11 11 ; .. •• • • ; • • • 11 •

•_. ".11 . •. .• :.: •••.. : .•• . ; i" . : .•.. . ·.' :·.t .· .. · I•

Ii'!! •• ., • •• •
.
• 11 11" • • ". ~ . .... a •

. ; ~ . ~ . -. ...... ::... ~,. _.I......

.. . -: .-:'•••
.. ....••. . • :~ .:.___~~!-: .. ~ , .
. ... ~ ••
•
•
-

I . a.
• • 11

• • : .,. . :
.... "
IIflj l l . . . .
111 1 ~ . •
:: · .. . 1
•
..
••
• •
•
:
•
.
I i.
: 'I .. .. . .. : . . ..... ,..
l •
• I' 6

_ ...
• •••

... .
11 e •• a • • • a

-:. 1 • • • • ~ • •: 11
•

. : •
•

• : eil • • • • 11 11 ell.1 11 11 . , 11 . ' ." . · . ·· 11 11 ... a

Fig. 16. The snapshot of simulation for pedestrians of uni-direction ßow: a = 0.5,
fast particles (light) and slow particles (dark).

the flows of slow particles and fast particles are sharply separated and each
flow moves homogeneously. In the intermediate density the bunching appears,
which is similar to the phenomena in trafik flow. This results in the formation
of the 'island' of jam cluster, in which shape is qualitatively different from the
case of granular. The jam of granular forms the 'band' perpendicular in the
direction of the flow, which is caused by the difference of angular-dependence
of sight, P((}mn) = 1 + COs(}mn(t). In the high density no structure appears
for too much particles.
Figure 17 shows the results of simulation for a counter flow of pedestri-
ans. In the low density several thin streams in the opposite directions are
420 Y. Sugiyama, A. Nakayama

formed. This is the specific properties for a counter flow of pedestrians. In

the intermediate density the streams become unstable and · grow wide and
move winding. But their stability are very sensitively dependent to initial
conditions. In the high density the jam islands are formed clearly, but several
particles find their way to escape from the jam island to free-moving region.

N = 150

..... . ... ..
N=200

I,, . . ...' .. .. , . . . . . .. .. •• 111 111 ••

.... . . :. ... .. .. . . .
. '
. . . . . . . . .... . ..
11
"
• " : : •
•
•
• 111 • •

..
.. 111 . : • • • • • .. • • •

• 111 .".

N=250

111 •
_. :••: .11 _. , : ....
. ;". .". " ..
••
:. ,.. . tI."
• • • • • ••• •
..
•• • I. : l1li • • -
• • .... .. • • 111

. :".: ..: ..... :.:.. . .

.. . .. . I.~ ·I .......··.:

• : .. . :l1li.:
. •~ 111 • •" •

Fig. 17. The snapshot of simulation for pedestrians of counter flow: a = 0.5, fast
particles (light) and slow particles (dark).

OV model is one of the successful model for describing the dynamics

of traffic flow. OV model can be extended to several directions. The 2-
dimensional OV models can develop to applications for the phenomena of
several kinds of cooperative transportation system. We hope the models can
provide the unified view of large dass of such systems.
 Freeway Traffic and Pedestrian 421

Acknowledgments

This work is partly supported by a Grant-in-Aid for Scientific Research (C)

(No.12650065) and (C) (No.13640409) ofthe Japanese Ministry ofEducation,
Science, Sports and Culture, and by Center for Integrated Research in Science
and Engineering in Nagoya University.

References
1. Chowdhury, D., Santen, L., Schadschneider, A.: Phys. Rep. 329 4-6, 199,
Statistical Physics of Vehicular Traffic and Some Related Systems (2000)
2. Helbing, D.: Rev. Mod. Phys.73, 1067, Traffic and related self-driven many
particle systems (2001)
3. Kerner, B. S., Rehborn, H.: Phys. Rev. E53, R4275; Kerner, B. S., Rehborn,
H., Aleksic, A.: in Traffic and Granular Flow '99, edited by Helbing, D., Her-
rmann, H. J., Schreckenberg, M. and Wolf, D. E. (Springer, Berlin 2000)
4. Bando, M., Hasebe, K, Nakayama, A., Shibata, A., Sugiyama, Y.: Japan J.
of Ind. and Appl. Math, 11, 203 (1994); Phys. Rev. E51, 1035 (1995); Bando,
M., Hasebe, K, Nakanishi, K, Nakayama, A., Shibata, A., Sugiyama, Y.: J.
Phys. I France 5, 1389 (1995).
5. Sugiyama, Y.: Wolf, D. E., Schreckenberg, M., Bachem, A.(Eds.), Trafiic and
Granular Flow'(World Scientific) 137 (1996); Schreckenberg, M., Wolf, D. E.
(Eds.), Traffic and Granular Flow '97 (Springer) 301 (1998); Computer Physics
Communications (in CCP99 conference) 121-122,399 (1999).
6. Koshi, M., Iwasaki, M., Ohkura, 1.: In Proceedings of the 8th International
Symposium on Transportation and Traffic Flow Theory, ed. by Hurdle, V.
F., Hauser, E., Stewart, G. N. (University of Toronto, Tronro, Ontario 1983)
ppA03
7. Koshi, M.: In Traffic and Granular Flow '01 , Fukui, M., Sugiyama, Y.,
Schreckenberg, M., Wolf, D. E.: (Eds.)(Springer) to be published.
8. Mitarai, N., Nakanishi, H.: Phys. Rev. Lett. 85, 1766-1769 (2000)
9. D. Helbing: Traffic and Granular Flow '97, M. Schreckenberg, D. E. Wolf
(Eds.) (Springer) 21 (1998); D. Helbing, I.J. Frakas, T. Vicsek: Traffic and
Granular Flow '99, Helbing, D., Herrmann, H. J., Schreckenberg, M., Wolf,
D. E.:(Eds.) (Springer)(1998) 245; Vicsek, T., Czir6k, A., Helbing, D.: 147
and references therein.
Testing Trame Flow Models

