Stochastic Interface problem for random internal

boundaries
N. Panda
Colorado State University
Collaborators: Don Estep, Clint Dawson, Troy Butler,
Lindley Graham, Simon Tavener,

Primary Collaborators
Measure-theoretic framework and analysis
Don Estep, Simon Tavener, and T.Butler

Applications
Clint Dawson, Lindley Graham

Software Development
Lindley Graham, Steve Mattis, and Scott Walsh (BET Development
Team)
Nishant Panda

N. Panda

UQ for random internal boundaries

2 / 38

Collaborative Research on Contaminant Transport

LANL

UQ

ls
de
Mo erics
m
Nu

For
m
An ulat
aly
sis ion

Application & UQ Goals

Model Development
V&V

CSU/UCD
Theoretical UQ Analysis
Methodology
Algorithms

N. Panda

UQ Algorithms
Error Analysis

UQ for random internal boundaries

UT-Austin
Numerical Implementation
Error Analysis
UQ

3 / 38

LANL Chromium Site(Mattis, Butler, Dawson, Estep

and Vesselinov

Vadose zone (~300 m)

Mortandad
Canyon

Sandia Canyon

Approximately 54,000
kg of Cr6+ released in
Sandia Canyon between
1956 and 1972
(uncertain).
Contaminant mass
distribution in the
subsurface is uncertain.

Single-screen aquifer
monitoring wells

Two-screen aquifer
monitoring wells

1000 ppb

Model predicted plume shape (~2012)

Cr6+ MCL 50 ppb

50 ppb

Current conceptual model for chromium migration

in the subsurface is supported by multiple lines of
N. Panda
UQ for random internal boundaries
evidence.

Contaminant source
location and mass flux
at the top of the regional
aquifer are unknown
due to complex 3D
pathways through the
vadose zone.
Limited remedial options
due to aquifer depth and
complexities in the
subsurface flow.
4 / 38

Inference for a deterministic model

Prediction Space

Observation Space

Predication
Functionals

eters
r Param

Fo rward Probl
em fo
r

Pr e

Solution Space

Physics Model

m fo
ble
Pro

dica
ti

on

rse
ve
In

Observation
Functionals

Space of Data and Parameters

N. Panda

UQ for random internal boundaries

5 / 38

The mathematical model

Ingredients
Compact parameter domain Rn
Model M (Y, ) with solution Y = Y () for
Quantities of interest (QoI) Q() = Q(Y ()) Rd
We assume that Q() is differentiable.
The specification of is critical and should be determined by physical
considerations.
The set of all possible QoI D = Q() Rd defines the observation
space.

N. Panda

UQ for random internal boundaries

6 / 38

The core deterministic inverse problem

Given a Q D, find with Q() = Q.

Q()

The solution is generally a set of values.

We call Q1 (Q()) a generalized contour.
N. Panda

UQ for random internal boundaries

7 / 38

Measure theory and inverse problems

The core deterministic inverse problem imposes significant structure

on the solution of the stochastic inverse problem.
Measure theory is designed to handle the set-valued inverses of a map
between measurable spaces in a natural way and is the basis for
rigorous probability theory.
Measure theory is ideally suited for the treatment of the stochastic
inverse problem for a deterministic model.

N. Panda

UQ for random internal boundaries

8 / 38

Measure theory ingredients

Measurable space
A specified domain X
A -algebra BX defining the collection of sets whose size can be
measured and the operations on those sets
(X, BX ) defines a measurable space.

Measure space
A procedure for computing the measure X of sets in the
algebra
(X, BX , X ) defines a measure space.

N. Panda

UQ for random internal boundaries

9 / 38

Measure theory ingredients

X
How do we compute the X -measure of this event?
N. Panda

UQ for random internal boundaries

9 / 38

Measure theory ingredients

X
We can approximate by using simpler sets from BX .
N. Panda

UQ for random internal boundaries

9 / 38

Measure theory ingredients

X
Only certain events in the approximation require refinement.
N. Panda

UQ for random internal boundaries

9 / 38

Inverse sensitivity analysis and generalized contours

The range of Q1 is not .

The range of Q1 is L whose individual points correspond to the
natural set-valued inverses of Q in that we call generalized contours.
Properties like well-posedness are posed in L not in .

From this perspective, the inability to distinguish between representors

in a set-valued solution is not ill-posedness.

