Probabilistic Elasto Plasticity and Its Applications in Finite Element Simulations of Stochastic Elastic-Plastic Boundary Value Problems

Probabilistic ElastoPlasticity and Its Application in Finite Element
Simulations of Stochastic ElasticPlastic Boundary Value Problems

By
KALLOL SETT
Bachelor of Civil Engineering (Jadavpur University, Calcutta, India) 1997
Master of Science in Civil Engineering (University of Houston, Houston, Texas) 2003
DISSERTATION
Submitted in partial satisfaction of the requirements for the degree of
DOCTOR OF PHILOSOPHY
in
Civil Engineering
in the
OFFICE OF GRADUATE STUDIES
of the
UNIVERSITY OF CALIFORNIA
DAVIS
Approved:
Committee in Charge
2007
-i-
Abstract
Probabilistic ElastoPlasticity and Its Application in Finite Element Simulations of

Stochastic ElasticPlastic Boundary Value Problems
by
Kallol Sett
Doctor of Philosophy in Civil Engineering
University of California, Davis
Professor Boris Jeremic, Chair
A computational framework has been developed for simulations of the behaviors of solids and
structures made of stochastic elasticplastic materials. Particular emphasis has been given
to soil, a highly nonlinear (elasticplastic) and highly uncertain material, and geotechnical
engineering applications.
Uncertain elasticplastic material properties are modeled as random fields, which,
in the governing partial differential equation of mechanics, appear as the coefficient term. A
spectral stochastic elasticplastic finite element method with FokkerPlanckKolmogorov
equation approach based probabilistic constitutive integrator is proposed for solution of this
nonlinear (elasticplastic) partial differential equation with stochastic coefficient. To this
end, the random field material properties are discretized, in both spatial and stochastic
dimension, into finite numbers of independent basic random variables, using Karhunen
-ii-
Loeve expansion. Those random variables are then propagated through the elasticplastic
constitutive rate equation using FokkerPlanck-Kolmogorov equation approach, to obtain
the evolutionary material properties, as the material plastifies. The unknown displacement
(solution) random field is then assembled, as a function of known basic random variables
with unknown deterministic coefficients, using polynomial chaos expansion. The unknown
deterministic coefficients of polynomial chaos expansion are obtained, by minimizing the
error of finite representation, by Galerkin technique.
The applicability of the developed methodology is demonstrated in obtaining the
probabilistic solutions of 1D static pushover test and response of 1D structure due to
sinusoidal base displacement. In addition, pure constitutive level simulations of von Mises,
Drucker-Prager, and Cam Clay material models are also shown. The results are verified
with analytical solution, where available, or Monte Carlo solution and good agreements
are obtained. Finally, the complete solution process, based on the developed computational framework, of a geotechnical engineering problem - seismic wave propagation through
stochastic elasticplastic soil - is illustrated using real-life soil data and with real earthquake
motion.
Professor Boris Jeremic

Dissertation Committee Chair
-iii-
To my wife
-iv-
Contents
List of Figures
ix
List of Tables
xvi
Acknowledgements
xvii
Motivation and Theoretical Background
1 Introduction
1.1 Hypothesis . . . . . .
1.2 Scope of Study . . . .
1.3 Summary of Contents
1.4 Original Features . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2 Probability Theory Background

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Basic Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Properties of Single Random Variable . . . . . . . . . . . . . . . . . . . . .
2.3.1 Probability Distribution Function . . . . . . . . . . . . . . . . . . . .
2.3.2 Cumulative Distribution Function . . . . . . . . . . . . . . . . . . .
2.3.3 Moments of Probability Distributions . . . . . . . . . . . . . . . . .
2.3.4 Characteristic Function and its relation with Moments and Cumulants
2.3.5 Relation between Moments and Cumulants . . . . . . . . . . . . . .
2.4 Properties of Two or More Random Variables . . . . . . . . . . . . . . . . .
2.4.1 Joint and Marginal Probability Distribution . . . . . . . . . . . . . .
2.4.2 Joint Cumulative Probability Distribution . . . . . . . . . . . . . . .
2.4.3 Conditional Probability Distribution . . . . . . . . . . . . . . . . . .
2.4.4 Dependency and Correlation between Random Variables . . . . . . .
2.4.5 Joint Characteristic Function . . . . . . . . . . . . . . . . . . . . . .
2.4.6 Some useful Properties of Two Random Variables . . . . . . . . . . .
2.5 Random Processes and Fields . . . . . . . . . . . . . . . . . . . . . . . . . .
-v-
2
2
8
9
10
11
11
11
13
13
15
15
18
20
21
21
26
27
28
30
30
32
2.6
II
2.5.1 Classification of Random Processes/Fields . . . .

2.5.2 Ergodicity . . . . . . . . . . . . . . . . . . . . . .
Stochastic Calculus . . . . . . . . . . . . . . . . . . . . .
2.6.1 Mean Square Convergence of a Random Function
2.6.2 Mean Square Derivatives . . . . . . . . . . . . .
2.6.3 Riemanian Stochastic Integral . . . . . . . . . . .
2.6.4 RiemannStieltjes Stochastic Integral . . . . . .
2.6.5 It
o Stochastic Differential Equation . . . . . . .
2.6.6 FokkerPlanckKolmogorov Equation . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Uncertain and Spatially Uncertain Material Properties
35
40
41
43
44
45
47
47
49
50
3 Characterization & Quantification of Uncertainties in Material Properties

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Classification of Uncertainties . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 Uncertainties in Soil Properties . . . . . . . . . . . . . . . . . . . . . . . . .
3.4 Uncertain Spatial Variability . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5 Probabilistic Geotechnical Site Characterization . . . . . . . . . . . . . . . .
51
51
52
53
58
61
4 Random Field Modeling of Uncertain Material Properties

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Basic Concept . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.1 Finite Scale Model . . . . . . . . . . . . . . . . . . . .
4.2.2 Fractal Model . . . . . . . . . . . . . . . . . . . . . . .
4.3 Example Estimation of Statistical Parameters . . . . . . . . .
64
64
65
68
71
72
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
III Material (Constitutive) Level Stochastic Simulation: Probabilistic

ElastoPlasticity
82
5 Probabilistic ElastoPlasticity: Theory
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 One Dimensional Development . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.1 Specialization of General Formulation to Particular Constitutive Laws
5.2.2 Solution Method of Probabilistic ElastoPlastic Equation . . . . . .
5.3 Three Dimensional Development . . . . . . . . . . . . . . . . . . . . . . . .
83
83
87
95
107
110
6 Probabilistic ElastoPlasticity: Numerical Examples and Verifications 116

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.2 Linear Elastic Shear Constitutive Behavior . . . . . . . . . . . . . . . . . . 119
6.3 ElasticPlastic Shear Constitutive Behavior with Mean Yield Criteria . . . 125
6.3.1 Drucker-Prager Associative Model . . . . . . . . . . . . . . . . . . . 126
6.3.2 Cam Clay Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
-vi-
6.4
ElasticPlastic Shear Constitutive Behavior with Probabilistic Yield Criteria 147

6.4.1 von Mises Associative Model . . . . . . . . . . . . . . . . . . . . . . 150
6.4.2 Drucker-Prager Associative Model . . . . . . . . . . . . . . . . . . . 155
IV Stochastic Simulations of Solids and Structures with Uncertain

Material Properties: Stochastic Finite Elements
159
7 Stochastic Finite Elements: Theory
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.2 Discretization of Governing Stochastic Partial Differential Equation .
7.2.1 Stochastic Discretization of Input Random Field . . . . . . .
7.2.2 Stochastic Discretization of Unknown Solution Random Field
7.2.3 Spatial Discretization of the Differential Operator . . . . . .
7.3 Stochastic Finite Element Formulation . . . . . . . . . . . . . . . . .
7.3.1 Non-Linear (ElasticPlastic) Formulation . . . . . . . . . . .
7.3.2 Dynamic Formulation . . . . . . . . . . . . . . . . . . . . . .
7.4 Post-Processor: Estimation of Response Statistics . . . . . . . . . . .
7.4.1 Mean and Autocovariance . . . . . . . . . . . . . . . . . . . .
7.4.2 Probability Density Function . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
8 Stochastic Finite Elements: Numerical Examples and Verifications

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2 One Dimensional Static Problem . . . . . . . . . . . . . . . . . . . . . .
8.2.1 Problem Statement and Formulation . . . . . . . . . . . . . . . .
8.2.2 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . .
8.3 One Dimensional Dynamic Problem . . . . . . . . . . . . . . . . . . . .
8.3.1 Problem Statement and Formulation . . . . . . . . . . . . . . . .
8.3.2 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . .
8.4 Seismic Wave Propagation through Stochastic ElasticPlastic Soil . . .
8.4.1 Problem Parameters . . . . . . . . . . . . . . . . . . . . . . . . .
8.4.2 Simulation Results and Discussions . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
160
160
164
165
174
178
179
182
187
189
190
191
.
.
.
.
.
.
.
.
.
.
194
194
195
195
203
211
211
215
217
219
221
Conclusions and Future Research Directions

229
Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
Future Research Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
Bibliography
241
Appendices
255
256
A.1 Derivation of Akl (Eq. (5.26)) . . . . . . . . . . . . . . . . . . . . . . . . . . 256
A.2 Derivation of B (Eq. (5.27)) . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
A.3 Derivation of KP (Eq. (5.28)) . . . . . . . . . . . . . . . . . . . . . . . . . . 261
-vii-
A.4 Derivation of Eq. (5.11): Ensemble average form of stochastic continuity

equation (Eq. (5.8)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
-viii-
262
List of Figures
1.1
1.2
2.1
2.2
2.3
2.4
2.5
2.6
3.1
3.2
4.1
4.2
4.3
4.4
4.5
4.6
Interpreted Effective Stress Strength Parameters at Opelika NGES (after

Mayne et al. [70]) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Soil: Inside Failure of Uniform MGM Specimen (after Swanson et al. [94])
Gaussian Probability Density Function of Friction Coefficient Random Variable (A) with mean = 0.3 and Standard Deviation = 0.05 . . . . . . . . . .
Gaussian Cumulative Density Function of Friction Coefficient Random Variable (A) with mean = 0.3 and Standard Deviation = 0.05 . . . . . . . . . .
Gaussian Joint Probability Density Function of Shear Modulus (G) and friction coefficient (A) with G = 2.5 MPa, (SD)G = 1.0 MPa, A = 0.3,
(SD)A = 0.05, and GA = 0 . . . . . . . . . . . . . . . . . . . . . . . . . .
Marginal PDF of friction coefficient (A) from joint PDF of shear modulus,
G, and friction coefficient, A (Fig. 2.3) . . . . . . . . . . . . . . . . . . . . .
Gaussian Joint Cumulative Density Function of Shear Modulus (G) and friction coefficient (A) with G = 2.5 MPa, (SD)G = 1.0 MPa, A = 0.3,
(SD)A = 0.05, and GA = 0 . . . . . . . . . . . . . . . . . . . . . . . . . .
A realization of Shear Modulus Random Field . . . . . . . . . . . . . . . . .
Graphical depiction of the deterministic, epistemic and aleatory uncertainty
related to geotechnical simulations. This is, in a sense, macro scale interpretation of Heisenberg uncertainty principle. . . . . . . . . . . . . . . . . . .
Measured values of mechanical properties of soil from Mexico city, typical
soft spot (after Baecher and Christian [6]) . . . . . . . . . . . . . . . . . . .
Shear modulus random field: Trend and residual around trend . . . . . . .
CPT Sounding locations . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
East-West soil profile interpreted from CPT soundings . . . . . . . . . . . .
Typical qT data: Borehole 1 sounding . . . . . . . . . . . . . . . . . . . . .
Maximum likelihood estimated Gauss-Markov autocovariance function along
with method of moment estimate (for borehole 1 sounding) . . . . . . . . .
Maximum likelihood estimated Gauss-Markov autocorrelation function along
with method of moment estimate (for borehole 1 sounding) . . . . . . . . .
-ix-
3
4
14
16
23
24
27
33
53
61
66
73
74
75
75
76
4.7
(a) Measured (at borehole 2) and (b) Simulated (finite scale Gauss-Markov
model) realizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.8 Deterministic Trend as obtained through global regression over 16 CPT
soundings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.9 Periodogram of borehole 1 sounding . . . . . . . . . . . . . . . . . . . . . .
4.10 Maximum likelihood estimated fractal (1/f -type noise with lower cut-off frequency) power spectral density function corresponding to borehole 1 sounding
4.11 Fractal (1/f -type noise with lower cut-off frequency) autocovariance function
for borehole 1 sounding . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1
5.2
5.3
Anticipated Influence of Material Fluctuations on Stress-Strain Behavior . .

Movements of Cloud of Initial Points, described by density (, 0), in the
-space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Spatial Discretization of the FokkerPlanckKolmogorov PDE . . . . . . . .
6.1
77
79
80
80
81
84
90
110
Approximation of Dirac delta initial condition (Eq. (5.52)) with a Gaussian

function of a zero mean and a standard deviation of 0.00001 MPa . . . . . . 119
6.2 Time (or strain) evolution of probability density function of shear stress for
elastic constitutive rate equation with random shear modulus obtained using
FPKE approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
elastic constitutive rate equation with random shear modulus obtained using
transformation method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.4 Comparison of contours of time (or strain) evolution of probability density
function for shear stress for elastic constitutive rate equation with random
shear modulus for FPKE solution and variable transformation method solution124
6.5 Comparison of evolution of mean and standard deviation of shear stress with
time (or shear strain) for elastic constitutive rate equation with random shear
modulus for FPKE solution and variable transformation method solution . 124
6.6 Effect of approximating function of Dirac delta initial condition: PDF of
stress at yield for different approximations of initial condition with actual
(variable transformation method) solution . . . . . . . . . . . . . . . . . . . 125
6.7 Initial condition for Fokker-Planck-Kolmorogov equation for probabilistic
simulation of post-yield region . . . . . . . . . . . . . . . . . . . . . . . . . . 130
Drucker-Prager elastic-plastic constitutive rate equation with random shear
modulus (Problem-I) (View obtained when one looks perpendicular to the
time/strain axis) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.9 Time (or strain) evolution of probability density function of shear stress
for Drucker-Prager elastic-plastic constitutive rate equation with random
shear modulus (Problem-I) (View obtained when one looks parallel to the
time/strain axis) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
6.10 Contour of time (or strain) evolution of probability density function for shear
stress for Drucker-Prager elastic-plastic constitutive rate equation with random shear modulus (Problem-I) . . . . . . . . . . . . . . . . . . . . . . . . . 133
-x-
6.11 Comparison of evolution of PDF of shear stress for Drucker-Prager elasticplastic linear hardening material model and extended linear elastic model
with random shear modulus . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.12 Comparison of evolution of mean and standard deviation of shear stress for
Drucker-Prager elastic-plastic constitutive rate equation with random shear
modulus (Problem-I), obtained from FPKE solution and Monte Carlo solution134
6.13 Initial condition for Fokker-Planck-Kolmorogov equation for probabilistic
simulation of Drucker-Prager post-yield region with random friction coefficient (Problem-II) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
6.14 Time (or strain) evolution of PDF of shear stress for Drucker-Prager elasticplastic constitutive rate equation with random friction coefficient (ProblemII) (only post-yield region is shown, note that the pre-yield region is deterministic) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.15 Contours of time (or strain) evolution of PDF for shear stress, along with
evolutions of mean and standard deviation of shear stress, for DruckerPrager elastic-plastic constitutive rate equation with random friction coefficient (Problem-II) (only post-yield region is shown, note that the pre-yield
region is deterministic) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.16 Comparison of evolutions of mean and standard deviation of shear stress for
Drucker-Prager elastic-plastic constitutive rate equation with random friction coefficient (Problem-II), obtained from FPKE solution and Monte Carlo
solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
6.17 Low OCR Cam Clay response with random normally distributed shear modulus (G): (a) Evolution of PDF and (b) Evolution of contours of PDF, mean,
mode, standard deviations, and deterministic solution of shear stress (12 )
with time (t)/shear strain (12 ) . . . . . . . . . . . . . . . . . . . . . . . . . 139
6.18 Comparison of FPK approach and Monte Carlo approach in obtaining low
OCR Cam Clay response with random normally distributed shear modulus
(G) in terms of evolution of mean and standard deviation of shear stress (12 )
with time (t)/shear strain (12 ) . . . . . . . . . . . . . . . . . . . . . . . . . 140
6.19 Low OCR Cam Clay response with random normally distributed shear modulus (G) and random normally distributed slope of critical state line (M ): (a)
Evolution of PDF and (b) Evolution of contours of PDF, mean, mode, standard deviations, and deterministic solution of shear stress (12 ) with time (t)
/shear strain (12 ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.20 Low OCR Cam Clay response with random normally distributed shear modulus (G), random normally distributed slope of critical state line (M ), and
random normally distributed overconsolidation pressure (p0 ): (a) Evolution
of PDF and (b) Evolution of contours of PDF, mean, mode, standard deviations, and deterministic solution of shear stress (12 ) with time (t)/shear
strain (12 ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
-xi-
6.21 Low OCR Cam Clay response with random normally distributed shear modulus (G) and random normally distributed overconsolidation pressure (p0 ):
(a) Evolution of PDF and (b) Evolution of contours of PDF, mean, mode,
standard deviations, and deterministic solution of shear stress (12 ) with time
(t)/shear strain (12 ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.22 Comparison of shear stresses at 1.62% shear strain obtained from low OCR
Cam Clay model with different degrees of randomnesses . . . . . . . . . . .
6.23 High OCR Cam Clay response with random normally distributed shear modulus (G) and random normally distributed slope of critical state line (M ): (a)
Evolution of PDF and (b) Evolution of contours of PDF, mean, mode, standard deviations, and deterministic solution of shear stress (12 ) with time (t)
/shear strain (12 ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.24 Comparison of FPK Approach and MonteCarlo Approach for High OCR
Cam Clay Response with Random Normally Distributed Shear Modulus (G)
and Random Normally Distributed Slope of Critical State Line (M ) in terms
of Evolution of Mean and Standard Deviation of Shear Stress (12 ) with Time
(t)/Shear Strain (12 ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.25 CDF of shear strength for von Mises model: (a) very uncertain case, (b)
fairly certain case. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.26 von Mises associative plasticity model with uncertain shear modulus and
shear strength (yield parameter): (a) Evolution of PDF of stress with strain
(PDF=10000 was used as a cutoff for surface plot) and (b) Contours of
evolution of stress PDF with strain. . . . . . . . . . . . . . . . . . . . . . .
6.27 von Mises elasticplastic model: (a) Shear modulus: very uncertain; shear
strength: fairly certain, (b) Shear modulus: fairly certain; shear strength:
very uncertain. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.28 CDF of yield stresses for Drucker-Prager model: (a) very uncertain and (b)
fairly certain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.29 Drucker-Prager associative elasticplastic model with uncertain shear modulus and frictional coefficient: (a) Evolution of probability density function
(PDF) of stress with strain (PDF=10000 was used as a cutoff for surface
plot) and (b) Contours of evolution of PDF with strain . . . . . . . . . . . .
6.30 Drucker-Prager elasticplastic model: (a) Shear modulus: very uncertain;
frictional coefficient: fairly certain, (b) Shear modulus: fairly certain; frictional coefficient: very uncertain. . . . . . . . . . . . . . . . . . . . . . . . .
7.1
7.2
7.3
7.4
KL eigenvalues of exponential covariance kernel having variance = 1000 kP a2

and correlation length = 1 m . . . . . . . . . . . . . . . . . . . . . . . . . .
KL eigenvectors of exponential covariance kernel having variance = 1000
kP a2 and correlation length = 1 m . . . . . . . . . . . . . . . . . . . . . . .
Exact exponential covariance kernel having variance = 1000 kP a2 and correlation length = 1 m (C(x1 , x2 ) = 2 e|x1 x2 |/r , 2 = 1000 kP a2 and r =
1.0 m) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
KL approximations (with estimated errors) of exponential covariance kernel
having variance = 1000 kP a2 and correlation length = 1 m . . . . . . . . .
-xii-
145
146
148
149
151
152
154
155
156
158
170
171
172
173
7.5
7.6
Exact exponential covariance kernel with variance = 1000 kP a2 and correlation length = 0.05 m (C(x1 , x2 ) = 2 e|x1 x2 |/r , 2 = 1000 kP a2 and r =
0.05 m) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
KL approximations (with estimated errors) of exponential covariance kernel
having variance = 1000 kP a2 and correlation length = 0.05 m . . . . . . . .
8.1
8.2
Schematic of static 1D soil column (shear beam) example . . . . . . . . . .

Mean and standard deviations of displacement at the top node of the soil
column, with linear elastic material model (with KL-dimension = 2, order of
PC = 2). Monte Carlo simulation is also shown . . . . . . . . . . . . . . . .
8.3 Mean and standard deviations of displacement at the top node of the soil
column, with von Mises elastic-plastic material model (with KL-dimension
= 2, order of PC = 2). Monte Carlo simulation is also shown . . . . . . . .
8.4 Comparison of PDF of top node displacements of the soil column: Elastic
versus von Mises elastic-plastic material model . . . . . . . . . . . . . . . .
8.5 Correlation length and KL dimension: Mean displacement along depth of
the 1D soil column with linear elastic material model, having very small
variance (COV = 1%) of shear modulus and very large ratio of correlation
length of shear modulus to domain length (= 100) . . . . . . . . . . . . . .
8.6 Correlation length and KL dimension: Standard deviation of displacement
along depth of the 1D soil column with linear elastic material model, having
very small variance (COV = 1%) of shear modulus and very large ratio of
correlation length of shear modulus to domain length (= 100) . . . . . . . .
8.7 Correlation length and KL dimension: Mean displacement along depth of
the 1D soil column with linear elastic material model, having very small
variance (COV = 1%) of shear modulus and very small ratio of correlation
length of shear modulus to domain length (= 0.0001) . . . . . . . . . . . . .
8.8 Correlation length and KL dimension: Standard deviation of displacement
along depth of the 1D soil column with linear elastic material model, having
very small variance (COV = 1%) of shear modulus and very small ratio of
correlation length of shear modulus to domain length (= 0.0001) . . . . . .
8.9 Variance and order of PC: Mean displacement along depth of the 1D soil
column with linear elastic material model, having large variance (COV =
20%) of shear modulus and ratio of correlation length of shear modulus to
domain length = 0.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.10 Variance and order of PC: Standard deviation of displacement along depth of
the 1D soil column with linear elastic material model, having large variance
(COV = 20%) of shear modulus and ratio of correlation length of shear
modulus to domain length = 0.1 . . . . . . . . . . . . . . . . . . . . . . . .
8.11 Variance and order of PC: Mean displacement along depth of the 1D soil
column with linear elastic material model, having very small variance (COV
= 1%) of shear modulus and ratio of correlation length of shear modulus to
domain length = 0.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
-xiii-
174
175
196
204
204
205
206
207
207
208
209
209
210
8.12 Variance and order of PC: Standard deviation of displacement along depth
of the 1D soil column with linear elastic material model, having very small
variance (COV = 1%) of shear modulus and ratio of correlation length of
shear modulus to domain length = 0.1 . . . . . . . . . . . . . . . . . . . . .
8.13 Schematic of dynamic 1D soil column example . . . . . . . . . . . . . . . .
8.14 Base displacement applied to the bottom node of the 1D soil column . . .
8.15 Visualization of 1D soil column with base displacement, as shown in Fig. 8.13,
as a soil column-stiff spring system . . . . . . . . . . . . . . . . . . . . . . .
8.16 Time evolution of mean of displacement at the top node of the 1D soil column, with linear elastic material model, due to sinusoidal base displacement
shown in Fig. 8.14 (with KL-dimension = 2, order of PC = 2) . . . . . . . .
8.17 Time evolution of mean of displacement at the top node of the 1D soil
column, with von Mises elasticplastic material model, due to sinusoidal base
displacement shown in Fig. 8.14 (with KL-dimension = 2, order of PC = 2)
8.18 Time evolution of standard deviation of displacement at the top node of the
1D soil column, with linear elastic material model, due to sinusoidal base
8.19 Time evolution of standard deviation of displacement at the top node of
the 1D soil column, with von Mises elasticplastic material model, due to
sinusoidal base displacement shown in Fig. 8.14 (with KL-dimension = 2,
order of PC = 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.20 Time evolution of mean standard deviation of displacement at the top node
of the 1D soil column, with linear elastic material model, due to sinusoidal
base displacement shown in Fig. 8.14 (with KL-dimension = 2, order of PC
= 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.21 Time evolution of mean standard deviation of displacement at the top
node of the 1D soil column, with von Mises elasticplastic material model,
due to sinusoidal base displacement shown in Fig. 8.14 (with KL-dimension
= 2, order of PC = 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.22 Base displacement applied to the bottom node of the 1D soil column: Modified 1938 Imperial Valley motion . . . . . . . . . . . . . . . . . . . . . . . .
column, with linear elastic material model, due to modified 1938 Imperial
Valley base displacement as shown in Fig. 8.22 (with KL-dimension = 2,
order of PC = 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
column, with von Mises elasticplastic material model, due to modified 1938
Imperial Valley base displacement as shown in Fig. 8.22 (with KL-dimension
= 2, order of PC = 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.25 Time evolution of standard deviation of displacement at the top node of the
1D soil column, with linear elastic material model, due to modified 1938
Imperial Valley base displacement as shown in Fig. 8.22 (with KL-dimension
= 2, order of PC = 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
-xiv-
210
212
212
213
215
216
216
217
218
218
221
222
222
223
8.26 Time evolution of standard deviation of displacement at the top node of

the 1D soil column, with von Mises elasticplastic material model, due to
modified Imperial Valley base displacement as shown in Fig. 8.22 (with KLdimension = 2, order of PC = 1) . . . . . . . . . . . . . . . . . . . . . . . .
8.27 Time evolution of mean standard deviation of displacement at the top node
of the 1D soil column, with linear elastic material model, due to modified
1938 Imperial Valley base displacement as shown in Fig. 8.22 (with KLdimension = 2, order of PC = 1) . . . . . . . . . . . . . . . . . . . . . . . .
8.28 Time evolution of mean standard deviation of displacement at the top
due to modified 1938 Imperial Valley base displacement as shown in Fig. 8.14
(with KL-dimension = 2, order of PC = 1) . . . . . . . . . . . . . . . . . .
8.29 Time evolution of coefficient of variation (COV) of displacement at the top
node of the 1D soil column, with linear elastic material model, due to modified 1938 Imperial Valley base displacement as shown in Fig. 8.22 (with
KL-dimension = 2, order of PC = 1) . . . . . . . . . . . . . . . . . . . . . .
8.30 Time evolution of coefficient of variation (COV) of displacement at the top
due to modified 1938 Imperial Valley base displacement as shown in Fig. 8.14
(with KL-dimension = 2, order of PC = 1) . . . . . . . . . . . . . . . . . .
8.31 Time evolution of probability density function (PDF) of displacement at the
top node of the 1D soil column, with linear elastic material model, due to
modified 1938 Imperial Valley base displacement as shown in Fig. 8.22 (with
KL-dimension = 2, order of PC = 1) . . . . . . . . . . . . . . . . . . . . . .
8.32 Time evolution of probability density function (PDF) of displacement at the
top node of the 1D soil column, with von Mises elasticplastic material
model, due to modified 1938 Imperial Valley base displacement as shown in
Fig. 8.22 (with KL-dimension = 2, order of PC = 1) . . . . . . . . . . . . .
-xv-
223
224
224
225
225
226
227
List of Tables
2.1
Classification of random processes based on state and parameter space . . .
3.1
Representative values of variabilities in consolidation parameters (after Baecher

and Christian [6]) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Representative values of variabilities in laboratory measured effective friction
angle (after Baecher and Christian [6]) . . . . . . . . . . . . . . . . . . . . .
Representative values of variabilities in some common in-situ soil properties
(after Phoon and Kulhawy [79]) . . . . . . . . . . . . . . . . . . . . . . . . .
Best fitting probability density functions (PDFs) for various soil properties
(after Lacasse and Nadim [56]) . . . . . . . . . . . . . . . . . . . . . . . . .
Representative values of testing errors of soil index properties (after Hammitt
[36], reproduced from the book by Baecher and Christian [6]) . . . . . . . .
Representative testing error of some laboratory tests that evaluates strength
properties (after Phoon and Kulhawy [79]) . . . . . . . . . . . . . . . . . . .
Representative testing error of some field tests (after Phoon and Kulhawy [79])
Transformation uncertainties of some common strength property correlations(after Phoon and Kulhawy [80]) . . . . . . . . . . . . . . . . . . . . . .
Representative scale of fluctuation of some common soil properties (after
Phoon and Kulhawy [79]) . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2
3.3
3.4
3.5
3.6
3.7
3.8
3.9
35
54
54
55
55
57
57
58
58
60
4.1
4.2
Maximum likelihood estimated constant mean Gauss-Markov model parameters 76

Maximum likelihood estimated constant mean fractal (1/f-type noise with
lower cut-off frequency) model parameters obtained using periodogram approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
81
8.1
Comparison of results (at top node) of FPKE-based spectral stochastic finite

element with direct spectral stochastic finite element, for 1D soil column
example, with linear elastic material . . . . . . . . . . . . . . . . . . . . . .
-xvi-
211
Acknowledgments
I would like to express my sincere gratitude to my advisor, Professor Boris Jeremic
for his guidance, advice, encouragement, and kindness all throughout my doctoral study. I
have been very fortunate to have an advisor like him. He gave me freedom to explore on
my own. However, with careful guidance and constructive criticism, he helped me to stay
focused. I am indebted to him for helping me to develop my critical thinking and research
skills.
My special gratitude goes to Professor M. Levent Kavvas. He has not only introduced me to stochastic methods, on which this dissertation is based, but also helped me to
develop my concept of the topic. Long discussions with him helped me to sort out technical
details of my work.
I would also like to extend my thanks to Professors Ross Boulanger and Niels
Grnbech Jensen for serving on my dissertation committee and taking time to critique my
work.
My work has benefited from conversations with the past and present members of
Computational Geomechanics Group. For their generous help, I would like to acknowledge
Jim Putnam, Guanzhou Jie, Zhao Cheng, Matthias Preisig, Mahdi Taiebat, and Alisa
Neeman.
Financial supports provided by the Department of Civil and Environmental Engineering (through Block Grant Living Allowance Fellowship) and the National Science
Foundation (through award # CMMI 0600766) are also gratefully acknowledged.
-xvii-
Part I
Motivation and Theoretical

Background
Chapter 1
Introduction
1.1
Hypothesis
Failure of geomaterial is generally preceded by formation of shear band and subse-
quent bifurcation of response. Shear bands form due to strain localization, which, according
to some recent studies (Carmeliet and De Borst [9]; Gutierrez and De Borst [34]), stems
from uncertain material non-uniformity
and speaking of natural geomaterials, they are
inherently uncertain and very non-uniform. Figure 1.1 shows variation of a typical soil
property, the undrained shear strength, measured with different field and laboratory tests,
with depth. Note large variability as a function of depth (inherent variability), as well
as with different test methods (testing uncertainty). Fig. 1.2 shows that even in a carefully prepared uniform laboratory specimen the failure is due to strain localization and
subsequent formation of shear bands.
1
though the usual practice in numerical simulation of shear band is to deterministically specifying the
imperfection in the numerical specimen and thus help in formation of shear band
Figure 1.1: Interpreted Effective Stress Strength Parameters at Opelika NGES (after Mayne
et al. [70])
The variabilities of in-situ soil properties are generally quite large. The usual
procedure in this case is to choose either a constant variation of properties or fit a smooth
curve (usually a straight line) to the properties of interest, for example undrained shear
strength as in the example above, with depth. This approach completely neglects variations,
which might have a large effect on response of this soil profile to, for example, seismic
excitation. The effects of neglecting the natural variability are not known a priori and the
usual remedy is to increase factors of safety in dealing with such (uncertain) soils. Book
by Lambe and Whitman [57] gives many more interesting, nonuniform profiles with quite
uncertain and nonuniform properties from various geotechnical sites.
In recent years, civil engineering practice, and in particular the geotechnical engi-
Figure 1.2: Soil: Inside Failure of Uniform MGM Specimen (after Swanson et al. [94])
4
5
neering practice, has seen an increasing emphasis on reliability. In geotechnical earthquake
engineering, for example, the earthquake ground motions are usually given with certain
probability of occurrence. However, whats missing is the actual quantification of this probability that come from uncertain and nonuniform soil layers. In other words, it is important
to discern how much of that uncertainty in ground motion is due to the uncertain and variable motion coming from the hypocenter and how much is due to the uncertain and variable
soil properties in top layers.
Hence the question that arises is how to account for this non-uniformity and uncertainty of soil parameters in analytical and numerical simulations (statics and dynamics)
of behaviors of dams, levees, shallow and deep foundations and other solids and structures
made of geomaterials. The best way to account for uncertainty is to quantify them and this
quest for quantifying uncertainties in response behavior lead to the development of probabilistic simulation tools. In addition, there exist several other practical reasons to pursue
this research project:
Modern building codes (regulations) are increasingly being based on reliability methods,
however the analyses are still largely (exclusively) deterministic. Other industries,
for example nuclear and offshore, are already using probabilistic analysis to a large
extent.
Financial decision making by object owners tends to lead towards the use of probabilistic
theories. For example, decisions on the best course of action for developing new objects
or upgrading and repairing existing objects are highly probabilistic. Crucial decisions
on the extent of work, financing and scheduling are also made using probability theory.
6
In contrast, the actual performance assessments (simulations of behavior) used in
designing new or in upgrading and repairing existing objects are still almost exclusively
deterministic.
Consistent development of a probabilistic framework for geotechnical simulations will provide a rational way to address our confidence (or lack thereof) in simulated behavior.
For example, probabilistic simulation tools will empower engineers to demonstrate
the need for more, uniform data on material properties, to develop novel site characterization techniques, and to design the geotechnical systems that (probably) achieve
best performance.
Further to the above practical factors related to civil engineering industry (with its fields
of geotechnical, structural, construction engineering) there are some wider concerns as well:
Societal needs: Society is demanding greater safety and reliability of infrastructure systems (structural, geotechnical). An important aspect of meeting this goal is development of performance assessment capabilities (numerical simulations on models).
Such probabilistic performance estimates of infrastructure systems (such as buildings,
bridges, dams, levees and lifeline networks) will greatly benefit the society in allowing
the public and decisions makers to appropriately assess the critical needs of economy,
safety, and usability for new infrastructure or upgradation of existing infrastructure
objects.
Complexity of Geotechnical media: The behavior of the soils and structures involves
complex physics of highly nonlinear, heterogeneous materials with uncertain constitutive properties. In particular, soils are known to have coefficients of variance that
7
are most of the time greater than 20 % (Phoon and Kulhawy [79, 80]). This large
uncertainty in soil properties renders any deterministic simulations almost useless,
unless large factors of safety are applied. However, use of large factors of safety is becoming unacceptable as it leads to design solutions that are not economical and many
times not even safe (e.g., Duncan [20]). Development of probabilistic simulation tools
will greatly improve our ability to consistently and efficiently perform simulations
of inelastic behavior of geomaterial solids and structures with non-homogeneous and
uncertain material parameters. In addition to that, it will also help in better understanding of the scale effects in nonhomogeneity of soils and, consequently will help
in performing better site characterizations.
Impracticality of Monte Carlo approach for large-scale probabilistic simulations:
While there are some recent works (e.g., Griffiths et al. [33], Paice et al. [78], Fenton
and Griffiths [25, 27]) on probabilistic simulations of stochastic soils, mostly on spatial
nonuniformity of material properties, it is computationally very expensive and in fact
impossible in any meaningful time to conduct necessary number of Monte Carlo runs
to obtain the complete probabilistic behavior in terms of probability density function.
2
This is especially true when the problem under consideration has more than one
material properties as random. Usually the number of random material properties in

any geotechnical problem is larger than one. For example, even in 1D both elastic
and inelastic
2
material properties are probabilistic. For a 3-D elasto-plastic model
The capability of any approach to obtain the complete probabilistic description in terms of probability
density function (and not, only the first few statistical moments e.g. mean and standard deviation) is very
important as geomaterials often fails at low probabilities and hence the tails of probability density function
are equally important as the first few statistical moments
3
the number of inelastic material properties depends on the type of elasticplastic model. The more
8
with 3 random variable material parameters (e.g., Youngs modulus E, friction angle
and cohesion c for simplest DruckerPrager model with no hardening) one needs
4
at least 10, 0003 = 1012 runs of computationally expensive deterministic model.
This is even more the case for elasto-plastic models with larger number of random
material parameters. In addition to this huge effort, there is an added burden of
post-processing the high volume of data. Reductions of number of statistically appropriate realizations are possible, but at the expense of increasing the error in Monte
Carlo simulations. For example, if one assumes that, the variables are correlated and
that statistically only 1, 000 realizations are appropriate for each of three variables,
the finite element model for given foundation or levee or lifeline network needs to be
solved a billion times (1, 0003 = 109 ) in order to obtain statistics and probabilities
of response. Due to this high computational cost in analysis and data processing,
the Monte Carlo approach is impractical for large-scale probabilistic simulation and
is used only for verification of other approaches.
1.2
Scope of Study
The main objective of this research is to develop a computational framework for
simulations of behaviors of solids and structures made of stochastic elasticplastic materials,

with particular emphasis on soil. More specifically the objectives are:
Probabilistic characterization and rational quantification of uncertainties in material

properties, which will act as input to the stochastic framework.
advanced the model, the more the number of (random) parameters
4
assuming statistically appropriate 10,000 realizations per random variable
9
Development of probabilistic elastoplasticity for constitutive (material) level simulation.
Development of a stochastic elasticplastic finite element method for solving boundary
value problems in simulating the behaviors of solids and structures made of stochastic
elasticplastic materials.
Examples to illustrate the applicability of the developed methodologies in solving
real-life geotechnical engineering problems.
1.3
Summary of Contents
This dissertation is divided into four parts. Part - I describes the motivation
and discusses the features of probability theory that have been used in the subsequent
chapters. Part - II deals with characterization and quantification of uncertainties in material
properties, with particular emphasis on soils. It also discusses the random field modeling
of spatially uncertain material properties and techniques to estimate the model parameters
from measured soil properties. Part - III develops a technique for probabilistic constitutive
simulation of elasticplastic materials with example simulations of various material models.
Part - IV builds upon the developments in Part - III in formulating stochastic elasticplastic
finite element method for stochastic solution of boundary value problems with example
simulations. In addition, Part - IV also illustrates the complete solution process, based on
the developed computational framework, of a geotechnical engineering problem with real-life
data.
10
1.4
Original Features
The intellectual merit of this research project is in the merging of state-of-the-art
probability theory with theory of elastoplasticity and subsequent development of a stochastic

elasticplastic finite element method. This is, to our knowledge, a unique endeavor with
minimal prior work in probabilistic elastoplasticity to rely on. Of particular importance
is the application of developed methodology to soils, which exhibit a high degree of nonlinearity and where material properties are highly uncertain.
The impact of the proposed research is expected to be much wider than just in the
area of geotechnical engineering. The phenomena of spatial variability and uncertainty in
material properties is present in all materials. The appropriate formulation and implementation that incorporate above phenomena into advanced numerical simulations will impact
mechanical, biomedical, materials, aerospace as well as other areas of civil engineering.
11
Chapter 2
Probability Theory Background