Elmar Brockfeld and Peter Wagner

German Aerospace Centre, Institute for Transportation

Research,Rutherfordstrasse 2, 12489 Berlin, Germany, URL: http://ivf.dlr.de

Abstract. The huge number of different trafIic flow models available ealls for an
attempt to c1assify, simplify and assess those models. In principle, this is simple to
aceomplish, sinee the principal algorithms for assigning quality measures such as
X2 are weIl known. However, due to a lack of public1y available data, a eommonly
agreed upon benchmark eould not be established up to now. This eontribution
shows (i) what ean be done with such a data-sets, onee they are available and (ii)
tell about efforts to change the far-from-optimal situation related to the data.

1 So Much Models, so Little Time

The current state of the art of modelling traffic flow is far from optimal.
There is a big market of different models of traffic flow that are thought to
describe traffic flow into more or less detail. They provide the playing ground
for theoreticians, each of which would like to convice the others, that (s)he
has gained the biggest understanding of them all, and that her jhis model de-
scribes the zoo of phenomena like jams, headway distributions, synchronized
flow, and even the stock-market [2].
It is true, that most of the models can be taken as a certain approximation
to the features observed in reallife, and that the models are in a certain kind
qualitatively correct. However, when it comes to more detailed questions it
starts to become difficult.
Not surprisingly, the result of the usual reviews of those models states
something like ''for this and that application, this model may be used", but
that's most ofthe time an excuse for being not exact. This small contribution
to computational physics tries to go a very small step ahead, by assigning
numbers for one given data-set and a number of different models.

2 Theory

Models may be classified into microscopic, mesoscopic and macroscopic, with

several subclasses (micro can be subdivided further into ODE, map, and cel-
lular automaton models, respectively). When it comes to comparison with
reality, most models are tested at most on the level of the so called fun-
damental diagram. With fundamental diagram the plot of the flow versus
density is meant, where both variables flow qi and density ki can be assigned

H. Emmerich et al. (eds.), Interface and Transport Dynamics

© Springer-Verlag Berlin Heidelberg 2003
 Testing Traffic Flow Models 423

in principle to an individual car. In this case, they are defined as:

1 LlXi
qi = - = - (1)
ri Vi
1
ki = - (2)
LlXi
i.e. the flow is defined either as the inverse of the time headway ri, or, al-
ternatively, as the quotient between front-bumper-to-front-bumper-distance
LlXi and the velo city Vi of a car. Unfortunately, most counting devices record
a temporal average of the velocity (v)T and the flow (q)T = LlN/T only,
where T is the averaging interval. In most cases the density can be inferred
only by using the questionable relationship (k)T = (q)T/(v)T. Any textbook
comes up with the much better formula (k)T = (q)T(~)T.
This is discussed at some length in order to make clear, that this did not
contain any idea about a specific model that may be hidden behind such a
plot.
Sometimes, the fundamental diagram is refered to as a model q = q( k) that
may serve as a fixed point model for an underlying microscopic car following
model. With fixed point it is meant, that if a car is following another car
that drives with constant speed V = Vlead, this following car will approach the
speed VJead and keep an optimal distance 9 = g(VJead). This correspondents
to a homogeneous situation on the road, where all cars drive with the same
speed and the same headway.
One of the interpretations of the fundamental diagram is shown in Fig. 1,
in this case there is an additional regime called synchronized flow, where cars
do not pick one particular fixed point. Instead, there are infinitely many
marginally stable states, for more details read the presentation of Boris
Kerner in this volume.
But compare this more or less theoretical fundamental diagram with an
empirical one. The scatter in the microscopic data is so big, that there didn't
seem any particular structure at all. Big question: can models applied suc-
cessfully to those data?