N. Panda

UQ for random internal boundaries

10 / 38

Describing the space of contour manifolds

Theorem
There exists a transverse parametrization (TP) representing L in .
N. Panda

UQ for random internal boundaries

11 / 38

Solutions on (L, BL )
Theorem
A probability measure PD on (D, BD ) corresponds to a unique
probability measure PL on (L, BL ).

L
N. Panda

UQ for random internal boundaries

12 / 38

Natural but not desirable

The inverse solution on (L, BL ) requires minimal assumptions.
But . . . the physically meaningful space is .
The ideal inferential target is a probability measure P on (, B ).
To construct P from PL requires answering two (related) questions.
How are events in B related to events in BL ?
I

Ans: Use a transverse product -algebra.

How are measures on (, B ) related to measures on (L, BL )?

I

Ans: Use the Disintegration Theorem.

N. Panda

UQ for random internal boundaries

13 / 38

Structure of measures on
The Disintegration Theorem implies that any P can be written
uniquely as the product of a marginal PL on a TP and conditional
probabilities {P` } on contours {C` } for ` L.
The Disintegration Theorem is like Fubinis theorem where P (A) is
written as an iterated integral involving P` and PL .

Theorem
Specifying P` on generalized contours corresponding to ` L
determines a unique probability measure on (, B ).

N. Panda

UQ for random internal boundaries

14 / 38

The Standard Ansatz

liz

ed

Co
nto
urs

ty A
Uniform Densi

G
lon g

en

er

The computational algorithm and BET code can treat any measure in
the Ansatz or work directly with the contour events.
N. Panda

UQ for random internal boundaries

15 / 38

Solution of the example under the Ansatz

Inverse density on L

N. Panda

UQ for random internal boundaries

Inverse density on

16 / 38

Simple function approximations of measures

Theorem
P can be approximated using simple functions.

P is approximated on a partition of taken as a subset of the

generating sets to B , BL , and {BC` , ` L}.
This results in a direct discretization of the iterated integral in the
Disintegration Theorem.

N. Panda

UQ for random internal boundaries

17 / 38

Approximations of events with random or regular

sampling

N. Panda

UQ for random internal boundaries

18 / 38

General measure-theoretic algorithmic outline

Step 1: Define Voronoi tessellation {Vj }N
j=1 of used to approximate
events in both B and C .
Step 2: Define approximation to PD (or its density) on (D, BD ) by
computation of PD (Ii ) for tessellation {Ii }M
i=1 of D.
Step 3: Use Ansatz to compute P (Vj ).
Step 4: Compute P (A) for events of interest A B using P (Vj ).
We let P,(N,M ) denote the approximate measure.

N. Panda

UQ for random internal boundaries

19 / 38

Applying the algorithm/identifying errors

1
9
8
7
6
5
4
3
2
1
0

For A B with (A) = 0, we estimate P (A).

The error in the -volume of a Voronoi coverage of Q1 (Ii ) affects P estimation.

N. Panda

UQ for random internal boundaries

20 / 38

Applying the algorithm/identifying errors

1

Theorem
As N, M , P,(N,M ) converges to P .

N. Panda

UQ for random internal boundaries

20 / 38

Multiple QoI
How does the geometric relationship between multiple QoI affect the
solution of the stochastic inverse problem at a specified sample size?

Geometrically distinct QoI

The component maps of Q are geometrically distinct (GD) if the
Jacobian of Q is full rank at every point in .

Theorem
If the component maps of Q are GD, then the generalized contours
exist as lower-dimensional manifolds and a TP exists as a
higher-dimensional manifold.

N. Panda

UQ for random internal boundaries

21 / 38

The geometry matters

We show the effect of maintaining or losing orthogonality in the
contours of two linear QoI component maps.
As contour events become more skewed, we observe significant
changes in the solution (accuracy and precision are lost).

P for different linear q1 () and q2 ()

maintaining orthogonality of the contours.
N. Panda

P for fixed q1 () and varying q2 () where at

one point q2 () = q1 ().

UQ for random internal boundaries

22 / 38

Quantifying the geometry of a QoI map

Different QoI maps may lead to

1

Different levels of accuracy in approximating P ;

Different -measures (i.e., sizes) of high probability events.

Quantifying the skewness is the first step in analysis of QoI.