2.1
Introduction
The intention of this chapter is to provide some background on probability theory.
Effort here has been made to outline only the mathematical tools that are used in the
subsequent chapters. This chapter develops mainly following the classnotes on Applied
Stochastic Methods in engineering by Professor M. Levant Kavvas [46], with appropriate
examples from mechanics. For a thorough outline of mathematical theory of probability, the
readers are encouraged to refer to standard mathematical texts on probability theory like
by Montgomery and Runger [72] or for advanced topics like stochastic calculus by Gardiner
[30].
2.2
Basic Definitions
The definitions are mainly following Montgomery and Runger [72]
12
Random Experiment: An experiment that can result in different outcomes, even though it
is repeated in the same manner every time, is called a random experiment.
Sample Space: The set of all possible outcomes of a random experiment is called the sample
space.
Probability of an Outcome: There are two schools of thoughts in interpreting probability of
an outcome - degree of belief interpretation and relative frequency interpretation. Degree of
belief interpretation is subjective in the sense that there is always a possibility that different
person will assign different probabilities to the same outcome. On the other hand relative
frequentists interpret probability based on the conceptual model of repeated replications of
any random experiment. According to them probability of an outcome is the limiting value
of the proportions of times any outcome occurs in n repetitions of a random experiment
as n increases beyond all bounds. For example, if the probability that the shear modulus
of a soil is equal to 2.5 MPa is 20%, the relative frequentists will interpret this as follows:
if we do many identical tests (using the same testing devices with the same method of
registering test data etc.) on the same soil, 20% of the tests will result in shear modulus of
the soil equal to 2.5 MPa. In mechanics, both interpretations of probability of an outcome
is important since for any material (and to a larger extent for soils) we have both natural
variability and knowledge uncertainty (discussed in details in next chapter (Chapter [3])).
Probability of an event A: Probability of an event A, denoted as P [A], is equal to sum of
the probabilities of the outcomes in A.
Random Variable: A function that assigns a real number to each outcome in the sample
13
space of a random experiment is called a Random Variable.
Realizations or Sample Values: The various values a random variable can take during a
random experiment are called Realizations or Sample Values.
2.3
Properties of Single Random Variable

The objective of this section is to describe the mathematical tools available to
analyze the complete probabilistic description of a single random variable.
2.3.1
Probability Distribution Function

For many distributions of probability over a sample space, there exists a function
fX (x) from which the probability P [E] of any event, E can be obtained by a summation of
the form:
P [E] =
fX (x)
(2.1)
For discrete sample spaces, the above function is called Probability Mass Function (PMF)
and for continuous sample spaces, is called a Probability Density Function (PDF).
By far the most common probability density function is the Gaussian or normal
probability density function, which, for a single random variable, say friction coefficient 1 ,
A is written as:
1
fA () = q
2(SD)2A
1
2
2 ( A )
2(SD)
A
e
defined as A = 2sin/ 3(3 sin), where is the friction angle of a material
(2.2)
14
where A and (SD)A are the mean and standard deviation of the friction coefficient random
variable (A). Fig. 2.1 shows the probability density function of friction coefficient, assuming
Gaussian distribution with mean and standard deviation of 0.3 and 0.05 respectively. From
the probability density function, the probability of an outcome, for example, the probability
of A being 0.25 in the above example
can be interpreted as:
P [A = 0.25] = P [0.245 < A < 0.255] =
0.255
fA ()d = 0.0483
0.245
P@AD
10
8
6
4
2
0.2
0.3
0.4
0.5
Figure 2.1: Gaussian Probability Density Function of Friction Coefficient Random Variable
(A) with mean = 0.3 and Standard Deviation = 0.05
This is somewhat counter-intuitive, because for a continuous random variable X and any outcome x,
P (X = x) = 0, as every point has zero width. However,in practice when a particular outcome is observed,
such as 0.25 in this example, the result can be interpreted as the rounded value of a particular outcome
that is actually in a range such as 0.245 < A < 0.255 in this example. Therefore, the probability that the
rounded value 0.25 is observed as the value for A is the probability that A assumes a value in the interval
[0.245, 0.255], which is not zero.
15
2.3.2
Cumulative Distribution Function

Cumulative distribution function is an alternate way of describing the probability
distribution of any random variable. For a continuous random variable, it is defined as:
F (x) = P (X x) =
f (u)du
for - < x <
(2.3)
where F (x) is the cumulative distribution function (CDF) and f () is the probability density
function. For example, for the Gaussian friction coefficient random variable as defined
previously, the CDF can be obtained by integrating the PDF and is shown in Fig. 2.2.
From the CDF of any random variable, the probability that any random variable (here, the
friction coefficient (A)) is less than equal to some value (say 0.3) can be obtained easily. In
this example:
P [A 0.3] = 0.5, which can be directly read from the CDF (Fig. 2.2)
2.3.3
Moments of Probability Distributions

It is mathematically convenient to represent probability distributions by their mo-
ments. The nth moment of a probability distribution about the origin is mathematically
represented as:
E(X ) =
xn fX (x)dx
for < x <
(2.4)
The above equation can be visualized as a weighted integral over the probability distribution.
In statistical terminology it is called an expectation operation. When n = 1, Eq. (2.4)
16
F@AD = P@A D
1
0.8
0.6
0.4
0.2
0.2
0.3
0.4
0.5
Figure 2.2: Gaussian Cumulative Density Function of Friction Coefficient Random Variable
(A) with mean = 0.3 and Standard Deviation = 0.05
transforms to the arithmetic average, which is called the mean or expected value and it is
mathematically represented as:
hXi = E(X) =
xfX (x)dx
for < x <
(2.5)
The expected value of a function h(X) of a continuous random variable is defined similarly
as:
hh(X)i = E[h(X)] =
h(x)fX (x)dx
for < x <
(2.6)
For example, if the elastic shear modulus (G) of a material is a normally distributed random
variable with a mean of 2.5 MPa and a standard deviation of 1 MPa, then the mean of
1D, elasticplastic shear modulus obeying von Mises associative plasticity (defined
3
derived in Subsection 5.2.1
as
17
G G2 /(G + cu / 3)), where cu = 0.3 (deterministic)) is the rate of evolution of internal

variable cu (unconfined compressive strength) with plastic strain) can be obtained as:
+
G2
G
1
G + 0.3
3
n
o
G2
2
=
G
0.398942 e0.5(2.5+G) dG
1
G + 0.3
3
= 0.161806 M P a
The nth moment about the mean is called the nth central moment and can be
represented mathematically as:
E[X E(X)]n =
[x E(X)]n fX (x)dx
for < x <
(2.7)
The most common central moment is the second central moment. It is called the variance
and is defined as:
V ar[X] = E[X E(X)]2

Z +
[x E(X)]2 fX (x)dx
=
=
Z +
for < x <
(2.8)
x2 fX (x)dx (E[X])2 for < x <
(2.9)
The square root of variance is called the standard deviation and the dimensionless ratio of
mean over standard deviation is called the coefficient of variation (COV).
18
Having defined variance, from Eqs. (2.6) and (2.9), one can obtain the variance of
a function of random variable. Thus, the variance of elasticplastic shear modulus obeying
von Mises associative plasticity can be obtained as:
G2
1
G + 0.3
3
+2
*
n
o
G2
G2
0.5(2.5+G)2
G
=
0.398942 e
dG G
1
1
G + 0.3
G + 0.3
3
3
2
= 0.00427604 M P a
V ar
G
and the COV of von Mises elasticplastic shear modulus as:
COV
G
2.3.4
G2
=
1
G + 0.3
3
u
u
u
u
uV ar
G
t
G2
0.3
G+
3
0.00427604
*
+
= 0.404
=
2
0.161806
G
G
1
G + 0.3
3
Characteristic Function and its relation with Moments and Cumulants

As with probability distribution function and cumulative distribution function,
characteristic function also completely characterizes the probabilistic behavior of a random

variable. The characteristic function of a random variable X is defined as:

X (u) = eiuX =
eiuX fX (x)dx
(2.10)
19
where X is continuous on some domain D on the real line. In case of continuous random
variable, taking the inverse Fourier transform, one can obtain the pdf, fX (x) of X:
1
2
fX (x) =
X (u)eiuX du
(2.11)
By expanding the characteristic function, x in terms of MacLaurin series:
X (u) = X (u)|u=0 + u
dX (u)
u2 d2 X (u)
|u=0 +
|u=0 +
du
2 du2
(2.12)
and evaluating the various derivatives using Eq. (2.10):
X (u)|u=0 =
fX (x)dx = 1
dX (u)
|u=0 = i
du
(2.13)
xfX (x)dx = i hXi
(2.14)
dj X (u)
|u=0 = ij
duj

xj fX (x)dx = ij X j
(2.15)
one can express the j th moment of X in terms of characteristic function of X as:
j
1 dj X (u)
X = j
|u=0
i
duj
(2.16)
The cumulants or semi-invariants of a random variable X are defined by the rela-
20
tionship:
X (u) = e
j=1
(iu)j
j
X
j!
(2.17)
Rearranging the above equation (Eq. (2.17)), the cumulants of any random variable (X)
can be written in terms of its characteristic function (X ) as:
1
dX (u)
|u=0
du
i
2
2 d X (u)
1
X
=
|u=0 2
2
du
i
hhXii =
(2.18)
(2.19)
hhX n ii =
2.3.5
1
dn X (u)
|u=0 n
n
du
i
(2.20)
Relation between Moments and Cumulants

Comparing between Eqs. (2.12) and (2.17), one can write:
n
P (iu)
hhX n ii
(iu)n
(iu)2
2
n=1
n!
X ++
hX n i = e
1+
2
n!

2
n
P (iu)n
P (iu)n
n
n
hhX ii
hhX ii
n=1
n=1
X
(iu)n
n!
n!
n
hhX ii +
+
+
= 1+
n!
2
6
n=1
(2.21)
Equating the coefficients of equal powers in Eq. (2.21), one can obtain the moments in terms
of cumulants as:
21
hXi = hhXii
(2.22)
2 2
X = X
+ X
(2.23)

X = X
+ 3 hhXii X 2 + hhXii3
(2.24)
It is interesting to note here that higher order moments are larger in magnitude
than the lower order moments. On the other hand, higher order cumulants are smaller in
magnitude than the lower order cumulants. Hence while dealing with series approximation,
it is always advantageous to go for the cumulant expansion method rather than the moment
expansion method.
2.4
Properties of Two or More Random Variables

In this section, the focus is mainly on complete probabilistic characterization of two
random variables (commonly known as bi-variate analysis). However, the concept described
here is applicable and can be easily extended to analysis of n number of random variables
(multi-variate analysis).
2.4.1
Joint and Marginal Probability Distribution

If X and Y are two random variables, the probability distribution that defines their
simultaneous behavior is called a Joint Probability Distribution. It is specified by providing

a method for calculating the probability that X and Y assume a value in any given region
of a 2D space. The most common joint probability distribution is Gaussian, which for two
continuous random variables, say shear modulus G (with mean G and standard deviation
22
(SD)G ) and friction coefficient A (with mean A and standard deviation (SD)A ) is defined
as:
fGA (G) =
2(SD)G (SD)A 1 2GA

8
9

<
(G G )2 2GA (G G )( A ) ( A )2 =
1
+
: 2(1 2 )
(SD)G (SD)A
(SDA )2 ;
(SD)2G
GA
e
(2.25)
where GA is the correlation (defined in Subsection 2.4.4) between the random variables G
and A. For G = 2.5 MPa, (SD)G = 1.0 MPa, A = 0.3, (SD)A = 0.05, and GA = 0,
the joint probability distribution between shear modulus and friction coefficient is shown in
Fig. 2.3. From the example joint PDF, the probability of any joint outcome, for example
G = 1.5 MPa and A = 0.25, can be interpreted as:
P [G = 1.5M P a, A = 0.25] = P [1.45 M P a < G < 1.55 M P a, 0.245 < A < 0.255]
Z 1.55 Z 0.255
2
2
3.1831e0.5(2.5+G) 200(0.3+) dGd
=
1.45
0.245
= 0.00117099
From the joint PDF, marginal probability density function of any continuous random variable e.g., the friction coefficient (A) in the above example can be found, by integrating the joint PDF over the set of all points for which A = :
23
3
fAG HGL
2
4
1
3
0
0.1
2
0.2
0.3
1
0.4
0.5 0
Figure 2.3: Gaussian Joint Probability Density Function of Shear Modulus (G) and friction
coefficient (A) with G = 2.5 MPa, (SD)G = 1.0 MPa, A = 0.3, (SD)A = 0.05, and
GA = 0
fA () =
=
fGA (G)dG
3.1831e0.5(2.5+G)
2 200(0.3+)2
dG
= 7.97885e200(0.3+)
which is plotted in Fig. 2.4. Similarly the marginal probability density function of shear
modulus (G) can be obtained as:
24
fG (G) =
=
fGA (G)d
3.1831e0.5(2.5+G)
2 200(0.3+)2
dG
= 0.098942e0.5(2.5+G)
(2.26)
P@AD
10
8
6
4
2
0.2
0.3
0.4
0.5
Figure 2.4: Marginal PDF of friction coefficient (A) from joint PDF of shear modulus, G,
and friction coefficient, A (Fig. 2.3)
From the joint PDF, the individual expected value (mean) and variance of any of
the random variables can be obtained. For example from a joint PDF (fXY (xy)) of X and
Y , the mean and variance of X can be obtained as:
E(X) =
Rx
xfXY (xy)dxdy
(2.27)
Ry
and
V ar[X] =
Rx
Ry
(x E(X))2 fXY (xy)dxdy
(2.28)
25
where, Rx and Ry are the domain over which the random variables X and Y are defined.
Similar to the case of single random variable, the expected value of a function of two random
variables (h(X, Y )) can be obtained from their joint PDF as:
E [h(X, Y )] =
Rx
h(x, y)fXY (xy)dxdy
(2.29)
Ry
For example, if the shear modulus and friction coefficient of a material are random
variables and their probabilistic behavior is jointly Gaussian with G = 2.5 MPa, (SD)G =
1.0 MPa, A = 0.3, (SD)A = 0.05, and GA = 0, then the expected value of elasticplastic
shear modulus obeying Drucker-Prager associative plasticity (defined
as, G G2 /(G +
9KA2 + I1 / 3), where K is bulk modulus (assume 3.33 MPa (deterministic)), I1 is the
first invariant of the stress tensor (assume 0.001 MPa (deterministic)) and is the rate of
evolution of friction coefficient (assume 5.5 (deterministic))) can be obtained as:
G2
1
G + 9KA2 + I1
3
Z Z
G2
G
=
G + 9 3.332 + 0.001 5.5
3
3.1831 e0.5(2.5+G)
2 200(0.3+)2
dGd
= 1.21684 M P a
and the variance of Drucker-Prager elasticplastic shear modulus can be calculated as:
4
derived in Subsection 5.2.1
26
G2
V ar
G
1
G + 9KA2 + I1
3
Z Z
G2
G
=
1
G + 9 3.332 + 0.001 5.5
3
3.1831 e0.5(2.5+G)
2 200(0.3+)2
G2
1
G + 9KA2 + I1
3
= 1.21684 M P a2
2.4.2
dGd
+2
Joint Cumulative Probability Distribution

Similar to the cumulative probability distribution of a single random variable,
joint cumulative probability distribution is another way of describing the simultaneous

probabilistic behavior of more than one random variables. For two continuous random
variables X and Y it is defined as:
F (xy) = P (X x, Y y)
Z x Z y
f (uv)dudv
=
for - < x < and - < y <
(2.30)
where F (xy) is the joint cumulative density function (joint cdf) and f ( ) is the joint
probability density function (joint pdf). For example, for the previously defined joint pdf
between shear modulus (G) and friction coefficient (A), the joint cdf can be obtained by
double integrating their joint pdf and is shown in Fig. 2.5. By plotting the joint cdf, the joint
27
probability that shear modulus (G) and friction coefficient (A) are less than certain desired
values can be easily interpreted. For example, from Fig. 2.5, P [G 2.5, A 0.3] = 0.5.
1
0.75
FHGL
0.5
5
4
0.25
0
0.1
3
2
0.2
0.3
1
0.4
0.5 0
Figure 2.5: Gaussian Joint Cumulative Density Function of Shear Modulus (G) and friction
coefficient (A) with G = 2.5 MPa, (SD)G = 1.0 MPa, A = 0.3, (SD)A = 0.05, and
GA = 0
2.4.3
Conditional Probability Distribution

Conditional probability distribution considers the fact that when two random vari-
ables are defined in a random experiment, knowledge of one can change (this is particularly
true, if the random variables are dependent on each other) the probabilities that we associate with the value of the other. For two continuous random variables, X and Y , the
28
conditional probability density function of Y given X = x is defined as:
fY |x (y) =
fXY (x, y)
fX (x)
for fX (x) 6= 0
(2.31)
For example, for the previously defined joint distribution between shear modulus (G) and
friction coefficient (A), the conditional probability distribution of friction coefficient (A),
given G = G can be obtained as:
fA|G () =
fGA (G, )
fG (G)
2
3.1831e0.5(2.5+G) 200(0.3+)
0.098942e0.5(2.5+G)2
= 7.97885e200(0.3+)
It may be noted that the the conditional probability distribution of friction coefficient
(A), given G = G is equal to the marginal pdf of friction coefficient (A) as obtained in
Section 2.4.1. This is logical as we assumed shear modulus (G) and friction coefficient (A)
are independent random variables i.e knowledge of one can not change the probabilities that
we associate with the value of the other.
2.4.4
Dependency and Correlation between Random Variables

In some random experiment, involving two random variables, knowledge of the
values of one do change the probabilities associated with the values for other. These type of
random variables are called dependent random variables. For dependent random variables,
29
X and Y :
fXY (xy) 6= fX (x)fY (y)
for all x and y
(2.32)
Covariance is a measure of only linear dependency5 between two random variables

and for two random variables X and Y it is defined as:
COV [X; Y ] = h(X hXi)(Y hY i)i = hXY i hXi hY i
(2.33)
The dimensionless counterpart of covariance is correlation. Correlation just scales

the covariance by the standard deviation of each variable. It is a dimensionless quantity
and hence can be used to compare linear relationships between two random variables with
different units. Correlation between two random variables X and Y is defined as:
XY = p
COV [X; Y ]
V ar[X]V ar[Y ]
(2.34)
Correlation between two random variables could be any number between -1 and +1. Zero
correlation means the random variables do not have linear relationship between them6 .
Negative correlation between two random variables means they are negatively correlated
i.e. if value of one variable increases, the value of other variable decreases. Similarly positive
correlation means if value of one variable increases, the value of other variable also increases.
5
if two random variables have nonlinear relationship between them, covariance might not be sensitive to
it i.e. nonlinearly related random variables could have zero covariance between them.
6
Here it is important to note that zero correlation doesnt mean, the random variables are independent.
It means just they do not have linear relationship between them. They could have nonlinear relationship
between them. Only for Gaussian distributions uncorrelated and independent are synonymous.
30
2.4.5
Joint Characteristic Function

Similar to the characteristic function defined for single random variable, one can
define joint characteristic function (Joint CF) for two or more random variables. For two
random variables X and Y , it is defined as:
i(uX+vY )
XY (uv) = e
Rx
ei(uX+vY ) fXY (xy)dxdy
(2.35)
Ry
From joint CF, one can obtain the joint pdf through inverse Fourier transform:
1
fXY (xy) =
(2)2
Rx
XY (uv)ei(uX+vY ) dudv
(2.36)
Ry
Hence joint CF also contains complete probabilistic information as the joint pdf. Similar to
the case of single random variable, from joint CF, the joint moments between two random
variables X and Y can be obtained as:
E
XkY m =
k+m
XY (uv)|u=v=0
ik+m uk v m
1
(2.37)
and the joint cumulants as:
hhXY ii =
2.4.6
1 2 ln(uv)
|u=v=0
i2 uv
(2.38)
Some useful Properties of Two Random Variables
Below are some useful properties of two random variables:

1. Expectation operator commutes with addition and subtraction operator: The expectation of sum of or difference between two random variables is equal to the sum of or
31
difference between the expected value of the random variables.
hX Y i = hXi hY i
(2.39)
For example, if both shear modulus(G) and bulk modulus (K) of a soil are random
variables, having means of 10 M P a and 12 M P a respectively, then the mean of sum
of shear modulus(G) and bulk modulus (K) can be calculated as:
hG + Ki = hGi + hKi = 10 + 12 = 22 M P a
2. If X and Y are two random variables and a and b are two deterministic variable, then
the variance of aX bY is given as:
V ar[aX bY ] = a2 V ar[X] + b2 V ar[Y ] 2abCOV [X; Y ]
(2.40)
For example, the shear modulus(G) and bulk modulus (K) as defined above have
variances of 1 M P a2 and 0.7 M P a2 respectively, and has a covariance of 0.1 between
them, then the variance of sum of shear modulus(G) and bulk modulus (K) can be
calculated as:
V ar[G + K] = V ar[G] + V ar[K] + 2COV [G; K] = 1 + 0.7 + 2 0.1 = 1.9 M P a2
3. Expectation operator commutes with multiplication operator, if and only if the random variables are uncorrelated. If X and Y are two correlated random variables, then
the expectation of their product is given as:
hXY i = hXi hY i + COV [X; Y ]
(2.41)
32
4. If X and Y are two independent 7 random variables, then the variance of XY is given
as:
V ar[XY ] = hXi2 V ar[Y ] + hY i2 V ar[X] + V ar[X]V ar[Y ]
2.5
(2.42)
Random Processes and Fields

Random Process/Field is a collection of random variables indexed from a contin-
uum. Hence, a random process/ field is a function of random variables. If the continuum is
time, it is called a random process and if the continuum is space, it is called a random field.
The set of possible values that a random process/field can take is called the state space and
the continuum, where the process/field takes place is called the parameter space.
For a random process/field, for a finite set of fixed locations in the continuum,
there exist a finite set (of same size as finite set of fixed locations) of random variables e.g.
in Fig. 2.6 , for the shear modulus random field, for the set of fixed locations (z1 , z2 , ...., zn )
along the depth, there exist a set of random variables (G(z1 ), G(z2 ), ...., G(zn )), each corresponding to one fixed location in the space continuum. In other words, shear modulus
random field is a random function G(
z ), whose probabilistic characteristics changes as a
function of location in the continuum (depth).
Some observations:
When this random function (G(

z )) is specialized to a fixed location in the continuum
(e.g., at position zn ), it results in a random variable (G(zn )).
7
For dependent random variables refer to Goodman [32]
33
Figure 2.6: A realization of Shear Modulus Random Field
When this random function (G(

z )) is specialized to one of its sample values, it results in a deterministic function of location z. This deterministic function is called a
realization of the random field (refer to Fig. 2.6)
The complete probabilistic characteristic of any random process/field are described

either by joint characteristic function or the joint probability density function of the nvariable random variable, which collectively form the random field/process. Most common
random field/process is where the random variables are normally distributed. For example,
34
if we assume the the shear modulus random field is normally distributed then the joint
characteristic function can be written as:
i
(u1 , u2 , ...., un ) = e
n
X
i=1
1 X
{hhG(zi )ii ui }
hhG(zi ), G(zj )ii
2
i,j=1
(2.43)
Also, after taking the inverse Fourier transform of , the joint probability density function
can be written as:
f (G(z1 ), ......., G(zn )) =
|K 2 |1/2 (2)n/2
n
1 X 1
K2 (i, j) {G(zi ) hG(zi )i} {G(zj ) hG(zj )i}

2
i,j=1
e
(2.44)
where, K 2 is the covariance matrix given as:
Cov [G(z1 ); G(z1 )] Cov [G(z1 ); G(z2 )]
Cov [G(z2 ); G(z1 )] Cov [G(z2 ); G(z2 )]
K2 =
Cov [G(zn ); G(z1 )] Cov [G(zn ); G(z1 )]

and |K 2 | is the determinant of K 2 .
Cov [G(z1 ); G(zn )]
Cov [G(z2 ); G(z1 )]
Cov [G(zn ); G(zn )]
(2.45)
35
2.5.1
Classification of Random Processes/Fields

Random processes/fields can be classified in many different ways - based on state
and parameter spaces, and/or memory, and/or location dependance. In the following they
are described in details:
Classification according to State and Parameter Space

Both the state and parameter space could be discrete or continuous. The following
table shows the classification of some well-known random processes based on state and
parameter space.
Table 2.1: Classification of random processes based on state and parameter space
Discrete State
Continuous State
Discrete Parameter
Markov Chain
Time Series
Continuous Parameter
Poisson Process
Stochastic Differential Equation
Classification according to Memory

A random process/field is said to have no memory or to be purely random or to
be -correlated, if for any fixed locations ti , tj T, i 6= j, the random variables X(ti ) and
X(tj ) are independent to each other. Sometimes -correlated processes/fields are also called
white noise processes/fields.
In the memory hierarchy, the Markov process comes next. It is a random process,
where given its present, the future is independent of its past. An initial value problem or a
36
random process X(t), where
dX(t)
= aX(t) + (t)
dt
(2.46)
X(0) = x0
(2.47)
with an initial condition:
where, (t) is a white noise process and the coefficient a is deterministic is an example of
Markov process.
Other random processes/fields having a memory structure longer than Markovian are called non-Markovian processes/fields. Non-Markovian processes/fields arise while
dealing with stochastic differential equations with random coefficients e.g., boundary value
problems with stochastic coefficients.
Classification according to Location Dependance

The concept of location dependance of a random process/field is very similar to the
concept of homogeneity and heterogeneity of a (deterministic) spatial variable e.g., shear
modulus of a material. In case of a random field (stochastic spatial variable), homogeneity
signifies invariance of the joint probability distributions to any shift in space:
f (G(x1 ), G(x2 ), , G(xn )) = f (G(x1 + l), G(x2 + l), , G(xn + l))
(2.48)
where the joint probability distribution (f (, , ....)) of shear modulus random field (G),
indexed in space(x) continuum is invariant to any displacement l in the continuum. It is
37
important to note that if this property (Eq. (2.48)) holds for probability distribution, it will
also be valid for any cumulants (hh, , , ii) and moments (h, , , i) of the random field
e.g.,
hhG(x1 ), G(x2 ), , G(xn )ii = hhG(x1 + l), G(x2 + l), , G(xn + l)ii
(2.49)
and,
hG(x1 ), G(x2 ), , G(xn )i = hG(x1 + l), G(x2 + l), , G(xn + l)i
(2.50)
Eqs. (2.48), (2.49), and (2.50) can be visualized nicely, if one substitutes l = x1 (i.e., when
x1 is taken as the origin):
f (G(x1 ), G(x2 ), , G(xn )) = f (G(0), G(x2 x1 ), , G(xn x1 )),
(2.51)
hhG(x1 ), G(x2 ), , G(xn )ii = hhG(0), G(x2 x1 ), , G(xn x1 )ii
(2.52)
hG(x1 ), G(x2 ), , G(xn )i = hG(0), G(x2 x1 ), , G(xn x1 )i
(2.53)
and,
i.e., for homogeneous random fields, the joint probability density function, the covariance
function, and the product moments depend on spatial position of the field relative to one
38
another and not on the absolute positions. All other random fields, which do not satisfy
the homogeneity condition (Eq. (2.48)) are called heterogeneous random fields.
Another important definition to note is the second-order homogeneous random
field, where only the mean and autocovariance function are invariant to any shift is space:
hG(x)i = constant
(2.54)
hhG(x1 ), G(x2 )ii = Cov [x1 x2 ] = Cov [x2 x1 ]
(2.55)
and,
where, Cov is the autocovariance function and
hhG(x), G(x)ii = V ar [G(x)] <
(2.56)
Therefore, if we assume that the previously defined shear modulus Gaussian random field
(Eq. (2.43) or Eq. (2.44)) is homogeneous, then the covariance matrix defined in Eq. (2.45)
simplifies to:
2
Cov [z2 z1 ]
Cov [z2 z1 ]
2
K2 =
Cov [zn z1 ] Cov [z2 zn ]
Cov [zn z1 ]
Cov [z2 zn ]
(2.57)
39
where, Cov[zi zj ] = Cov[G(0); G(zi zj )] and 2 = V ar [G(0)], i, j = 1,....,n.
The concept of location dependance as discussed above for random fields is also
applicable to random processes e.g., earthquake displacement time history. However, for
random processes the terminologies are stationary (equivalent to homogeneity of random
field) and non-stationary (equivalent to heterogeneity of random field).
Classification according to Direction Dependance

This classification is only applicable to random fields and the concept is similar to
the concept of isotropy and anisotropy of a (deterministic) spatial variable. Mathematically
speaking, a 3-D random field (e.g., shear modulus of soil), defined in spherical coordinates,
is considered isotropic if the the joint cumulative distribution functions are invariant to
directional shifts in space:
F [G(k , k , r1 ), , G(k , k , rn )] = F [G(j , j , r1 ), , G(j , j , rn )]
(2.58)
for any k 6= j and any k 6= j . Also if the above equation (Eq. (2.58)) is valid, the
following statements about mean and cumulants are also valid:
hG(i , i , r)i = hG(j , j , r)i
(2.59)
hhG(i , i , r1 ), G(i , i , r2 )ii = hhG(j , j , r1 ), G(j , j , r2 )ii
(2.60)
and,
40
All other random fields, which do not satisfy isotropy condition (Eq. (2.58)) are called
anisotropic random fields. When a random field is homogeneous and isotropic, one can
easily derive, using Eqs. (2.49) and (2.60), that its covariance function (hh (xj ), (xk )ii)
is a real, deterministic scalar function of absolute distance between two random variables:
hhG(xi ), G(xj )ii = g(||xi xj ||)
(2.61)
In this context one should note that, though not many studies were undertaken to examine
direction dependances of soil properties random fields, it can be hypothesized that they are
anisotropic in nature due to layered nature of soils. Though it is site-specific, soils generally
exhibit a stronger correlation in the horizontal direction (cf. Fenton and Griffiths [25]).
2.5.2
Ergodicity
Loosely stated ergodicity of a random process/field suggests that the complete
probabilistic characteristics of a random field/process can be obtained from one realization

of the field/process. It is important to note that ergodicity of random field/process implies
homogeneity (stationarity) e.g., the heterogeneous random field z(x) = Y x + (x), where
Y is a random variable and (x) is an autocorrelated random field, is non-ergodic. This is
because in one realization of the field, there is only one value of Y . So to get the complete
probabilistic characteristics of z(x), one needs many realizations.
Mathematically speaking, a time-space homogeneous stochastic process (V (t, s), t, s
T xR3 ) is said to be ergodic relative to some domain D if for every g(V (t, s)) in D, the time-
41
space average is equal to the ensemble average with probability one (cf. Kavvas [46]):
) = hg(V (t, s))i
g(V
(2.62)
where, g represents the time-space average function g of the stochastic process V (t, s) in
the domain D.
A concise but very nice description of the concept of ergodicity, as it is applicable
to soil properties random field, is given by Baecher and Christian [6]. They debated, citing
Cressie [13], that though the sufficient condition for ergodicity of a Guassian random field
z(x), lim|||| Cz () = 0, where Cz () is the autocovariance function of the random field
z(x), can be checked empirically by inspecting the sample moments, the method can not be
philosophically correct as as all the common autocovariance functions used in geotechnical
soil modeling obey this condition. Citing Christakos [11], their conclusion was: since it is
very difficult to verify ergodicity for soil properties random field, they must be considered
a falsifiable hypothesis, and best way to prove it is through successful application of soil
properties random field.
2.6
Stochastic Calculus
Classical theory of stochastic calculus has been developed for random processes.
Though most of the developments can also be applied to random fields, there is no such
unified theory for random fields. Stochastic differential equations can be broadly classified
into two categories SDE with stochastic forcing and SDE with stochastic coefficients.
Most of the classical theory deals with SDE with stochastic forcings. There exists some approximate methods for the other kind of SDE i.e., SDE with stochastic coefficients. Though
42
SDE with stochastic coefficients is the main focus of this research, they will not be discussed
in this section. They will be discussed in particular case by case basis in the subsequent
chapters. This chapter will mainly outline the classical theory of stochastic calculus, which
is considered as foundations for developing approximate solution of SDE with stochastic
coefficient.
One of the first stochastic differential equations solved was that of Brownian motion, which is governed by Langevin equation:
dv
= F (x) v + (t)
dt
(2.63)
where the velocity (v) of a Brownian particle of mass m is expressed as a sum of viscous force
(which is proportional to the particle velocity (v)), systematic interaction force (F (x)) due
to intermolecular and intramolecular interactions, and a white noise term ((t)) arising due
to collision of atoms. Einstein [22] in his one the famous 1905 papers solved the equation in
terms of probability density of particle displacement (f (x)) and showed that the probability
density of particle displacement (f (x)) obeys a simple diffusion equation of the form:
f (x, t)
2 f (x, t)
=D
t
x2
(2.64)
where D is the diffusion coefficient. Later Fokker [28] in 1914, Planck [81] in 1917, and
Kolmogorov [52] in 1931, all working independently, showed that the addition of external
forces (such as gravity, elastic or magnetic attraction) into the equation of Brownian motion
results in probability density of particle displacement (f (x)) obeying an advection-diffusion
equation, which is known as Fokker-Planck-Kolmogorov (FPK) equation.
43
The classical approach to the second-order analysis of stochastic processes was
first proposed by Wiener (1923), who developed the relationship between autocorrelation
function and spectral density function of a stochastic process and that paved the way to the
solution of general stochastic differential equations (SDEs). In recognition of his work, his
construction of Brownian motion is often referred as Wiener process. Later It
o developed
theory of stochastic integration. His first paper on stochastic integration was published in
1944 [41].
In the following, effort has been made to outline the stochastic calculus in a systematic way - starting with mean square differentiation, and then integration and then
solution of SDEs. Finally, the Fokker-Planck-Kolmogorov equation and its relation with It
o
SDE.
2.6.1
Mean Square Convergence of a Random Function

A sequence of random variables {Xn } converges in mean square (m.s.) to a random
variable X as n if:
lim ||Xn X|| = 0

lim (Xn X)2 = 0
(2.65)
Hence, the necessary and sufficient condition for the existence of a m.s. limit X to the
sequence Xn , in terms of autocovariance function of the random process can be written as:
lim
n,n

Xn Xm X 2 = 0
lim
n,n

hXn Xm i = X 2
(2.66)
i.e., if the autocovariance function of any random process converges to a second-order moment then the random process has a m.s. limit.
44
2.6.2
Mean Square Derivatives
A Regular Random Function8 (RRF) X(t), t T, has a m.s. derivative X(t)

at t
if 9 :

1

lim [X(t + ) X(t)] X(t) = 0
0
(2.67)
A necessary and sufficient condition for m.s. differentiability of a RRF X at time t is that the
partial derivatives of its correlation function, ((t, s)/t, (t, s)/s, and 2 (t, s)/ts)
should exist at t=s. The properties of mean square (m.s.) derivatives of a RRF are:
If RRFs X1 (t) and X2 (t), t T , are m.s. differentiable at t T , then:
d
dX1 (t)
dX2 (t)
[aX1 (t) + bX2 (t)] = a
+b
dt
dt
dt
(2.68)
If RRF X(t) is m.s. differentiable at t and the deterministic function f(t) is differentiable in ordinary sense at t, then:
df (t)
dX(t)
d
[f (t)X(t)] =
X(t) + f (t)
dt
dt
dt
(2.69)
If Z(t) = X(t), where is a deterministic function of RRF X(t), then:

dZ(t)
d(X(t))
dX
=
=
dt
dt
X dt
(2.70)
Mean square differentiation and expectation operator commute:

8
9

p X1 (t1 ) q X2 (t2 )
r Xn (tn )
...
trn
tp1
tq2

p
q

=
...
hX1 (t1 ) X2 (t2 ) ... Xn (tn )i
t1
t2
tn
defined as random functions which have finite mean squares

In comparison, a deterministic function x(t) has a regular derivative at t if
lim
[X(t + ) X(t)] X(t)

=0
(2.71)
45
Hence for the mean:

dX(t)
dt
d
hX(t)i
dt
(2.72)
and for the covariance:

2.6.3
p X(t) q X(s)
tp
sq
p
q
hX(t)X(s)i =
p
q
(t, s)
(2.73)
Riemanian Stochastic Integral

Riemanian stochastic integral is the integral of a stochastic function with respect
to a deterministic integration measure. The properties of mean square (m.s.) Riemann

integral are:
Mean square continuity of a RRF X(t) on [a,b] implies m.s. Riemann integrability of
X(t) on [a,b]
If the m.s. Riemann integrals of X(t) and Z(t) exist on [a,b], then
(R)
[X(t) + Z(t)]dt = (R)

a
X(t)dt + (R)
a
Z(t)dt
(2.74)
If X(t) is m.s. continuous on [a,t] T, then

Y (t) = (R)
X(u)du
a
is m.s. continuous on T for all tT.