3 Testing Dynamical Models

Quite in general, any dynamical system can be written as:

(j(t) = J(q(t), t;p) + Lgii(q, t)ri(t) i = 1, ... ,n (3)

j

where the structure of the noise is given by

(ri(t) = 0, (ri(t)· rk(t + 1'» = 2dikJ(r)
with J(r) the noise correlation function, if J(r) = der) it is a white noise
process, and a coloured noise process otherwise. Here, the phase-space vari-
able q is given by q = {(Xi, Vi, . •. ) h=l, ... ,N, and the variable p denotes a set
424 E. Brockfeld, P. Wagner

2500

2000

€
....
......
1500
C'"
~ 1000
o
:;:::
500

O T------r-----.------.-----~--~~
o 0.2 0.4 0.6 0.8 1
occupancy p/Pjam [1]

Fig. 1. Theoretical fundamental diagram. Shown is the line of virtual homogeneous

states, and the illusive state of synchronized ßow.

140

120

100

I >
80

60

40

20
0 0.5 1.5 2 2.5 3
t " 1/q [8]

Fig. 2. Empirical fundamental diagram. It looks unfamiliar, since in this case the
speed is plotted versus the time-headway, which is the inverse ofthe fiow and can be
understood as a scaled distance (distance divided by speed, see 1). The advantage:
plotted here are only measured values, not computed ones like the density.

of parameters. (Of course, in case of a time-discrete dynamics anything can

be carried over by just replacing ci by the finite difference (q(t+h) -q(t)/h).)
 Testing Traflic Flow Models 425

In principle, any car i may have its own set of parameters, which is true but
would lead to a nightmare of parameters to determine. But note, that this
is not just an excuse: eertain parameters like the desired acceleration or the
maximum speed are definitely different from driver to driver, and even for
the same driver from time to time. Nevertheless, having a description such
as those above at hand is physieist's dream, however, in most instanees only
a small part of a system ean actually be observed:

Q(t) = G(q(t» + ro(t).

(Here, G(q(t» is some map from the high-dimensional state space of the
system into the low dimensional of the measurement, and ro(t) is the mea-
surement noise.) Not surprisingly, in many cases, knowing Q is not enough
to infer either the system's state q or it's dynamies J, g. This is especially
true for trafik, and obtaining really good ideas about J and 9 for traffie flow
eonstitutes an interesting research program in its own right.
Nevertheless, some work can still be done if a eomparatively bad model is
at hand, or even none at all, and this is what will be reported next. If there
is a model at hand, two things ean be done:
- use a Kaiman filter to obtain q and the set of free parameters p, or,
alternatively,
- fit the parameters p direetly to the data, by some more or less sophisti-
eated method.
In principle, it ean be tried to infer the model directly from the data, an
approach whose first results are reported in [1]. This is a different and very
interesting avenue of research, however this contribution will stay with the
simplest thing to be done, fitting the parameters direetly. (The Kaiman apo.
proach is a fairly eomplex one if it comes to testing microscopic models, which
are the main foeus of this eontribution.)
Suppose that a eertain data-set should be reproduced by a given model.
That means that an appropriate error measure, for instanee:

e = ..!!(I-=.;qe=m~p_----7q8=im::;I::...) (4)
(qemp)

should be minimized by searching for a set of parameters that minimizes e.

(In this equation, qemp are the observed values and qsim are the simulated
values. This measure gives an easy to understand error measure, the relative
absolute error.) The fitting itself ean be done by starting a simulation with
a given set of parameters, obtaining e and looking then for a new set of
parameters p' by means of one of the following algorithms:
- guessing (that is really done, and it is sometimes not as bad as it sounds),
- Nelder-Mead type algorithm (amoeba in Numerieal Recipes),
- simulated annealing techniques,
426 E. Brockfeld, P. Wagner

- MJD Powell's COBYLA (implemented in IBM's NAO),

or any other means of doing a non-linear optimization when derivatives are
not available. Of course, the business of doing such an optimization is an
interesting and still fertile field of research in its own right, which would
deserve another presentation as weH. For an quick idea, visit [13].

4 The Trouble with the Data

Let's go! (i) Getting some data, using one of the algorithms above, this con-
tribution will stick to NeIder Mead, and report the results. However, it turns
out that it is not that simple to obtain meaningful data-sets, and even if
one such data-set is at hand, sharing the data with others is not possible or
not allowed. In order to remedy this situation, we are going to record them
by our own devices and publish them on our homepage ivf.dlr.de/clearing.
Successively we will add (stay tuned!):
- single car data from various locations,
- data from car-following experiments,
- video-based data, and their extracted trafik data,
- probably some aerial data,
- FCD-Data, recorded in Berlin (NDA required!).
together with some tests we are going to perform with those data and a
number of models whose description is publicly available. Do you have a
good idea what to measure? Please tell! And: please tell what you have done
with the data. Watch out there for other goodies as weIl, e.g. the open source
trafik simulation project SUMO (www.sourceforge.org/sumo).

5 Daganzo's Data
Here is an example, and what to do with it. Some years ago, Carlos Daganzo
recorded a nice data-set [4] which will be used in the following to perform
model testing. Along a one-Iane road where overtaking was not possible,
eight observers were positioned, each of them equipped with a lap-top to
record the times a certain car passes him/her. One lead car that drove several
times along this road provided the initial car number, therefore trafik fiow
is recorded by so called N(t)-curves where N is the number of cars that
passed a certain position up to a certain time t. Typical recording errors are
of the order of 5 cars out of 600 by which the counts of the different observers
differed. In the following a data-set has been used that has been deliberately
corrected for those small mistakes, we will make it available on our homepage
as weIl. The time accuracy can be assumed as of the order of the one second
resolution that is reported in Daganzo's data. The Fig. 3 shows a sketch of
the recoding site, in the following some additional informations are given:
 Testing Traflic Flow Models 427
observer '2 observer .4 observer'6 observer .8

• ' I T.' I I'

observer '1 observer '3 observer '5 observer H

Fig. 3. Sketch of the Pablo Dam Road where the data used in this article have been
recorded.