We connect skewness to the Jacobian of Q restricted to L, JQ|L .

N. Panda

UQ for random internal boundaries

23 / 38

From accuracy to precision: -measures of Q1 (B)

We want to design experiments that both

lead to accurate quantification of uncertainties, and
reduce uncertainties.

By reducing uncertainties, we mean that high probability events in D

correspond to events in with small -measure.

N. Panda

UQ for random internal boundaries

24 / 38

 -measure of contour events: Linear example

Consider square invertible matrix J, and let Q() = J.
Standard linear algebra results give
(Q1 (B)) = (J 1 (B)) = D (B)| det(J 1 )| =

D (B)
D (B)
= Qn
.
| det J|
k=1 k

(Q1 (B)) = 0.0769

D (B) = 0.04

N. Panda

UQ for random internal boundaries

25 / 38

 -measures of contour events for nonlinear maps

When Q is nonlinear, different choices of B may lead drastically
different values of (Q1 (B)).
D

N. Panda

UQ for random internal boundaries

26 / 38

 -measures of inverse events for nonlinear maps

The value of (Q1 (B)) depends on the sensitivity of Q near values

of that map into B, i.e., on JQ|L ().
(Q1 (B)) (J1
(i) (B)) =

D (B)
D (B)
= Qn
.
| det J(i) |
k=1 ik

Q
We let M (JQ|L ()) = ( nk=1 ik )1 denote how the measure of an
event B DB is scaled by Q1 locally in .

N. Panda

UQ for random internal boundaries

27 / 38

Setting up the random interface problem

A motivating example comes from soil characterization research soil
dynamics
paper http://dx.doi.org/10.1016/j.soildyn.2016.04.010

N. Panda

UQ for random internal boundaries

28 / 38

Setting up the random interface problem

1

The boundaries separating the various the various rock types are
uncertain.

Seismic imaging gives a reference interface based on imaging

stress waves.

N. Panda

UQ for random internal boundaries

29 / 38

A simple scenario
T = (xT , yT )

T ype 1

T ype 2

AL

AR

B = (xB , yB )

Figure: A single interface problem

N. Panda

UQ for random internal boundaries

30 / 38

A simple scenario

Two distinct physical types are separated by a single linear

interface which is uncertain

The interface, a line in this simple case, is determined completely

by the two parameters B , T .

Having some knowledege of the domain of B and T , the goal is

to find the most probable configuration of the interface.

So, how do we specify a probability model?

N. Panda

UQ for random internal boundaries

31 / 38

Some other ideal scenario

c

TYPE 2

Figure: A polygonal region

we have two distinct parameter types where the second type is

described by a rhombus. In this case, we need 4 points to completely
specify the region.

N. Panda

UQ for random internal boundaries

32 / 38

Some other ideal scenarios

TYPE 2

Figure: A polygonal region where TYPE 2 is inside TYPE 1.

we have two distinct parameter types where the second type is

embedded in the first type. In this case, we need 4 points in R2 to
completely specify the region.

N. Panda

UQ for random internal boundaries

33 / 38

Motivation for a probabilistic model

The polygonal regions are completely specified by the boundary

points.

Uncertainty on the ranges of boundary points correspond to

uncertainty in the interface

The space of all boundary points corresponds naturally to the

physical space of polygonal interface.

N. Panda

UQ for random internal boundaries

34 / 38

The probabilistic model

N. Panda

UQ for random internal boundaries

35 / 38

The probabilistic model

Choose k points from a parametrizing set S.

This defines a subset k R(d1)k .

Each point in k corresponds to a piecewise linear interface

function.

k corresponds to the physical space of polygonal interface Vk .

Now we can put a measure structure on k (k , Bk.k , L )

k

N. Panda

UQ for random internal boundaries

36 / 38

Forward and Inverse Stochastic Problem for random

interface

N. Panda

UQ for random internal boundaries

37 / 38

The BET Python Package

Open-source package for measure-theoretic inversion:

B
E
T

https://github.com/UT-CHG/BET
Documentation includes several introductory examples:
http://ut-chg.github.io/BET/

Main Development Team

Lindley Graham (FSU), Steve Mattis (UT-Austin), Scott Walsh (UCD),
and Troy Butler (UCD)
Questions?

N. Panda

UQ for random internal boundaries

38 / 38