Leibnitz rule: If X(t) is m.s. Riemann integrable on T, then the m.s. derivative of
Z(t) = (R)
X(u)du
a
46
exists at all points t of m.s. continuity of X(t) and
dZ(t)
= X(t)
dt
when X(t) is m.s. continuous at tT.
Mean square integration by parts: If X(t) is m.s. differentiable on T, the deterministic
function g(t,s) is continuous on TT, g(t, s)/s exists on TT and if
Z(t) = (R)
g(t, s)
a
dX(s)
ds
ds
then,
Z(t) =
g(t, s)X(s)|ta
(R)
t
a
g(t, s)
X(s)ds
s
(2.75)
Fundamental theorem of m.s. calculus: If dX(t)/dt is m.s. Riemann integrable, for

tT the,
Z(t) = X(t) X(0) =
dX(u)
du
du
for (0,t)T
Mean of a m.s. Riemann integral:If the m.s. Riemann integral

Z(t) = (R)
X(u)du
0
exists for a RRF X(t), tT, then

hZ(t)i =
hX(u)i du
(2.76)
Correlation function of m.s. Riemann integral: If the m.s. Riemann integral

Z(t) = (R)
X(u)du
0
exists for a RRF X(t), tT, then

hZ(t)Z(s)i =
Z tZ
0
X X(u, v)dudv
0
(2.77)
47
2.6.4
RiemannStieltjes Stochastic Integral

Stochastic RiemannStieltjes integral arises while integrating with respect to a
stochastic integration measure e.g., Wiener process(W (t))
10 .
For example the following is
a RiemannStieltjes Stochastic Integral:
F (X(s), s)dW (s)
(2.78)
where F may be interpreted as a deterministic function of a stochastic function X(s) or

it may be interpreted as a stochastic function F (s). Solution of this type of equation,
especially when they are of multiplicative type, is the backbone of It
o calculus and the It
o
evaluation of this type of integral will be discussed in the following subsection.
2.6.5
It
o Stochastic Differential Equation
The following equation:
dxt = a(xt , t)dt + b(xt , t)dWt
(2.79)
where Wt is the Weiner process, is called It

o type equation, if for all t and t0 , the solution
10
Wiener process is a Gaussian process, having stationary, independent increments. It has the following
properties:
hW (t)|w0 ,t0 i = w0
V ar[W (t)] = [W (t) w0 ]2 = 2 (t t0 )

h[W (t) W (s)]|w0 ,t0 i = 0
[W (t) W (s)]2 |w0 ,t0 = 2 (t s)
Cov[W (t); W (s)] = h(W (t) w0 )(Ws w0 )i = 2 min(t t0 , s t0 )
48
xt obeys the stochastic integral equation:
xt = xt0 +
a(xs , s)ds + I
t0
b(xs , s)dWs
(2.80)
t0
On the r.h.s of Eq. (2.80), the first integral is of Riemann type and the second integral is of
multiplicative RiemannStieltjes type. One may note that I precedes the second integral.
That means It
o evaluation of the integral. The It
o evaluation of RiemannStieltjes integral
involves substitution of Weiner process (Wt ) by the dependent variable (xt ) and evaluation
R
of the integral ( b(t, xt )dt), which is a standard Reimanian integral) and assigning the value
of the integral to a function, g(t, xt ) and finding corresponding dg(t, xt ) from It

o formula,
which in 1-D takes the form:
dg(t, xt ) =
and then recognizing
evaluation of:
g
g
1 2g
(dxt )2
dt +
dxt +
t
xt
2 x2t
(2.81)
b(t, Wt )dWt in the expression for dg(t, xt ). For example the It

o
Ws dWs
0
can be obtained by first substituting Ws by xs . Then the integral g can be evaluated as:
g=
1
xdx = x2
2
and then from 1-D It

o formula (Eq. (2.81)):
d
1 2
x
2
1
= 0 + xdx + (dx)2
2
49
Now substituting W for x:
d
1 2
W
2
1
1
= 0 + Ws dWs + (dWs )2 = Ws dWs +
2
2
(since (dWs )2 = ds)
therefore,
I
2.6.6
t
0
1 2 t 1
W 0
Ws dWs =
2
2
ds =
0
1
1
Wt2 W02 t
2
2
FokkerPlanckKolmogorov Equation
It
o SDE, when transformed into probability density space, becomes a Fokker
PlanckKolmogorov equation. For example the It

o SDE in Eq. (2.79), in probability density
space takes the form:
1 2 2
p(x, t|x0 ,t0 ) =
[a(x, t)p(x, t|x0 ,t0 )] +
b (x, t)p(x, t|x0 ,t0 )
2
t
x
2 x
(2.82)
which is a Fokker-Planck-Kolmogorov equation. The advantage of solving the FokkerPlanck-Kolmogorov equation is that it is linear partial differential equation, even if the
equivalent It
o SDE in real space is nonlinear (e.g., Eq. (2.79)).
50
Part II
Uncertain and Spatially Uncertain

Material Properties
51
Chapter 3
Characterization & Quantification

of Uncertainties in Material
Properties
3.1
Introduction
Over the years engineers have developed several strategies for dealing with uncer-
tainty. They include: by ignoring or by being conservative or by using observational

method or by quantifying the uncertainty (cf. Baecher and Christian [12]). Quantification or mathematical description of uncertainty is usually done within the framework of
probability theory, although fuzzy sets (cf. Zadeh [105]), convex models (cf. Elishakoff [8]),
and interval arithmetic (cf. Moore [73]) have also been used in the past to describe uncertainty mathematically. In this chapter, based on an extensive literature review, character-
52
ization and quantification of material uncertainties are discussed with particular emphasis
on soil.
3.2
Classification of Uncertainties
Uncertainties can be broadly classified into two categories - aleatory and epis-
temic. Aleatory uncertainty is inherent variation associated with the physical system of
the environment (e.g., variation in external excitation, material properties etc.). It is also
known as irreducible uncertainty, variability and stochastic uncertainty. Epistemic uncertainty describes potential deficiency in any phase of the modeling process that is due to lack
of knowledge (e.g., poor understanding of mechanics etc.). It is also known as reducible
uncertainty, model form uncertainty and subjective uncertainty
Figure 3.1 illustrates my interpretation of the boundaries between deterministic,
and epistemic and aleatory uncertainty in relation to probabilistic simulations in geotechnical engineering. The only hard limit, as physics teaches us, is a macro scale interpretation of
the Heisenberg uncertainty principle (Heisenberg [37]), which stipulates that at the particle
level (particle of soil) one cannot obtain deterministic values for position and momentum at
the same time resulting from some deterministic or stochastic loading. The other boundaries
are not rigid. For example, epistemic uncertainty is sometimes traded=off for aleatory uncertainty as the mathematical tools for dealing with aleatory uncertainty is well developed.
In doing so, one doesnt reduce the total uncertainty but assumes that the total uncertainty
is irreducible, though some of it might be reducible. On the other hand, one could move
the boundary between the deterministic and epistemic uncertainty to the right (refer to
53
Figure 3.1: Graphical depiction of the deterministic, epistemic and aleatory uncertainty
related to geotechnical simulations. This is, in a sense, macro scale interpretation of Heisenberg uncertainty principle.
Fig. 3.1), by knowing any random variable (e.g., soil property) better e.g., by performing
more boreholes.
3.3
Uncertainties in Soil Properties

In recent years, with technological advancements, our ability to measure soil prop-
erties has expanded considerably. A wide varieties of laboratory and field tests are being
used in research and practice for evaluation of soil properties. However, due to different
types of variabilities, the measured soil properties show a lot of variation depending upon
the property and test method. Over the years researchers have quantified and collected
typical variations1 of different soil properties, ranging from consolidation parameters (refer
to Table 3.1), laboratory measured strength properties (refer to Table 3.2) to in-situ prop1
Total variations, includes natural variability, testing and transformation error
54
Table 3.1: Representative values of variabilities in consolidation parameters (after Baecher
and Christian [6])
Parameter Type of Soil
COV(%)
Cc , Cr
Bangkok Clay
20
Various
25-20
Dredge Spolis
35
Gulf of Mexico Clay
25-28
Cv
Ariake Clay
10
Singapore Clay
17
Bangkok Clay
16
Table 3.2: Representative values of variabilities in laboratory measured effective friction
angle (after Baecher and Christian [6])
Soil type
COV (%)
Various soils
9
Clay
40
Alluvial
16
Sands
2-5
Tailings
5-20
erties (refer to Table 3.3). The statistical distributional properties, i.e., which are best-fit
probability density functions to different soil properties, has been suggested by Lacasse
and Nadim [56] (Table 3.4). One should note that the tabulated statistical properties are
mainly for illustrative purposes, as the variabilities vary from site to site. In an ideal case
one should plan to get the site specific statistical properties from detailed soil investigation
for each site. However, the above tables give tentative idea that how much variability we
should expect in a particular soil property. Also in small projects (where very few soil tests
are planned) they act us general guidelines. This aspect is discussed in more details in
Section 3.5.
The uncertainties associated with soil properties are generally attributed to three
primary sources: (1) the natural variability of soil deposits; (2) testing error; and (3)
55
Table 3.3: Representative values of variabilities

(after Phoon and Kulhawy [79])
Test type Property
Soil type
CPT
qT
Clay
qc
Clay
qc
Sand
VST
su
Clay
SPT
N
Clay and Sand
DMT
A Reading
Clay
A Reading
Sand
B Reading
Clay
B Reading
Sand
ID
Sand
KD
Sand
ED
Sand
PMT
pL
Clay
pL
Sand
EP M T
Sand
in some common in-situ soil properties

Mean
0.5-2.5
0.5-2
0.5-30
5-400
10-70
100-450
60-1300
500-880
350-2400
1-8
2-30
10-50
400-2800
1600-3500
5-15
Units
M N/m2
M N/m2
M N/m2
kN/m2
blows/ft
kN/m2
kN/m2
kN/m2
kN/m2
M N/m2
kN/m2
kN/m2
M N/m2
COV(%)
< 20
20-40
20-60
10-40
25-50
10-35
20-50
10-35
20-50
20-60
20-60
15-65
10-35
20-50
15-65
Table 3.4: Best fitting probability density functions (PDFs) for various soil properties (after
Lacasse and Nadim [56])
Soil property
Soil type
Mean
COV(%) PDF
Cone resistance
Sandy clay
LN
Clay
N/LN
Undrained shear strength Clay (triaxial)
5-20
LN
Clay (index su )
10-35
LN
Clayey silt
5-15
N
Ratio su /v 0
Clay
5-15
N/LN
Plastic Limit
Clay
0.13-0.23
3-20
N
Liquid Limit
Clay
0.30-0.80
3-20
N
3
Submerged unit weight
All soils
5-11 kN/m
0-10
N
Friction angle
Sand
2-5
N
Void ratio, Porosity
All soils
7-30
N
Overconsolidation ratio
Clay
10-35
N/LN
56
transformation error (cf. Baecher and Christian [6]). The first source of uncertainty is of
aleatory type, which represents the natural randomness of a property and it can not be
reduced or eliminated. In this context one may note that in principle the soil properties
are not random but uncertain. One may not know the soil properties at every point in the
deposit, however, they are knowable and hence, by definition the variability in soil property
is of epistemic type. But one transfers the uncertainty in soil deposit into aleatory column
because it is easier to handle mathematically. By trading off epistemic uncertainties with
aleatory uncertainties one doesnt lessen the total uncertainties but creates an irreducible
amount of uncertainties in the prediction. The other two sources of uncertainties are of
also of epistemic type, which represents a range of values that can be reduced by collecting more information, improving measurement methods and/or improving the calculation
methods (cf. Lacassae and Nadim [56]). Testing uncertainty is due to imperfections of an
instrument or of a method to register a quantity and can be evaluated from data provided
by the manufacturer or from laboratory tests. The standard way to measure the testing
uncertainty is to make repeated measurements of the same property of the same soil and
then characterize how much the property varies (cf. Baecher and Christian [6]). Hammitt
[36] (refer to Table 3.5) reported the testing uncertainties of the index properties tests, from
a series of comparative laboratory tests conducted by the American Council of Independent
Laboratories. But for destructive tests direct measure of testing uncertainty poses a problem as the specimen is destroyed as part of the test and the only way to directly measure
the testing error for destructive tests is to repeat the test on similar (because of inherent
soil variability) and not same soil. Phoon and Kulhawy [79] quantified the testing uncer-
57
Table 3.5: Representative values of testing errors of soil index properties (after Hammitt
[36], reproduced from the book by Baecher and Christian [6])
Highly Plastic Soil
Medium Plastic Soil
Low Plastic Soil
Type of test
Mean Std. Deviation Mean Std. Deviation Mean Std. Deviation
LL
54.3
5.4
32.7
2.3
27
1.7
PL
22.2
3.4
22.4
2.8
23.6
2.4
PI
32
5.7
10.4
3.6
3.8
2.1
Specific Gravity 2.63
0.115
2.66
0.060
2.69
0.054
Table 3.6: Representative testing error of some laboratory tests that evaluates strength
properties (after Phoon and Kulhawy [79])
Property
su (Triaxial Compression) (kN/m2 )
su (Direct Shear) (kN/m2 )
su (Lab. Vane Shear) (kN/m2 )
(Triaxial Compression) (0 )
(Direct Shear) (0 )
(Direct Shear) (0 )
Soil
type
Clay, Silt
Clay, Silt
Clay
Clay, Silt
Clay, Silt
Sand
No. of
data groups
11
2
4
5
2
No. of tests
per group
13
15
10
11
26
Mean
125
119
29
19.1
33.3
32.7
COV (%)
19
20
13
24
13
14
tainties for some common laboratory and field tests (refer to Tables 3.6 and 3.7), assuming
that the testing error can be logically determined directly by analyzing the variation of the
results obtained by a representative group of soil testing companies performing the same
test on nominally identical soil samples. Marosi and Hiltunen [66], recently, quantified
the testing uncertainty of SASW method. An indirect measure of quantifying the testing
uncertainty is described in Chapter 4.
Transformation uncertainty is introduced when field or laboratory measured soil
properties (e.g., SPT N-value) are transformed to design soil properties (e.g., Youngs modulus). Phoon and Kulhawy [80] described some of the common transformation equations
and reported their transformation uncertainties. Following Phoon and Kulhawy [80], some
of the strength properties correlation uncertainties are tabulated in Table 3.8.
58
Table 3.7: Representative testing error of some field tests (after Phoon and Kulhawy [79])
Test
Standard Penetration Test
Mechanical Cone Penetration Test
Electric Cone Penetration Test
Vane Shear Test
Dilatometer Test
Pressuremeter Test, prebored
Self-boring Pressuremeter Test
COV (%)
15-45
15-25
5-15
10-20
5-15
10-20
15-25
Table 3.8: Transformation uncertainties of some common strength property correlations(after Phoon and Kulhawy [80])
Type of transformation
CPT to su (CIUC)
CPT to su (UU)
SPT to su (UU)
CPT to (TC)
PI to (constant volume)
SPT to E (PMT)
3.4
transformation equation
su = (0.0789 + )(qT v0 )
su = (0.0512 + )(qT v0 )
su = 0.29pa N0.72 10
qc
p
a
= 17.6 + 11log10
r v0 +
pa
sin
() = 0.8 + 0.094loge (P I) +
log10
E
Pa
= log10 (19.3) + 0.63log10 N +
Transformation
uncertainty (SD )
0.0276
0.014
0.15
2.80
0.070
0.37
Uncertain Spatial Variability

For all materials, uncertainty is a spatial property, rather than a point property.
This is particularly true for soils. Soils are mainly formed by some weathering process,
e.g., river flow etc., which transport the soil particles to their present location. And the
spatial uncertain property of any soil is function of the weathering process through which
it was formed. Under the framework of probability theory, uncertain spatial variability
of soil deposit is modeled as a random field, where the information about autocovariance
59
function and correlation length or scale of fluctuation are also needed in addition to mean
and COV to characterize the uncertainties up to second-order. The concept of random field
modeling of soil properties is little counter-intuitive. In principle there is nothing random
about any soil property at any site. We all know that we could excavate any entire site
and obtain the soil properties at each point of the 3-D continuum. However, that would be
totally impractical. So we have to live with soil properties at some specific points (borehole
locations) and we have to predict (statistical prediction, because deterministic might not
be appropriate) the soil properties at the intermediate locations. Thats what exactly the
random field modeling of soil property does. In random field modeling, the soil property
is modeled by a set of correlated random variables indexed in the space continuum. So
at each borehole location, the measured soil property becomes a realization of the random
field. Based on a number (the more the better) of realizations i.e., borehole measurements,
the autocovariance function and the correlation length of the random field soil property
can be evaluated. There are different methods for estimating the correlation length. They
are discussed in Chapter 4. Phoon and Kulhawy [80], based on an approximate method
(correlation length = 0.8 x (the average distance between intersections of the fluctuating
property and its trend function )), reported typical scale of fluctuation of some common soil
properties (refer to Table 3.9) It can be seen that in Table 3.9 the vertical and horizontal
scale of fluctuation has been reported separately. This is because, due to formation by
layers, soil is best described as cross-anisotropic material. It can also be seen from the
Table 3.9 that the horizontal correlation length of soil tends to be higher than vertical
correlation length. This is self-explainable considering the layering formation of soil.
60
Table 3.9: Representative scale of fluctuation of some common soil properties (after Phoon
and Kulhawy [79])
Number of
Property
Soil type
studies
Scale of fluctuation (m)
Vertical fluctuation
su
Clay
5
0.8-6.1
qc
Sand, Clay
7
0.1-2.2
qT
Clay
10
0.2-0.5
su (VST)
Clay
6
2-6.2
N
Sand
1
2.4
wn
Clay, Loam
3
1.6-12.7
wL
Clay, Loam
2
1.6-8.7
Clay, Loam
2
2.4-7.9
Horizontal fluctuation
qc
Sand, Clay
11
3-80
qT
Clay
2
23-66
su (VST)
Clay
3
46-60
wn
Clay
1
170
Quantification of soil correlation length can sometimes be very tricky as soil properties vary at many scales. Inspected closely, soil can change significantly within a few
meters relative to the few meters considered e.g., variation within homogeneous deposit,
as shown in Fig. 3.2 Typical correlation length for soil depends on the distance over which
it is estimated. Sampling soil properties every 5 cm over 2 m will likely yield a correlation
length of about 20 cm. On the other hand sampling every 1 km over 1000 km will likely
yield an estimate of about 200 km. [33]. Thus soil can have lingering correlations over
entire continent. Hence the domain size over which the correlation is sought is significant
and hence the ratio of domain size (a) over correlation length (b) is more important than
the absolute value of correlation length. Typical a/b ratio found in literature for soil varies
between 1/10 and 1/5 (cf. Griffiths et al. [33]).
61
Figure 3.2: Measured values of mechanical properties of soil from Mexico city, typical soft
spot (after Baecher and Christian [6])
3.5
Probabilistic Geotechnical Site Characterization

For complete statistical characterization of field soil for a geotechnical project a
large amount of data must be collected. This is especially true for accurate estimation of
soil covariance structure. However, due to limited budgets for most of the projects, only a
small number of data are collected.
Lets suppose for a particular geotechnical site enough data are available for a site-
62
specific statistical analysis. For this case, the approach to probabilistic site characterization
could be to estimate the mean and variance from site specific data and the covariance
structure from a more detailed analysis of similar sites as the mean and the variance of
a soil property can usually be established reasonable accurately with only a statistically
small (but large physically) number of samples. However, for accurate estimation of entire
covariance structure one needs a vast amount of data which is usually not achievable in
todays projects due to budget restrictions. The idea behind using correlation structure
from similar sites stems from the fact that the correlation structure is external - more
related to the formation process of soil; that is, the correlation between soil properties at
two disjoint points will be related to where the materials making up the soil at two points
originated and to common weathering processes experienced at the two points (c.f. Fenton
[23]). Such an external 3-D random field correlation model was established by Fenton [24]
by statistically analyzing an extensive spatially distributed set of cone penetration tests
(CPT) gathered at the site of the Oslo Main Airport, Gardermoen, located to the north of
the city of Oslo.
On the other hand where the soil data at a particular site are too limited for any
meaningful statistical analysis, only the mean value could be obtained from the site specific
data and the COV could be assumed from published data for similar type of soil e.g.,
from the tabulated data in Section 3.3, which is based on extensive literature review. The
correlation structure could be assumed as before i.e. from a detailed study on a similarly
formed soil site or could be assumed from the general guideline (refer to Table 3.9).
For the cases where limited data are available to carry out statistical analysis,
63
a more robust estimate of geotechnical variability can be obtained by combining the sitespecific data with general guidelines using Bayesian updating techniques.
64
Chapter 4
Random Field Modeling of

Uncertain Material Properties
4.1
Introduction
In the last chapter, we have noted some valuable works in quantifying uncertainties
in material properties for geotechnical engineering applications. Most of the materials,

in particular soils are spatially correlated because of the formation process. Hence, the
uncertain material properties are generally best quantified as random fields. The typical
COVs and correlation lengths of some important soil properties have been tabulated in the
last chapter. In absence of any site-specific data, they can be used as a general guideline.
In this chapter, focus has been made in describing the different available techniques
in modeling soil properties as random field from site-specific data. This chapter mainly
builds on following the works of DeGroot and Baecher [18] and Fenton [23, 24]. For a detailed
65
conceptual understanding, the readers are encouraged to refer to books by Christakos [11],
Vanmarcke [97], and Baecher and Christian [6].
4.2
Basic Concept
Uncertain spatial variation is generally described as a sum of a deterministic trend
(T (X)) and uncertain residual ((X)) about the trend:

Y (X) = T (X) + (X)
(4.1)
Uncertain residual ((X)) is composed mainly of two components - true spatial uncertainty
(r (X)) and measurement error/testing uncertainty (n (X)). Hence Eq. (4.2) can be rewritten as:
Y (X) = T (X) + r (X) + n (X)
(4.2)
The uncertain spatial variation of shear modulus (G(z)) as defined in Chapter 2 (refer to
Fig. 2.6) can then be mathematically written as:
G(z) = G(z)
+ R(z)
= G(z)
+ Gr (z) + Gn (z)
(4.3)
(4.4)
which is graphically shown in Fig. 4.1, R(z) being the total residual spatial uncertainty
about the trend, Gr (z) being the true residual spatial uncertainty of shear modulus about
the trend and Gn (z) being the measurement error of the testing process. The deterministic
trend is usually estimated by regression analysis. The zero mean residual random field
(R(z)) is completely defined by its covariance (autocovariance) structure. For homogeneous
(stationary) fields, where the covariances are function of the lag (separation) distance r
66
Figure 4.1: Shear modulus random field: Trend and residual around trend
between random variables, the autocovariance structure can be conveniently described by

a function, known as autocovariance function (C(r)), of lag distance r between the random
variables. The autocovariance function (C(r)) can be expressed as:
C(r) = Cr (r) + Cn (r)
(4.5)
where Cr (r) is the autocovariance function of true residual spatial uncertainty and Cr (r)
is the autocovariance function of measurement error. True spatial uncertainty is generally
modeled as zero mean stationary but correlated random field. On the other hand, the
67
measurement error, which is a systematic error, is best modeled (cf. DeGroot and Baecher
[18]) as zero mean stationary and uncorrelated random field - autocovariance function is a
spike (equal to measurement error variance) at r = 0 and zero elsewhere.
The estimation of this spatial covariance model is the main focus of this chapter. Broadly, the spatial covariance models can be classified into two categories - finite
scale (short memory, finite correlation length) or fractal (long memory, infinite correlation
length). In general, for most of the materials, the residual covariance is high at small separation distance and it reduces rapidly as the separation distance increases. Hence correlation
length, which is defined as the area under the correlation function (cf. Fenton [23]), is
considered finite. In practice, some common finite scale autocovariance models are:
Exponential (first-order Markov):
C(r) = 2 er/r0
(4.6)
Squared-Exponential:
C(r) = 2 e(r/r0 )
(4.7)
Spherical:

C(r) = 2 1 3r/2r0 + r3 /2r03
(4.8)
C(r) = 2 (1 rr0 )
(4.9)
C(r) = 2 (Sin r)/r
(4.10)
Triangular (binary noise):
Sine (band-limited white noise):
68
Linear-Exponential (second-order Markov):
(1)
(2)
C(r) = 2 (1 + r0 r)er/r0
(4.11)
where 2 is the random field variance and r0 is the correlation length. For soils, the most
commons are exponential and square exponential (cf. DeGroot and Baecher [18]). One
the other hand, fractal model has very large correlation length. In other words, for fractal
model correlation remains significant over a very large distance. This is particularly true
for soils, especially in the horizontal direction. This is because of its formation process the eroded constitutive elements of the same rock is carried over by river to a very long
distance (cf. Fenton [23, 24]).
4.2.1
Finite Scale Model

The most common method of estimation of autocovariance structure of finite scale
models is method of moments (cf. Montgomery and Runger [72], Rose and Smith [84]).
For the shear modulus random field, the covariance function C(r)
can be obtained using
method of moments estimator as (DeGroot and Baecher [18], Fenton [23, 24], Baecher and
Christian [6]):
nj+1
1 X
C(rj ) =
(G(zi ) G(z))(G(z
i+j1 ) G(z))
n
j=1,2,...,n
(4.12)
i=1
and the form of autocovariance function is selected (usually one of the six described above)
by fitting the function to the method of moment estimate by inspection. One may note
the use of hat with the estimator. This is because the estimate (which estimates the
population statistics based on sample statistics) itself is a random variable, having its own
69
mean and standard deviation, with the mean usually reported. The correlation coefficient
can be obtained by normalizing the covariance as:
(rj ) =
j)
C(r
C(0)
(4.13)
One may note that the covariance estimates computed using Eq. (4.12) comprise of both
true spatial covariance as well as measurement error covariance. In separating the true
spatial covariance and measurement error covariance, DeGroot and Baecher [18] proposed
that the magnitude of the fitted C(r) function at r=0 represents the variance of true spatial
variability of the material property random field (G(z)) and the measurement error variance
can be indirectly estimated as1 :
Cn (r)|r=0 = C(r)|
r=0 Cr (r)|r=0
(4.14)
The method of moment estimator of autocovariance function (C(r))

is consistent
but only asymptotically unbiased. Also, there is a subjectivity to this approach as it involves
fitting a function to the method of moment estimates.
Another widely popular method, which has asymptotically many desirable properties and which doesnt require subjectivity of fitting function by inspection, is the method
of maximum likelihood (cf. Montgomery and Runger [72], Rose and Smith [84]). According
to maximum likelihood method in the space domain (Mardia and Marshall [65], DeGroot
and Baecher [18], Fenton [23, 24], Baecher and Christian [6]), one can estimate the unknown
parameters of a random field model (mean, variance, and correlation length) by assuming a
joint distribution (say Gaussian) with an assumed covariance function model (say first-order
1
It might be noted that covariance function of measurement noise is a spike (equal to the variance of
measurement error) at r=0 and zero elsewhere.
70
Markov) and then maximizing the likelihood (or log-likelihood) of observing the spatial data
under the assumed joint distribution. Mathematically, the log-likelihood function can be
written, e.g., for the normally distributed shear modulus random field as defined previously,
as:
1
n
L(G|) = ln(2) ln||
2
2
GG
T

()1 G G
2
2G
(4.15)
where, G = [G(z1 ), G(z2 ), ....., G(zn )]T is the given set of observations (realization of the
shear modulus random field), is the assumed correlation matrix, which is a function
of lag distance and correlation length, r0 (e.g., for first-order Markov correlation model
2 , r0 ]T is the vector of unknown parameters, G
is the spaij = e|zi zj |/r0 ), = [G,
G
tially varying deterministic trend (usually modeled as a lower-order polynomially (usually
= z, where = [1, 1 , ..., q ] is a vector of the unknown
linearly) varying with depth: G
coefficients of the polynomial and z is a n x q coordinate matrix of the sample locations, q
being the degree of the assumed polynomial and n being the number of sampling location),
2 is the variance of shear modulus random field.
G
The maximum of Eq. (4.15), in principle. can be found by differentiating with

respect to the unknown parameter vector () and setting it to zero. But there are some
computational issues (discussed in details by Mardia and Marshall [65] and Fenton [23]).
The maximum likelihood estimators are asymptotically normally distributed, unbiased. They are also weakly consistent with a covariance structure equal to the inverse of
Fisher information matrix (cf. Rose and Smith [84], DeGroot and Baecher [18])
71
4.2.2
Fractal Model
Fenton [23] showed that the finite scale model estimators show large discrepancy
between estimated and true correlation length when the material properties correlation
length is close to the domain length and proposed use of fractal method in dealing with
those cases. However, the problem with fractal model (e.g., 1/f-type noise process with
power spectral density P () ) is that it has physically unrealizable infinite variance.
When 0 < 1, the infinite variance contributions come from high frequencies (so the
field becomes homogeneous) and when > 1, the infinite variance contributions come from
low frequencies (so the field becomes non-homogeneous).
In rendering the fractal model useful for practical applications Fenton [23] suggested use of either upper cut-off frequency (when 0 < 1) or lower cut-off frequency
(when > 1) or both (when = 1). In addition to making the random field variance finite,
the cut-off frequencies make the field homogeneous. Among different available methods
Fenton [23] suggested periodogram approach in estimating the statistical parameters of the
fractal model, which is defined in power spectral density space as:
P () =
P0
0<
(4.16)
The log-likelihood of seeing the periodogram estimates Pj = P (j ), j = 1, 2, ..., k, where

k = (n 1)/2, and j = 2j/D, D being the domain length and n being the number of
measured locations is:
L(P |) = k lnP0 +
k
X
j=1
ln j
k
1 X
j Pj
P0
j=1
(4.17)
72
where = [P0 , ] is the unknown parameter vector, which is estimated by maximizing the
log-likelihood equation (Eq. (4.17) as follows:
k
1X
P0 =
Pj j
k
(4.18)
j=1
and the estimate for by solving the following equation:

j=1 Pj j ln
k
1 X
Pj j
k
j=1
Pk
4.3
k
X
ln j = 0
(4.19)
j=1
Example Estimation of Statistical Parameters

In this section, the CPT data [99], collected by the USGS Western Earthquake
Hazards Team for miscellaneous field and project investigations in Alameda County, CA,
is analyzed for vertical spatial variability to a depth of 13.5m (with top 1m removed, so
effective soil depth is 12.5m). 16 CPT soundings was considered over an area of approximately 7 KM 2 as shown in Fig. 4.2. The site is sloping from east to west. However, for
illustration purpose the site is considered horizontal. The subsoil to a depth of 13.5m is
mostly composed of soft clay with lenses of stiff clay and sand. The soil classifications at
the sounding locations are arranged from east to west and is shown in Fig. 4.3.
The vertical spatial variability of CPT tip resistance (qT ) has been modeled as 1D homogeneous random field, using both finite scale (Gauss-Markov) and fractal (1/f-type
noise) model as described in Section 4.2. A typical sounding (measurement at borehole 1)
of qT is shown in Fig. 4.4.
73
SN Coordinate (m)
3500
1
2
3000 4
3
2500
5
2000
7
8
1500
10
11
1000
13
14
12
15
500
17
16
18
250
750
1250
2000
WE Coordinate (m)
Figure 4.2: CPT Sounding locations
The method of moment estimated and finite scale Gauss-Markov maximum likelihood autocovariances for Borehole 1 sounding (refer to Fig. 4.4) are shown in Fig. 4.5.
The autocorrelation, which is the normalized form of autocovariance, is shown in Fig. 4.6
as estimated using both the methods from borehole 1 sounding.
Maximum likelihood estimated constant mean Gauss-Markov model statistical parameters for all the 16 soundings are tabulated in Table 4.1, along with the mean and
standard deviations of the estimates. A simulated realization of the resulting random field
(using the mean values of the estimates) is shown in Fig. 4.7(b). It was simulated using the
Figure 4.3: East-West soil profile interpreted from CPT soundings

74
75
q (MPa)
T
15
30
Depth (m)
4
6
8
10
12
Figure 4.4: Typical qT data: Borehole 1 sounding
ML estimate
Autocovariance (MPa^2)
Method of momemt estimate

20
10
2
10
10
12
Lag Distance (m)
20
Figure 4.5: Maximum likelihood estimated Gauss-Markov autocovariance function along
with method of moment estimate (for borehole 1 sounding)
76
ML estimate
Method of moment estimate
Autocorrelation
0.75
0.5
0.25
2
0.25
10
12
Lag Distance (m)
0.5
0.75
Figure 4.6: Maximum likelihood estimated Gauss-Markov autocorrelation function along

with method of moment estimate (for borehole 1 sounding)
Table 4.1: Maximum likelihood estimated

Estimated Variance
Borehole Method of Maximum
No.
Moment
Likelihood
1
25.078
24.94
2
6.87
7.2
3
3.76
4.42
4
15.22
14.74
5
11.72
9.56
6
27.52
29.54
7
11.71
4.25
9
6.41
5.93
10
20.06
23.18
11
48.19
44.92
12
10.69
3.53
13
66.13
57.43
14
44.92
42.16
15
31.6
31.82
16
23.44
25.09
18
93.97
82.08
Mean
SD
-
constant mean Gauss-Markov model parameters

Interpreted Variance
True
Testing
Correlation
Spatial
Error
Mean
Length
24.94
0.138
6.41
0.53
7.2
0
4.8
0.6
4.42
0
4.42
0.21
14.74
0.48
3.97
0.72
9.56
2.16
3.19
0.39
29.54
0
6.39
0.33
4.25
7.46
2.31
0.41
5.93
0.48
3.9
0.38
23.18
0
4.3
0.26
44.92
3.27
4.15
1.51
3.53
7.16
2.2
0.39
57.43
8.7
7.4
0.92
42.16
2.76
7.04
0.51
31.82
0
6.3
0.59
25.09
0
6.01
0.42
82.08
11.89
7.06
1.64
25.67
2.78
4.99
0.61
22.22
3.86
1.69
0.41
77
Cholesky decomposition of covariance matrix as follows (cf. Mardia and Marshall [65]):
qT = hqT i + (L)qT Z
(4.20)
where hqT i is the mean vector (assumed constant with depth in this example), (K 2 )qT =
(L)qT (L )qT is the Cholesky decomposition of the covariance matrix ((K 2 )qT ), and Z is
the zero mean unit variance normally distributed random matrix (Z N (0, I), I being
the identity matrix). Fig. 4.7(a) shows one of the measured realizations. Note that it is
not expected to be identical, except having same statistical nature, as they are merely two
possible realizations of the same random field.
q (MPa)
T
40
20
Depth (m)
Depth (m)
20
q (MPa)
T
6
8
6
8
10
10
12
12
(a)
40
(b)
Figure 4.7: (a) Measured (at borehole 2) and (b) Simulated (finite scale Gauss-Markov
model) realizations
For comparison, the same set of data (16 CPT soundings over approximately
78
7 KM 2 ) was analyzed using fractal method as described in Subsection 4.2.2. It was assumed
that the random field mean and variance was calculated by other methods e.g. a constant
(or any polynomially varying with depth) mean deterministic trend could be extracted
using global regression analysis over 16 soundings as shown in Fig. 4.8 along with measured
data. And the field variance could be estimated by method of moment using Eq. (4.12) as
tabulated in Table 4.1.
The periodogram for the borehole 1 sounding is shown in Fig. 4.9 and corresponding the 1/f-type noise model with lower cut-off frequency is shown in Fig. 4.10.
The
statistical parameters (0 , , and P0 ) required to define the 1/f-type noise model with
lower cut-off frequency given as:
P0 /0
P () =
P0 /
if 0 < 0
(4.21)
if > 0
was estimated using maximum likelihood technique (as discussed in Subsection 4.2.2) and
is tabulated in Table 4.2 for all the 16 CPT soundings, along with the mean and standard deviation of the estimates. The equivalent correlation lengths (computed as r0 =
P0 /
2 0 , refer to Fenton [24]) are also tabulated. Having defined the fractal model with
cut-off frequency(ies), the autocovariance function can be easily computed using the WeinerKhintchine relationship as (cf. Fenton [24]):
C(r) =
=
Z0 0
0
P ()cos(r)d
P0
cos(r)d +
0
P0
cos(r)d
(4.22)
The autocovariance function as computed using Eq. (4.22) and borehole 1 estimates is
plotted in Fig. 4.11.
79
q
0
20
(MPa)
40
Depth (m)
10
12
Figure 4.8: Deterministic Trend as obtained through global regression over 16 CPT
soundings
80
Spectral Power
10
1
0.1
0.01
0.1
10
100
Frequency (rad/sec)
Figure 4.9: Periodogram of borehole 1 sounding
Spectral Power
10
1
0.1
0.01
0.1
10
100
Frequency (rad/sec)
Figure 4.10: Maximum likelihood estimated fractal (1/f -type noise with lower cut-off frequency) power spectral density function corresponding to borehole 1 sounding
81
Table 4.2: Maximum likelihood estimated constant mean fractal (1/f-type noise with lower
cut-off frequency) model parameters obtained using periodogram approach
BH
0 (rad/sec)
P0
Equivalent Correlation Length, r0 (m)
1
2.08
1.78 19.42
0.66
2
3.13
2.09 12.51
0.52
3
9.36
1.27 1.49
0.07
4
1.33
1.76 8.18
1.01
5
1.66
1.64 6.37
0.73
6
2.04
1.39 10.42
0.43
7
0.69
1.77 3.85
1.98
9
2.99
1.86
7.7
0.48
10
4.39
1.52 15.19
0.24
11
1.14
2.06 28.62
1.41
12
1.65
2.14 10.14
1.01
13
1.93
2.13 74.46
0.86
14
1.42
1.64 22.22
0.86
15
1.91
1.8 23.85
0.73
16
2.02
1.59 13.36
0.58
18
1.09
2.02 52.29
1.44
Mean
2.43
1.78 19.38
0.81
SD
2.06
0.26 19.13
0.48
Autocovariance (MPa^2)
25
20
15
10
5
10
12
Lag Distance (m)

Figure 4.11: Fractal (1/f -type noise with lower cut-off frequency) autocovariance function
for borehole 1 sounding
82
Part III
Material (Constitutive) Level