- The totallength of the one-lane road is about 6.4 km,

- there is a trafik light at the end, controlling the flow into another highway,
- a speed-limit of about 25 mjs is imposed on this road,
- two days of data have been recorded, each of them from 7 to 9 am.
The trafIic light at the end causes a congestion, when demand rose during
the morning rush-hour. This can be seen most easily in the Fig. 4 where the
travel times on segment one and the last segment six are plotted. Only six
segments have been used, since the data from observer # 5 are partly missing,
so they were discarded alltogether in order to simplify the analysis.

1000

-.
__ "_"

Fig. 4. Travel time as function of the number of seconds that have passed since
starting of the record. It can be seen very clearly how the small jams created by
the traflic light travel backwards. Note however, that they merge more or less into
one big jam wave that starts to flatten out at the first observer.

6 The Contest
6.1 Building a Reference Model
How weH can models describe this situation? In order to have some model
for comparison, a very simple reference model will be constructed. Since the
data are from Carlos Daganzo, a variant of his cell transmission model will
be used [5]. This is basicaHy a very simple queueing model, which takes care
of continuity & capacity constraints. The model is specified further by:
428 E. Brockfeld, P. Wagner

- The road is divided into small cells of length 13 cars (100 m), where any
cell has a maximum occupancy (Ni), an actually occupancy ni and a
maximum flow out Crnax.
- Cars can pass to the next cell only if there is enough room, and the
number of cars that pass from one cell i to the next one i + 1 during
one time-step are computed as LlN(i -+ i+ 1) = min{vn,cmax ,w(Ni+1 -
ni+1}'
- The size of a time-step h is just the minimum travel time, Le. h =
Li/vmax •
In order to compare this model directly to the data, some more microscopic
details have been added. After some try and error, it converged to the fol-
lowing set of rules:
- Upon entering the cell, any car i gets assigned an assumed leaving time
t~xit = t1n + tmin, where tmin is the minimum travel time.
- Each cell is modelled as a short first-in-first-out queue, where the capacity
is implemented by noting that a flow is just an inverse time headway.
Therefore, cars can leave only if the car that has left the cell before is at
minimum T ~ l/cmax away, where Cmax is the maximum possible flow.
- The entry into the next cell can be delayed if space is scarce there. This
is being done by adding a penalty time before allowing to leave cell i that
is proportional to ni+l.
This cell transmission and its microscopic counterpart are well-known for the
fact that jams are not stable, fortunately they decay fairly slowly. In order
to render the jam-waves stable, the entry into the next cell has to be delayed
a little if n>ni+l> compared to the approach above [15].
In the presentation given at the comphys02-conference, a travel time func-
tion has been used that was a TIlllction of the number of cars in the cello
Surprisingly, this basically has no effect on the outcome of a simulation of
the queue-model, therefore it is not used anymore. It should be admitted,
that there still a better way to translate cell transmission into a microscopic
queueing model can be found, see for instance [14].
This simple model has four parameters Tmin, Cmax , a and ß, where a and
ß are just the proportionality constants in the computation of the penalty
term, Tpen = ani+l, the same for ß instead of a.

6.2 Technicalities
In order to actually run a simulation, some technical problems have to be
addressed. These are basically the boundary conditions, which are compli-
cated by the fact that the observers have not recorded the speeds. So, for the
inflow, velocities were set to the so called safe velo city of the corresponding
microscopic model, which may cause in principle travel times that are too
short. The outflow condition was modelIed at the position of observer # 8 by
 Testing Trafik Flow Models 429

a fictitious traffic light that is switched by the number of cars that have left
the road. The light is on red if n~~~) ~ n~:~p), and it switches to green once
this inequality is false. This forces the system to the correct boundary con-
dition based upon the number of cars, obviously, this is not the best possible
solution, since the models have to deal with the discontinuity of switching
the traffic light.
The reference model, it turns out, gives reasonable results. Best values
obtained by the NeIder Mead algorithm was about 18 % error for the travel
times between the observers, some more details are discussed in the figure
caption of Fig. 5.

... 0:: f:.::::c::··'·::::r::::::,··::::r:::::::::·:::r::.:::r:::::1

•
0.25 ~ ...... ~..... ····j·······~······t····(····~·······~······j
~ 0..2 1-.... +.... ; ; ; i .: i .:

i ~!··: IIIJ !. ,
1 2 3 4 5 6
1
segment numbfi

Fig. 5. Left figure: travel time as function of time for the simulation, compared to
the data. It can be seen, that even the simple cell transmission model fits the data
quite weIl. Right figure: The distribution of the errors for the different segments. It
can be seen, that the second segment has the largest error, a pattern that is robust
for all of the other models.

6.3 The Rest of the Pack

Interestingly, no model could do much better. Even the much more sophisti-
cated models did not yield an error rate substantially better than the 17 %
of the cell transmission model. Since a winner must be named, we do that
in the following, note however, that the differences to the second best model
is very small and may subject to change: optimization is sometimes more an
art than a science, and 1 % decrease in the error rate may be achieved e.g. by
making minor changes to the simulation set-up. The model that performed
best is, we are really sorry about that, our own. (Remember what we told in
the introduction!) It is a variant of the Gipps-model [12], invented some years
aga by Stefan Krauß [9]. The basic idea of the model is still fairly simple: in
order to avoid crashes, brake realistically, and try to keep the speed always
smaller than a certain maximum safe velocity Vsafe . If braking happens to
occur with constant deceleration, then Vsafe can be computed exactly:
(5)
430 E. Brockfeld, P. Wagner

400
dala -
SK -
CT -
350

300

.•
i: 250
E

1200
150

100

50
0 1000 2000 3000 .000 5000 5000 7000 8000
tim.(.)

Fig.6. Travel times for the SK-model below on the first segment, in comparison
with the data and the best values obtained by the cell transmission model. Note,
that there are very few differences in the performance of the two models.

Note, that this model is, like the CA-model, a kind of time-discrete optimum
velocity model (OVM), this time however with an optimal velocity that ad-
ditionally depends on the speed of the leading vehicle. It can be understood
as the special case of an optimal velocity model differential equation which
has been time-discretized with the (too big) time step h = 7:

7V = Vsafe - V (6)

Unfortunately, even this result came with a drawback: the optimal parameters
of the model found need a very large acceleration of around a ~ 1.8m/s2
which are not very realistic, and which are different from the parameters
Krauß has found [8].

6.4 Finally, the Rest (Preliminary Results!!!)

Below, the results of all the models tested so far are shown. We re-iterate, that
these results are preliminary, for the reasons stated above. The abbreviations
used in the plots are translated as follows: IDM = intelligent driver model [11],
CA = cellular automaton model [7], SK is the Gipps-like model mentioned
above, CT is the cell transmission model above, VDR = velo city dependent
randomization, a variant of the CA-model [16] with so called slow-to-start
rule, OVM = optimal velo city model [10] (see also [15] for the reference to
a similar model). We failed up to now to implement the recent model of
 Testing Traffic Flow Models 431

Kerner [3] without any bugs (which is by no means surprisingly, given its
enormous complexity), the same is true for the new CA-model with braking
lights [6]. The results will be reported elsewhere. Note, that our best value
obtained so far is 16.3 %, but note also that only one of us (EB) obtained
this result. By using a different program, the other author (PW) gets 16.9 %
error only for this model.

.g
.!
l!! 15 _ ....

CI) 10 _ ....

5-····

0-····

Fig.1. Comparison of the different models.

7 Conclusions & Further Plans

Physicists can measure certain numbers with an accuracy better than 10-12 ,
but in this case the models just display, that either there are no good models
actually available, or that there is an unexplainable rest of about 20 % in the
behaviour of the drivers that is completely random, or far beyond our current
description. It remains to be seen whether this is true for other data-sets and
other models as weIl. Note, that the situation provided here was a very simple
one, that favoured the ceIl transmission model. In other set-ups, it probably
isn't as good as in this application.
Another point is that we have to utilize the second day of data as weIl
just to discuss how weIl a given model is capable of generalizing. In order to
improve the models, a number of suggestions will close this contribution:

- add some psycho facts, e.g. about recognition, learning, comfort, plan-
ning,
- people are sometimes planning their behaviour,
432 E. Brockfeld, P. Wagner

- acceleration usually adopts on time-scales of 1. .. 2 second => use a-

models?,
- people drive sloppy, a phenomena that is hard to model.

Acknowledgements
We would like to thank Carlos Daganzo for putting this beautiful data-set
on his web-site. Additionally, we would like to thank Chris Cassir, Georg
Hertkorn, Boris Kerner, Kai Nagel, Michael Schreckenberg and Richard
Woesler for discussions, and the Federal German Ministry of Education and
Research for financial support.