Stochastic Simulation:
Probabilistic ElastoPlasticity
83
Chapter 5
Probabilistic ElastoPlasticity:
Theory
5.1
Introduction
Advanced elastoplasticity based constitutive models, when properly calibrated,
are very accurate in capturing important aspects of material behavior within continuum.
But all materials, in particular geomaterials (soil, rock, concrete, powder, bone etc.) behaviors are uncertain due to inherent spatial uncertainties and testing and transformation
uncertainties (as discussed in details in Part II of this dissertation). These uncertainties in
material properties, needed for calibrating constitutive models, could outweigh the advantages gained by using advanced constitutive models. For example, Fig. 5.1 shows a schematic
of anticipated influence of material uncertainties on a bilinear elastic-plastic stress-strain
behavior. Depending on uncertainties in material property(ies) and interaction between
84
them, the behavior of the same material could be very different. This could be even more
complicated for non-linear materials.
Figure 5.1: Anticipated Influence of Material Fluctuations on Stress-Strain Behavior
The uncertainties in material properties are inevitable in real life and the best way
to deal them with is to account for them in our modeling and simulation. In traditional
deterministic constitutive modeling, material models are calibrated against set of experimental data. Though the experimental data generally exhibit some statistical distribution,
the models are usually calibrated against the mean of the data and the uncertainties with
respect to the mean are neglected. Hence, when these constitutive models are used for
further modeling (e.g., for modeling the behavior of solids and structures made with those
materials), the uncertainties in material properties are lost from the simulation results.
The strategy for propagating uncertainties through governing differential equa-
85
tions can be broadly classified into two categories - stochastic differential equation (SDE)
with random forcing and SDE with random coefficient. For SDE with random forcing, when
the governing equation is of It
o type (refer to Subsection 2.6.5, for details refer to Gardiner
[30]), highly developed mathematical theory exists - the solution is a Markov process and the
probability density of the solution obeys a Fokker-Planck-Kolmogorov (FPK) partial differential equation (refer to Subsection 2.6.6, for details refer to Gardiner [30]). The advantage
of the FPK equation is that it transforms the original nonlinear SDE in real space into a
linear deterministic PDE in probability density space. On the other hand, for SDEs with
random coefficients, which is of immediate interest of this study, approximate numerical
methods (e.g., perturbation method (cf. Klieber and Hien [51])) are very popular especially
for nonlinear problems. Monte Carlo method, which is based on law of large numbers, is
also very popular. It is carried out sequentially by generating randomized parameters and
using them as input into a set of deterministic models. This set of models is then used in
a multitude of simulations to determine the value of desired response function. Finally the
statistics of the response variable are quantified. The advantage of Monte Carlo method is
that accurate solutions can be obtained for any problem (either linear or nonlinear) whose
deterministic solution (either analytical or numerical) is known. The major drawback of
Monte Carlo method is that it is computationally very expensive. This is even more the
case for problems where no closed-form solution exists for solving the deterministic problem. On the other hand, perturbation method, applicable to both linear and non-linear
stochastic problems, uses Taylor series expansion with respect to the mean and considers
first few terms of the expansion. Inherent to the Taylor series expansion, regular pertur-
86
bation methods often exhibit closure problems (cf. Kavvas [47]), where information on
higher-order moments are necessary to solve for lower-order moments. Also, the regular
perturbation approach is applicable only to small fluctuations in the state variable since
the linearization approximation fails when the input parameters exhibit large coefficient of
variations (COVs) (cf. Matthies et al. [69]).
First attempt to propagate uncertainties through elasticplastic constitutive equations considering random Youngs modulus was published only recently, e.g., Anders and
Hori [1]. They took perturbation expansion at the stochastic mean behavior and considered
only the first term of the expansion. In computing the mean behavior they took advantage
of bounding media approximation. Though this method doesnt suffer from computational
difficulty associated with Monte Carlo method for problems having no closed-form solution,
it inherits closure problem and the small COV requirements for the material parameters. Furthermore, with bounding media approximation, difficulty arises in computing the
mean behavior when one considers uncertainties in internal variable(s) and/or direction(s)
of evolution of internal variable(s).
Recently, Kavvas [47] obtained a generic EulerianLagrangian form of FPK equation, exact to second-order, corresponding to any nonlinear ordinary differential equation
with random coefficients and random forcings. The FPK equation approach doesnt suffer
from the drawbacks of Monte Carlo method and perturbation technique. In this chapter,
the development by Kavvas [47] is applied in obtaining probabilistic solution for a general, incremental elasticplastic constitutive equation with random coefficient. The solution
methodology is designed with several applications in mind, namely to
87
obtain probabilistic stressstrain behavior from spatial average form (upscaled form)
of constitutive equation, when input uncertain material properties to the constitutive
equation are random fields; and
obtain probabilistic stress-strain behavior from point-location scale constitutive equation, when input uncertain material properties to the constitutive equation are random
variables.
5.2
One Dimensional Development

In this section, one-dimensional general formulation of probabilistic elastoplasticity
is shown first, followed by its specialization in obtaining particular (obeying particular

elastoplasticity model) point-location scale constitutive behaviors and solution methodology of the resulting equation. Governing equations for probabilistic solutions of von Mises,
Drucker-Prager, and Cam Clay models have been derived. In addition, the governing equation for probabilistic solution of linear elastic constitutive equation has also been derived
as a special case of general nonlinear derivation.
The most general, incremental (rate) form of spatial-average elastic-plastic constitutive equation can be written as
dij (xt , t)
dkl (xt , t)
= Dijkl (xt , t)
dt
dt
where the stiffness tensor Dijkl (xt , t) can be either elastic or elastic-plastic:
(5.1)
88
Dijkl =
el
Dijkl
when
U f el
el
D
Dijmn
mn pq pqkl
el
Dijkl
f el U
f
r
rs rstu tu q
f < 0 (f = 0 df < 0)
(5.2)
when
f = 0 df = 0
el is the elastic stiffness tensor and D ep is the elasticplastic stiffness tensor, f

where Dijkl
ijkl
is the yield function, which is a function of stress (ij ) and internal variables (q ), U is the
potential function (also a function of stress and internal variables). The internal variables
(q ) could be scalar (for perfectly-plastic and isotropic hardening models), second-order
tensor (for translational and rotational kinematic hardening) or fourth-order tensor (for
distortional hardening) or combinations of the above. The same classification applies to
the direction of evolution of internal variables (r ). Therefore, the most general form of
constitutive rate equation in terms of its parameters can be written as
dij (xt , t)
dkl (xt , t)
el
= ijkl (ij , Dijkl
, q , r ; xt , t)
dt
dt
(5.3)
el ) and/or internal variables (q )

Due to randomnesses in elastic constants (Dijkl
and/or rate of evolution of internal variables (r ) the material stiffness operator ijkl in
Eq. (5.3) becomes stochastic and hence Eq. (5.2) becomes a set of linear/non-linear ordinary differential equations with stochastic coefficients. On the other hand, the randomness
in forcing term (kl ) (e.g., seismic loading), Eq. (5.3) becomes a set of linear/non-linear
ordinary differential equations with stochastic forcing. In general, randomnesses in material properties and forcing function, Eq. (5.3) becomes a set of linear/non-linear ordinary
89
differential equation with stochastic coefficients and stochastic forcing.
In order to gain better understanding of the effects of random material parameters
and forcing on response, focus is shifted from a general 3-D case to a 1-D case. Focusing
on 1-D behavior, the Eq. (5.3) can be written as
d(xt , t)
d(xt , t)
= (, Del , q, r; xt , t)
dt
dt
(5.4)
which is a non-linear ordinary differential equation with stochastic coefficient and stochastic
forcing. The right hand side of Eq. (5.4) is replaced with the function as
(, Del , q, r, ; x, t) = (, Del , q, r; xt , t)
d(xt , t)
dt
(5.5)
so that now Eq. (5.4) can be written as
(xt , t)
= (, Del , q, r, ; x, t)
t
(5.6)
(x, 0) = 0
(5.7)
with initial condition,
In the above equation (Eq. (5.6)), can be considered to represent a point in the
-space and hence it can be said that Eq. (5.6) determines the velocity for the point in the
-space. This may be visualized, from the initial point, and given initial condition 0 , as
a trajectory that describes the corresponding solution of the non-linear stochastic ordinary
differential equation (ODE) (Eq. (5.6)). Considering a cloud of initial points (refer to
Fig. 5.2), described by a density (, 0) in the -space and movement of the points dictated
90
by Eq. (5.6), the phase density of (x, t) varies in time according to a continuity equation
which expresses the conservation of all these points in the -space.
Figure 5.2: Movements of Cloud of Initial Points, described by density (, 0), in the -space
This continuity equation can be expressed in mathematical terms, using Kubos

stochastic Liouville equation (cf. Kubo [55]):
((x, t), t)
= [(x, t), Del (x), q(x), r(x), (x, t)].[(x, t), t]

t
(5.8)
with an initial condition,
(, 0) = ( 0 )
(5.9)
where is the Dirac delta function and Eq. (5.9) is the probabilistic restatement of the
original deterministic initial condition (Eq. (5.7)). Then by using Van Kampens Lemma
91
(cf. Van Kampen [95]), one can write
h(, t)i = P (, t)
(5.10)
where, the symbol hi denotes the expectation operation, and P (, t) denotes evolutionary
probability density of the state variable of the constitutive rate equation (Eq. (5.4)).
In order to obtain the deterministic probability density (P (, t)) of the state variable, , it is necessary to obtain the deterministic partial differential equation (PDE) of
the -space mean phase density h(, t)i from the linear stochastic PDE system (Eqs. (5.8)
and (5.9)). This necessitates the derivation of the ensemble average form of Eq. (5.8) for
h(, t)i. This ensemble average can be derived as (for detailed derivation refer to Appendix
(Section A.4)):
h((xt , t), t)i

=
t

el
((xt , t), D (xt ), q(xt ), r(xt ), (xt , t))

Z t
d Cov0 ((xt , t), Del (xt ), q(xt ), r(xt ), (xt , t));
((xt , t ), Del (xt ), q(xt ), r(xt )(xt , t )
Z t
h

h((xt , t), t)i
i h((x , t), t)i

t
((xt , t ), D (xt ), q(xt ), r(xt ), (xt , t ))
el
(5.11)
to exact second order (to the order of the covariance time of ). In Eq. (5.11), Cov0 [] is
92
the time ordered covariance function defined by
Cov0 [(x, t1 ), (x, t2 )] = h(x, t1 )(x, t2 )i h(x, t1 )i h(x, t2 )i
(5.12)
By combining Eqs. (5.11) and (5.10) and rearranging the terms the following
EulerianLagrangian form of Fokker-Planck-Kolmogorov equation (FPKE) (cf. Kavvas [47])
yields:
P ((xt , t), t)
=
t

el
((xt , t), D (xt ), q(xt ), r(xt )(xt , t))

Z t
((xt , t), Del (xt ), q(xt ), r(xt )(xt , t))
;
d Cov0
+

((xt , t ), D (xt ), q(xt ), r(xt ), (xt , t ) P ((xt , t), t)

Z t
2
d Cov0 ((xt , t), Del (xt , t), q(xt , t), r(xt , t), (xt , t));
2
0

el
1 ((xt , t ), D (xt ), q(xt ), r(xt ), (xt , t ))
P ((xt , t), t)
el
(5.13)
to exact second order. This is the most general relation for probabilistic behavior of inelastic
(nonlinear) 1-D stochastic constitutive rate equation. The solution of this deterministic
linear FPKE (Eq. (5.13)) in terms of the probability density P (, t) under appropriate
initial and boundary conditions will yield the PDF of the state variable of the original
1-D non-linear stochastic constitutive rate equation (Eq. (5.4)). It is important to note that
while the original equation (Eq. (5.4)) is non-linear, the FPKE (Eq. (5.13)) is linear in terms
of its unknown, the probability density P (, t) of the state variable . This deterministic
93
linearity, in turn, provides significant advantages in the solution of the probabilistic behavior
of the constitutive rate equation (Eq. (5.4)).
One should also note that Eq. (5.13) is a mixed Eulerian-Lagrangian equation
since, while the real space location xt at time t is known, the location xt is an unknown.
If one assumes small strain theory, one can relate the unknown location xt from the
known location xt by using the strain rate, (=d/dt) as,
d =
=
xt xt
xt
(5.14)
or, rearranging
xt = (1
)xt
(5.15)
Once the evolutionary probability density function P (, t) is obtained it can be

used to obtain the evolutionary statistical moments of state variable () by usual expectation
operation (refer to Subsection 2.3.3) e.g., the evolutionary mean by:
h(t)i =
(t)P ((t))d(t)
(5.16)
Another interesting aspect of this development, but could be possibly of mathematical interest, is to obtain the equivalent It
o stochastic differential equation (refer to
Subsections 2.6.5, for more details refer to Gardiner [30]) corresponding to the general
FPKE (Eq. (5.13)). Using the equivalency between It
o stochastic differential equation and
FPKE (cf. Gardiner [30]) one can obtain the equivalent It
o form:
94

el
d(x, t) =

Z t
((xt , t), Del (xt ), q(xt ), r(xt ), (xt , t))
d Cov0
+
;

((xt , t ), D (xt ), q(xt ), r(xt ), (xt , t )) dt
el
+ b(, t)dW (t)
(5.17)
where,
b (, t) = 2

((xt , t ), D (xt ), q(xt ), r(xt ), (xt , t ))
el
(5.18)
and, dW (t) is an increment of Wiener process W(t) (refer to Subsections 2.6.4 and 2.6.5, for
more details refer to Gardiner [30]) with hdW (t)i = 0. It is also interesting to note that all
the stochasticities of the original equation (Eq. (5.4)) are lumped together in the last term
(Wiener increment term) of the right-hand-side of Eq. (5.17). But the problem in solving for
the statistical moments of the state variable () using the It
o form e.g., say the evolution
of the mean of the state variable (), which can be written mathematically as (taking
advantage of the independent increment property of the Wiener process (hdW (t)i = 0), for
details refer to Kavvas [47]):
95
hd(x, t)i
dt

el
=

Z t
((xt , t), Del (xt ), q(xt ), r(xt ), (xt , t))
;
d Cov0
+

((xt , t ), D (xt ), q(xt ), r(xt ), (xt , t )) (5.19)
el
is the non-linear stochasticity (note that state variable also appears within () on the righthand-side of Eq. (5.19) and is random) in the resulting equation. No analytical treatment
is available for dealing with this type of problem. There exist approximate numerical
method e.g., perturbation approach (cf. Anders and Hori [2]) but the closure problem
will appear. Also, due to linearization approximation using Taylor series expansion, the
error in perturbation approximation is a function of COV, which for soil is usually quite
large.
5.2.1
Specialization of General Formulation to Particular Constitutive

Laws
Having obtained the relation for probabilistic behavior of 1-D inelastic (nonlinear)
constitutive rate equation with stochastic coefficient and stochastic forcing in most general
form, in this subsection the general relation will be specialized to four particular types of
point-location scale shear constitutive behavior: a) 1-D linear elastic, b) 1-D Drucker-Prager
associative elasticplastic linear hardening c) 1-D von Mises associative elasticplastic linear
hardening, and d) 1-D Cam Clay elasticplastic.
96
Probabilistic Linear Elastic Shear Constitutive Equation
The 1-D point-location scale linear elastic shear constitutive rate equation can be
written as,
d12
d12
=G
dt
dt
(5.20)
so that the function , defined in Eq. (5.5), takes the form as

=G
d12
dt
(5.21)
and hence, considering both the shear modulus, G and the strain rate, d12 /dt(t) as random,
one can substitute Eq. (5.21) in Eq. (5.13) to obtain the particular Fokker-Planck equation
for the probabilistic behavior of 1-D point-location scale linear elastic shear constitutive
rate equation as,
P (12 (t), t)
=
t

d12
(t)
G
12
dt

Z t
d12
d12
G
(t) ; G
(t ) P (12 (t), t)
d Cov0
+
12
dt
dt
0
Z t

d12
d12
2
d Cov0 G
(t); G
(t ) P (12 (t), t)
+
2
dt
dt
12
0
(5.22)
The first random process in the covariance term of the first coefficient on the r.h.s
of above equation (Eq. (5.22)) is independent of 12 and noting that covariance of zero with
97
any random process is zero, one can further simplify the above equation as,
P (12 (t), t)
=
t

d12
(t) P (12 (t), t)

G
12
dt

Z
t
d12
d12
2
d Cov0 G
(t); G
(t ) P (12 (t), t)
+
2
dt
dt
12
0
(5.23)
Given the random shear modulus and shear strain rate random process, with appropriate initial and boundary conditions, Eq. (5.23) will predict the evolution of probability
density function of shear stress with time following 1-D linear elastic shear constitutive rate
equation. Once the evolution of PDF of shear stress with time is obtained, one can integrate it to obtain the evolution of mean and variance of the shear stress random process
with time. It should be noted that there exist several other exact methods (e.g., cumulant
expansion method) to obtain probabilistic behavior of stochastic processes driven by linear
ordinary differential equations. The main objective of FPKE approach presented here is
to deal with non-linear stochastic processes (e.g., elastic-plastic constitutive rate equation)
and the linear elastic case is obtained as a special case of that non-linear process.
Probabilistic Drucker-Prager Associative ElasticPlastic Shear Constitutive Equation

For materials obeying Drucker-Prager yield criteria (without cohesion), the yield
surface can be represented as:
f=
J2 I1 = 0
(5.24)
where J2 = 21 Sij Sij is the second invariant of the deviatoric stress tensor Sij = ij ij kk /3,
98
and I1 = ii is the first invariant of the stress tensor, and , an internal variable, is a function
of friction angle of any material and is given by = 2 sin()/(
(3)(3 sin)), where is
the friction angle of any material (cf. Chen and Han [10]). By assuming associative flow
rule, i.e. by assuming that the yield function is the same as the plastic potential function,
one can write:
f
U
=
ij
ij
(5.25)
and hence one can expand parts of the tangent constitutive tensor given in Eq. (5.2) to read
(for detailed derivation refer to Appendix (Sections A.1 and A.2):

f
I1
I1
I1
1l 1k +
2l 2k +
3l 3k
2G
I1
11
22
33

I1
2
cd kl
K G
3
cd

2
J2
f
J1
ik jl + K G
ab kl
2G
ij
3
ab
J2
f el
Akl =
D
=
pq pqkl
+
+
(5.26)
and,
f el f
B=
D
rs rstu tu

)
I1 2
I1 2
I1 2
=
2G
+
+
11
22
33

2 #

I1
2
ij
+
K G
3
ij

2 #
2 "
J2
f
2
J2 J2
+ K G
ij
+
2G
ij ij
3
ij
J2

f
I1
2 "
(
(5.27)
where, K and G are the elastic bulk modulus and the elastic shear modulus respectively. By
99
further assuming that the evolution of internal variable is a function of equivalent plastic
strain1 , epeq = 2/3epij epij one can write (for detailed derivation, refer to Appendix (Section A.3))
KP =
f
1 f d f
rn =
p
qn
3 deeq J2
(5.28)
It should be noted that if material properties are assumed to be random, the resulting
stress tensor will also become random and hence the derivatives of the stress invariants with
respect to stress tensor (ij ) will become differentiations with random measures. Therefore,
differentiations appearing in Eqs. (5.26), (5.27), and (5.28) can not be carried out in an
ordinary sense. The tangent constitutive tensor in Eq. (5.2) then becomes

when f < 0 (f = 0 df < 0)

Gik jl + Gil jk + K G ij kl
Dijkl =
(5.29)

Aij Akl
2
when f = 0 df = 0
3
B + KP
where tensor Aij and scalars B and KP are defined by Eqs. (5.26), (5.27), and (5.28)
respectively. The above tangent constitutive tensor (Eq. (5.29)) is the most general form
for a Drucker-Prager isotropic linear hardening material model obeying associated plasticity.
Focusing attention on one dimensional point-location scale shear constitutive relationship
between 12 and 12 , using Eq. (5.29), for Drucker-Prager material, one can simplify the
function (, D, q, r, ; x, t) as defined in Eq. (5.5) (on Page 89) to read
1
This is a fairly common assumption, e.g., Chen and Han [10]
100
d12
dt
2

J2
f
2
4G
=
J2 12 d12
B + KP
dt
when
f < 0 (f = 0 df < 0)
(5.30)
when
f = 0 df = 0
and hence, considering both the material properties (shear modulus G, bulk modulus K,
friction angle , and rate of change of friction angle with plastic strain (plastic slope)
) and the strain rate d12 /dt(t) as random, one can substitute as derived above in
Eq. (5.13) to obtain the particular Fokker-Planck equation for the probabilistic behavior
of Drucker-Prager associative linear hardening type 1-D point-location scale elastic-plastic
shear constitutive rate equation. In particular, two cases are recognized, one for elastic
(preyield) behavior of material (f < 0 (f = 0 df < 0))
P (12 (t), t)
=
t

d12
G
(t) P (12 (t), t)
12
dt

Z
t
d12
d12
2
d Cov0 G
(t); G
(t ) P (12 (t), t)
+
2
dt
dt
12
0
(5.31)
One may note that this is the same equation as Eq. (5.23). In addition to that, the case of
elasticplastic behavior (f = 0 df = 0) is described by the following probabilistic equation
+
+

P (12 (t), t)
d12
ep
=
(t)
(G (t))
t
12
dt

Z t
d12
d12
ep
ep
G (t)
(t) ; G (t )
(t ) P (12 (t), t)
d Cov0
12
dt
dt
0

Z t
d12
d12
2
ep
ep
d Cov0 G (t)
(t); G (t )
(t ) P (12 (t), t) (5.32)
2
dt
dt
12
0
101
where Gep (a) is defined as probabilistic elasticplastic kernel and is introduced to shorten
the writing (but will have other uses later)
Gep (a) = G
4G2
and a assumes values t or t .
2
f
J2 (a)
J2 (a) 12 (a)
B(a) + KP (a)
(5.33)
One important aspect to note that the solution of P (12 , t), governed by Eq. (5.32),
would involve assumption of deterministic values of 12 in the given 12 -domain and solving
for probability densities (P ) for those deterministic 12 s. Hence the differentiations appearing in the coefficient terms of the FPK partial differential equation (Eq. (5.32)), within the
probabilistic elasticplastic kernel Gep (a) (i.e. Eq. (5.33)), are for deterministic values of
12 and hence those differentiations can be carried out in an ordinary sense. After carrying
out the deterministic differentiations, the probabilistic elasticplastic kernel becomes:
ep
G (a)|12 deterministic =
G2
G
G + 9K2 + 13 I1 (a)
(5.34)
which, after substitution, result in simplification of the FPK equation (Eq. (5.32)). Further
simplification is possible by noting that the first random process in the covariance term of the
first coefficient on the r.h.s of the equation (Eq. (5.32)) is independent of 12 . Furthermore,
since the covariance of zero with any random process is zero, the FokkerPlanck-Kolmogorov
102
equation (Eq. (5.32)) is further simplified as
P (12 (t), t)
=
t

d12
ep
G (t)
(t) P (12 (t), t)
12
dt
Z t

d12
d12
2
ep
ep
d Cov0 G (t)
(t); G (t )
(t ) P (12 (t), t) (5.35)
+
2
dt
dt
12
0
where the probabilistic elasticplastic kernel Gep (a) is given by the Eq. (5.34).
Probabilistic von Mises Associative ElasticPlastic Shear Constitutive Equation

Probabilistic behavior of 1D von Mises associative shear constitutive model, which
is governed by the following yield surface:
f=
p
J2 cu = 0
(5.36)
where cu is the shear strength, is very similar to the probabilistic behavior of 1D Drucker
Prager shear constitutive model. Following the derivation as shown for Drucker-Prager
constitutive equation and here also recognizing two different probabilistic equations one
for preyield elastic region and the other for postyield elasticplastic region the FPKE
corresponding to 1D von Mises associative shear constitutive model can be written as:
103
when
f < 0 (f = 0 df < 0), then
P (12 (t), t)
=
t

d12
(t) P (12 (t), t)
G
12
dt

Z t
2
d12
d12
+
d Cov0 G
(t); G
(t ) P (12 (t), t)
2
dt
dt
12
0
and when f = 0 df = 0, then

d12
P (12 (t), t)
ep
(G (t))
=
(t)
t
12
dt

Z t
d12
d12
ep
ep
G (t)
(t) ; G (t )
(t ) P (12 (t), t)
d Cov0
+
12
dt
dt
0

Z t
d12
2
d12
ep
ep
(t); G (t )
(t ) P (12 (t), t) (5.37)
+ 2
d Cov0 G (t)
dt
dt
12
0
where, Gep , when 12 is deterministic is given as:
Gep = G
G2
1
G + cu
3
where cu is the rate of evolution of cu with plastic shear strain.
(5.38)
104
Probabilistic Cam Clay ElasticPlastic Shear Constitutive Equation
For classical Cam Clay material model, the yield surface (as well as the potential
surface) is given as:
q2
f = f(p, q) = p2 p0 p + 2 = 0
M
(5.39)
where M represents the slope of the critical state line in the p q space and the stress
invariants p and q are given as:
p=
q=
11 + 22 + 33
kk
=
3
3
(5.40)

3
1
2
2
2 1/2
+ 623
+ 631
sij sij = (11 22 )2 + (22 33 )2 + (33 11 )2 + 612
2
2
(5.41)
Now, working on the components of Eq. 5.2 in p q space (similar to the I1 J2

space derivations for the cases of Drucker-Prager and von Mises constitutive equation, as
derived in Sections A.1 and A.2 in the Appendix), one can obtain:
f el
D
=
pq pqkl

f
2
p
p
p
p
2G
1l 1k +
2l 2k +
3l 3k + K G
cd kl
p
11
22
33
3
cd

f
q
2
q
+
(5.42)
2G
ik jl + K G
ab kl
q
ij
3
ab
Akl =
and
105
f el f
D
=
rs rstu tu
"
(

)

2 #
2
p
p 2
p 2
2
p 2
f
+ K G
ij
2G
+
+
p
11
22
33
3
ij

2 "
2 #
q
f
2
q q
+ K G
ij
+
2G
(5.43)
q
ij ij
3
ij
B=
where, K and G are the elastic bulk and shear modulus, respectively. For the Cam Clay
formulation the material parameter p0 represents an internal variable and its evolution
depends on the plastic volumetric strain. Hence, the plastic modulus KP becomes (again,
can be derived in a very similar way in pq space as derived in Section A.3 (in the Appendix)
for I1 J2 space for applications to Drucker-Prager and von Mises constitutive equations):
KP =
f
p0 f
p0 = (1 + e0 )
p0
p
(5.44)
Substituting Eqs. (5.42), (5.43), and (5.44) into (5.2) one may obtain the most
general form of Cam Clay tangent constitutive tensor as:
Dijkl =


Aij Akl
2
3
B + KP
when
f < 0 (f = 0 df < 0)
(5.45)
when
f = 0 df = 0
where Aij , B, and KP are defined by Eqs. (5.42), (5.43), and (5.44) respectively. Eq. (5.45),
when specialized to 1-D shear behavior (12 12 relationship), can be written as:
106
D1212 =
Gep
when
f < 0 (f = 0 df < 0)
(5.46)
when
f = 0 df = 0
where Gep , as defined below, is the elasticplastic shear modulus obeying Cam Clay material
model and hence, the Fokker-Planck-Kolmogov equation for 1-D shear Cam Clay behavior,
recognizing two equations - one for the pre-yield behavior and the other for the post yield
behavior, can be written as:
when
f < 0 (f = 0 df < 0), then
P (12 (t), t)
=
t

d12
(t) P (12 (t), t)
G
12
dt

Z t
2
d12
d12
+
d Cov0 G
(t); G
(t ) P (12 (t), t)
2
dt
dt
12
0
and when f = 0 df = 0, then

d12
P (12 (t), t)
ep
(G (t))
=
(t)
t
12
dt

Z t
d12
d12
ep
ep
G (t)
(t) ; G (t )
(t ) P (12 (t), t)
d Cov0
+
12
dt
dt
0

Z
t
d12
d12
2
ep
ep
d Cov0 G (t)
(t); G (t )
(t ) P (12 (t), t) (5.47)
+ 2
dt
dt
12
0
107
where, Gep , when 12 is deterministic is given as:
Gep |12 deterministic =

G
5.2.2
3 2
(1 + e0 )p p 122
pM
2

G2
2
36 4 12
M
(5.48)

2
1 + e0
312
G
2
+ 18 4 12 +
pp0 p
M
pM 2
Solution Method of Probabilistic ElastoPlastic Equation

The general solution method for FPKE can be found in any standard text books
(e.g., Gardiner [30], Risken [83]). In this subsection, based on the above references, a simple
finite difference based numerical solution scheme for FPKE is developed for probabilistic
solution of constitutive equation. Computational efficiency of the numerical solution process
has not been analyzed in this study. There exist several numerically efficient solution scheme
for FPKE (e.g., Langtangen [58], Masud and Bergman [68] etc.). But applicability of those
in probabilistic solution of constitutive equation is left as future research.
Initial and Boundary Conditions

The FPKE describing the probabilistic behavior of any general constitutive rate
equation (Eq. (5.13)) can be written in the following general form:
P (12 , t)
t
12
=
12
=
12

2
P (12 , t)N(2)
P (12 , t)N(1) +
2
12

P (12 , t)N(2)
P (12 , t)N(1)
12

(5.49)
(5.50)
(5.51)
108
where, N(1) and N(2) are coefficients2 of the FPKE and represent the expressions within the
curly braces of the first and second terms respectively on the righthandside of Eq. (5.13)
These terms are called the advection (N(1) ) and diffusion (N(2) ) coefficients as the form of
Eq. (5.51) closely resembles advectiondiffusion equation (cf. Gardiner [30]). The symbol
in Eq. (5.51) can be considered to be the probability current since Eq. (5.51) is basically
a continuity equation and the state variable of the equation is probability density.
Introducing initial and boundary conditions, one can solve Eq. (5.51) for evolution of probability densities of 12 with evolution of time. The initial condition could be
deterministic or stochastic depending upon the type of problem. For probabilistic behavior
of linear elastic constitutive rate equation (Eq. (5.23)), one can assume that all the probability mass at time t = 0 is concentrated at 12 = 0 or at some constant value of 12 if
there were some initial stresses to begin with (e.g., overburden pressure on a soil mass). In
mathematical term, this translates to,
P (12 , 0) = (12 )
(5.52)
where, () is the Dirac delta function. For the postyield behavior of any probabilistic
elastic-plastic constitutive rate equation (e.g., Eq. (5.35) for post-yield Drucker-Prager behavior), there will be a distribution of 12 , corresponding to the solution of the pre-yield
probabilistic behavior to begin with.
This probability mass, dictated by Eq. (5.13), will advect and diffuse into the
domain of the system throughout the evolution of the simulation. Since it is required
that the probability mass within the system is conserved i.e. no leaking is allowed at
2
Placement of indices in brackets, even if they are not dimensional variables (like i or j or k), means that
there is no summation over repeated indices.
109
the boundaries, a reflecting barrier at the boundaries will be the preferred choice. In
mathematical term, one can express this condition as (cf. Gardiner [30]):
(12 , t)|AtBoundaries = 0
(5.53)
In theory, the domain of stress could be from to so that boundary conditions are
then
(, t) = (, t) = 0
(5.54)
With these initial and boundary conditions, the probabilistic differential equation
(with random material properties and random strain) for elastoplasticity, specialized to
any particular material model, can be solved for probability densities of shear stress (12 )
as it evolves with time/shear strain (12 ).
Numerical Scheme
The FokkerPlanckKolmogorov PDE is first semi-discretized into a set of simultaneous ODE systems and then solved by using an ODE solver. To this end, differentiating
by parts the right hand side of Eq. (5.49) yields:
P
t
N(1)
2 N(2)
P N(2)
P
2P
P
+ N(2) 2 + 2
+P
2
12
12
12 12
12
12
!

2 N(2) N(1)
N(2)
P
2P
+
N(1) +
2
2 N(2)
12
12
12
12
12
= N(1)
= P
Then, using central difference discretization, as shown in Fig. 5.3

one can write semi-discretized form of Eq. (5.55) at any intermediate node i as:
(5.55)
110
Figure 5.3: Spatial Discretization of the FokkerPlanckKolmogorov PDE
P (i)
t
P (i1)
(i)
N(1)
(i)
(i)
N(2)
(i)
N(1)
(i)
N(2)
(i)
2 N(2)
P (i)
+2
2 2
12 12
12
12
12
(i)
(i)
N
N(2)
1
(2)
+
+
(5.56)
2
12
12
12
2
12
212
(i)
N(1)
(i+1)
+P
212
N(2)
which forms an initial value problem in the time dimension.
5.3
Three Dimensional Development

The one-dimensional development as presented in Section 5.2 is general enough to
extend the derivation to three-dimension. In this section, the probabilistic elastoplasticity

as presented in Section 5.2 is in cast three-dimensional form. We will start with the most
general form of 3D elasticplastic constitutive rate equation as presented in Eq. (5.2). The
111
following equation writes Eq. (5.2) more explicitly:
dij (xt , t)
dkl (xt , t)
ep
el
= Dijkl
(ij , Dijkl
, f, U, q , r ; xt , t)
dt
dt
(5.57)
By denoting all the random material parameters by a parameter tensor as:
i
h
el
Dijkl = Dijkl
, f, U, q , r
(5.58)
and introducing a random operator tensor, ij , one can write Eq. (5.57) as,
dij (x, t)
= ij (ij , Dijkl , kl ; x, t)
dt
(5.59)
ij (x, 0) = ij 0
(5.60)
with an initial condition,
In Eq. (5.59), the stress tensor ij can be considered to represent a point in a 9-dimensional
stress ()-space and hence Eq. (5.59) determines the velocity for the point in this space.
Similar to the one-dimensional case, it is possible to visualize this, by imagining an initial
point in a 9-dimensional stress space, given by its initial condition ij 0 , with a trajectory
starting out that describes the corresponding solution of the non-linear stochastic ordinary
differential equation (ODE) system (Eq. (5.59)). Let us now consider a cloud of initial
points, described by a density (ij , 0) in the -space, and with movements of these points
dictated by Eq. (5.59), the phase density of ij (x, t) varies in time according to a continuity
equation which expresses the conservation of all these points in the -space.
112
Expressing this continuity equation in mathematical terms, one obtains Kubos
stochastic Liouville equation (cf. Kubo [55]):
(ij (x, t), t)
=
mn (mn (x, t), Dmnpq (x), pq (x, t)) (ij (x, t), t)
t
mn
(5.61)
with initial condition,
(ij , 0) = (ij ij 0 )
(5.62)
where () is the Dirac delta function. Eq. (5.62) is the probabilistic restatement in the phase space of the original deterministic initial condition (Eq. (5.60)). One can then apply
Van Kampens Lemma (Van Kampen [95]) to obtain:
h(ij , t)i = P (ij , t)
(5.63)
where, the symbol hi denotes the expectation operation, and P (ij , t) denotes evolutionary
probability density of the state variable tensor ij of the constitutive equations.
Therefore, in order to obtain the multivariate probability density function (PDF),
P (ij , t), of the state variable tensor ij , it is necessary to obtain the deterministic partial
differential equation (PDE) of the -space mean phase density h(ij , t)i from the linear
stochastic PDE system (Eqs. (5.61) and (5.62)). This necessitates the derivation of the
ensemble average form of Eq. (5.61) for h(ij , t)i, which can be derived as:
113

h(ij (xt , t), t)i
=
mn (mn (xt , t), Dmnrs (xt ), rs (xt , t))
t
mn

Z t
d Cov0 mn (mn (xt , t), Dmnrs (xt ), rs (xt , t));
0

ab (ab (xt , t ), Dabcd (xt ), cd (xt , t )
h(ij (xt , t), t)i
ab
Z t
d Cov0 [mn (mn (xt , t), Dmnrs (xt ), rs (xt , t));

+
mn
0

h(ij (xt , t), t)i
ab (ab (xt , t ), Dabcd (xt ), cd (xt , t ))]
(5.64)
ab
to exact second order (to the order of the covariance time of ). In Eq. (5.64), Cov0 [] is
the time ordered covariance function defined by,
Cov0 [mn (x, t1 ), ab (x, t2 )] = hmn (x, t1 )ab (x, t2 )i hmn (x, t1 )i hab (x, t2 )i
(5.65)
By combining Eq. (5.64) with Eq. (5.63) and rearranging the terms yields the following
Eulerian-Lagrangian form of FokkerPlanckKolmogorov equation (FPKE) (cf. Kavvas
[47]):

P (ij (xt , t), t)
=
t
mn

Z t
;
d Cov0
+
ab
0

ab (ab (xt , t ), Dabcd (xt ), cd (xt , t ) P (ij (xt , t), t)

Z t
2
+
mn ab
0

ab (ab (xt , t ), Dabcd (xt ), cd (xt , t ))
P (ij (xt , t), t)
(5.66)
114
to exact second order. The solution of this deterministic linear FPKE (Eq. (5.66)) in terms
of P (ij , t) under appropriate initial and boundary conditions will yield the multivariate
PDF of the state variable tensor ij .
Utilizing the connection between FPKE and It
o SDE (cf. Gardiner [30]), similar to
the one-dimensional development, the It
o SDE corresponding to Eq. (5.66) can be written
as:
dij (x, t) =

Z t

ij (ij (xt , t), Dijkl (xt ), kl (xt , t))
ij (ij (xt , t), Dijkl (xt ), kl (xt , t)) +
d Cov0
;
ab
0

ab (ab (xt , t ), Dabcd (xt ), cd (xt , t )) dt + Cijm (K , t)dWm (t)
(5.67)
where,

d Cov0 ij (ij (xt , t), Dijkl (xt ), kl (xt , t));
Cijm (ij , t)Cabm (ab , t) = 2
0

(5.68)
Z
and, dWi (t) is an increment of the vector Wiener process Wi (t) having the following properties:
hdWi (t)i = 0
and,
(5.69)
115
hdWi (t)dWj (t)i = ij dt for = t