References
1. Kriso, S., Friedrich, R., Peinke, J., Wagner, P.: Reconstruction of dynamic equa-
tions for trafiic flow, submitted to Physica A (2002).
2. Helbing, D., Batic, D., Schoenhof, M., and Treiber, M.: Modelling Widely Scat-
tered States in 'Synchronized' TrafIic Flow and Possible Relevance for Stock
Market Dynamics, http://de.arxiv.org/abs/cond-mat/0108548 (2001).
3. Kerner, B.S. and Osipov, X.: Cluster Effect in Initially Homogeneous Trafiic
Flow, 35 L31 - L43 (2002)
4. The data of Carlos Daganzo are publicly available on his web-site
www.ce.berkeley.edu;-daganzo.html
5. Daganzo, C. F.: The cell transmission model: a simple dynamical representation
of highway trafiic, Transp. Res. B 28 269 (1994).
6. Knospe, W., Santen, L., Schadschneider, A., and Schreckenberg, M.: Towards a
realistic microscopic description of highway traflic, J. Phys. A33, L477 (2000).
7. Nagel, K., and Schreckenberg, M.: J. Physique I 2 2221 (1992).
8. Krauß, S.: Microscopic modelling of traflic flow: Investigation of Collision Free
Vehicle Dynamics, University of Cologne, 1998.
9. Krauß, S., Wagner, P., and Gawron, C.: Metastable states in a microscopic model
of trafiic flow, Phys. Rev. E, 55 5597 - 5605 (1997).
10. Bando, M, Hasebe, K., Nakayama, A., Shibata, A., and Sugiyama, Y.: Dynam-
ical model of traffic congestion and numerical simulation, Phys. Rev. E, 51 1035
- 1042 (1995).
11. Treiber, M., Hennecke, A., Helbing D.: Derivation, Properties, and Simulation
of a Gas-Kinetic-Based, Non-Local Traflic Model, Physical Review E, 59 239 -
253 (1999).
12. Gipps, P. G.: A behavioural car following model for computer simulation,
Transp. Res. B 15 105 - 111 (1981).
13. Surf to http://plato.la.asu.edu/guide.html.
14. Heidemann, D.: A Queueing theory model of nonstationary traflic flow, Trans-
portation Science 35, 405 - 412 (2001).
15. Newell, G. F.: Theories of instability in dense highway traffie, J. Oper. Res. Soe.
Japan, 5, 9 - 54 (1962).
16. Barlovic, R., Santen, L., Schadschneider, A., and Schreckenberg, M.: Metastable
States in Cellular Automata for Traflic Flow, Europ. Journ. Phys. B 5, 793 (1998)
Editorial Policy
§1. Volumes in the following three categories will be published in LNCSE:
i) Research monographs
ii) Lecture and seminar notes
iii) Conference proceedings

Those considering a book which might be suitable for the series are strongly advised
to contact the publisher or the series editors at an early stage.
§2. Categories i) and ii). These categories will be emphasized by Lecture Notes in
Computational Science and Engineering. Submissions by interdisciplinary teams of
authors are encouraged. The goal is to report new developments - quickly, inform-
ally, and in a way that will make them accessible to non-specialists. In the evaluation
of submissions timeliness of the work is an important criterion. Texts should be well-
rounded, weIl-written and reasonably self-contained. In most cases the work will
contain results of others as weIl as those ofthe author(s). In each case the author(s)
should provide sufficient motivation, examples, and applications. In this respect,
Ph.D. theses will usually be deemed unsuitable for the Lecture Notes series. Proposals
for volumes in these categories should be submitted either to one of the series editors
or to Springer-Verlag, Heidelberg, and will be refereed. A provisional judgment on the
acceptability of a project can be based on partial information about the work:
a detailed outline describing the contents of each chapter, the estimated length, a
bibliography, and one or two sampIe chapters - or a first draft.A fmal decision whether
to accept will rest on an evaluation of the completed work which should include
- at least 100 pages of text;
- a table of contents;
- an informative introduction perhaps with some historical remarks which should be
accessible to readers unfamiliar with the topic treated;
- a subject index.
§3. Category iii). Conference proceedings will be considered for publication provided
that they are both of exceptional interest and devoted to a single topic. One (or more)
expert participants will act as the scientific editor(s) of the volume. They select the
papers which are suitable for inclusion and have them individually refereed as for a
journal. Papers not closely related to the central topic are to be excluded. Organizers
should contact Lecture Notes in Computational Science and Engineering at the
planning stage.
In exceptional cases some other multi-author-volumes may be considered in this
category.
§4. Format. Only works in English are considered. They should be submitted in
camera-ready form according to Springer-Verlag's specifications. Electronic material
can be included if appropriate. Please contact the publisher. Technical instructions
and/or TEX macros are available via hup://www.springer.de/math/authors/help-
momu.html. The macros can also be sent on request.
General Remarks
Lecture Notes are printed by photo-offset from the master-copy delivered in camera-
ready form by the authors. For this purpose Springer-Verlag provides technical
instructions for the preparation of manuscripts. See also Editorial Policy.
Careful preparation of manuscripts will help keep production time short and ensure
a satisfactory appearance of the fmished book.
The following terms and conditions hold:
Categories i), ii), and iii):
Authors receive 50 free copies of their book. No royalty is paid. Commitment to
publish is made by letter of intent rather than by signing a formal contract. Springer-
Verlag secures the copyright for each volume.
For conference proceedings, editors receive a total of 50 free copies of their volume for
distribution to the contributing authors.
All categories:
Authors are entitled to purchase further copies of their book and other Springer
mathematics books for their personal use, at a discount of 33,3 % directly from
Springer-Verlag.

Addresses:
Timothy J. Barth DirkRoose
NASA Ames Research Center Department of Computer Science
NAS Division Katholieke Universiteit Leuven
Moffett Field, CA 94035, USA Celestijnenlaan 200A
e-mail: [email protected] 3001 Leuven-Heverlee, Belgium
e-mail: [email protected]
Michael Griebel
Institut für Angewandte Mathematik Tamar Schlick
der Universität Bonn Department of Chemistry
Wegelerstr. 6 Courant Institute of Mathematical
D-53115 Bonn, Germany Sciences
e-mail: [email protected] NewYork University
and Howard Hughes Medical Institute
David E. Keyes
251 Mercer Street
Computer Science Department
New York, NY 10012, USA
Old Dominion University
e-mail: [email protected]
Norfolk, VA 23529-0162, USA
e-mail: [email protected] Springer-Verlag, Mathematics Editorial IV
Tiergartenstrasse 17
Risto M. Nieminen
D-69121 Heidelberg, Germany
Laboratory of Physics
Tel.: *49 (6221) 487-8185
Helsinki University of Technology
e-mail: [email protected]
02150 Espoo, Finland
http://www.springer.de/math/
e-mail: [email protected]
peters.html
Lecture Notes
in Computational Science
and Engineering

Vol.l D. Funaro, Spectral Elements for Transport-Dominated Equations. 1997.