= 0 for 6= t
(5.70)
116
Chapter 6
Probabilistic ElastoPlasticity:
Numerical Examples and
Verifications
6.1
Introduction
An elastic-plastic constitutive law is a set of linear/non-linear ordinary differential
equations (ODEs) and relates rate of stress with the rate of strain through linear/nonlinear
material modulus and is of the form,
dij (t)
ep dkl (t)
= Dijkl
dt
dt
(6.1)
ep
where Dijkl
could be linear or a non-linear function of ij . Constitutive (deterministic)
simulation generally involves obtaining stress response of any material, which is a function
117
of material properties and constitutive model, with strain. For the particular case, when
the material is linear elastic, Eq. (6.1) simplifies to a set of linear algebraic equations of the
form,
el
ij = Dijkl
kl
(6.2)
el is the elastic moduli tensor. In Eq. (6.2), if we assume the elastic material pawhere Dijkl
el ) is random, the probabilistic constitutive simulation of any material
rameter tensor (Dijkl
obeying Eq. (6.2) (i.e., probabilistic stress response with strain) can be easily easily carried
out by variable transformation method (cf. Montgomery and Runger [72]), because of linear
algebraic simplicity of Eq. (6.2). But difficulty arises for probabilistic simulations of nonlinear constitutive models, as they involve probabilistic solution of nonlinear ODEs with
random coefficients, whose most general form is given by Eq. (6.1). To avoid this problem
of non-linear stochasticity in constitutive equation, in the last chapter (Chapter 5), the general upscaled form of constitutive equation was transformed into an EulerianLagrangian
form of FPKE (Eq. (5.13)) in the probability density space. The advantage of writing
the constitutive equation in probability density space is that the resulting FPKE is linear
and deterministic PDE, which is much easier to deal with than the nonlinear stochastic
ODE in the real space. Also in the last chapter, the general EulerianLagrangian FPKE
(Eq. (5.13)) was then specialized to the particular cases of point-location scale linear elastic (Eq. (5.23)), von Mises associative linear hardening type (Eq. (5.37)), Drucker-Prager
associative linear hardening type (Eqs. (5.31) and (5.32)), and Cam Clay elastic-plastic
(Eq. (5.47)) constitutive rate equations to show the applicability of the general formulation
118
in obtaining particular constitutive behaviors. A simple finite difference based numerical
solution scheme is In this chapter the solution process of those particular FPEKs will be
presented, along with their verification using either variable transformation method (linear
elastic case) or Monte Carlo method (non-linear elastic-plastic case).
In all the following examples, for sake of simplicity, we assumed a constant strain
rate and hence d12 /dt terms in the coefficients of Eqs. (5.23), (5.37), (5.31), (5.32), and
(5.47) can be substituted by a constant numerical value for the entire simulation of the
evolution of probability densities of stress. One may note that the FPKE (Eq. (5.13))
describes the evolution of probability densities of stress with time and strain rate describes
the evolution of strain with time and combining the two one can obtain the evolution of
probability densities of stress with strain. Time has been brought in this simulation as
an intermediate dimension to help in solution process and hence the numerical value of
strain rate could be any arbitrary value, which will cancel out once one converts the time
evolution of probability densities of stress to strain evolution of probability densities of
stress. For simulation of all the three example problems, we assumed the arbitrary value
of strain rate as 0.054/second. It may also be noted that, eventhough in theory the shear
stress (12 ) ranges from - to +, for simulation purpose the domain between 0.1 MPa
and +0.1 MPa is chosen. This choice is based upon the material properties of the example
problems, and by considering the practical range of shear stress (12 ). In addition to that,
the Dirac delta function used in initial conditions of preyield (elastic) probabilistic FPKEs
are numerically approximated using a Gaussian function with zero mean and standard
deviation of 0.00001 MPa, as shown in Fig. 6.1.
119
ProbDensityHStressL
40000
30000
20000
10000
0.0005
0.001
0.0015
Stress HMPaL
0.002
Figure 6.1: Approximation of Dirac delta initial condition (Eq. (5.52)) with a Gaussian
function of a zero mean and a standard deviation of 0.00001 MPa
While this approximation introduces some error in solution (PDF of shear stress) the error
quickly disappears as soon as the solution moves away from the initial condition as will be
shown in the following examples.
6.2
Linear Elastic Shear Constitutive Behavior

Problem Definition
Assuming linear elastic constitutive rate equation is applicable and given that
shear modulus (G) is normally distributed at a point- location scale with mean of 2.5
MPa and standard deviation of 0.707 MPa, calculate the evolution of probability density
function of shear stress (12 ) with shear strain (12 ) for a displacement-controlled test with
deterministic shear strain increment. The other parameters are considered deterministic
120
and are as follows: Poissons ratio () = 0.2, confining pressure (I1 ) = 0.03 MPa.
Estimation of Advection and Diffusion Coefficients
Substituting the values of deterministic and random material properties and the
strain rate, one can obtain the coefficients N(1) and N(2) of the FPEs as
N(1)

d12
=
G
dt
d12
hGi
=
dt
= 0.27 M P a/s
N(2)
d12
d V ar G
=
dt
0
2

d12
V ar[G]
= t
dt
Z
= 0.0058t (M P a/s)2
Results and Discussion
Figs. 6.2 and 6.3 show the evolution of PDF of shear stress with time and shear
strain for materials obeying linear elastic constitutive law having random shear modulus
obtained using FPKE approach and transformation method (cf. Montgomery and Runger
[72]) respectively. The contours of evolution of PDFs were compared in Fig. 6.4 and the
evolutions of mean and standard deviations of stress were compared in Fig. 6.5. It can be
seen from the above figures that, though the FPKE approach predicted the mean behavior
exactly, it slightly overpredicted the standard deviation. This is because of the approximation used to represent the Dirac delta function, which was used as the initial condition for
the FPKE. One may note that at 12 = 0, the probability density of shear stress at 12 =
121
0 should theoretically be i.e. all the probability mass should theoretically be concentrated at 12 = 0 and would be best described by a Dirac delta condition but for numerical
simulation of FPKE that Dirac delta initial condition was approximated with a Gaussian
function with a mean of zero and a standard deviation of 0.00001 MPa as shown in Fig. 6.1.
This initial error in standard deviation advected and diffused into the domain during the
simulation of evolution process. This error could be minimized by better approximating
the Dirac delta initial condition (but at the expense of computational cost). The effect of
approximating the initial condition on the PDF of shear stress at 12 =0.0426 % is shown
in Fig. 6.6. In this figure the actual PDF at 12 =0.0426 %, obtained using transformation
method, was compared with the PDFs at 12 =0.0426 % obtained using FPE approach with
three different approximate initial conditions - all having zero mean but standard deviations
of 0.01 MPa, 0.005 MPa and 0.00001 MPa.
Verification of developed numerical solution plays crucial role in building confidence in the solution (cf. Oberkamp [77]). For the linear elastic constitutive rate equation
the verification was done by a transformation method of random variables (cf. Montgomery
and Runger [72]) as for rate-independent linear elastic case one can simplify the 1-D shear
constitutive equation to a linear algebraic equation of the form,
12 = G12 = u(G, 12 )
(6.3)
Using the definition of strain rate, one can write the above equation in terms of time, t as,
12 = G(0.054t) = v(G, t)
(6.4)
where, 0.054/second is the arbitrary strain-rate assumed for the example problems. Hence,
according to the transformation method of random variables (cf. Montgomery and Runger
122
10000
0.0027
7500
Prob Density
Strain (%)
5000
0.0426
2500
0
0.004
0.002
0.003
0.004
Time (Sec)
0.002
0.006
0.001
0.00789
Stress (MPa)
Figure 6.2: Time (or strain) evolution of probability density function of shear stress for elastic constitutive rate equation with random shear modulus obtained using FPKE approach
[72]), given the continuous random variable, shear modulus (G), with PDF g(G) and
Eqs. (6.3) or (6.4) as one-to-one transformations between the values of random variables
of G and 12 , one can obtain the the PDF of shear stress (12 ), P (12 ) as,
P (12 ) = g(u1 (12 , 12 )) |J|
(6.5)
which will allow for predicting the evolution of PDF of 12 with 12 or,
P (12 ) = g(v 1 (12 , t)) |J|
(6.6)
Eq. (6.6) will predict the evolution of PDF of 12 with t. In Eqs. (6.5) and
(6.6) G = u1 (12 , 12 ) or G = u1 (12 , t) are the inverse functions of 12 = u(G, 12 ) or
12 = v(G, t) respectively and J = du1 (12 , 12 )/d12 or J = dv 1 (12 , t)/d12 are their
123
respective Jacobians of transformations.
0.0027
Strain(%)
0.0426
10000
7500
5000
Prob Density
2500
0
0.004
0.002
0.003
0.004
Time (Sec)
0.002
0.006
0.001
0.00789
Stress (MPa)
Figure 6.3: Time (or strain) evolution of probability density function of shear stress for
elastic constitutive rate equation with random shear modulus obtained using transformation
method
One may also note that finer approximation of initial condition necessitates finer
discretization of stress domain close to 12 = 0. The finite difference discretization scheme
adopted here uses same fine discretization uniformly all throughout the entire domain,
though it is not required. To properly capture the approximate initial condition as shown in
Fig. 6.1, the stress domain between -0.1 MPa and +0.1 MPa were discretized with a uniform
step size of 0.000005 MPa and hence using total of 40,000 nodes. This not only requires large
computational effort but also very memory sensitive. An adaptive discretization technique
will be a better approach for this problem. But it is left as a future research.
124
Variable Transformation
0
Strain (%)
FokkerPlanck Equation
0
0.0426
0.003
Strain (%)
0.003
0.0426
70.250.
250.
10.
Stress (MPa)
Stress (MPa)
500.
0.002
500.
70.
0.001
0.002
500.
250.
0.001
70.
10.
10.
0
0
0.004
0.008
0.004
Time (Sec)
0.008
Time (Sec)
Figure 6.4: Comparison of contours of time (or strain) evolution of probability density
function for shear stress for elastic constitutive rate equation with random shear modulus
for FPKE solution and variable transformation method solution
Strain (%)
0.0426
0.0025
Std. Deviation Lines

(FokkerPlanck)
Stress (MPa)
0.002
0.0015

(Variable Transformation)
0.001
Mean Line
(FokkerPlanck)
0.0005
Mean Line
(Variable Transformation)
0.002
0.004
0.006
0.008
Time (Sec)
Figure 6.5: Comparison of evolution of mean and standard deviation of shear stress with
time (or shear strain) for elastic constitutive rate equation with random shear modulus for
FPKE solution and variable transformation method solution
125
ProbDensity
600
500
Actual (Variable Transformation)
400
With Finer Approximation
300
200
With Fine Approximation

With Crude Approximation
100
0.01 0.005
0.005
Stress (MPa)
0.01
0.015
Figure 6.6: Effect of approximating function of Dirac delta initial condition: PDF of stress
at yield for different approximations of initial condition with actual (variable transformation
method) solution
6.3
ElasticPlastic Shear Constitutive Behavior with Mean

Yield Criteria
In the last chapter, during the development of probabilistic behavior of constitu-
tive equations, it was stated that the complete probabilistic solution of any elasticplastic
constitutive equation e.g., DruckerPrager associative linear hardening type elasticplastic
constitutive equation would involve solutions of two FPKEs one for the pre-yield elastic
behavior (Eq. (5.31) and the other for the post-yield elasticplastic behavior (Eq. (5.32).
But the question arises, where to make the shift or in other words what is the yield criteria
in a probabilistic sense.
126
One possible way could be to define a mean yield criteria or in other words shift
from elastic FPKE to elasticplastic FPKE when the mean of elastic solution exceeds mean
of yield stress. In mathematical term this translates to:
if
or, if
hf i < 0 (hf i = 0 d hf i < 0)

hf i = 0 d hf i = 0
then use elastic FPKE
then use elasticplastic FPKE
This above defined mean yield criteria, though doesnt consider complete probabilistic yielding of materials, could be very useful, if not enough data is available to statistically characterize the yield strength completely. The complete probabilistic yielding of materials is
considered in the next section (Section 6.4).
6.3.1
Drucker-Prager Associative Model
Problem Definition
Problem - I
Assuming linear elastic Drucker-Prager associative plastic isotropic linear hardening type constitutive rate equation is applicable and given that shear modulus (G) is
normally distributed at a point-location scale with a mean of 2.5 MPa and a standard deviation of 0.707 MPa, calculate the evolution of probability density function of shear stress
(12 ) with shear strain (12 ) for a displacement-controlled test with deterministic shear
strain increment. The other parameters are considered deterministic and are as follows:
Poissons ratio () = 0.2, confining pressure (I1 ) = 0.03 MPa, yield parameter () = 0.071,
plastic slope ( ) = 5.5.
127
Problem - II
Assuming linear elastic Drucker-Prager associative plastic isotropic linear hardening type constitutive rate equation is applicable and given that and given that the friction
coefficient () is normally distributed at a point-location scale with a mean of 0.52 and a
standard deviation of 0.1, calculate the evolution of probability density of shear stress (12 )
with shear strain (12 ) for a displacement-controlled test with deterministic shear strain
increment. The other parameters are considered deterministic and are as follows: shear
modulus (G) = 2.5 MPa, Poissons ratio () = 0.2, confining pressure (I1 ) =0.03 MPa,
plastic slope ( ) = 5.5.
Estimation of Advection and Diffusion Coefficients
Substituting the values of deterministic and random material properties and the
strain rate, one can obtain the coefficients N(1) and N(2) of the FPKEs as
For Problem I For pre-yield linear elastic case, the coefficients N(1) and N(2) will
be the same as those of linear elastic case as computed in Section 6.2. For post-yield
elastic-plastic they can be estimated as follows:
+
*
2
d12
G
N(1) =
dt
2
G + 9K + I1
3
+
*
2
d12
G
=
G
1
dt
G + 9K2 + I1
3
= 0.147 M P a/s
N(2) = t
d12
dt
2
V ar
G
= 0.00074t (M P a/s)2
G2
G + 9K2 + I1
3
128
For Problem II For post-yield elastic-plastic simulation the coefficients N(1) and
N(2) can be estimated as follows:
G
N(1) =
d12
dt
d12
dt
2
G2
1
G + 9K2 + I1
3
G2
d12
dt
1
G + 9K2 + I1
3
= 0.2365 M P a/s
N(2) = t
V ar
G
= 0.0001t (M P a/s)2
G2
1
G + 9K2 + I1
3
One may note that for Problem II, since the shear modulus is deterministic, the
pre-yield elastic case is deterministic.
Results and Discussions
Problem-I
The solution of this problem involves solution of two FPKEs, one corresponding to
the pre-yield elastic part and the other corresponding to the post-yield elastic-plastic part.
The elastic part of this problem is identical to linear elastic case as shown in Section 6.2.
The initial condition for the post-yield elastic-plastic part of the problem is random and is
shown in Fig. 6.7. It may be noted that this initial condition corresponds to the PDF of
shear stress (P (12 )) at yield obtained from the solution of FPKE of the pre-yield elastic
part. A view of the surface of time and shear strain evolution of PDF of shear stress for both
elastic and elastic-plastic part, obtained when one looked perpendicular to the time/strain
axis, is shown in Fig. 6.8. Another view to the surface, obtained when one looked parallel to
129
the time/strain axis, is shown in Fig. 6.9. One may note that the yielding of this material
occurred at t=0.00789 second (which is equivalent to 12 = 0.0426 %). The contours of
time/strain evolution of PDF of shear stress along with the time/strain evolutions of mean
and standard deviations of shear stress are shown in Fig. 6.10. It can be seen from Fig. 6.10
that, as expected, the evolution of mean of shear stress changed slope after the material
yielded. Another interesting aspect to note is that the relative slope of the evolutions of
standard deviations of shear stress with respect to the evolution of mean of shear stress. The
relative slope in the pre-yield elastic region increased at a higher rate during the evolution
process when compared with that in the post-yield elastic-plastic zone. In other words,
in the evolution process the post-yield elastic -plastic constitutive rate equation did not
amplify the initial uncertainty as much as the pre-yield elastic constitutive rate equation
did. This can be easily viewed from Fig. 6.11 where the post-yield elastic-plastic evolution
of PDF of shear stress was compared with fictitious extended elastic evolution of PDF.
Comparing the PDF of shear stress at 12 =0.0804 % (which is equivalent to t = 0.01489s),
one can conclude that the variance of predicted elastic-plastic shear stress is much smaller
(i.e., prediction is less uncertain) as compared to the same if the material were modeled as
completely elastic.
Fig. 6.12 compares the evolution of means and standard deviations of predicted
shear stress obtained, using FPKE approach and transformation method (pre-yield behavior)/Monte Carlo approach (post-yield behavior). Though in the pre-yield response the
FPKE approach overpredicted the evolution of standard deviations because of reasons discussed earlier, in the post-yield response, it matched closely at regions further from yielding
130
600
Prob Density
500
400
300
200
0.00213 MPa
100
0.001
0.001
0.003
0.005
Stress (MPa)
Figure 6.7: Initial condition for Fokker-Planck-Kolmorogov equation for probabilistic simulation of post-yield region
region. The difference at/or close to yielding region attributes to the fact that the initial
condition for solution of post-yield elastic-plastic FPKE was obtained from the solution of
pre-yield elastic FPKE. One way to better predict the overall probabilistic elastic-plastic
behavior, is to obtain the pre-yield elastic behavior through transformation method and
then use FPKE approach to predict post-yield elastic-plastic behavior.
Problem-II
In this problem the pre-yield linear elastic part is deterministic but at yield there
is a distribution in shear stress due to assumed distribution in friction coefficient, . The
distribution in shear stress corresponds to the PDF of the random variable I1 , where I1 is
the first invariant of the stress tensor (can be visualized as confining stress) and is assumed
131
0.0108
Strain (%)
0.0804
2000
Prob Density
1000
0.004
0
0.002
0.005
Stress (MPa)
Time (Sec)
0.01
0
0.015
Figure 6.8: Time (or strain) evolution of probability density function of shear stress
for Drucker-Prager elastic-plastic constitutive rate equation with random shear modulus
(Problem-I) (View obtained when one looks perpendicular to the time/strain axis)
to be deterministic. This PDF of shear stress at yield was assumed to be the initial condition
for the solution of post-yield elastic-plastic FPKE and is shown in Fig. 6.13.
The time/strain evolution of PDF of shear stress is shown in Fig. 6.14 and the
contours of evolution of PDF along with mean and standard deviation of shear stress is
shown in Fig. 6.15. Looking at Fig. 6.15 and comparing the slopes of evolution of mean
and standard deviation, one can conclude that the elastic-plastic evolution process didnt
amplify the initial uncertainty in yield strength much. The initial (at yield) probability
density function of shear stress just advected into the domain during the elastic-plastic
evolution process without diffusing much. Fig. 6.14 clearly shows this advection process.
The evolution of mean and standard deviations of shear stress obtained from FPE approach
132
0.0108
0.005
Strain (%)
Time (Sec)
0.01
0.0804
0.015
2000
Prob Density
1000
0.004
0.002
0
Stress (MPa)
Figure 6.9: Time (or strain) evolution of probability density function of shear stress
(Problem-I) (View obtained when one looks parallel to the time/strain axis)
was compared with those obtained from Monte Carlo simulation and is shown in Fig. 6.16.
6.3.2
Cam Clay Model

For Cam Clay constitutive law, we begin by presenting the results of elastic and
elasticplastic probability density functions for shear stress when only one material parameter of the model is random. Fig. 6.17(a) shows the evolution of probability density
function (PDF) of shear stress with time/shear strain for a low overconsolidation ratio
(OCR) CamClay material with normally distributed random shear modulus G (mean
value of 2.5 MPa and a standard deviation of 0.5 MPa). The other material properties
are considered deterministic and are given as follows: slope of critical state line M = 0.6;
overconsolidation pressure p0 = 0.2 MPa; applied confinement pressure p = 0.1 MPa; slopes
133
0
Strain (%)
0.0804
0.008
0.006
20.
Stress (MPa)
70.
0.004
250.
500.
20.
70.
Mean Line
500.
250.
250.
500.
0.002
70.
20.
500.
250.
70.
20.
0
0
0.004
0.008
0.012
0.01489
Time (Sec)
Figure 6.10: Contour of time (or strain) evolution of probability density function for shear
stress for Drucker-Prager elastic-plastic constitutive rate equation with random shear modulus (Problem-I)
of the unloadingreloading and normal compression lines in e ln(p ) space = 0.05 and
= 0.25, respectively; and initial void ratio e0 = 2.18.
As mentioned earlier, the strain is assumed deterministic (strain driven algorithm)
while the strain rate is assumed constant (d12 /dt in Eq. (5.31)). The numerical value of
assumed constant strain rate can be substituted in Eq. (5.31) for the entire simulation of
the evolution of PDF. As mentioned earlier at the beginning of this section, the FPKE
134
0.0426
Strain (%)
0.0804
600
Plastic Linear
Hardening
400
Extended
Elastic
200
0
0.008
0.006
0.008
0.01
0.004
0.012
Time (Sec)
0.002
Stress (MPa)
0.014
0.01489
Figure 6.11: Comparison of evolution of PDF of shear stress for Drucker-Prager elasticplastic linear hardening material model and extended linear elastic model with random
shear modulus
0.0804
Strain (%)
(FokkerPlanck)
Stress (MPa)
0.003

(Actual)
0.002
0.001
Mean Line
(Actual)
0
0.004
0.008
Mean Line
(FokkerPlanck)
0.012
Time (Sec)
0.01489
Figure 6.12: Comparison of evolution of mean and standard deviation of shear stress
(Problem-I), obtained from FPKE solution and Monte Carlo solution
135
Probability Density
120
100
80
60
40
0.0156 MPa
20
0.01
0.01
0.02
0.03
Stress (MPa)
Figure 6.13: Initial condition for Fokker-Planck-Kolmorogov equation for probabilistic simulation of Drucker-Prager post-yield region with random friction coefficient (Problem-II)
(Eq. (5.31)) describes the evolution of PDF of shear stress (12 ) with time (t), while the
shear strain rate (d12 /dt) describes the evolution of shear strain (12 ) with time (t) also.
Combining the two, the evolution of PDF of shear stress (12 ) with shear strain (12 ) is
easily obtained. Time has been brought in this simulation as an intermediate dimension to
help in solution process, and hence, the numerical value of shear strain rate (d12 /dt) could
be any arbitrary value. This value of shear strain rate will cancel out once one converts
the evolution of PDF of shear stress (12 ) as a function of time, to evolution of PDF of
shear stress (12 ) as a function of shear strain (12 ). For simulation purpose we assumed
the arbitrary value of shear strain rate (d12 /dt) as 0.054/second.
Once the evolution of PDF of shear stress (12 ) is obtained, the evolution of mean,
136
0.3115
0.0577
0.1
Strain (%)
Time (Sec)
0.15
1.08
100
0.2
50
0.02
0.04
0.06
Prob Density
Stress (MPa)
Figure 6.14: Time (or strain) evolution of PDF of shear stress for Drucker-Prager elasticplastic constitutive rate equation with random friction coefficient (Problem-II) (only postyield region is shown, note that the pre-yield region is deterministic)
mode and standard deviations of shear stress were calculated by standard operations on the
PDF. The results are shown in Fig. 6.17. In particular, Fig. 6.17(a) shows the evolution
(surface) of PDF of shear stress (12 ) with time (t)/shear strain (12 ). The evolution of PDF
of shear stress is given in more detail in Fig. 6.17(b) showing contours of the evolutionary
PDF of shear stress (12 ) and also the evolution of mean, mode, and standard deviations of
shear stress. The deterministic solution, obtained using the mean value of shear modulus
(G) as fixed, deterministic input is also shown. As can be seen from Fig. 6.17(b), the
standard deviation of shear stress increases in elastic regime (before approx. 0.056 s) until
the yield point and then decreases as the solution approaches the critical state line. In
other words, close to critical state the uncertainty in shear modulus (G) didnt have much
effect on the uncertainty in shear stress as it had close to yielding. But on the other hand
137
0.324
1.08
Strain (%)
0.07
0.01
0.06
0.1
Stress (MPa)
0.05
10
Mean Line
0.04
10
1
0.1
0.03
0.01
0.02
0.01
0
0.06 0.08
0.1
0.12 0.14 0.16 0.18
0.2
Time (Sec)
Figure 6.15: Contours of time (or strain) evolution of PDF for shear stress, along with
evolutions of mean and standard deviation of shear stress, for Drucker-Prager elastic-plastic
constitutive rate equation with random friction coefficient (Problem-II) (only post-yield
region is shown, note that the pre-yield region is deterministic)
if one follows the evolution of deterministic solution one could see that eventhough the
deterministic solution coincides with mean and mode for pre-yield (elastic) behavior, it
deviates for post-yield behavior, suggesting that the deterministic solution is neither the
mean nor the most probable solution for post-yield response.
When compared with Monte Carlo simulation, shown in Fig. 6.18, it can be seen
that FPKE approach predicted the mean behavior exactly but it slightly over-predicted
138
0.324
Strain (%)
(Monte Carlo)
0.05
Stress (MPa)
1.08
0.04
Mean Line
(Monte Carlo)

(FokkerPlanck)
0.03
Mean Line
(FokkerPlanck)
0.02
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
Time (Sec)
Figure 6.16: Comparison of evolutions of mean and standard deviation of shear stress for
Drucker-Prager elastic-plastic constitutive rate equation with random friction coefficient
(Problem-II), obtained from FPKE solution and Monte Carlo solution
the standard deviation behavior. This difference is attributed to the approximation used
to represent the initial deterministic condition 12 = 0 (Dirac delta function). One may
note that at 12 = 0, the probability of shear stress (12 ) at 12 = 0 should theoretically
be 1 i.e., all the probability mass should theoretically be concentrated at 12 = 0 and
would be best described by a Dirac delta condition. However, for numerical simulation of
FPKE, that Dirac delta initial condition was approximated with a Gaussian function of
mean zero and standard deviation of 0.00001 M P a as shown in Fig. 6.1. This initial error
in standard deviation advected and diffused into the domain during the simulation of the
evolution process. This error could be minimized by better approximating the Dirac delta
initial condition (but at the expense of computational cost).
Next few examples examine results (evolution of PDF of shear stress) by the
introduction of uncertainties in other material parameters (other than shear modulus G).
139
0
Strain (%)
1.62
400
ProbDensity
0.04
200
0.03
0
0.02
0.1
Stress (MPa)
0.01
Time (Sec)
0.2
0.3
(a)
0
1.62
Strain (%)
Mean
0.04
Deterministic
Std. Deviations
Stress (MPa)
0.03
0.02
1
0.01
Mode
0.1
0.005
0.05
0.1
0.15
0.2
0.25
0.3
Time (s)
(b)
Figure 6.17: Low OCR Cam Clay response with random normally distributed shear modulus
(G): (a) Evolution of PDF and (b) Evolution of contours of PDF, mean, mode, standard
deviations, and deterministic solution of shear stress (12 ) with time (t)/shear strain (12 )
140
1.62
Strain (%)
0.04
Std. Deviations (FokkerPlanck)
Stress (MPa)
0.03
Std. Deviations (MonteCarlo)
0.02
Mean (MonteCarlo)
0.01
Mean (FokkerPlanck)
0.05
0.1
0.15
0.2
0.25
0.3
Time (s)
Figure 6.18: Comparison of FPK approach and Monte Carlo approach in obtaining low
OCR Cam Clay response with random normally distributed shear modulus (G) in terms of
evolution of mean and standard deviation of shear stress (12 ) with time (t)/shear strain
(12 )
In examples where we use more than one random material property (see next examples) they
are assumed uncorrelated and independent. The presented development can also deal with
corelated random variables, as described by FokkerPlanckKolmogorov (FPK) equation
5.66. The advection and diffusion coefficients N(1) and N(2) (in Eq. (5.50)) need to be
recomputed accordingly for any correlation between random soil properties.
Additional uncertainty is introduced for the slope of critical state line M in terms
141
of a normal distribution with a mean value of 0.6 and a standard deviation of 0.1. The shear
modulus G is again treated as random, as before. The resulting PDF of shear stress 12 is
now exhibiting a non-symmetric distribution in the post-yield (elasticplastic) region. This
nonsymmetry is evident from different post-yield evolution of mean and mode of shear
stress 12 as seen in Fig. 6.19(b). The nonsymmetry is also evident from the trace of
the surface of PDF of shear stress at time (t) = 0.3s (or shear strain (12 ) = 1.62%)
in Fig. 6.19(a). Another interesting aspect to note (refer to Fig. 6.19(b)) is that for this
assumed combination of material properties (random G and random M ) the post-yield
deterministic shear stress (12 ) under-predicted the post-yield most probable (mode) shear
stress (12 ) but over-predicted the mean shear stress (12 ).
The effect of addition of further uncertainty into the system, in terms of random
normally distributed overconsolidation pressure (p0 ) with a mean value of 0.2 MPa and
standard deviation of 0.07 MPa, in addition to previously introduced random shear modulus
(G) and random slope of critical state line (M ), is shown in Fig. 6.20. It can be seen by
comparing Figs. 6.19 and 6.20, the additional uncertainty in overconsolidation pressure
(p0 ) didnt affect much the probabilistic response of shear stress (12 ) when compared
to its response to random normally distributed shear modulus (G) and random normally
distributed slope of critical state line (M ) (Fig. 6.19). That is, the responses in Figs 6.19
and 6.20 are quite similar, at least qualitatively while quantitatively there are some small
differences.
The uncertainty in overconsolidation pressure (p0 ) didnt even have much effect
on the probabilistic response (PDF) of shear stress (12 ) when considered separately with
142
0
Strain (%)
1.62
400
0.04
ProbDensity 200
0.03
0.02
Stress (MPa)
0.1
0.01
Time (Sec)
0.2
0.3
(a)
Strain (%)
1.62
0.04
Mode
Std. Deviations
Stress (MPa)
0.03
50
0.02
30
10
0.01
Deterministic
Mean
0
0
0.05
0.1
0.15
0.2
0.25
0.3
Time (s)
(b)
(G) and random normally distributed slope of critical state line (M ): (a) Evolution of PDF
and (b) Evolution of contours of PDF, mean, mode, standard deviations, and deterministic
solution of shear stress (12 ) with time (t) /shear strain (12 )
143
0
Strain (%)
1.62
400
ProbDensity
0.04
200
0.03
0.02
Stress (MPa)
0.1
0.01
Time (Sec)
0.2
0
0.3
(a)
0
Strain (%)
1.62
0.04
Mode
Stress (MPa)
0.03
50
0.02
30
10
0.01
Deterministic
Mean
0
0
0.05
0.1
0.15
0.2
0.25
0.3
Time (s)
(b)
(G), random normally distributed slope of critical state line (M ), and random normally
distributed overconsolidation pressure (p0 ): (a) Evolution of PDF and (b) Evolution of
contours of PDF, mean, mode, standard deviations, and deterministic solution of shear
stress (12 ) with time (t)/shear strain (12 )
144
random normally distributed shear modulus (G). This can be seen by comparing Fig. 6.21,
where the probabilistic behavior of shear stress (12 ) for low OCR cam clay model with randomly distributed shear modulus (G) and random normally distributed overconsolidation
pressure (p0 ) was presented, and Fig. 6.17. Similar to the case where the shear modulus was
the only random parameter (Fig. 6.17), here also the post-yield deterministic shear stress
(12 ) overpredicted the post-yield mean and post-yield most probable (mode) shear stress
(12 ).
The probabilistic Cam Clay elasticplastic responses (low OCR) with different
degrees of randomnesses (as presented above) were compared in Fig. 6.22. In that figure
the PDFs of shear stress (12 ) at shear strain (12 ) = 1.62% are shown and compared
with the deterministic value of shear stress obtained by choosing mean value (which in
this case is also mode as we used normal distribution) of material properties. It is very
interesting to note that the deterministic value of shear stress (12 ) is different from the
most probable (mode) value of shear stress (12 ) for all the cases. It either under-predicted
or over-predicted the most probable (mode) value.
Presented methodology is applicable to both hardening (as shown in examples
above) and softening material response. A high OCR CamClay material sample was
analyzed. Following parameters were used: random normally distributed shear modulus
(G) with mean of 2.5 MPa and standard deviation of 0.5 MPa; random normally distributed
slope of critical state line (M ) with mean of 0.6 and standard deviation of 0.1; deterministic
overconsolidation pressure p0 = 0.8 MPa; deterministic applied confinement pressure p =
0.02 MPa; deterministic slopes of the unloadingreloading and normal compression lines
145
3.24
0.6
0.4
Time (Sec)
Strain (%)
0.2
400
ProbDensity
200
0
0.04
0.03
0.02
0.01
0 Stress (MPa)
(a)
0
Strain (%)
1.62
0.04
Deterministic
Mode
Stress (MPa)
0.03
0.02
Mean
50
30
Std. Deviations
0.01
10
0
0
0.05
0.1
0.15
0.2
0.25
0.3
Time (s)
(b)
(G) and random normally distributed overconsolidation pressure (p0 ): (a) Evolution of PDF
and (b) Evolution of contours of PDF, mean, mode, standard deviations, and deterministic
solution of shear stress (12 ) with time (t)/shear strain (12 )
146
500
PDF of Shear Stress
Deterministic Value = 0.02906 MPa

400
300
Random G only
Random G and Random p0
200
Random G and Random M

Random G, Random M, and Random p0
100
0.01
0.02
0.03
0.04
Shear Stress (MPa)
Figure 6.22: Comparison of shear stresses at 1.62% shear strain obtained from low OCR
Cam Clay model with different degrees of randomnesses
in e ln(p ) space = 0.05 and = 0.25, respectively; deterministic initial void ratio
e0 = 2.18. The (arbitrary) value of strain rate is still the same, d12 /dt = 0.054/s. The
resulting PDF of shear stress is shown in Fig. 6.23(a). Fig. 6.23(b) gives a more detailed
view of results, in terms of evolution of contours of PDF, mean, mode, standard deviation
and deterministic solution of shear stress (12 ). Similar to the low OCR simulation with
random G and random M , the high OCR simulation also yielded a symmetric distribution
of PDF shear stress (12 ) in the pre-yield (elastic) regime and a non-symmetric distribution
of PDF shear stress (12 ) in the post-yield (elastic-plastic) regime. Similarly to probabilistic
response for low OCR (with random G and random M ), the deterministic shear stress (12 )
coincided with the mean and most probable (mode) shear stress in the pre-yield (elastic)
regime but deviated in the post-yield non-linear elastic-plastic regime. The deterministic
shear stress (12 ) under-predicted the mean shear stress (12 ) in the entire post-yield regime
147
but over-predicted the most probable (mode) shear stress (12 ) in the region close to the
yield point, though close to critical state line it under-predicted the most probable (mode)
shear stress (12 ).
Verification of probabilistic simulations for high OCR sample were again done using
MonteCarlo approach. Fig 6.24 compares the evolution of mean and standard deviation
of shear stress (12 ) obtained using FPKE approach and MonteCarlo approach. It can be
seen that FPKE approach slightly overpredicted the standard deviation behavior because
of the reasons discussed earlier. That is, the FPKE approach is somewhat wider since the
initial condition was not a Dirac delta function, but rather an approximation. Again, this
deviation in initial condition (from Dirac delta function) can be controlled by choosing an
initial normal distribution for PDF of shear stress with smaller standard deviation, but this
will increase computational cost in solving the resulting finite difference system of equations.
6.4
ElasticPlastic Shear Constitutive Behavior with Probabilistic Yield Criteria

The above examples assumed mean yield criteria and needed solutions of two
el and N el and the other for

equations one for the elastic (preyield) region, governed by N(1)
(2)
ep
ep
the elasticplastic (postyield) region, governed by N(1)
and N(2)
for complete simulation
of probabilistic constitutive behavior. However, if yield surface (yield point in 1D) is

uncertain, uncertainty propagates into separation of elastic and elasticplastic regions. That
is, depending on the degree of uncertainty of yield surface, stress points can only have certain
probability of being in elastic or elasticplastic state.
148
2
0.3
0.2
Time (s)
Strain (%)
0.1
400
Prob
Density
200
0
0.02
Stress (MPa)
0.01
(a)
0
Strain (%)
2.0
0.03
0.025
Mean
Stress (MPa)
0.02
Std. Deviations
0.015
0.01
100
0.005
50
Deterministic
Mode
30
0
0
0.05
0.1
0.15
0.2
0.25
0.3
0.37
Time (s)
(b)
Figure 6.23: High OCR Cam Clay response with random normally distributed shear modulus (G) and random normally distributed slope of critical state line (M ): (a) Evolution
of PDF and (b) Evolution of contours of PDF, mean, mode, standard deviations, and
deterministic solution of shear stress (12 ) with time (t) /shear strain (12 )
149
0
Strain (%)
2.0
0.03
0.025
Std. Deviation (FokkerPlanck)
Std. Deviation (MonteCarlo)
Stress (MPa)
0.02
0.015
0.01
0.005
Mean (MonteCarlo)
Mean (FokkerPlanck)
0
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Time (s)
Figure 6.24: Comparison of FPK Approach and MonteCarlo Approach for High OCR
Cam Clay Response with Random Normally Distributed Shear Modulus (G) and Random
Normally Distributed Slope of Critical State Line (M ) in terms of Evolution of Mean and
Standard Deviation of Shear Stress (12 ) with Time (t)/Shear Strain (12 )
A solution to this problem of uncertain yielding is to assign weights to the elastic

el and N ep ) and diffusion (N el and N ep ) coefficients based
and elasticplastic advection (N(1)
(2)
(1)
(2)
on the cumulative probability density function (CDF) of the yield function (or stress y in
1D) random variable. That combined (weighted) equation, which now has both elastic and
elasticplastic coefficient, appropriately weighted, can then be used to obtain the complete
constitutive behavior with equivalent advection and diffusion coefficients. In other words,
while solving the FPK partial differential equation, for each stress () in the stress domain,
probability weight will be assigned to the elastic and elasticplastic advection and diffu-
150
sion coefficients corresponding to that stress. Mathematically, the equivalent (weighted)
eq
eq
advection and diffusion coefficients (N(1)
and N(2)
) can be written as:
eq
ep
el
N(1)
() = (1 P [y ])N(1)
+ P [y ]N(1)
(6.7)
eq
ep
el
N(2)
() = (1 P [y ])N(2)
+ P [y ]N(2)
(6.8)
where (1 P [y ]) represents the probability of material being elastic, while P [y ]

represents the probability of material being elasticplastic.
For example, for a given CDF of yield stress (shown in Fig. 6.25(a)), the probability
of yielding happening at = 0.0012 MPa1 is P [y ( = 0.0012 MPa)] = 0.8 so that the
equivalent advection and diffusion coefficients are
6.4.1
eq
N(1)
|=0.0012
MP a
ep
el
= (1 0.8)N(1)
+ 0.8N(1)
eq
|=0.0012
N(2)
MP a
ep
el
= (1 0.8)N(2)
+ 0.8N(2)
von Mises Associative Model

The yielding probability weights assigned to the advection (N1 ) and diffusion (N2 )
coefficients are based on the cumulative density function (CDF) of the yield stress, which in
this case corresponds to CDF of the shear strength (cu )). The CDF of yield stress is shown
in Fig. 6.25.
Fig. 6.26(a) shows the evolution of probability density function (PDF) of shear
stress with respect to shear strain for probabilistic von Mises associative plasticity model.
The shear modulus (G) is modeled as normally distributed random variable with a mean
1
Mathematically we will write this probability as P [y ( = 0.0012 MPa])
151
F@Sy D = P@Sy y D
1
0.8
(b)
(a)
0.6
0.4
0.2
0.00005 0.0001 0.00015
y HMPaL
Figure 6.25: CDF of shear strength for von Mises model: (a) very uncertain case, (b) fairly
certain case.
of 2.5 MPa and coefficient of variation (COV) of 20 %. The shear strength (cu ) is also
assumed to be normally distributed with a mean of 0.0001 MPa and COV of 20 %. The
hardening parameter (cu , representing the rate of evolution of cu with plastic strain) is
considered deterministic with an assumed value of 0.3 MPa. The contours of the evolution
of PDF of shear stress with shear strain are shown in Fig. 6.26(b), along with mean, mode
(most probable solution), and the standard deviations. The deterministic solution obtained
using mean values of shear modulus (2.5 MPa) and shear strength (0.0001 MPa) is also
shown. It is interesting to note the smooth response for probabilistic von Mises elasto
plastic model. On the other hand, the deterministic von Mises elasticplastic model shows
an expected sharp change in stiffness at the boundary between elastic and elasticplastic
solutions. In addition to that, the most likely stress response (mode) is different than the
mean and the deterministic stresses. This difference in mode, mean and deterministic stress
responses observed for linear hardening material model is a novel feature that deserves some
152
P[Stress]
10000
8000
6000
4000
2000
0
0
0.0004
0.0216
Stra 0.0324
in (%
0.0432
)
0.0001
0.054
(M
Str
0.0108
ess
0.0002
Pa
0.0003
(a)
0.0003
Stress (MPa)
0.00025
Mode
Mean
0.0002
0.00015
0.0001
0.00005
Std. Deviations
Deterministic
Solution
0
0
0.0108 0.0216 0.0324 0.0432 0.054

Strain (%)
(b)
Figure 6.26: von Mises associative plasticity model with uncertain shear modulus and shear
strength (yield parameter): (a) Evolution of PDF of stress with strain (PDF=10000 was
used as a cutoff for surface plot) and (b) Contours of evolution of stress PDF with strain.
153
attention. This more realistic yielding response contrasts earlier observation of equivalence
of mode, mean and deterministic solutions for probabilistic elastoplasticity with mean
stress yielding (refer to Subsection 6.3.1). Apparently in the case solved here, involving
probabilistic yielding, the difference between mode, mean and deterministic solutions is
present even for linear elastic hardening cases (which will involve perfectly plastic material
model as well).
The probabilistic solution for one of the limiting cases, where shear modulus is
very uncertain (statistical properties are same as the above example), while shear strength
is fairly certain (mean of 0.0001 MPa and COV of 1 %) is shown in Fig. 6.27(a). The CDF
of the fairly certain yield stress is shown in Fig. 6.25. Note that, since yielding is fairly
certain, the mean, mode and deterministic behaviors are very similar in the yielding region.
The other limiting case that was considered has a fairly certain shear modulus
(mean of 2.5 MPa and COV of 1 %), however shear strength is very uncertain (mean of
0.0001 MPa and COV of 20 %). The simulation result is shown in Fig. 6.27(b). Note that,
as shear modulus is fairly certain, the standard deviations lines almost match the mean
and mode response line. The response is still probabilistic, if very sharp (in stress PDF
strain space) so instead of showing contours (most of which overlap) shown are only the
mean, mode, standard deviation and the deterministic response. The certainty of response
as shown in Fig. 6.27(b) is attributed to the lack of shear strength parameter cu (which is
ep
ep
the only very uncertain parameter here) in either of the coefficients N(1)
and N(2)
. It is
important to note, however, that the deterministic response is still quite a bit different from
mode and mean response.
154
0.0003
Stress (MPa)
0.00025
Deterministic
Solution
0.0002
0.00015
Std. Deviations
0.0001
Mode
0.00005
Mean
0
0
0.0108 0.0216 0.0324 0.0432 0.054

Strain (%)
(a)
Mode
0.00025
Stress (MPa)
0.0002
Mean
Std. Deviations
0.00015
0.0001
Deterministic Solution
0.00005
0.0108
0.0216 0.0324
Strain (%)
0.0432
0.054
(b)
Figure 6.27: von Mises elasticplastic model: (a) Shear modulus: very uncertain; shear
strength: fairly certain, (b) Shear modulus: fairly certain; shear strength: very uncertain.
155
6.4.2
Drucker-Prager Associative Model

Similar to the von Mises case, presented are results of probabilistic Drucker-Prager
material model simulation, for the case where both shear modulus and yield parameter
(frictional coefficient, ) are considered as very uncertain. The assumed mean and COV of
normally distributed shear modulus (G) are 2.5 MPa and 20 %, respectively. The frictional
coefficient () has a mean and COV of 0.1 and 20 %, respectively. The other parameters
needed for DruckerPrager probabilistic elasticplastic simulations, were assumed deterministic and were as follows: the bulk modulus K = 3.33 MPa, the rate of evolution of
with plastic strain = 5.5, and the confinement pressure I1 = 0.01 MPa. For a given
confinement pressure (I1 ) and (probabilistic) frictional coefficient (), the CDF of a yield
stress (y = I1 ) is calculated and shown in Fig. 6.28(a).
F@Sy D = P@Sy y D
1
0.8
(b)
(a)
0.6
0.4
0.2
y HMPaL
0.00005 0.0001 0.00015
Figure 6.28: CDF of yield stresses for Drucker-Prager model: (a) very uncertain and (b)
fairly certain
Fig. 6.29(a) shows the evolution of PDF of shear stress with strain. Same results
are presented as contours of evolution of PDF of shear stress with shear strain along with
the mean, mode, and the deterministic solution in Fig. 6.29(b).
156
P [Stress]
10000
8000
6000
4000
2000
0
0
0.0004
0.0001
Str
0.0108
0.0216
Stra 0.0324
in (%
0.0432
)
ess
(M
0.0002
Pa
)
0.0003
0.054
(a)
0.00035
Mean
0.0003
Mode
Stress (MPa)
0.00025
0.0002
0.00015
Std. Deviations
0.0001
0.00005
0.0108
0.0216
0.0324
0.0432
0.054
Strain (%)
(b)
Figure 6.29: Drucker-Prager associative elasticplastic model with uncertain shear modulus
and frictional coefficient: (a) Evolution of probability density function (PDF) of stress with
strain (PDF=10000 was used as a cutoff for surface plot) and (b) Contours of evolution of
PDF with strain
157
The limiting cases, where shear modulus (G) is very uncertain (COV = 20 %),
while frictional coefficient () is considered fairly certain (COV = 5 %) is shown in Figs. 6.30(a).
This case is to be contrasted with the case where shear modulus (G) is fairly certain (COV =
1 %), while frictional coefficient () is very uncertain (COV = 20 %), shown in Figs. 6.30(b).
158
0.0003
Mean
Stress (MPa)
0.00025
0.0002
Mode
0.00015
0.0001
Deterministic
Solution
0.00005
Std. Deviations
0
0
0.0108 0.0216 0.0324 0.0432 0.054

Strain (%)
(a)
0.00035
Mean
Mode
0.0003
Stress (MPa)
0.00025
0.0002
0.00015
Std. Deviations
0.0001
0.00005
0.002
0.0216 0.0324 0.0432
0.054
Strain (%)
(b)
Figure 6.30: Drucker-Prager elasticplastic model: (a) Shear modulus: very uncertain;
frictional coefficient: fairly certain, (b) Shear modulus: fairly certain; frictional coefficient:
very uncertain.
159
Part IV
Stochastic Simulations of Solids

and Structures with Uncertain
Material Properties: Stochastic
Finite Elements
160
Chapter 7
Stochastic Finite Elements: Theory

7.1
Introduction
In mechanics, simulations of static or dynamic behaviors of solids and structures
involve solutions of boundary value problems, which comprise of the equilibrium equation,
A = (t),
(7.1)
together with the strain compatibility equation,

Bu = ,
(7.2)
= D,
(7.3)
and the constitutive equation,
along with a set of additional restraints, called the boundary conditions. In Eqs. (7.1)
(7.3), is generalized stress, (t) is generalized force that can be time (t) dependent, u
is generalized displacements, is generalized strain and A, B, and D are operators which
could be linear or non-linear.
161
Rigorous mathematical theory has been developed for problems where the only
random parameter is the external force (t). In this case, the probability distribution function (PDF) of the response variable will satisfy FPK partial differential equation (cf. Soize
[92]). With appropriate initial and boundary conditions the FPK PDE can be solved for
PDF of response variable. The numerical solution method of FPK equation corresponding
to structural dynamics problems was described by number of researchers (e.g., Langtangen
[58], Masud and Bergman [68]).
The other extreme case, which is of main interest of this dissertation, is when the
stochasticity of the system is purely due to operator uncertainty. Exact solution of the problems with stochastic operator was attempted by Hopf [39], using characteristic functional
approach. Later, Lee [59] applied the methodology to the problem of wave propagation in
random elastic media and derived a FPK equation, satisfied by the characteristic functional
of the random wave field. This characteristic functional approach is very complicated for
linear problems and becomes even more intractable (and possibly unsolvable) for nonlinear
problems with irregular geometries and boundary conditions.
Monte Carlo simulation technique is an alternative to analytical solution of partial
differential equation with stochastic coefficient. Nice descriptions of different aspects of
formulation of Monte Carlo technique for stochastic mechanics problem was described in
a state-of-the-art report edited by Schueller [85]. Monte Carlo method is very popular
tool with the advantage that accurate solution can be obtained for any problem whose
deterministic solution (either analytical or numerical) is known. Monte Carlo technique has
been used by several researchers ( Paice et al. [78], Popescu et al. [82], Mellah et al. [71],
162
DeLima et al. [15], Koutsourelakis et al. [53], Griffiths et al. [33], Nobahar [76], Fenton
and Griffiths [26, 27]) in obtaining probabilistic solution of geotechnical boundary value
problems. However, the major disadvantage of Monte Carlo analysis is the repetitive use
of the deterministic model until the solution variable become statistically significant. The
computational cost associated with it could be very high especially for nonlinear problems
with multiple uncertain material properties.
The difficulty in analytical solution and the high computational cost associated
with Monte Carlo technique lead to the development of numerical methods for the solution
of stochastic differential equation with random coefficient. For stochastic boundary value
problems, stochastic finite element method (SFEM) is the most popular. There exist several
formulations of SFEM, among which perturbation (Kleiber and Hien [51], Der Kiureghian
and Ke [19]; Mellah et al. [71], Gutierrez and De Borst [34]) and spectral (Ghanem and
Spanos [31], Keese and Matthies [50], Xiu and Karniadakis [103], Debusschere et al. [17],
Anders and Hori [2]) methods are very popular. A nice review on advantages and disadvantages of different formulations of SFEM was provided by Matthies et al. [69]. Mathematical
issues regarding different formulations of SFEM was addressed by Deb et al. [16] and by
Babuska and Chatzipantelidis [5]. However, most of the formulations are for linear elastic
problems. Though there has been some published work on geometric nonlinear problems
(Liu and Der Kiureghian [62], Keese and Matthies [50], Keese [49]), there exist only few
published papers on material nonlinear (elasticplastic) problems with uncertain material
parameters.
The major difficulty in extending the available formulations of SFEM to general
163
elasticplastic problem is the high nonlinear coupling in the elasticplastic constitutive
rate equation. First attempt to propagate uncertainties through elasticplastic constitutive
equations considering random Youngs modulus was published only recently, e.g., Anders
and Hori [1, 2]. They took perturbation expansion at the stochastic mean behavior and
considered only the first term of the expansion. In computing the mean behavior they
took advantage of bounding media approximation. Although this method doesnt suffer
from computational difficulty associated with Monte Carlo method for problems having no
closed-form solution, it inherits closure problem and the small coefficient of variation
requirements for the material parameters. Closure problem refers to the need for higher
order statistical moments in order to calculate lower order statistical moments (cf. Kavvas
[47]). The small COV requirement claims that the perturbation method can be used (with
reasonable accuracy) for probabilistic simulations of solids and structures with uncertain
properties only if their COV < 20 % (cf. Sudret and Der Kiureghian [93]). For soils and
other natural materials, COVs are rarely below 20 % (refer to Chapter 3). Furthermore,
with bounding media approximation, difficulty arises in computing the mean behavior when
one considers uncertainties in internal variable(s) and/or direction(s) of evolution of internal
variable(s). These difficulties, associated with probabilistic simulation of the elasticplastic
constitutive equation, prevent application of SFEM to general geomechanics problems.
In this Chapter, the spectral approach (cf. Ghanem and Spanos [31]) is used
in formulation of stochastic finite element method. The FPKE approach, developed in
Part III, is used for constitutive level integrations. Both the formulation and numerical
solution scheme are discussed.
164
7.2
Discretization of Governing Stochastic Partial Differential Equation

The governing partial differential equation in mechanics can be mathematically
written, combining Eqs. (7.1), (7.2) and (7.3), as:
(x)u(x) = (x)
(7.4)
where (x) is a linear/non-linear differential operator, (x) is the external force and u(x) is
the response. If the material properties are uncertain, in Eq. (7.4) becomes a stochastic
linear/non-linear differential operator and as a result the response, u becomes a random
field. Splitting the stochastic operator () into a deterministic part (L) and a random part
(), whose coefficients are zero-mean random fields, one can write Eq. (7.4) as:
[L(x) + (x, )] u(x, ) = (x)
(7.5)
Further, if one split the input material properties random field into a deterministic trend
and zero-mean uncertain residual about trend (refer to Chapter 4, Section 4.2), D(x, ) =
D(x)
+ R(x, ), one can write Eq. (7.5) as:
h
i
L1 (x)D(x)
+ L2 (x)R(x, ) u(x, ) = (x)
(7.6)
where L1 (x) and L2 (x) are deterministic differential operators. For stochastic finite element
formulation, Eq. (7.5) needs to be discretized in both stochastic and spatial dimensions,
specifically:
165
1. The stochastic discretization of the zero-mean fluctuating part of the input material
properties random field (R(x, )) into finite number of independent basic random
variables.
2. The stochastic discretization of the (unknown) response random field (u(x, )) in
functions of the above basic random variables.
3. The spatial discretization of the differential operators (L1 (x) and L2 (x)).
In the following subsections, the above discretization techniques are described in details.
7.2.1
Stochastic Discretization of Input Random Field

For stochastic finite element formulation, the zero-mean uncertain residual of in-
put random field material properties (R(x, ) in Eq. (7.6)) need to be discretized in finite
number of independent discrete random variables. This is similar to the deterministic finite
element formulation, where continuous functions are represented by a finite set of nodal
point parameters. Different discretization techniques
are available. However, all are some
kind of series representations of the form:

R(x, ) =
m
X
i (x)i ()
(7.7)
i=1
where R(x, )
is the zero-mean uncertain residual of any material property random field,
i (), i = 1 to m, are finite number of random variables, and i (x), i = 1 to m, are

deterministic spatial functions. The discretization scheme depends on correlation length of
the input random field. The distance between two adjacent random variables should be
1
2
for details on different techniques, refer to Keese [49]

In R(x, ), x denotes spatial nature of the variable and denotes the stochastic nature of the variable
166
small enough to capture the correlation structure of the random field. On the other hand,
too fine discretization will result in highly correlated random variables and subsequently a
nearly singular correlation matrix.
Some of the series representation techniques, for example Kriging technique [54]
and Karhunen-Loève expansion ([45], [63], [64], [44]), avoid the above problem by implicitly
taking into account the covariance structure. Karhunen-Loève (KL) expansion has some
other useful properties too. KL-expansion is an optimal expansion in the sense that it
reduces the number of stochastic dimension. Each discretized independent random variable
adds one extra dimension to the problem. Hence, the ability to represent the input random
field efficiently with few number of random variables is a very desirable property of the
discretization technique, especially from computational tractability point of view. KLexpansion also minimizes error due to finite representation of the series.
Due to the above desirable properties, in this dissertation, KL-expansion is used
in formulation of stochastic finite element method. In this section the main features of
Karhunen-Loève expansion is discussed. For a detailed description of KL-expansion in
the context of random fields, the readers are encouraged to refer standard text books, for
example by Vanmarcke [97], to Christakos [11], Ghanem and Spanos [31].
Any random field material properties, 3 , D(x, ), having expected value D(x)
and
autocovariance function C(x1 , x2 ) can be expanded in a series as follows (cf. Karhunen [45],
Loève [63, 64], Kac and Siegert [44]):
3
Assuming Gaussian. For treatment of non-Gaussian random fields refer to Keese [49]
This expansion was independently developed by Karhunen [45] in 1947, Loève [63, 64] in 1948, and Kac
and Siegert [44] in 1947
4
167
D(x, ) = D(x)
+ R(x, )
= D(x)
+
n n ()fn (x)
(7.8)
(7.9)
n=1
where, R(x, ) is the zero-mean uncertain residual about the mean and n and fn (x) are the
eigenvalues and eigenvevtors respectively of the covariance kernel. The random variables,
n () are mutually independent, zero mean and have unit variance random variables. After
solving for eigenvalues and eigenvectors of the covariance kernel, the random variables can
be obtained as:
1
n () =
n
R(x, )fn (x)dx
(7.10)
The eigenvalues and the eigenvectors of the autocovariance kernel are given by the following
integral equation, also known as Fredholm equation of the second kind,
C(z1 , z2 )fn (z1 )dz1 = n fn (z2 )
(7.11)
Analytical solution of the above equation Eq. (7.11) is possible only for some
particular covariance functions with regular geometries. The analytical solution procedures
for exponential (first-order Markov) and triangular (binary noise) covariance kernels with
regular geometries were described by Ghanem and Spanos [31]. Book by Van Trees [96]
discussed analytical solution techniques of many more different covariance functions.
However, for non-standard covariance kernels (for example, fractal covariance function, as discussed in Chapter 4) and for problems with irregular geometries, numerical approximations are necessary. The use of numerical approximation though, was a subject of
168
little debate in the past. Some researchers had reservations that approximate solution required more terms than analytical solution (that means increased stochastic dimension) and
hence representation of random field by KL expansion should be limited to cases where analytical solutions were known. Li and Der Kiureghian [61] opined that KL method reduced
to shape function method if the Fredholm equation (Eq. (7.11)) was solved numerically.
However, Keese [49] argued against the above claim that the non-availability of analytical
solution technique limited the efficiency of KL-expansion. Keese [49], citing Atkinson [4]
and Hackbusch [35], stated that the numerical solution converged to the analytical solution
for finer and finer approach spaces and suggested that eigenvectors be obtained on a sufficiently fine mesh and then be interpolated to a coarser mesh for the representation of the
stochastic field.
Among several available numerical techniques 5 , Galerkin type procedure is popular (cf. Ghanem and Spanos [31], Huang et al. [40]). Galerkin type procedure involves
approximation of eigenfunctions of the covariance kernel by linear combinations of some basis functions, usually the standard deterministic finite element shape functions, as follows:
fk (z) =
n
X
dik Ni (z)
(7.12)
i=1
where dik are the unknown coefficients of the k th eigenfunction (fk ) and Ni are the shape
functions. The unknown coefficients can then be obtained by minimizing the error of finite
representation by Galerkin technique. The error due to finite representation of Eq. (7.12)
5
For a detailed review of different available techniques on numerical solution of Fredholm equation of the
second kind, refer to Atkinson [4]
169
can be written by substituting Eq. (7.12) into Eq. (7.11) as:
n =
n
X
dik
i=1

C(z1 , z2 )Ni (z2 )dz2 n Ni (z1 )
Z
(7.13)
Setting the error in Eq. (7.13) to be orthogonal to the approximating space (basis function
space), Eq. (7.13) can be written as:
n
X
i=1
dik
Z Z
D
C(z1 , z2 )Ni (z2 )Nj (z1 )dz2 dz1 n
Ni (z)Nj (z)dz1 = 0
(7.14)
The above equation (Eq. (7.14)) can be written in matrix form as:
B
d
Ad =
(7.15)
C(z1 , z2 )Ni (z2 )Nj (z1 )dz2 dz1
(7.16)
where,
Aij =
Bij =
Z Z
ZD
Ni (z)Nj (z)dz1
(7.17)
ij = ij i
(7.18)
One may note that Eq. (7.15) represents a generalized eigenvalue problem, which can be
solved numerically for d by open source or commercially available implementation. In this
dissertation, the KL eigenvalue problem is solved using LAPACK [3].
Figures 7.1 and 7.2 show first six eigenvalues and eigenvectors of an exponential
covariance kernel having a variance of 1000 kP a2 and a correlation length of 1 m. These
eigenvalues and eigenvectors were obtained from Galerkin type numerical solution (as described above) of Fredholm equation of the second kind (Eq. (7.11)). In this context it may
170
be noted that eigenvalues obtained using Galerkin type procedure are a lower bound of the
correspondingly numbered exact eigenvalues (cf. Ghanem and Spanos [31]). Further, accuracy in estimating eigenvalues using Galerkin type procedure is better than that achieved
for eigenfunctions (cf. Ghanem and Spanos [31]).
Eigenvalue
700
600
500
400
300
200
100
2
Index
Figure 7.1: KL eigenvalues of exponential covariance kernel having variance = 1000 kP a2

and correlation length = 1 m
Figure 7.3 shows the exact exponential covariance having a variance of 1000 kP a2
and a correlation length of 1 m (as defined above). Figure 7.4 shows KL-approximation of
the covariance kernel with up to 6 terms. The KL-approximation of the covariance kernel
is calculated as:
C(x1 , x2 ) =
M
X
k fk (x1 )fk (x2 )
(7.19)
k=1
where M is the term after which the KL-expansion is truncated. In Figure 7.4, the average
truncation errors
6
are also shown. The average error, due to truncation of the series after
include errors due to numerical solution also
171
Eigenvector
1
0.5
0.2
0.4
0.6
0.8
Depth HmL
-0.5
-1 First Eigenvector
Second Eigenvector
Third Eigenvector
Fourth Eigenvector
Fifth Eigenvector
Sixth Eigenvector
Figure 7.2: KL eigenvectors of exponential covariance kernel having variance = 1000 kP a2

and correlation length = 1 m
M terms, is calculated as:
M =
Z Z "
D
2 |x1 x2 |/r
M
X
k=1
k fk (x1 )fk (x2 ) dx1 dx2
(7.20)
It may be observed from Figure 7.4, that with two terms one may represent the
covariance kernel with sufficient accuracy. This is the most powerful property of KLexpansion. As each KL-term (or random variable, in general) adds one extra (stochastic)
dimension to the problem, it is always desirable, from computational tractability point of
view, to minimize the number of random variables in a problem. In general, number of KL
172
1
0.8
0.6
x2
0.4
0.2
nce
Covaria
(kPa^2)
0
1000
800
600
400
0
0.2
0.4
0.6
0.8
x1
Figure 7.3: Exact exponential covariance kernel having variance = 1000 kP a2 and correlation length = 1 m (C(x1 , x2 ) = 2 e|x1 x2 |/r , 2 = 1000 kP a2 and r = 1.0 m)
terms needed to represent any covariance kernel with sufficient accuracy depends on the
ratio of correlation length to the domain length. The smaller the ratio, the more terms are
needed. It is plausible as, when the correlation length is close to the domain length i.e.,
when the random variables in the domain are highly correlated, a single random variable
can be used to represent the entire domain. On the other hand, when random variables are
loosely correlated or in other words, the ratio of correlation length to domain length is small,
one needs more random variables to represent the domain. For example, Figure 7.6 shows
the KL-approximations of an exponential covariance kernel where the ratio of correlation
length to domain length is 0.05 (the exact covariance kernel is shown in Figure 7.5). As
can be observed from Figure 7.6, at least 4 KL-terms are necessary to represent the kernel
of Figure 7.5 with sufficient accuracy.
173
1-term approximation
Estimated Error = 84.8625 kPa2
1
0.8
0.6
x2
0.4
2-terms approximation
1
0.8
0.6
x2
0.4
0.2
0.2
0
700
600
500
0
0
750
700
650
600
1
0.2 0.4 0.6 0.8

x1
1
0.8
0.6
x2
0.4
0.2 0.4 0.6 0.8

x1
1
0.8
0.6
x2
0.4
0.2
0.2
0
800
700
600
0
0
800
700
600
1
0.2 0.4 0.6 0.8

x1
1
0.8
0.6
x2
0.4
1
0.8
0.6
x2
0.4
0.2
0.2
0
800
700
600
0
0
800
700
600
0
0.2 0.4 0.6 0.8

x1
0.2 0.4 0.6 0.8

x1
0.2 0.4 0.6 0.8

x1
Figure 7.4: KL approximations (with estimated errors) of exponential covariance kernel

having variance = 1000 kP a2 and correlation length = 1 m
174
1
0.8
0.6
x2 0.4
0.2 0.4 0.6 0.8
Covaria
nc e
(kPa^2)
0.2
1000
750
500
250
0
1
x1
Figure 7.5: Exact exponential covariance kernel with variance = 1000 kP a2 and correlation
length = 0.05 m (C(x1 , x2 ) = 2 e|x1 x2 |/r , 2 = 1000 kP a2 and r = 0.05 m)
A convergence study on KL-expansion was carried out by Huang et al. [40]. They
concluded that, in addition to the ratio of correlation length to domain length, the number
of KL-terms needed to represent any covariance kernel with given accuracy also depended
on the smoothness of the covariance kernel.
7.2.2
Stochastic Discretization of Unknown Solution Random Field

The response (u(x, )) of Eq. (7.5) is also a random field. Hence, for formulation
of stochastic finite element, u(x, ) also needs to the discretized in the stochastic dimension.
Assuming that the solution random field u(x, ) is a second-order random field, one can
expand it by KL-expansion of the form:
u(x, ) =
L
X
j=1
ej j ()bj (x)
(7.21)
175
1-term approximation
1
0.8
0.6
x2
0.4
0.2 0.4 0.6 0.8

x1
1
0.8
0.6
x2
0.4
0.2
0.2
0
200
150
100
50
0
1
0
300
200
100
0
0
1
0.8
0.6
x2
0.4
0.2 0.4 0.6 0.8

x1
1
0.8
0.6
x2
0.4
0.2
0.2
0
300
200
100
0
0
0
750
500
250
0
1
0.2 0.4 0.6 0.8

x1
1
0.8
0.6
x2
0.4
0.2 0.4 0.6 0.8

x1
1
0.8
0.6
x2
0.4
0.2
0.2
0
750
500
250
0
0
0
750
500
250
0
1
0.2 0.4 0.6 0.8

x1
0.2 0.4 0.6 0.8

x1
Figure 7.6: KL approximations (with estimated errors) of exponential covariance kernel

having variance = 1000 kP a2 and correlation length = 0.05 m
176
But the problem with KL-expansion of response random field (u(x, )) is that its covariance
function is not known a priori. Further, it is also not known whether the solution (u(x, ))
will be a Gaussian random field and hence nothing can be said about whether the set ()
is a Gaussian vector. Therefore, Eq. (7.21), in its present form is of very little use, other
than providing a way of separating the random part from the spatial part.
Hence, to circumvent this problem, noting that the solution random field will
involve some functional of known random variables and unknown deterministic functions of
the form,
u(x, ) = [i (), x]
(7.22)
where [, ] is some nonlinear functional of its arguments, an expansion, which will involve
a basis of known random variables with deterministic coefficients, is sought. The deterministic coefficients can be found by minimizing some norm of the error resulting from a
finite representation. This is similar to Fourier series solution of deterministic differential
equations where series coefficients are determined by satisfying some optimality criterion.
Ghanem and Spanos [31] used polynomial chaos (cf. Wiener [100]) as such basis.
The second-order random variables () can be expanded in polynomial chaos expansion as
(cf. Ghanem and Spanos [31]):
(j)
j () = ai0 0 +
(j)
ai1 1 (i1 ) +
i2
i1 X
X
X
i1 =1 i2 =1 i3 =1
(j)
ai1 i2 2 (i1 , i2 )
i1 =1 i2 =1
i1 =1
i1
X
X
(j)
ai1 i2 i3 3 (i1 , i2 , i3 ) + ..........
(7.23)
177
where, ai1 ......ip are deterministic constants, independent of and p (i1 , ......, ip ) is the pth
order polynomial chaos. The polynomial chaoses can be constructed using the following
equation (cf. Ghanem and Spanos [31]):
(7.24)
1 T
i1 .....in

e 2
polynomial, can be written, for writing conve-
n (i1 , ...., in ) = (1)n

Eq. (7.23), truncating after P th
nience, as follows:
j () =
P
X
i=0
(j)
where, i
(j)
i i [{r }]
(7.25)
and i [{r }] are identical to ai1 ......ip and p (i1 , ......, ip ) respectively. In Eq. (7.25),
P denotes the total number of polynomial chaoses used in the expansion, excluding the
zeroth order term. Given the number of terms, M used in KL-expansion of the input random field, and the order of homogeneous chaos, p, the total number of polynomial chaoses,
P may be determined by the following equation (cf. Ghanem and Spanos [31]):
P =1+
p
s1
X
1 Y
(M + p)
s!
s=1
(7.26)
r=0
Substituting Eq. (7.25) for k (), Eq. (7.21) becomes:
u(x, ) =
P
L X
X
j=1 i=0
(j)
i i [{r }]cj (x)
(7.27)
where,
cj (x) = ej bj (x)
(7.28)
178
Changing the order of summation in Eq. (7.27) one can write:
u(x, ) =
P
X
i=0
P
X
i=0
i [{r }]
L
X
(j)
i cj (x)
(7.29)
j=1
i [{r }]di (x)
(7.30)
L
X
(7.31)
where,
di (x) =
(j)
i cj (x)
j=1
7.2.3
Spatial Discretization of the Differential Operator

In finite element setting, for spatial discretization of deterministic differential op-
erators, shape function method (cf. Zienkiewicz and Taylor [106], Bathe [7]) is very popular.
In traditional deterministic finite element method, shape function method involves expanding the response function along a finite dimensional basis (usually linear), called the shape
function. Using the same principle, in case of stochastic finite element formulation, the
spatial component (di (x), in Eq. (7.29)) of the response random field can be expanded as
follows:
di (x) =
N
X
n=1
where lm are the shape functions.
dni ln (x)
(7.32)
179
7.3
Stochastic Finite Element Formulation

Expanding R(x, ) and u(x, ) in Eq. (7.6) by KL- and PC-expansion, one can
write:
"
L1 (x)D(x)
+
M
X
L2 (x)
m m ()fm (x)
m=1
P
X
i=0
i [{r }]di (x) = (x)
(7.33)
Further, expanding the spatial component of the response random field (di (x)) by shape
function method, one can write Eq. (7.33) as:
"
L1 (x)D(x)
+
M
X
L2 (x)
m m ()fm (x)
m=1
P
X
i=0
i [{r }]
N
X
dni ln (x) = (x)
(7.34)
n=1
Eq. (7.34) can be rearranged as follows:
N
P X
X
i=0 n=1
"
i [{r }]L1 (x)D(x)l

n (x) +
M
X
m=1
m ()i [{r }] L2 (x) m fm (x)ln (x) dni = (x)

(7.35)
Galerkin type procedure can be applied to Eq. (7.35) to solve for the unknown
coefficients (dmi ) of the PC-expansion of the response random field. To this end, one may
multiply both sides of Eq. (7.35) by j [{r }] lk (x) and integrate to yield:
N
P X
X

Z
lk (x)dx +
L1 (x)ln (x) D
dni i [{r }] j [{r }]
i=0 n=1
M
X
m=1
m () i [{r }] j [{r }]
L2 (x)ln (x)
#
Z
o
m fm (x) lk (x)dx = j [{r }]
(x)lk (x)dx
D
(7.36)
180
where, k = 1, ..., N and j = 0, ..., P . Eq. (7.36) can be written in a familiar form as:
N
P X
X
i=0 n=1
"
i [{r }] j [{r }]
Knk
M
X
m=1
m () i [{r }] j [{r }]
Kmnk
dni = j [{r }]k

(7.37)
where K , the deterministic component of the generalized stiffness matrix, is:
Knk
=
lk (x)dx
L1 (x)ln (x) D
(7.38)
K , the stochastic component of the generalized stiffness matrix, is:
=
Kmnk
L2 (x)ln (x)
o
m fm (x) lk (x)dx
(7.39)
and k , the generalized force vector, is:
k =
(x)lk (x)dx
(7.40)
Taking expectation on both sides of Eq. (7.37) and noting that:
hi [{r }]j [{r }]i = ij h [{r }] [{r }]i
(7.41)
one can write Eq. (7.37) as:
N
X
n=1
h [{r }] [{r }]i Kmn

dni +
N
P X
X
j=0 n=1
dnj
M
X
k=1
= hm i [{r }]i
hk ()i [{r }]j [{r }]i Kmnk
(7.42)
181
Normalizing the polynomial chaoses and denoting:
cijk = hk ()i [{r }]j [{r }]i
(7.43)
one can simplify Eq. (7.43) as:
N
X
n=1
Kmn
dni
P
N X
X
dnj
n=1 j=0
M
X
k=1
cijk Kmnk
= hFm i [{r }]i
(7.44)
Eq. (7.44) can be solved for the unknown coefficients of the PC. Then, using PC-expansion,
the solution (displacement) random field can be constructed as follows:
u(x, ) =
P
X
i=0
i [{r }]()
N
X
dni ln (x)
(7.45)
n=1
Eq. (7.44) can be written in a familiar matrix form as:
K
=
(7.46)
is the generalized stiffness matrix of order (N P ) X (N P ), u

where K
is the generalized
is the generalized force vector of order
displacement vector of order (N P ) X 1, and
(N P ) X 1. One may also note that the coefficient tensor (cijk ), needed to generate the
generalized stiffness matrix (K), is the expected value of product of PC-basis functions and
hence is problem independent. The components of the coefficient tensor (cijk ) can easily be
pre-calculated symbolically (for example, using Mathematica [101]).
182
7.3.1
Non-Linear (ElasticPlastic) Formulation

For elasticplastic (nonlinear) problems, Eq. (7.46) needs to be solved incremen-
tally. To this end, one can write Eq. (7.46) in force-residual form as:
r(
u, ) = 0
(7.47)
where, u
are the generalized degrees of freedom and is the control parameter. Differentiating Eq. (7.47), one can write the rate form of force-residual equation (Eq. (7.47))
as,
ri =
ri
ri
=0
u j +
uj
(7.48)
Denoting, the variation of vector r w.r.t the components of u

, while keeping fixed, as the
tangent stiffness matrix
(S):
ri
= Sij
uj
(7.49)
and denoting the variation of negative of vector r w.r.t , while keeping u

fixed, as the load
vector
(
q ):
7
8
ri
= qi
sometimes called the Jacobian, especially in the mathematics literatures

sometimes called the control vector, especially in the mathematica literatures
(7.50)
183
one can write Eq. (7.48) in vector form as,
r = Su
q = 0
(7.51)
At regular points of u
space, the tangent stiffness matrix is nonsingular and
hence one can solve the force-residual rate equation for u
as,

u
= S1 q
(7.52)
or,
u
=
d
u
= S1 q = v
d
(7.53)
where, v = S1 q is the incremental velocity vector. Eq. (7.53), which is a set of nonlinear
ODE can be solved numerically for u
with a set of initial condition (at = 0, u
=u
0 ) using
either pure incremental method (forward Euler method) or incrementaliterative method
(Newton method). In this dissertation, only the pure incremental method is considered and
is the described in the following.
Pure Incremental Method

According to forward Euler method (pure incremental method, no iteration),
knowing the solution of u
at the nth step (
un ) one can obtain the solution at (n + 1)th
step (
un+1 ) as,
u
n+1 = u
n + n u
n
(7.54)
184
or,
un = u
n+1 u
n = Sn1 qn n = vn n
(7.55)
In finite element implementation, generally vn is solved first using Sn vn = qn and then is

substituted into Eq. (7.55) to solve for
un . The generalized tangent stiffness matrix (Sn )
needs to be re-evaluated at each step because the constitutive properties of the material
change as the material plastifies. Hence in our stochastic finite element formulation, as
the material plastifies, at each integration point (Gauss point) of the generalized stiffness
in Eq. (7.38) and
in Eq. (7.46)), the material properties (D
matrix (K
m fm (x), where
the subscript m denotes the KL-mode, in Eq. (7.39)) change obeying the elastic-plastic
constitutive rate equation, which in most general form can be written as (Eq. (5.1) of
Chapter 5 is re-written for convenience):
dij (xt , t)
dkl (xt , t)
= Dijkl (xt , t)
dt
dt
(7.56)
where the material properties tensor Dijkl (xt , t) can be either elastic or elastic-plastic:
Dijkl =
el
Dijkl
U f el
el
Dijmn
D
mn pq pqkl
el
Dijkl
f el U
f
r
rs rstu tu q
when
f < 0 (f = 0 df < 0)
(7.57)
when
f = 0 df = 0
el is the elastic material properties tensor and D ep is the elasticplastic material

where Dijkl
ijkl
properties tensor, f is the yield function, which is a function of stress (ij ) and internal vari-
185
ables (q ), U is the potential function (also a function of stress and internal variables). The
internal variables (q ) could be scalar (for perfectly-plastic and isotropic hardening models),
second-order tensor (for translational and rotational kinematic hardening) or fourth-order
tensor (for distortional hardening) or combinations of the above. The same classification
applies to the direction of evolution of internal variables (r ).
Eq. (7.56), when written in probability density space, becomes a FokkerPlanck
Kolmogorov equation, which, in the most general form, was derived in Chapter 5 as (refer
to Eq. (5.66)):

P (ij (xt , t), t)
=
t
mn

Z t
;
d Cov0
+
ab
0

ab (ab (xt , t ), Dabcd (xt ), cd (xt , t ) P (ij (xt , t), t)

Z t
2
+
mn ab
0

P (ij (xt , t), t)
(7.58)
One may note that in Eq. (7.58), t is the pseudo-time of the constitutive rate equation
(Eq. (7.56)) and ij is a random operator tensor, which is a function of stress tensor (ij ),
material properties tensor (Dijkl ), and strain tensor (kl ). Eq. (7.58) can be written in a
compact form as follows:

o
P (ij , t)
n
=
N(2)mnab P (ij , t)
N(1)mn P (ij , t)
t
mn
ab
(7.59)
where N(1) and N(2) are advection and diffusion coefficient respectively. The advection
186
and diffusion coefficients are function of statistical properties of material parameters and
strain as well as the type of constitutive model. The solution procedure of Eq. (7.59) for
different constitutive models and material parameters was described in details in Chapter 6.
In our stochastic finite element formulation, at each integration point (Gauss point), the
above equation (Eq. (7.59)) can be solved for probability density of stress at each KL-space,
including the mean-space, with corresponding (random) material properties at those spaces
and (random) strain increment. For example, at the k th global load increment step, the
spatially discretized material parameters are (D(x))

k and ( m m ()fm (x))k , where the
subscript m denotes the KL-mode. The spatially discretized incremental strain random
field at the k th global load increment step can be calculated, by differentiating Eq. (8.12)
as follows:
((x, ))k =
P
X
i=0
i [{r }]()
N
X
n=1
dni Bn (x)
(7.60)
where, Bn (x) is the derivative of shape function.