X, 211 pp. Softcover. ISBN 3-540-62649-2
Vol. 2 H. P. Langtangen, Computational Partial Differential Equations. Numer-
ical Methods and Diffpack Programming. 1999. XXIII, 682 pp. Hardcover.
ISBN 3-540-65274-4
Vol. 3 W. Hackbusch, G. Wittum (eds.), Multigrid Methods V. Proceedings of
the Fifth European Multigrid Conference held in Stuttgart, Germany, October
1-4,1996. 1998. VIII, 334 pp. Softcover. ISBN 3-540-63133-X
Vol. 4 P. Deuflhard, J. Hermans, B. Leimkuhler, A. E. Mark, S. Reich, R. D. Skeel
(eds.), Computational Molecular Dynamics: Challenges, Methods, Ideas. Proceed-
ings of the 2nd International Symposium on Algorithms for Macromolecular Mod-
elling, Berlin, May 21-24, 1997. 1998. XI, 489 pp. Softcover. ISBN 3-540-63242-5
Vol. 5 D. Kröner, M. Ohlberger, C. Rohde (eds.), An Introduction to Recent Devel-
opments in Theory and Numerics for Conservation Laws. Proceedings of the Inter-
national School on Theory and Numerics for Conservation Laws, Freiburg / Litten-
weiler, October 20-24, 1997. 1998. VII, 285 pp. Softcover. ISBN 3-540-65081-4
Vol. 6 S. Turek, Efficient Solvers for Incompressible Flow Problems. An Algorith-
mic and Computational Approach. 1999. XVII, 352 pp, with CD-ROM. Hardcover.
ISBN 3-540-65433-X
Vol. 7 R. von Schwerin, Multi Body System SIMulation. Numerical Methods,
Algorithms, and Software. 1999. XX, 338 pp. Softcover. ISBN 3-540-65662-6
Vol. 8 H.-J. Bungartz, F. Durst, C. Zenger (eds.), High Performance Scientific and
EngineeringComputing. Proceedings ofthe International FORTWIHR Conference
on HPSEC, Munich, March 16-18, 1998. 1999. X, 471 pp. Softcover. 3-540-65730-4
Vol. 9 T. J. Barth, H. Deconinck (eds.), High-Order Methods for Computational
Physics. 1999. VII, 582 pp. Hardcover. 3-540-65893-9
Vo1.to H. P. Langtangen, A. M. Bruaset, E. Quak (eds.), Advances in Software
Tools for Scientific Computing. 2000. X, 357 pp. Softcover. 3-540-66557-9
Vol.u B. Cockburn, G. E. Karniadakis, c.-W. Shu (eds.), Discontinuous Galerkin
Methods. Theory, Computation and Applications. 2000. XI, 470 pp. Hardcover.
3-540-66787-3
Vol. 12 U. van Rienen, Numerical Methods in Computational Electrodynamics.
Linear Systems in Practical Applications. 2000. XIII, 375 pp. Softcover.
3-540-67629-5
Vol.13 B. Engquist, 1. Johnsson, M. Hammill, F. Short (eds.), Simulation and Vi-
sualization on the Grid. Parallelldatorcentrum Seventh Annual Conference, Stock-
holm, December 1999, Proceedings. 2000. XIII, 301 pp. Softcover. 3-540-67264-8
Vol.14 E. Dick, K. Riemslagh, J. Vierendeeis (eds.), Multigrid Methods VI. Pro-
ceedings of the Sixth European Multigrid Conference Held in Gent, Belgium,
September 27-30, 1999. 2000. IX, 293 pp. Softcover. 3-540-67157-9
Vol.15 A. Frommer, T. Lippert, B. Medeke, K. Schilling (eds.), Numerical Chal-
lenges in Lattice Quantum Chromodynamics. Joint Interdisciplinary Workshop of
John von Neumann Institute for Computing, Jülich and Institute of Applied Com-
puter Science, Wuppertal University, August 1999. 2000. VIII, 184 pp. Softcover.
3-540-67732-1
Vol. 16 J. Lang, Adaptive Multilevel Solution ofNonlinear Parabolic PDE Systems.
Theory, Algorithm, and Applications. 2001. XII, 157 pp. Softcover. 3-540-67900-6
Vol. 17 B.1. Wohlmuth, Discretization Methods and Iterative Solvers Based on Do-
main Decomposition. 2001. X, 197 pp. Softcover. 3-540-41083-X
Vol.18 U. van Rienen, M. Günther, D. Hecht (eds.), Scientific Computing in Elec-
trical Engineering. Proceedings of the 3rd International Workshop, August 20-23,
2000, Warne münde, Germany. 2001. XII, 428 pp. Softcover. 3-540-42173-4
Vol.19 I. Babuska, P. G. Ciarlet, T. Miyoshi (eds.), Mathematical Modeling and
Numerical Simulation in Continuum Mechanics. Proceedings of the International
Symposium on Mathematical Modeling and Numerical Simulation in Continuum
Mechanics, September 29 - October 3,2000, Yamaguchi, Japan. 2002. VIII, 301 pp.
Softcover. 3-540-42399-0
Vol. 20 T. J. Barth, T. Chan, R. Haimes (eds.), Multiscale andMultiresolution Meth-
ods. Theory and Applications. 2002. X, 389 pp. Softcover. 3-540-42420-2
Vol. 21 M. Breuer, F. Durst, C. Zenger (eds.), High Performance Scientific and En-
gineering Computing. Proceedings of the 3rd International FORTWIHR Confer-
ence on HP SEC, Erlangen, March 12-14, 2001. 2002. XIII, 408 pp. Softcover.
3-540-42946-8
Vol. 22 K. Urban, Wavelets in Numerical Simulation. Problem Adapted Construc-
tion and Applications. 2002. Xv, 181 pp. Softcover. 3-540-43055-5
Vol. 23 1. F. Pavarino, A. Toselli (eds.), Recent Developments in Domain Decom-
position Methods. 2002. XII, 243 pp. Softcover. 3-540-43413-5
Vol. 24 T. Schlick, H. H. Gan (eds.), Computational Methods for Macromolecules:
Challenges and Applications. Proceedings of the 3rd International Workshop on
Algorithms for Macromolecular Modeling, New York, October 12-14,2000. 2002.
IX, 504 pp. Softcover. 3-540-43756-8
Vol. 25 T. J. Barth, H. Deconinck (eds.), Error Estimation andAdaptive Discretiza-
tion Methods in Computational Fluid Dynamies. 2003. VII, 344 pp. Hardcover.
3-540-43758-4
Vol. 26 M. Griebel, M. A. Schweitzer (eds.), Meshfree Methods for Partial Differ-
ential Equations. 2003. IX, 466 pp. Softcover. 3-540-43891-2
Vol. 27 S. Müller, Adaptive Multiscale Schemes for Conservation Laws. 2003. XlV,
181 pp. Softcover. 3-540-44325-8
Vol. 28 C. Carstensen, S. Funken, W. Hackbusch, R. H. W. Hoppe, P. Monk (eds.),
Computational Electromagnetics. Proceedings of the GAMM Workshop on "Com-
putational Electromagnetics", Kiel, Germany, January 26-28,2001. 2003. X, 209 pp.
Softcover. 3-540-44392-4
Vol. 29 M. A. Schweitzer, A Parallel Multilevel Partition of Unity Method for EI-
liptic Partial Differential Equations. 2003. V, 194 pp. Softcover. 3-540-00351-7
Vol. 30 T. Biegler, O. Ghattas, M. Heinkenschloss, B. vanBloemen Waanders (eds.),
Large-Scale PDE-Constrained Optimization. 2003. approx. 344 pp. Softcover.
3-540-05045-0
Vol. 31 M. Ainsworth, P. Davies, D. Duncan, P. Martin, B. Rynne (eds.) Topics in
Computational Wave Propagation. Direct and Inverse Problems. 2003. VIII,
399 pp. Softcover. 3-540-00744-X
Vol. 32 H. Emmerich, B. Nestler, M. Schreckenberg (eds.) Interface and Transport
Dynamics. Computational Modelling. 2003. Xv, 432 pp. Hardcover. 3-540-40367-1
Vol. 33 H. P. Langtangen, A. Tveito (eds.) Advanced Topics in Computational Par-
tial Differential Equations. Numerical Methods and Diffpack Programming. 2003.
XIX, 659 pp. Softcover. 3-540-01438-1

Texts in Computational Science

and Engineering
Vol. 1 H. P. Langtangen, Computational Partial Differential Equations. Numeri-
cal Methods and Diffpack Programming. 2nd Edition 2003. XXVI, 855 pp. Hard-
cover. ISBN 3-540-43416-X
Vol. 2 A. Quarteroni, F. Saleri, Scientific Computing with MATLAB. 2003. IX,
257 pp. Hardcover. ISBN 3-540-44363-0

For further information on these books please have a look at our mathematics
catalogue at the following URL: http://www.springer.de/math/ index. html