For advanced global level solution schemes such as Newton-type incrementaliterative scheme, this probability density of stress at each KL-space could be used to
compute the internal forces. However, in this dissertation, only pure incremental solution scheme is considered. The pure incremental solution scheme only needs information
in Eq. (7.38)
on the new tangent material properties at each KL-space (for example, D
and
m fm (x), where the subscript m denotes the KL-mode, in Eq. (7.39)) to increment
forward. The new tangent material properties at each KL-space can be calculated from the
probability density of stress at each KL-space. To this end, one can write the constitutive
187
rate equation (Eq. (7.56)), after taking expectation on both sides and assuming the rate of
change of mean stress as -correlated, as:
hij (t)i
=
t

kl (t)
kl (t)
Dijkl (t)
+ V ar Dijkl (t)
t
t
(7.61)
In Eq. (7.39), the expected value of rate of change of mean stress can be easily computed
from the probability density function of stress by standard integration. Having known the
l.h.s of Eq. (7.39), the pseudo-time evolution of mean and variance behaviors of product of
tangent material modulus and pseudo strain rate can be obtained using a least square type
procedure. The statistical properties of the tangent material modulus can then be easily
de-coupled, knowing the statistical properties of pseudo strain rate.
7.3.2
Dynamic Formulation
The stochastic finite element formulation as presented above can easily be extended
to solve dynamic problems (both linear elastic and elasticplastic) with deterministic forcings. For dynamic problems, the governing equilibrium equation takes the following form:
u
u
(t) + C u(t)
+ K
M
(t) = (t)
(7.62)
, C,
and K
are the generalized mass, damping, stiffness matrices respectively.
where M
is the
, u,
and u
u
are the generalized acceleration, velocity, and displacement vectors.
generalized external force vector. Similar to the elasticplastic static problem, for dynamic
problems also, the governing equation (Eq. (7.62)) needs to be solved incrementally. Force-
188
residual form of Eq. (7.62) can be written as:
(t), u(t),
r(u
u
(t), t) = 0
(7.63)
There exists several techniques for solving above type of equation (Eq. (7.63)).
For details on available techniques, the readers are encouraged to refer to any standard
finite element methods textbook (for example, Bathe [7]). In the following the Newmarks
method (cf. Bathe [7]), which is used in the subsequent chapters, is outlined.
The Newmark Method

and acFor solution of the generalized displacements (t+t u
), velocities (t+t u),
celerations (t+t u
) at time t + t, following the dynamic equilibrium equation (Eq. (7.62)),
Newmarks constant-average-acceleration scheme [75] assumes the following:
t+t

+ (1 ) t u
+ t+t u
t
u =t u

1
t+t
t
t
+
u
= u
+ u t +
tu
2
(7.64)
t+t
t2
(7.65)
where the parameters and control integration accuracy and stability. Bathe [7] suggested
to use 0.50 as minimum value of and 0.25(0.5 + )2 as the minimum value of . From
Eqs. (7.64) and (7.65), one can obtain expressions for
t+t u
and
t+t u,
in terms of
t+t u
,
which when substituted in Eq. (7.62), yields the following equation:
ef f
K
t+t
u
=
t+t
ef f
(7.66)
189
ef f is given as:
where, the effective stiffness matrix, K
ef f = K
+ a0 M
+ a1 C
K
(7.67)
ef f is given as:
where a0 = 1/(t2 ) and a1 = /(t). The effective force vector,
t+t

ef f + M
a0 t u
+ a3 t u
+ C a1 t u
+ a5 t u
ef f = t
+ a2 t u
+ a4 t u
(7.68)
where a2 = 1/(t), a3 = (1/2) 1, a4 = (/) 1, and a5 = 0.5t((/) 2). After solving

for
t+t u
) and generalized velocities

from Eq. (7.66), the generalized accelerations (t+t u
at t + t can be computed as:

(t+t u)
t+t
= a0
u
t+t
= t u
+ a6
u
t+t

a3
u
t u
a2 t u
t
+ a7
u
t+t
(7.69)
(7.70)
where a6 = t(1 ) and a7 = t.
7.4
Post-Processor: Estimation of Response Statistics

The spectral stochastic finite element formulation 9 , presented in the previous
section, solves for unknown coefficients of the PC-expnasion. The solution random field
can then be constructed using the PC-expansion as shown in Eq. (7.45). However, for
engineering applications, the statistical properties of the solution random field are of interest. All common statistics of engineering importance, for example moments (Ghanem
9
Eq. (7.46) for linear elastic problems, Eq. (7.47) for nonlinear elasticplastic problems, and Eq. (7.63)
for dynamic problems
190
and Spanos [31], Anders and Hori [1, 2]), cumulants (Ghanem and Spanos [31]), reliability
index/probability of failure (Sudret and Der Kiureghian [93]), probability density function
(Ghanem and Spanos [31], Xiu and Karniadakis [102]) and cumulative density function (Yu
[104]), can be calculated from the PC-expanded solution random field. In fact, certain properties of the PC-basis functions vastly simplifies, as can be seen in the following subsections,
the response statistics estimation process. In the following subsections estimation procedures for only the most common statistics, namely mean, autocovariance and probability
density function, from the PC-expanded random field are discussed.
7.4.1
Mean and Autocovariance

Taking expectation on both sides of Eq. (7.45), one can easily calculate the mean
behavior of the solution random field as:
hu(x, )i =
=
P
X
i=0
P
X
i=0
hi [{r }]()i
N
X
dni ln (x)
(7.71)
n=1
hi [{r }]()i di (x)
(7.72)
Using the properties of PC-basis functions, Eq. (7.72) can be further simplified. Except
the 0th order, which is 1, the PC-basis functions have zero-mean. Hence, Eq. (7.72) can be
simplified as:
hu(x, )i = d0 (x)
(7.73)
191
The covariance matrix of the solution random field can be computed as:
K2u =
P
X
i=0
+* P
+
X
hi [{r }]()i di (x)
hj [{r }]()i dj (x) d0 (x)dT0 (x)
(7.74)
j=0
Noting that the PC-basis functions are orthogonal to each other, one can simplify Eq. (7.74)
as:
K2u =
P
X
k=1
hk [{r }]() k [{r }]()i dk (x)dTk (x)
(7.75)
where the variances of the problem-independent PC-basis functions can easily be precalculated analytically using any symbolic operation software (for example Mathematica
[101]).
7.4.2
Probability Density Function

In theory, probability density function at the finite element nodes could analytically
be calculated using variable transformation method (cf. Montgomery and Runger [72]), as at
each node the solution is obtained as a function of random variables (PC-basis functions).
However, difficulty arises in analytical computations for high-dimensional PC-expansions
(cf. Keese [49]). To circumvent this problem, Ghanem and Spanos [31] proposed to obtain
the approximate probability density function using Edgeworth expansion
10
[21]. Other
methods, available in the literature, to approximately calculate the probability density

function from PC-expanded solution random field include estimation by histogram (Yu
[104]), method of distribution (Xiu and Karniadakis [102]), and first-order reliability method
10
Also known as Gram-Charlier expansion
192
(Sudret and Der Kiureghian [93]). In this dissertaion, the Edgeworth exapnsion method,
as proposed by Ghanem and Spanos [31] is used to approximately calculate the probability
density function at the finite element nodes and is outlined below.
Estimation of Probability Density Function using Edgeworth Expansion

The PC-expanded spatially discretized solution random field, at each finite element
node, can be written as:
u() =
P
X
k [{()}]d
k
(7.76)
k=0
where dk , k = 0, ..., P are the deterministic coefficients of the PC-expansion. Multiplying
both sides of Eq. (7.76) by p [{()}]

and averaging throughout, one can write Eq. (7.76)
as:
u()p [{()}]
= dp
(7.77)
Writing the expected value on the l.h.s of Eq. (7.77) in terms of (P+1)-dimensional joint
probability distribution function of u() and (),

gu(),()
, Eq. (7.77) can be re-written as:
...
g
du d = dp
u p [{}]
(u, )
u(),()
(7.78)
Expanding the joint probability distribution function of u() and (),

gu(),()
, in a multidimensional truncated Edgeworth series [21] and substituting Eq. (7.76) for u, one can write
193
Eq. (7.78) as:
P
X
k=0
dk
P
+1
X
p=0
bp
...
m [{}]
p [{u, }]

du d = dk
k [{}]
(u, )
u(),()
(7.79)
where bp , p = 0, ..., P +1 are the unknown coefficients of the Edgeworth series. u(),()
(u, )
is the (P+1)-dimensional joint Gaussian probability distribution function of u() and (),
which is completely characterized by the variance of the solution (displacement) and known
PC coefficients. One may note that, Eq. (7.79) is a system of P equations with P + 1
unknowns. The additional equation may be obtained by assuming the vector to be jointly
normal. Mathematically, this condition can be written as:
gu(),()
du = ()
(7.80)
One may solve Eqs. (7.79) and (7.80) simultaneously to obtain the unknown coefficients
(bp ) of the Edgeworth series. Solving for the coefficients (bp ), the (P+1)-dimensional joint
can be expressed in a
probability distribution function of u() and (),
gu(),()
(u, ),
truncated Edgeworth series as:
=
gu(),()
(u, )
P
+1
X
bp p [{u, }]
(u, )
u(),()
(7.81)
p=0
The Marginal probability distribution of the solution (u()) can then be obtained by integrating the joint probability distribution over the set of all points for which u() = u.
Mathematically,
gu() (u) =
gu(),()
d
(7.82)
194
Chapter 8
Stochastic Finite Elements:

Numerical Examples and
Verifications
8.1
Introduction
In this chapter, the stochastic finite element method proposed in Chapter 7, is
applied to obtain probabilistic solutions of 1D static shear loading of a soil column, 1D

dynamic response of a soil column subjected to a sinusoidal base displacement, and the 1D
dynamic response of a soil column subjected to an earthquake displacement time series,
with soil properties obtained from real CPT measurements. The governing (stochastic) differential equations of the problems are derived first, using energy principle. The differential
equations are then discretized, using the techniques discussed in Section 7.2, in both spatial
195
and stochastic dimensions and the governing stochastic finite element system of equations
are obtained using the Galerkin type procedure described in Section 7.3. The statistics
(mean, standard deviation, and probability density function) of the solutions are finally obtained using methods described in Section 7.4. Open source FORTRAN/C/C++ libraries
are used for numerical computations. LAPACK [3] is used for KL eigen analysis. UMFPACK [14] is used to solve the global level generalized degrees of freedom stochastic finite
element system of equations. SUNDIALS [38] is used to solve the semi-discretized FPK
partial differential equations at the constitutive level. For evaluation of high-dimensional
integration, in estimating the advection and diffusion coefficients of FPK partial differential
equations corresponding to elasticplastic constitutive models, GNU Scientific Library [29]
is used. Readily available open source code (LMFIT [74]) for Levenberg-Marquardt ([60],
[67]) type least square regression analysis is used in estimating the evolution of statistical properties of material properties from the evolutionary statistical properties of internal
stress (refer to Eq. (7.61)).
8.2
8.2.1
One Dimensional Static Problem

Problem Statement and Formulation
Problem Statement
A 1-D soil column of length L (= 1.0m) and area A (= 0.01m2 ) is subjected to a
deterministic static shear force (varied from 0 to 0.1 N) at the top, with its base fixed.
The schematic is shown in Fig. 8.1. The shear modulus of the soil column is a Gaussian
(= 2.5 kPa), variance 2 (= 0.15 kP a2 ), and covariance
random field with mean value G
196
Figure 8.1: Schematic of static 1D soil column (shear beam) example
function C(x1 , x2 ) (assumed exponential with a correlation length r = 0.3 m). Determine
the complete statistical description of the displacement at the top node of the soil column.
Two different material models linear elastic and von Mises associative elasticplastic with
a deterministic shear strength, cu (= 5 Pa) and rate of evolution of shear strength with
plastic strain, cu (= 2 kPa) are considered. One may note that the material properties
assumed in this example for the shear beam are not very realistic for soils. The intention of
this example is only to verify the mathematical formulation and computer implementation.
A more realistic example for soil is shown in Section 8.4, where the material properties are
obtained from real CPT measurements.
197
SFEM Formulation
The strain energy, U stored in the soil column is given as:
1
U= A
2
(x)(x)dx
(8.1)
where (x) and (x) represent the shear stress and shear strain as a function of the location
on the soil column. If one considers random transverse displacement be u(x,), then one
can write:
(x, ) =
u(x, )
x
(8.2)
and
(x, ) = G(x, )(x, ) = G(x, )
u(x, )
x
(8.3)
Substituting Eqs. (8.2) and (8.3), one can write Eq. (8.1) as:
1
U= A
2
G(x, )
L
u(x, )
x
2
dx
(8.4)
Similarly, one can define the potential energy due to external force as:
U = A
u(x, )dx
(8.5)
Expanding the shear modulus random field using KL-expansion and the displacement random field using PC-expansion, one can write Eq. (8.4) as:
198
" P
#!
X
i [{r ()}]di (x)
x
L
i=0
!
P
M
X
X
j [{r ()}]dj (x) dx

k () k hk (x)
G(x)
+
1
U= A
2
(Z
(8.6)
j=0
k=1
and Eq. (8.5) as:
U =A
Z X
P
i [{r ()}]di (x) dx
(8.7)
L i=0
As in deterministic finite element, the function di (x) in Eqs. (8.6) and (8.7) can
be expanded with deterministic shape functions as:
di (x) =
N
X
dmi lm (x)
(8.8)
m=1
where lm is the shape function. Substituting Eq. (8.8), one can write Eq. (8.6) as:
" P
#!
N
X
X
dmi lm (x)
i [{r ()}]
x
L
m=1
i=0
!
Q
M
P
X
X
X
k () k hk (x)
G(x)
+
j [{r ()}]
dnj ln (x) dx (8.9)
1
U= A
2
(Z
j=0
k=1
n=1
and Eq. (8.7) as:
U = A
Z X
P
L i=0
i [{r ()}]
N
X
dmi lm (x) dx
(8.10)
m=1
The unknown coefficients of the PC-expansion (dmi ) can be found by minimizing

the total potential energy. To this end, one can minimize the total potential energy (U U )
199
of the system as:
(U U )
=0
dnj
(8.11)
Substituting Eqs. (8.9) and (8.10), one can write Eq. (8.11) as:
N
P X
X
dnj i [{r ()}]j [{r ()}]
j=0 n=1
G(x)
+
Z
L
M
X
!
k () k hk (x)
k=1
ln (x)
x
#
Z
lm (x) dx = i [{r ()}] lm (x)dx

x
L
(8.12)
Introducing:
Bn (x) =
ln (x)
x
(8.13)
and,
k (x) =
k hk (x)
(8.14)
N
P X
X
j=0 n=1

Z
dnj i [{r ()}]j [{r ()}]

Bn (x)G(x)B
m (x)dx
M
X
k=1
k ()
Bn (x)k (x)Bm (x)dx

L
)#
= i [{r ()}]
lm (x)dx
L
(8.15)
200
Taking expectation on both sides of Eq. (8.15) and further introducing deterministic component of generalized stiffness matrix,
Kmn =
Bn (x)G(x)B
m (x)dx,
(8.16)
stochastic component of generalized stiffness matrix,
Kmnk =
Bn (x)k (x)Bm (x)dx,
(8.17)
generalized force vector,
Fm =
lm (x)dx,
(8.18)
and coefficient tensor,
cijk = hk ()i [{r ()}]j [{r ()}]i ,
(8.19)
N
X
n=1
Kmn dni +
N X
P
X
n=1 j=0
dnj
M
X
k=1
cijk Kmnk = hFm i [{r ()}]i
(8.20)
Eq. (8.20) is the spectral stochastic finite element equilibrium equations for the given problem. Eq. (8.20) can be solved for the unknown coefficients of the PC-expansion. For
linear elastic case, Eq. (8.20) can be solved in one step. However, for elasticplastic case,
Eq. (8.20) needs to be solved incrementally, as the constitutive properties (G(x)

and k (x)
in Eqs. (8.13) and (8.14) respectively) at the integration points, will evolve as the material
201
plastifies. The constitutive equation, when transformed to the probability density space,
can be written as:
P ( (t))
P ( (t))
2 P ( (t))
= N(1)
+ N(2)
t
(8.21)
where N(1) and N(2) are advection and diffusion coefficient respectively. Eq. (8.21) can be
solved at each integration point for each KL-space, including the mean space. The values
of N(1) and N(2) will be different for different KL-spaces. For example, for the mean space
they can be calculated, assuming mean yield criteria (cf. Section 6.3) 1 , as:
N(1)
N(2)

()
*
+
()
1 t
cu
G+
3

()
V ar G
()
V
ar
1 t
G + cu
3
when
hf i < 0 (hf i = 0 d hf i < 0)

(8.22)
when
hf i = 0 d hf i = 0
when
hf i < 0 (hf i = 0 d hf i < 0)

(8.23)
when
hf i = 0 d hf i = 0
Similarly, for the k th KL-space, N(1) and N(2) can be calculated as:
1
For, probablistic yield criteria, the equivalent advection and diffusion coefficients need to be used (refer
to Eqs. (6.7) and (6.8)
202
k
N(1)
k
N(2)

()
k k ()
*
+
())
()
k
k
k k ()
1 t
k k () + cu
3
when
hf i < 0 (hf i = 0 d hf i < 0)

(8.24)
when

()
V
ar
()
k k
(k k ())2
()
()
V
ar
k k
1
t
k k () + cu
3
hf i = 0 d hf i = 0
when
hf i < 0 (hf i = 0 d hf i < 0)

(8.25)
when
hf i = 0 d hf i = 0
One may note that the incremental shear strain () appearing in N(1) and N(2) (refer to
Eqs. (8.22) to (8.25)) is also a random variable. It can be obtained from the stochastic
finite solution of the last step as a function of PC-basis function as follows:
(x, ) =
#
" P
N
X
X
dmn ln (x)
m [{r ()}]
x
n=1
m=0
P
X
m=0
m [{r ()}]
N
X
dmn Bn (x)
(8.26)
n=1
One may also note that the time, t appearing in Eqs. (8.21) to (8.25) is the pseudo-time of
constitutive rate equation. Using the respective N(1) and N(2) , Eq. (8.21) can be solved for
pseudo-time evolution of probability density of internal shear stress (P ( ))) at each integration point of each KL-space, which then can be used to obtain the pseudo-time evolved
constitutive properties in each KL-space, to be used for next global level load increment
step. The least square regression procedure proposed in Chapter 7 is used to obtain the
203
pseudo-time evolution of constitutive properties from probability density of internal shear
stress (P ( (t))).
8.2.2
Simulation Results
Figs. 8.2 and 8.3 show the mean displacements, along with the standard deviations
of the displacements at the top node, obtained using SFEM and Monte Carlo technique,
for linear elastic and von Mises elasticplastic mean yielding material model respectively.
The probability densities of the top node displacements, calculated using Edgeworth series
expansion (refer to Section 7.4), of the soil column, corresponding to 0.1 N shear load,
for linear elastic and von Mises elastic-plastic material models are compared in Fig. 8.4.
One may observe that the uncertainties in displacement reduced once the material plastfies.
This is because of the predictor-corrector type explicit form of the elasticplastic von Mises
modulus. Both mean modulus and uncertainties with respect to the mean, following explicit
form of von Mises material model, reduce as the material plastifies. This observation is in
line with the pure constitutive level observation of the probabilistic behavior of DruckerPrager material model (refer to Fig. 6.11).
One may note that the results shown in Figs. 8.2 to 8.4 are obtained using two
KL-terms (i.e., using first two eigenmodes of the covariance kernel of the random field shear
modulus) and second order PC expansion (i.e., using six PC terms, refer to Eq. (7.26)). The
smaller the ratio of correlation length to domain length, the more KL-terms are needed.
This is demonstrated in Figs. 8.5 to 8.8, where the effect of increasing the number of KLterms (while keeping the order of PC as constant (= 2)) on mean and standard deviation of
displacement along the depth of the 1D soil column, with linear elastic material model is
204
Mean (MC)
0.1
Mean (SFEM)
Load (N)
0.08
0.06
Standard Deviations (MC)
0.04
Standard Deviations (SFEM)
0.02
1
2
3
4
5
Displacement at Top Node (mm)
Figure 8.2: Mean and standard deviations of displacement at the top node of the soil
column, with linear elastic material model (with KL-dimension = 2, order of PC = 2).
Monte Carlo simulation is also shown
0.1
Mean (SFEM)
Load (N)
0.08
0.06
Mean (MC)
Standard Deviations (MC)
0.04
Standard Deviations (SFEM)
0.02
Displacement at Top Node (mm)

Figure 8.3: Mean and standard deviations of displacement at the top node of the soil
column, with von Mises elastic-plastic material model (with KL-dimension = 2, order of
PC = 2). Monte Carlo simulation is also shown
Probability Densities of Displacement at Top Node
205
0.002
von Mises ElasticPlastic

400
300
Elastic
200
100
0.002
0.004
0.006
0.008
0.01
Dispalcement at Top Node (m)
Figure 8.4: Comparison of PDF of top node displacements of the soil column: Elastic versus
von Mises elastic-plastic material model
shown. Two extreme ratios of correlation length of shear modulus to domain length - one
very large (= 100) and other very small (= 0.0001) - were considered. In both the cases, the
variance of shear modulus was assumed very small (assumed COV = 1%), to minimize the
effect of number of PC terms. One may observe that, though the mean solution is almost
independent of the number of KL-terms in both the cases, the standard deviation of the
solution converges with increase in KL terms. However, the percentage difference in two
KL-terms solution and six KL-terms solution of standard deviation of displacement is much
larger for the case, where the ratio of correlation length to domain length was smaller. One
may also observe that the both the means and standard deviations of displacements along
the depth of soil column, estimated with same number of KL terms, is higher for the case
206
where the ratio of correlation length of shear modulus to domain length is higher.
KL = 6
Mean of Displacement (mm)
KL = 4
3
KL = 2
Order of PC = 2
0.2
0.4
0.6
0.8
Depth (m)
Figure 8.5: Correlation length and KL dimension: Mean displacement along depth of the
1D soil column with linear elastic material model, having very small variance (COV =
1%) of shear modulus and very large ratio of correlation length of shear modulus to domain
length (= 100)
PC-expansion, on the other hand, expands the stochastic dimension of the solution
random field as a function of stochastic components of the input random field material
properties. Hence, the larger the variance of input random field, the more number of PC
terms are needed for accurate solution. This is demonstrated in Figs. 8.9 to 8.12, where the
effect of increasing the order of PC (while keeping the KL dimension constant (= 4)) on
mean and standard deviation of displacement along the depth of the 1D soil column, with
linear elastic material model is shown. Two cases - one with realistic COV (= 20%) and the
other with very small COV (= 1%) - were considered. In both cases, the ratio of correlation
2
This supports the fact that the usual procedure in geotechnical engineering of modeling material parameters as random variables (and not random fields and hence neglecting the scale effects) overestimates
the solution (displacement or force or factor of safety).
Standard Deviation of Displacement (mm)
207
KL = 6
0.12
KL = 4
0.1
KL = 2
0.08
0.06
0.04
0.02
Order of PC = 2
0.2
0.4
0.6
0.8
Depth (m)
Figure 8.6: Correlation length and KL dimension: Standard deviation of displacement along
depth of the 1D soil column with linear elastic material model, having very small variance
(COV = 1%) of shear modulus and very large ratio of correlation length of shear modulus
to domain length (= 100)
KL = 6
KL = 4
KL = 2
2
Order of PC = 2
0.2
0.4
0.6
0.8
Depth (m)
Figure 8.7: Correlation length and KL dimension: Mean displacement along depth of the
1D soil column with linear elastic material model, having very small variance (COV =
1%) of shear modulus and very small ratio of correlation length of shear modulus to domain
length (= 0.0001)
208
0.03
KL = 6
0.025
KL = 4
0.02
KL = 2
0.015
0.01
0.005
Order of PC = 2
0.2
0.4
0.6
0.8
Depth (m)
Figure 8.8: Correlation length and KL dimension: Standard deviation of displacement along
depth of the 1D soil column with linear elastic material model, having very small variance
(COV = 1%) of shear modulus and very small ratio of correlation length of shear modulus
to domain length (= 0.0001)
length of shear modulus to domain length was kept constant (assumed ratio = 0.1). One
may observe from Figs. 8.9 to 8.12 that, for the case of high variance, both the mean and
the standard deviation of the displacement along the depth of the 1D soil column slowly
converges as the order of PC increases. However, for the case of small variance, second-order
PC expansion was enough for accurate solution.
As an additional verification of the proposed FPKE-based spectral stochastic finite
element, in Table 8.1, the direct (without going into the FPKE-based constitutive integrator
3)
solution of displacement at top node of the 1D soil column, with linear elastic material
model, is compared with the FPKE-based spectral stochastic finite element solution. Six
steps of load increment are shown and as can be observed from Table 8.1, both the solution
3
For linear elastic material, going through the constitutive integrator is redundant, as constitutive properties do not change with load
209
Order of PC = 1
Order of PC = 2
Order of PC = 3
3
2
KL dimension = 4
0.2
0.4
0.6
0.8
Depth (m)
Figure 8.9: Variance and order of PC: Mean displacement along depth of the 1D soil
column with linear elastic material model, having large variance (COV = 20%) of shear
modulus and ratio of correlation length of shear modulus to domain length = 0.1
Order of PC = 3
1.5
1.25
Order of PC = 2
Order of PC = 1
0.75
0.5
0.25
KL dimension = 4
0.2
0.4
0.6
0.8
Depth (m)
Figure 8.10: Variance and order of PC: Standard deviation of displacement along depth of
the 1D soil column with linear elastic material model, having large variance (COV = 20%)
of shear modulus and ratio of correlation length of shear modulus to domain length = 0.1
210
Order of PC = 4
Order of PC = 3
Order of PC = 2
Order of PC = 1
KL dimension = 4
0.2
0.4
0.6
0.8
Depth (m)
Standard Deviation of Dispolacement (mm)
Figure 8.11: Variance and order of PC: Mean displacement along depth of the 1D soil
column with linear elastic material model, having very small variance (COV = 1%) of shear
modulus and ratio of correlation length of shear modulus to domain length = 0.1
0.05
Order of PC = 4
Order of PC = 3
0.04
Order of PC = 2
Order of PC = 1
0.03
0.02
0.01
KL dimension = 4
0.2
0.4
0.6
0.8
Depth (m)
Figure 8.12: Variance and order of PC: Standard deviation of displacement along depth of
the 1D soil column with linear elastic material model, having very small variance (COV
= 1%) of shear modulus and ratio of correlation length of shear modulus to domain length
= 0.1
211
Table 8.1: Comparison of results (at top node) of FPKE-based spectral stochastic finite
element with direct spectral stochastic finite element, for 1D soil column example, with
linear elastic material
Load Step
0
1
2
3
4
5
6
Mean of solution
Direct (mm) FPKE (mm) Error (%)
0
0
0
2.015
2.015
0
4.03
4.00
0.7
6.04
5.95
1.4
8.06
7.89
2.1
10.00
9.84
1.7
12.09
11.88
1.7
Standard deviation of solution

Direct (mm) FPKE (mm) Error (%)
0
0
0
0.2005
0.2005
0
0.40
0.39
0.7
0.60
0.59
1.4
0.80
0.78
2.1
1.00
0.98
1.7
1.20
1.21
0.6
match closely.
8.3
8.3.1
One Dimensional Dynamic Problem

Problem Statement and Formulation
Problem Statement
A 1-D soil column of length L (= 1.0m) and area A (= 0.01m2 ) is subjected to a
sinusoidal base displacement. The schematic is shown in Fig. 8.13. The base displacement
is shown in Fig. 8.14. The shear modulus of the soil column is a Gaussian random field with
(= 2.5 kPa), variance 2 (= 0.15 kP a2 ), and covariance function C(x1 , x2 )
mean value G
(assumed exponential with a correlation length r = 0.3 m). The density of the soil is 20
kN/M 3 . Neglecting damping, determine the time evolution of mean and standard deviation
of the displacement at the top node of the soil column. Two different material models
linear elastic and von Mises associative elasticplastic with a deterministic shear strength,
212
Figure 8.13: Schematic of dynamic 1D soil column example
Displacement (mm)
10
10
Time (s)
Figure 8.14: Base displacement applied to the bottom node of the 1D soil column
213
cu (= 5 Pa) and rate of evolution of shear strength with plastic strain, cu (= 2 kPa) are
considered. Similar to the static example, here also the assumed material properties for
the soil column may not be very realistic for soils. This example is presented only to show
the applicability of the developed computational framework to dynamic problems. A more
realistic example for soil is presented in Section 8.4.
SFEM Formulation
The given problem can be visualized as a soil column with a stiff spring, having
a large spring constant Ks , attached at the bottom node of the soil column, as shown in
Fig. 8.15. One can write the strain energy stored in the soil column-spring system, following
Figure 8.15: Visualization of 1D soil column with base displacement, as shown in Fig. 8.13,
as a soil column-stiff spring system
214
the derivation in Subsection 8.2.1 of static soil column example, as:
1
U= A
2
G(x, )
L
u(x, t, )
x
2
dx + Ks
(8.27)
and, the potential energy due to external force, assuming the reaction force of the spring
as (t) (= (Ks ) x (base displacement (t))) as:
U = A
(t)u(x, )dx A
u(x, t, )dx
(8.28)
One may recognize that the second term on the r.h.s of Eq. (8.28) is the inertia term,
where is the mass density per unit length of the soil column. Similar to the derivation in
Subsection 8.2.1 of static soil column example, one can expand the shear modulus random
field using KL-expansion, the stochastic components displacement and acceleration random
fields using PC-expansion, and the spatial components of displacement and acceleration
random fields using shape functions, and minimize the total energy of the soil columnspring system to obtain the following governing equation:
N
X
n=1
Mmn dni +
N
X
n=1
Kmn dni +
N X
P
X
n=1 j=0
dnj
M
X
k=1
cijk Kmnk = hFm i [{r ()}]i
(8.29)
where the mass matrix, Mmn is given as:
Mmn =
Bn (x)mBm (x)dx
(8.30)
where, m (= A) is the mass per unit length of the soil column. However, for simplification,
in this example lumped mass matrix is used. The deterministic (K ) and stochastic (K )
215
components of the generalized stiffness matrix and the generalized force vector (F ) are
given in Eqs. (8.20), (8.21), and (8.22) respectively. However, for this soil column-spring
system, the spring stiffness (Ks ) needs to be added to the elements, corresponding to the
bottom nodes, of the deterministic (K ) and stochastic (K ) components of the generalized

stiffness matrix. Eq. (8.29) is solved incrementally using Newmarks method, as outlined in
Subsection 7.3.2, with = 0.65 and = 0.33.
8.3.2
Simulation Results
The time evolution of mean of displacements at the top node of the 1D soil column
due to sinusoidal base displacement, as shown in Fig. 8.14, are shown in Figs. 8.16 and 8.17
for linear elastic and von Mises elastic-plastic material model respectively. The evolution of
2
1.5
1
0.5
1
0.5
1
Time (s)
Figure 8.16: Time evolution of mean of displacement at the top node of the 1D soil column,
with linear elastic material model, due to sinusoidal base displacement shown in Fig. 8.14
(with KL-dimension = 2, order of PC = 2)
standard deviation of displacements for linear elastic and von Mises elastic-plastic material
model are shown in Fig. 8.18 and 8.19 respectively.
Figs 8.20 and 8.21 show the time
216
1.5
0.5
Time (s)
with von Mises elasticplastic material model, due to sinusoidal base displacement shown
in Fig. 8.14 (with KL-dimension = 2, order of PC = 2)
1.2
1
0.8
0.6
0.4
0.2
1
Time (s)
Figure 8.18: Time evolution of standard deviation of displacement at the top node of the
1D soil column, with linear elastic material model, due to sinusoidal base displacement
shown in Fig. 8.14 (with KL-dimension = 2, order of PC = 2)
Standard Deviation of Displacvement (mm)
217
0.7
0.6
0.5
0.4
0.3
0.2
0.1
1
Time (s)
1D soil column, with von Mises elasticplastic material model, due to sinusoidal base
evolution of mean standard deviation of displacement at the top node of the 1D soil
column, with linear elastic and von Mises elasticplastic material model respectively.
One may note that the displacements shown in Figs. 8.16 to 8.21 are absolute
displacements and not relative to the base. In Figs. 8.16 to 8.21, one may also note that both
the mean and the standard deviation reduced for elasticplastic model once the material
plastified, for similar reason to that given for the static example (refer to Subsection 8.2.2).
8.4
Seismic Wave Propagation through Stochastic Elastic

Plastic Soil
In this section, the applicability of the computational tools, developed in Chap-
ters 5 and 7 of this dissertation, are demonstrated on a real life problem in geotechnical
218
Standard Deviations
Displacement (mm)
Mean
2
1
Time (s)
Figure 8.20: Time evolution of mean standard deviation of displacement at the top
node of the 1D soil column, with linear elastic material model, due to sinusoidal base
Standard Deviations
Mean
Displacement (mm)
1.5
1
0.5
0.5
Time (s)
Figure 8.21: Time evolution of mean standard deviation of displacement at the top node
of the 1D soil column, with von Mises elasticplastic material model, due to sinusoidal
base displacement shown in Fig. 8.14 (with KL-dimension = 2, order of PC = 2)
219
engineering. The effect of uncertainties in soil properties on the seismic wave propagation
through elasticplastic soil is investigated. To this end, modified 1938 Imperial Valley earthquake displacement is applied to the base of a 1D soil column. The earthquake motion
is then propagated through the stochastic elasticplastic soil to the top of the soil column.
It is assumed that the earthquake motion is deterministic and all the stochasticities are
in the soil parameter. Soil is modeled as a vertical random field. The parameters for the
soil random field are obtained following similar analysis to that described in Part II of this
dissertation. The time evolution of displacement at the top of the soil column is presented
in terms of mean, standard deviation and probability density function. In the following
sections, the problem parameters are described first, followed by simulation results.
8.4.1
Problem Parameters
The soil column is assumed to represent top (after removing 1 m. top soil) 12.5
m. layer of the Alameda County subsoil (refer to Figs. 4.2 and 4.3). The vertical soil
variability of the 1D soil column was calculated from CPT data [99], collected by the
USGS Western Earthquake Hazards Team and was tabulated in Table 4.1. The CPT tip
resistance (qT ) was modeled as a vertical Gaussian random field. One may note that a lognormal random field would be more realistic for soil properties as it would prevent unrealistic
negative modulus. However, for simplicity 4 , in this example the cone tip resistance (qT ) is
assumed Gaussian 5 . Assuming finite scale model is applicable, the average (over 16 CPT
4
The purpose of this example is just to show the complete solution process based on the developed
computational framework.
5
The developed computational framework is general enough to handle any arbitrary distribution as input.
However, the input needs to be calculated accordingly. The procedure for modeling soil properties as lognormal random field was described by Fenton [23, 24].
220
soundings, covering an area of approximately 7 km2 ) random field parameters, obtained
using maximum likelihood estimate (Table 4.1), are as follows: hqT i = 4.99 M P a, Var[qT ]
= 25.67 M P a2 , and Correlation length of qT = 0.61 m. The testing error was estimated as
2.78 M P a2 . The random field model for the vertical CPT tip resistance, is then transformed
to obtain the shear modulus random field of the soil column. The transformation equation,
proposed by Weiher and Davis [98], is used for this purpose. According to Weiher and Davis
[98], the cone tip resistance can be transformed to elastic shear modulus according to the
following equation:
G
= 2.9qT
(1 )
(8.31)
where, G, , and qT are shear modulus, Poissons ratio, and cone tip resistance respectively.
Assuming = 0.2 (deterministic) and COV of transformation uncertainty = 5%, one may
obtain the parameters of the shear modulus random field, as: hGi = (2.32)(4.99) = 11.57
M P a, Var[G] = [(2.322 )(25.67) + 2.78 + ((0.05)(5.26))2 ] = 142.32 M P a2 , and Correlation
length of G = 0.61 m. One may note that the second additive term in the above variance
term is the variance of testing uncertainty and the third additive term is the variance of
transformation uncertainty. One may also note that the correlation length is independent of
the transformation, as the transformation equation is a point equation. The shear strength 6
(cu ) and rate of evolution of shear strength with plastic strain (cu ) of the soil are assumed to
be 0.636 kPa and 0.5 MPa, both deterministic, respectively. The motion, which is applied
to the base of the soil column is modified a 1938 Imperial Valley seismic motion, whose
6
Assumed constant with depth for simplicity. Also, little lower than realistic value is assumed for the
deterministic shear strength to make sure the soil plastifies once the slightest motion develops. This is to
help in zooming in the effects of soil plasticity on uncertainty propagation
221
displacement time history is shown in Fig. 8.22.
Displacement HmmL
30
20
10
0.5
1.5
Time HsL
-10
-20
-30
Figure 8.22: Base displacement applied to the bottom node of the 1D soil column: Modified
1938 Imperial Valley motion
8.4.2
Simulation Results and Discussions

In the following, the statistical properties of the displacement at the top node of
the soil column, obtained for two different material models - linear elastic and von Mises
elasticplastic - are presented in number of ways. Figs 8.23 and 8.24 show the time evolutions of mean displacements. The time evolutions of standard deviation of displacements
are shown in Figs 8.25 and 8.26 Figs 8.27 and 8.28 show the time evolutions of mean
standard deviation of displacements. The time evolutions of coefficient of variation (COV)
of displacements are shown in Figs 8.29 and 8.30 One may note that the spikes in the
evolution of COV plots are due to the definition of COV, which is standard deviation over
mean. Hence, when the means tend to zero, the COVs tend to infinity. The time evolutions
of probability density function (PDF) of displacements are shown in Figs 8.31 and 8.32 In
222
Figs. 8.31 and 8.32, for clarity of the figures, the PDFs of displacements at the top of the
soil column are shown only at few selective locations on the time axis.
Mean of DisplacementHmmL
15
10
5
0.5
1.5
Time HsL
-5
-10
-15
with linear elastic material model, due to modified 1938 Imperial Valley base displacement
as shown in Fig. 8.22 (with KL-dimension = 2, order of PC = 1)
Mean of DisplacementHmmL
10
0.5
1.5
Time HsL
-10
-20
with von Mises elasticplastic material model, due to modified 1938 Imperial Valley base
displacement as shown in Fig. 8.22 (with KL-dimension = 2, order of PC = 1)
223
Std. Deviation of DisplacementHmmL

15
12.5
10
7.5
5
2.5
0.5
1.5
Time HsL
1D soil column, with linear elastic material model, due to modified 1938 Imperial Valley
base displacement as shown in Fig. 8.22 (with KL-dimension = 2, order of PC = 1)
Std. Deviation of DisplacementHmmL
8
6
4
2
0.5
1.5
Time HsL
1D soil column, with von Mises elasticplastic material model, due to modified Imperial
Valley base displacement as shown in Fig. 8.22 (with KL-dimension = 2, order of PC = 1)
224
DisplacementHmmL
30
20
10
0.5
1.5
Time HsL
-10
-20
-30
of the 1D soil column, with linear elastic material model, due to modified 1938 Imperial
Valley base displacement as shown in Fig. 8.22 (with KL-dimension = 2, order of PC = 1)
DisplacementHmmL
20
10
0.5
1.5
Time HsL
-10
-20
-30
of the 1D soil column, with von Mises elasticplastic material model, due to modified 1938
Imperial Valley base displacement as shown in Fig. 8.14 (with KL-dimension = 2, order of
PC = 1)
225
COV
20
17.5
15
12.5
10
7.5
5
2.5
0.5
1.5
Time HsL
Figure 8.29: Time evolution of coefficient of variation (COV) of displacement at the top
node of the 1D soil column, with linear elastic material model, due to modified 1938
PC = 1)
COV
20
17.5
15
12.5
10
7.5
5
2.5
0.5
1.5
Time HsL
Figure 8.30: Time evolution of coefficient of variation (COV) of displacement at the top
node of the 1D soil column, with von Mises elasticplastic material model, due to modified
1938 Imperial Valley base displacement as shown in Fig. 8.14 (with KL-dimension = 2,
order of PC = 1)
226
)
0.03
nt (m 0.015
eme
isplac
0.015
Probability Density of Displacement
0.03
1500
1000
500
5
1.7
0
1.5
1.2
2.0
Time (s)
0
1.0
0.7
0
0.5
5
0.2
0.0
Figure 8.31: Time evolution of probability density function (PDF) of displacement at the
top node of the 1D soil column, with linear elastic material model, due to modified 1938
PC = 1)
As can be observed from the above figures (Figs. 8.23 to 8.32), the coefficient
of variation of the displacement at top node reduced as the material plastified and remained constant. The coefficient of variation reduced because of the predictor-corrector
type relationship of the elasticplastic shear modulus for von Mises constitutive equation.
227
nt (m
me
lace
Disp
Probability Density of Displacement
0.015
0.03
1500
0.03
0.015
1000
500
0
2.0
1.7
5
1.5
Time (s)
1.2
1.0
5
0.7
0.5
5
0.2
0.0
Figure 8.32: Time evolution of probability density function (PDF) of displacement at the
top node of the 1D soil column, with von Mises elasticplastic material model, due to
modified 1938 Imperial Valley base displacement as shown in Fig. 8.22 (with KL-dimension
= 2, order of PC = 1)
It remained constant due to bilinear nature of the von Mises constitutive equation. For
truly nonlinear constitutive models, for example Cam Clay material model, the uncertainty
in the system might evolve. At the constitutive level, this evolution of uncertainty, as the
material plastifies, can be observed in Figs. 6.17 to 6.21, where the post-yield uncertainty
228
either decreased (Fig. 6.17) or increased (Figs. 6.19, 6.20 and 6.21), depending upon the
nonlinear dynamics of the constitutive equation and effect the of more than one uncertain
soil parameters. However, the effects of nonlinear dynamics of truly nonlinear constitutive
models and more than one uncertain soil parameters on the finite element solutions of the
behaviors of solids and structures are left as future research directions.
229
Conclusions and Future Research

Directions
Summary and Conclusions
In this dissertation a computational framework was developed for simulations of
behaviors of solids and structures made of stochastic elasticplastic materials. Particular
emphasis was given to soil, a highly nonlinear (elasticplastic) and highly uncertain material, and geotechnical engineering applications. The different types of uncertainties in
material properties, which would be the input to the stochastic framework, were characterized first. Then, based on an extensive literature review, those different types of material
uncertainties, in particular soil uncertainties, were quantified. Spatial variability of material
properties were modeled as random fields. Techniques for estimating statistical parameters
of random field models from test data were also presented. The propagation of these uncertainties through the governing partial differential equation of elasticplastic solids hinges
on two techniques merging of state-of-the-art probability theory with theory of elasto
plasticity, in development of probabilistic elastoplasticity and subsequent development of
230
stochastic elasticplastic finite element method developed (original development) as part
of this dissertation.
For probabilistic constitutive simulations, a second-order accurate analytical expression, in the form of Fokker-Planck-Kolmogorov equation (FPKE), was obtained for
solution of any general elasticplastic constitutive rate equation with random material
properties. With appropriate initial and boundary conditions the FPKE could be solved
for probability density of stress response, following any constitutive rate equations, given
random material parameters and deterministic/random strain. The advantage of writing
the constitutive equation in FPKE form and solving the constitutive rate equation in the
probability density space was that the FPKE was a linear deterministic partial differential
equation, where as the constitutive equation in real space was a nonlinear stochastic ordinary differential equation. Though, the FPKE, corresponding to general constitutive rate
equation, was not amenable to analytical solution, the deterministic linearity vastly simplified the numerical solution process. Also the FPKE approach didnt suffer from closure
problem associated with perturbation technique or high computational cost associated with
Monte Carlo simulation.
For simulations of solids and structures made with uncertain elasticplastic materials, the FPKE-based constitutive level development was combined with the spectral
approach of stochastic finite element method in formulation of stochastic elasticplastic
finite element method (SEPFEM). To this end, the random field material properties were
discretized, in both spatial and stochastic dimension, into finite numbers of independent
basic random variables, using KarhunenLoeve expansion. Those random variables were
231
then propagated through the elasticplastic constitutive rate equation using FokkerPlanckKolmogorov equation approach, to obtain the evolutionary material properties, as the material plastified. The unknown displacement (solution) random field was then assembled,
as a function of known basic random variables with unknown deterministic coefficients,
using polynomial chaos expansion. The unknown deterministic coefficients of polynomial
chaos expansion were obtained, by minimizing the error of finite representation, by Galerkin
technique.
The applicability of the developed methodology was demonstrated in obtaining
the probabilistic solutions of 1D static pushover test and response of 1D structure due to
sinusoidal base displacement. In addition, pure constitutive level simulations of von Mises,
Drucker-Prager, and Cam Clay material models were also shown. The results were verified
with analytical solution, where available or Monte Carlo solution and good agreements
were obtained. In addition, the developed computational framework was used in analyzing
seismic wave propagation through stochastic elasticplastic soil, using real-life soil data and
with real earthquake motion. In the following, a detailed summary and main findings of
each part of this dissertation is presented:
Part - I described the motivation and discussed the features of probability theory that
had been used in the subsequent chapters. The properties of single as well as two
or more random variables, in terms of moments and cumulants were outlined, with
appropriate examples from mechanics. Also introduced was the theory of random
fields/processes, which then was explored in details in Part - II, in modeling spatial
variability of field soil properties. The classical theory of stochastic calculus was also
232
outlined in providing the mathematical foundation to the Fokker-Planck-Kolmogorov
equation developed in Part - III for probabilistic constitutive simulation.
In Part - II, the different types of uncertainties in material properties were characterized.
Particular emphasis was given on soils, which exhibit high degree of uncertainty. Based
on an extensive literature review, the uncertainties associated with soil properties,
namely the natural variability of soil deposit, testing uncertainty and transformation
uncertainty, were quantified, based on soil type, testing method and type of transformation equation. In the absence of enough data for a meaningful statistical analysis,
those could be used as a general guideline. On the other hand, when sufficient data
are available, the statistical parameters for the random field model of the soil property
could be estimated directly from the data. To this end, two different random field
modeling techniques, namely the finite scale model and the fractal model, were discussed, with example estimations, using maximum likelihood estimation technique, of
cone tip resistance (qT ) random field parameters from CPT data collected in Alameda
County, CA.
Part - III dealt with constitutive level probabilistic simulation. In obtaining probabilistic solution of constitutive equation, the up-scaled form of any general constitutive
rate equation, with stochastic material parameters, was written in probability density
space. It was shown that the constitutive rate equation, when written in probability
density space, became a Fokker-Planck-Kolmogorov equation (FPKE). The FPKE
corresponding to the constitutive rate equation was a second-order accurate linear deterministic partial differential equation. This deterministic linearity of FPKE, when
233
compared to nonlinear stochasticity of constitutive rate equation in real space, greaty
simplifies the solution process. Also, being second-order accurate, the FPKE didnt
suffer from closure problem as associated with perturbation technique. Further, it
didnt need repetitive solution of computationally expensive deterministic equation,
as associated with Monte Carlo technique. In addition, the form of the general equation was like a template and it was shown that the probabilistic description of constitutive solution following any particular type of constitutive equation could easily be
tailor-made from the general expression.
To demonstrate the applicability of FPKE approach in obtaining the probabilistic
solution of constitutive rate equation, the general form of FPKE, corresponding to
up-scaled form of any general constitutive rate equation was then specialized to 1
D point-location scale von Mises, Drucker Prager and Cam Clay material models,
which are then solved with appropriate initial and boundary conditions. Reflective
type boundary conditions were found suitable. Initial condition could be deterministic
(Dirac delta function) or stochastic. The solution strategy of FPKE was demonstrated
assuming mean yield criteria as well as more realistic probabilistic yield criteria. The
solution assuming mean yield criteria involved solving two FPKEs one corresponding to preyield elastic equation and the other corresponding to postyield elastic
plastic equation for complete probabilistic solution of constitutive rate equation.
On the other hand, solution assuming probabilistic yielding involved solution of one
equation, with equivalent advection and diffusion coefficients, obtained based on the
cumulative density function of the yield stress. The FPKE was solved numerically
234
by semi-discretizing the PDE in the stress domain and then solving the resulting system of ODE simultaneously. Probabilistic solution of constitutive rate equation with
mean yield criteria showed sharp transition between pre-yield and post-yield regions,
whereas solution with probabilistic yield criteria showed a more realistic smooth transition. It was also shown that for truly nonlinear models like Cam Clay, the mean and
mode of the solution could vary greatly from the deterministic solution and post-yield
uncertainty in the system could either increase or decrease, depending upon the type
of material model and effect of uncertainties in more than one material parameters.
Good agreements were obtained between FPKE solutions and analytical solutions,
where available or Monte Carlo solutions.
Part - IV build upon the developments in Part - III in formulating stochastic elasticplastic
finite element method (SEPFEM) for stochastic solution of boundary value problems.
To this end, the input material random fields were discretized, in both spatial and
stochastic dimensions, into finite number of independent basic random variables, using
KarhunenLoeve (KL) expansion, which basically represent the random field by eigenmodes of its covariance kernel. It was shown that the number of KL-terms needed to
represent any random field with sufficient accuracy depended on the ratio of correlation length to domain length. The smaller the ratio, the more terms were needed and
vice versa. However, even for a very small ratio of correlation length to domain length,
only the first few eigenmodes were enough to represent the random field with sufficient accuracy. This was a very desirable property of KL-expansion as each random
variable added one extra dimension to the problem. These independent random vari-
235
ables were then propagated through the constitutive equation, using FPKE approach
as developed in Part - III of this dissertation, to obtain the statistical properties of
internal stresses and subsequently the evolutionary material parameters, as the material plastified. The displacement (solution) random field was then assembled, as
a function of known basic random variables with unknown deterministic coefficients,
using polynomial chaos expansion.
The stochastic finite element equilibrium equation, a system of algebraic equations,
which was obtained by Galerkin technique by minimizing the error of finite representation, was solved for the unknown deterministic coefficients of polynomial chaos
expansion either by pure incremental method (for static problems) or by Newmark
method (for dynamic problems). To demonstrate the applicability, the developed
SEPFEM was applied in obtaining probabilistic solution of 1D static pushover test
and response of 1D structure due to sinusoidal base displacement. Probabilistic solutions were presented in terms of mean, standard deviation, and probability density
function of displacement. SEPFEM solutions were verified with Monte Carlo solutions
and good agreements were obtained.
Further, the complete solution process, based on the developed computational framework, of a geotechnical engineering problem was illustrated with real-life data. To
this end, probabilistic solution of seismic wave propagation through stochastic elastic
plastic soil was analyzed, with real-life soil data and real earthquake motion. Top 12.5
m (after removing 1 m top soil) soil of Alameda County, CA was modeled as a 12.5
m deep 1D soil column. The spatial variability of the soil column was modeled as
236
random field. The maximum likelihood estimated random field statistical parameters
were obtained by analyzing CPT cone tip resistance (qT ) data, assuming finite scale
modeling was applicable. The testing error was also indirectly estimated from the
data analysis. The cone tip resistance was then transformed into elastic shear modulus using an available transformation equation. In the absence of available data, the
COV of transformation uncertainty was assumed to be 5%. The soil was assumed to
be a von Mises material, with deterministic shear strength (cu ) and deterministic rate
of evolution of shear strength with plastic strain (cu ). Modified 1938 Imperial Valley
seismic displacement was applied at the base of the soil column. The displacement
at the top of the soil column were estimated using the computational tools developed
in Chapters 5 and 7 of this dissertation. The probabilistic solution at top were presented in terms of evolutionary mean, standard deviation, mean standard deviation,
COV, and probability density function of displacement. Both elastic and von Mises
elasticplastic solutions were presented. It was shown that for von Mises model with
deterministic cu and cu , the post-yield uncertainty of the system reduced, as the material plastified and remained constant for the entire postyield region. However, this
reduction in post-yield uncertainty is model dependent. The post-yield uncertainty
could evolve, depending on type of model or effect of uncertainties in more than one
material (soil) parameters. This is especially true for truly nonlinear models, for
example Cam Clay material model. The postyield evolution of uncertainty for Cam
Clay material model was shown by pure constitutive simulations in Part - III of this
dissertation.
237
The developments presented in this doctoral dissertation were published as three
journal and and three peer-reviewed conference papers. One more journal paper is in review
and another one is in preparation. In the following, the publications out of this doctoral
dissertation are listed:
Sett, K., Jeremic, B., and Kavvas, M. L. The role of nonlinear hardening/ softening in
probabilistic elastoplasticity. International Journal for Numerical and Analytical
Methods in Geomechanics 31, 7 (June 2007), 953-975
Sett, K., Jeremic, B., and Kavvas, M. L. Probabilistic elasto-plasticity: Solution and
verification in 1D. Acta Geotechnica 2, 3 (September 2007). In press (published
online in the Online F irst section).
Jeremic, B., Sett, K., and Kavvas, M. L. Probabilistic elasto-plasticity: Formulation in
1D. Acta Geotechnica 2, 3 (September 2007). In press (published online in the
Online F irst section).
Sett, K., and Jeremic, B. On probabilistic yielding of materials. Communications in
Numerical Methods in Engineering (2007). In review.
Sett, K., and Jeremic, B. Spectral stochastic elastic-plastic finite element formulation for
geomechanics applications. International Journal for Numerical Methods in Engineering (2007). In preparation.
Jeremic, B., and Sett, K. The influence of uncertain material parameters on stress-strain
response. In Geomechanics II: Testing, Modeling, and Simulation (Proceedings of the
Second-U.S. Workshop on Testing, Modeling, and Simulation in Geomechanics, held
238
in Kyoto, Japan from September 8-10, 2005) (August 2006), P. V. Lade and T. Nakai,
Eds., Geotechnical Special Publication No, 156, American Society for Civil Engineers,
pp. 132-147.
Sett, K., and Jeremic, B. Uncertain soil properties and elasticplastic simulations in geomechanics. In Probabilistic Applications in Geotechnical Engineering (Proceedings
of Geo-Denver 2007: New Peaks in Geotechnics, held in Denver, CO from February
18-21, 2007) (2007), K. K. Phoon, G. A. Fenton, E. F. Glynn, C. H. Juang, T. F.
Griffiths, T. F. Wolff, and L. Zhang, Eds., Geotechnical Special Publication No. 170,
American Society for Civil Engineers, pp. 1-11.
Sett, K., and Jeremic, B. Stochastic ground motions. In Proceedings of Geo- Congress
2008: The Challenge of Sustainability in the Geoenvironment, to be held in New
Orleans, LA from March 9-12, 2008 (2008), American Society for Civil Engineers.
Accepted (to appear as Geotechnical Special Publication).
Future Research Directions

This dissertation achieved its objective of development of a computational framework for simulations of the behaviors of solids and structures made of stochastic elastic
plastic materials. Though, emphasis was given on soils and geotechnical engineering applications, the impact of the findings of this dissertation research is expected to be much wider
than just in the area of geotechnical engineering. The phenomena of spatial variability and
uncertainty in material properties is present in all materials. The appropriate formulation
and implementation that incorporate above phenomena into advanced numerical simula-
239
tions will impact mechanical, biomedical, materials, aerospace as well as other areas of civil
engineering. During the course of this research, some other related topics are identified as
future research directions. They are itemized in the following:
The complete solution process, based on the developed computational framework,

of a geotechnical earthquake engineering problem was presented in this dissertation.
However, it is suggested to apply the developed methodology for site response analysis of any instrumented site and validate the methodology with down hole array
measurements.
The numerical solution algorithm adopted in this dissertation for solution of non
linear stochastic finite element equilibrium equation was load controlled pure incremental method. However, for more accurate solution, especially for problems with
truly nonlinear constitutive models, it is suggested to use incrementaliterative type
global (finite element) level solution algorithm. To this end, the probability densities
of internal stresses obtained at each Gauss integration point could be used to estimate
the statistical properties of the internal force.
In this dissertation, at each Gauss integration point, the FokkerPlanckKolmogorov
equations were solved by semi-discretizing the PDE in the stress domain by method of
lines. It is suggested to investigate the applicability of adaptive technique, for example
reduced order modeling or Krylov subspace technique, in solving the the FPK PDE, as
this will reduce the computational loads at the Gauss integration points significantly.
In this dissertation, probabilistic solution methodology of any general 3D elastic
240
plastic constitutive rate equation was developed and subsequently a 3D formulation
of stochastic elasticplastic finite element method was presented. However, implementation of the developed formulation and examples were restricted to 1D. 3D
implementation of the presented general formulation will have much wider applications in the field of geotechnical engineering.
The stochastic elasticplastic finite element method presented in this dissertation assumed small deformation elastoplasticity. Similar formulations for large deformations
and coupled behaviors will have more interesting applications in the field of geotechnical earthquake engineering, especially in probabilistic simulations of liquefaction and
lateral spreading.
241
Bibliography
[1] Anders, M., and Hori, M. Stochastic finite element method for elasto-plastic body.
International Journal for Numerical Methods in Engineering 46 (1999), 18971916.
[2] Anders, M., and Hori, M. Three-dimensional stochastic finite element method for
elasto-plastic bodies. International Journal for Numerical Methods in Engineering 51
(2000), 449478.
[3] Anderson, E., Bai, Z., Bischof, C., Blackford, S., Demmel, J., Dongarra,
J., Du Croz, J., Greenbaum, A., Hammarling, S., McKenney, A., and
Sorensen, D. LAPACK Users Guide, third ed. Society for Industrial and Applied Mathematics, Philadelphia, PA, 1999.
[4] Atkinson, K. E. The Numerical Solution of Integral Equations of the Second Kind.
Cambridge University Press, Cambridge, 1997.
[5] Babuska, I., and Chatzipantelidis, P. On solving elliptic stochastic partial
differential equations. Computer Methods in Applied Mechanics and Engineering 191
(2002), 40934122.
242
[6] Baecher, G. B., and Christian, J. T. Reliability and Statistics in Geotechnical
Engineering, second ed. Wiley, West Sussex PO19 8SQ, England, 2003. ISBN 0-47149833-5.
[7] Bathe, K.-J. Finite Element Procedures. Prentice Hall, New Jersy, 1996.
[8] Ben-Haim, Y., and Elishakoff, I. Convex Models of Uncertainty in Applied
Mechanics. Elsevier, Amsterdam, 1990.
[9] Carmeliet, J., and De Borst, R. Stochastic approches for damage evolution in
standard and non-standard continua. International Journal for Solids and Structures
32 (1995), 11491160.
[10] Chen, W. F., and Han, D. J. Plasticity for Structural Engineers. Springer-Verlag,
1988.
[11] Christakos, G. Random Field Models in Earth Sciences. Academic Press, San
Diego, 1992.
[12] Christian, J. T. Geotechnical engineering reliability: How well do we know what
we are doing? Journal of Geotechnical and Geoenvironmental Engineering 130, 10
(October 2004), 9851003.
[13] Cressie, N. A. C. Statistics for Spatial Data. John Wiley & Sons., New York, 1991.
[14] Davis, T. A. Algorithm 832: UMFPACK, an unsymmetric-pattern multifrontal
method. ACM Transactions on Mathematical Software 30, 2 (June 2004), 196199.
243
[15] De Lima, B. S. L. P., Teixeira, E. C., and Ebecken, N. F. F. Probabilistic and
possibilistic methods for the elastoplastic analysis of soils. Advances in Engineering
Software 132 (2001), 569585.
[16] Deb, M. K., Babuska, I. M., and Oden, J. T. Solution of stochastic partial
differential equations using Galerkin finite element techniques. Computer Methods in
Applied Mechanics and Engineering 190 (2001), 63596372.
[17] Debusschere, B. J., Najm, H. N., Matta, A., Knio, O. M., and Ghanem,
R. G. Protein labeling reactions in electrochemical microchannel flow: Numerical
simulation and uncertainty propagation. Physics of Fluids 15 (2003), 22382250.
[18] DeGroot, D. J., and Baecher, G. B. Estimating autocovariance of in-situ soil
properties. Journal of Geotechnical Engineering 119, 1 (January 1993), 147166.
[19] Der Kiureghian, A., and Ke, B. J. The stochastic finite element method in
structural reliability. Journal of Probabilistic Engineering Mechanics 3, 2 (1988),
8391.
[20] Duncan, J. M. Factors of safety and reliability in geotechnical engineering. ASCE
Journal of Geotechnical and Geoenvironmental Engineering 126, 4 (April 2000), 307
316.
[21] Edgeworth, F. Y. The law of error. Cambridge Philosophical Society 20 (1905),
3666 and 113141.
244
[22] Einstein, A. On the motion required by the molecular kinetic theory of heat
of small particles suspended in a stationary liquid. Annalen der Physik 17 (1905),
549562.
[23] Fenton, G. A. Estimation of stochastic soil models. Journal of Geotechnical and
Geoenvironmental Engineering 125, 6 (June 1999), 470485.
[24] Fenton, G. A. Random field modeling of CPT data. Journal of Geotechnical and
Geoenvironmental Engineering 125, 6 (June 1999), 486498.
[25] Fenton, G. A., and Griffiths, D. V. Probabilistic foundation settlement on
spatially random soil. Journal of Geotechnical and Geoenvironmental Engineering
128, 5 (May 2002), 381390.
[26] Fenton, G. A., and Griffiths, D. V. Bearing capacity prediction of spatially
random c soil. Canadian Geotechnical Journal 40 (2003), 5465.
[27] Fenton, G. A., and Griffiths, D. V. Three-dimensional probabilistic foundation settlement. Journal of Geotechnical and Geoenvironmental Engineering 131, 2
(February 2005), 232239.
[28] Fokker, A. D. Die mittlere energie rotierender elektrischer dipole im strahlungsfeld.
Annalen der Physik 43 (1914), 810820.
[29] Galassi, M., Davies, J., Theiler, J., Gough, B., Jungman, G., Booth, M.,
and Rossi, F. GNU Scientific Library Reference Manual, revised and updated second ed. Network Theory Limited, Bristol, United Kingdom, 2006.
245
[30] Gardiner, C. W. Handbook of Stochastic Methods for Physics, Chemistry and the
Natural Science, third ed. Springer:Complexity. Springer-Verlag, Berlin Heidelberg,
2004.
[31] Ghanem, R. G., and Spanos, P. D. Stochastic Finite Elements: A Spectral Approach. Dover Publications, Inc., Mineola, New York, 2003.
[32] Goodman, L. A. On the exact variance of products. Journal of American Statistical
Association 55, 292 (December 1960), 708713.
[33] Griffiths, D. V., Fenton, G. A., and Manoharan, N. Bearing capacity of
rough rigid strip footing on cohesive soil: Probabilistic study. Journal of Geotechnical
and Geoenvironmental Engineering 128, 9 (2002), 743755.
[34] Gutierrez, M. A., and De Borst, R. Numerical analysis of localization using
a viscoplastic: Influence of stochastic material defects. International Journal for
Numerical Methods in Engineering 44 (1999), 18231841.
[35] Hackbusch, W.
Integralgleichungen, Theorie und Numerik.
B. G. Teubner,
Stuttgart, 1995.
[36] Hammitt, G. M. Statistical analysis of data from a comparative laboratory test
program sponsored by ACITL. U.S. Army Waterways Experiment Station, Vicksburg,
MS, 1966.
[37] Heisenberg, W. Uber

den anschaulichen inhalt der quantentheoretischen kinematik
und mechanik. Zeitschrift f
ur Physik 43 (1927), 172198. English translatation: J. A.
246
Wheeler and H. Zurek, Quantum Theory and Measurement, Princeton Univ. Press,
1983, pp. 62-84.
[38] Hindmarsh, A. C., Brown, P. N., Grant, K. E., Lee, S. L., Serban, R.,
Shumaker, D. E., and Woodward, C. S. SUNDIALS: SUite of Nonlinear and
DIfferential/ALgebraic equation Solvers. ACM Transactions on Mathematical Software 31, 3 (2005), 363396.
[39] Hopf, E. Statistical hydromechanics and functional calculus. Journal of rational
Mechanics and Analysis 1 (June 1952), 87123.
[40] Huang, S. P., Quek, S. T., and Phoon, K. K. Convergence study of the truncated Karhunen-Loève expansion for simulation of stochastic processes. International
Journal for Numerical Methods in Engineering 52 (2001), 10291043.
, K. Stochastic integral. In Proceedings of Imperial Academy of Tokyo (1944),
[41] Ito
vol. 20, pp. 519524.
, B., and Sett, K. The influence of uncertain material parameters on stress[42] Jeremic
strain response. In Geomechanics II: Testing, Modeling, and Simulation (Proceedings
of the Second-U.S. Workshop on Testing, Modeling, and Simulation in Geomechanics,
held in Kyoto, Japan from September 8-10, 2005) (August 2006), P. V. Lade and
T. Nakai, Eds., Geotechnical Special Publication No, 156, American Society for Civil
Engineers, pp. 132147.
247
, B., Sett, K., and Kavvas, M. L. Probabilistic elasto-plasticity: For[43] Jeremic
mulation in 1D. Acta Geotechnica 2, 3 (September 2007). In press (published online
in the Online F irst section).
[44] Kac, M., and Siegert, A. J. F. An explicit representation of a stationary Gaussian
process. Annals of Mathematical Statistics 18 (1947), 438442.
[45] Karhunen, K. Uber

lineare methoden in der wahrscheinlichkeitsrechnung. Ann.
Acad. Sci. Fennicae. Ser. A. I. Math.-Phys., 37 (1947), 179.
[46] Kavvas, M. L. Applied Stochastic Methods in Engineering (ECI 266) classnotes.
Lecture Notes, University of California, Davis, 1993.
[47] Kavvas, M. L. Nonlinear hydrologic processes: Conservation equations for determining their means and probability distributions. Journal of Hydrologic Engineering
8, 2 (March 2003), 4453.
[48] Kavvas, M. L., and Karakas, A. On the stochastic theory of solute transport by
unsteady and steady groundwater flow in heterogeneous aquifers. Journal of Hydrology
179 (1996), 321351.
[49] Keese, A. A review of recent developments in the numerical solution of stochastic partial differential equations (stochastic finite elements). Scientific Computing
2003-06, Deapartment of Mathematics and Computer Science, Technical University
of Braunschweig, Brunswick, Germany, 2003.
[50] Keese, A., and Matthies, H. G. Efficient solvers for nonlinear stochastic problem.
In Proceedings of the Fifth World Congress on Computational Mechanics, July 7-12,
248
2002, Vienna, Austria (http://wccm.tuwien.ac.at/publications/Papers/fp81007.pdf,
2002), H. A. Mang, F. G. Rammmerstorfer, and J. Eberhardsteiner, Eds.
[51] Kleiber, M., and Hien, T. D. The Stochastic Finite Element Method: Basic
Perturbation Technique and Computer Implementation. John Wiley & Sons, Baffins
Lane, Chichester, West Sussex PO19 1UD , England, 1992.
[52] Kolmogorov, A. N. Uber

die analytischen methoden in der wahrscheinlichkeitsrechnung. Math. Ann. 104 (1931), 415458.
[53] Koutsourelakis, S., Prevost, J. H., and Deodatis, G. Risk assesment of an
interacting structure-soil system due to liquefaction. Earthquake Engineering and
Structural Dynamics 31 (2002), 851879.
[54] Krige, D. G. A statistical approach to some mine valuations and allied problems
at the Witwatersrand. Masters thesis, University of Witwatersrand, Johannesburg,
South Africa, 1951.
[55] Kubo, R. Stochastic Liouville equations. J. Math. Phys. 4, 2 (1963), 174183.
[56] Lacasse, S., and Nadim, F. Uncertainties in characterizing soil properties. In
Uncertainty in Geologic Environment: From Theory to Practice, Proceedings of Uncertainty 96, July 31-August 3, 1996, Madison, Wisconsin (1996), C. D. Shackelford
and P. P. Nelson, Eds., vol. 1 of Geotechnical Special Publication No. 58, ASCE, New
York, pp. 4975.
[57] Lambe, W. T., and Whitman, R. V. Soil Mechanics, SI Version. John Wiley and
Sons, 1979.
249
[58] Langtangen, H. A general numerical solution method for Fokker-Planck equations
with application to structural reliability. Probabilistic Engineering Mechanics 6, 1
(1991), 3348.
[59] Lee, L. C. Wave propagation in a random medium: A complete set of the moment equations with different wavenumbers. Journal of Mathematical physics 15, 9
(September 1974), 14311435.
[60] Levenberg, K. A method for the solution of certain non-linear problems in least
squares. Quart. Appl. Math. 2 (1944), 164168.
[61] Li, C.-C., and Der Kiureghian, A. Optimal discretization of random fields.
Journal of Engineering Mechanics 119, 6 (1993), 11361154.
[62] Liu, P. L., and Der Kiureghian, A. A finite element reliability of geometrically
nonlinear uncertain structures. Journal of Engineering Mechanics 117, 8 (1991),
18061825.
`ve, M. Fonctions aleatories du second ordre. Supplement to P. Levy, Processus
[63] Loe
Stochastic et Mouvement Brownien, Gauthier-Villars, Paris, 1948.
`ve, M. Probability Theory, Volume II, fourth ed. Graduate Texts in Mathematics,
[64] Loe
Vol.46. Springer-Verlag, 1978.
[65] Mardia, K. V., and Marshall, R. J. Maximum likelihood estimation of models
for residual covariance in spatial regression. Biometrika 71, 1 (1984), 135146.
250
[66] Marosi, K. T., and Hiltunen, D. R. Characterization of spectral analysis of
surface waves shear wave velocity measurement uncertainty. Journal of Geotechnical
and Geoenvironmental Engineering 130, 10 (October 2004), 10341041.
[67] Marquardt, D. An algorithm for least-squares estimation of nonlinear parameters.
SIAM Journal of Applied Mathematics 11 (1963), 431441.
[68] Masud, A., and Bergman, L. A. Application of multi-scale finite element methods to the solution of the Fokker-Planck equation. Computer Methods in Applied
Mechanics and Engineering 194, 1 (April 2005), 15131526.
[69] Matthies, H. G., Brenner, C. E., Bucher, C. G., and Soares, C. G. Uncertainties in probabilistic numerical analysis of structures and soilds - stochastic finite
elements. Structural Safety 19, 3 (1997), 283336.
[70] Mayne, P. W., Brown, D., Vinson, J., Schneider, J. A., and Finke,
K. A. Site characterization of piedmont residual soils at the NGES, Opelika, Alabama. In National Geotechnical Experimentation Sites (NGES) (2000), J. Benoit
and A. Lutenegger, Eds., Geotechnical Special Publication No. 93, Geo-Institute of
ASCE, ASCE, Reston, VA, pp. 160185.
[71] Mellah, R., Auvinet, G., and Masrouri, F. Stochastic finite element method
applied to non-linear analysis of embankments. Probabilistic Engineering Mechanics
15 (2000), 251259.
251
[72] Montgomery, D. C., and Runger, G. C. Applied Statistics and Probability for
Engineers, third ed. John Wiley & Sons, 605 Third Avenue, New York, NY 10158,
2003.
[73] Moore, R. Methods and Applications of Interval Analysis. SIAM, Philadelphia,
1979.
, J. J., Garbow, B. S., and Hillstrom, K. E. User guide for MINPACK-1.
[74] More
Tech. Rep. ANL-8-74, Argonne National Laboratory, Argonne, IL, 1980. C translation
by Steve Moshier; Code available at https://sourceforge.net/projects/lmfit.
[75] Newmark, N. M. A method of computation for structural dynamics. ASCE Journal
of Engineering Mechanics Division 85 (1959), 6794.
[76] Nobahar, A. Effects of Soil Spatial Variability on Soil-Structure Interaction. Doctoral dissertation, Memorial University, St. Johns, NL, 2003.
[77] Oberkampf, W. L., Trucano, T. G., and Hirsch, C. Verification, validation
and predictive capability in computational engineering and physics. In Proceedings
of the Foundations for Verification and Validation on the 21st Century Workshop
(Laurel, Maryland, October 22-23 2002), Johns Hopkins University / Applied Physics
Laboratory, pp. 174.
[78] Paice, G. M., Griffiths, D. V., and Fenton, G. A. Finite element modeling
of settlement on spatially random soil. Journal of Geotechnical Engineering 122, 9
(1996), 777779.
252
[79] Phoon, K.-K., and Kulhawy, F. H. Characterization of geotechnical variability.
Canadian Geotechnical Journal 36 (1999), 612624.
[80] Phoon, K.-K., and Kulhawy, F. H. Evaluation of geotechnical property variability. Canadian Geotechnical Journal 36 (1999), 625639.
[81] Planck, M. Ueber einen satz der statistichen dynamik und eine erweiterung in der
quantumtheorie. Sitzungberichte der Preussischen Akadademie der Wissenschaften
(1917), 324341.
[82] Popescu, R., Prevost, J. H., and Deodatis, G. Effects of spatial variability on
soil liquefaction: Some design recommendations. Geotechnique 47, 5 (1997), 1019
1036.
[83] Risken, H. The Fokker-Planck Equation: Methods of Solutions and Applications,
second ed. Springer Series in Synergetics. Springer-Verlag, Berlin Heidelberg, 1989.
[84] Rose, C., and Smith, M. D. Mathematical Statistics with Mathematica. Springer
Texts in Statistics. Springer-Verlag, New York, 2002.
eller, G. I. A state-of-the-art report on computational stochastic mechanics.
[85] Schu
Probabilistic Engineering Mechanics 12, 4 (1997), 197321.
, B. On probabilistic yielding of materials. Communications
[86] Sett, K., and Jeremic
in Numerical Methods in Engineering (2007). In review.
253
, B. Spectral stochastic elastic-plastic finite element formu[87] Sett, K., and Jeremic
lation for geomechanics applications. International Journal for Numerical Methods in
Engineering (2007). In preparation.
, B. Uncertain soil properties and elasticplastic simulations
[88] Sett, K., and Jeremic
in geomechanics. In Probabilistic Applications in Geotechnical Engineering (Proceedings of Geo-Denver 2007: New Peaks in Geotechnics, held in Denver, CO from February 18-21, 2007) (2007), K. K. Phoon, G. A. Fenton, E. F. Glynn, C. H. Juang, T. F.
Griffiths, T. F. Wolff, and L. Zhang, Eds., Geotechnical Special Publication No. 170,
American Society for Civil Engineers, pp. 111.
, B. Stochastic ground motions. In Proceedings of Geo[89] Sett, K., and Jeremic
Congress 2008: The Challenge of Sustainability in the Geoenvironment, to be held in
New Orleans, LA from March 9-12, 2008 (2008), American Society for Civil Engineers. Accepted (to appear as Geotechnical Special Publication).
, B., and Kavvas, M. L. Probabilistic elasto-plasticity: Solution
[90] Sett, K., Jeremic
and verification in 1D. Acta Geotechnica 2, 3 (September 2007). In press (published
online in the Online F irst section.
, B., and Kavvas, M. L.
[91] Sett, K., Jeremic
The role of nonlinear harden-
ing/softening in probabilistic elastoplasticity. International Journal for Numerical

and Analytical Methods in Geomechanics 31, 7 (June 2007), 953975.
[92] Soize, C. The Fokker-Planck Equation for Stochastic Dynamical Systems and its
Explicit Steady State Solutions. World Scientific, Singapore, 1994.
254
[93] Sudret, B., and Der Kiureghian, A. Stochastic finite element methods and reliability: A state of the art report. Technical Report UCB/SEMM-2000/08, University
of California, Berkeley, 2000.
[94] Swanson, R., Alshibli, K., Frank, M., Costes, N., Sture, S., Batiste, S.,
, B. Mechanics of granular materials at low effective
Lankton, M., and Jeremic
stresses. In EOS Transactions, American Geophysical Union (1998), vol. 79 (17),
p. S332. poster.
[95] Van Kampen, N. G. Stochastic differential equations. Phys. Rep. 24 (1976), 171
228.
[96] Van Trees, H. L. Detection, Estimation and Modulation Theory, Part 1. Wiley,
New York, 1968.
[97] Vanmarcke, E. Random Fields: Analysis and Synthesis. The MIT Press, Cambridge, Massachusetts, 1983.
[98] Weiher, B., and Davis, R. Correlation of elastic constants with penetration resistance in sandy soil. International Journal of Geomechanics 4, 4 (December 2004),
319329.
[99] Western
etration
Earthquake
testing
(CPT)
Hazard
data
Team
of
of
Alameda
USGS.
County,
Cone
pen-
California.
http://earthquake.usgs.gov/regional/nca/cpt/data/?map=alameda.
[100] Wiener, N. The homogeneous chaos. American Journal of Mathematics 60 (1938),
897936.
255
[101] Wolfram Research Inc. Mathematica Version 5.0. Wolfram Research Inc., Champaign, Illonois, 2003.
[102] Xiu, D., and Karniadakis, G. E. Modeling uncertainty in flow simulations via
generalized polynomial chaos. Journal of Computational Physics 187 (2003), 137167.
[103] Xiu, D., and Karniadakis, G. E. A new stochastic approach to transient heat conduction modeling with uncertainty. International Journal of Heat and Mass Transfer
46 (2003), 46814693.
[104] Yu, Y. Coupling of ansys with a stochastic finite element solver and visualization
of results. Masters thesis, Technische Universitat Braunschweig, Institut f
ur Wissenschaftliches Rechnen, Braunschweig, Germany, 2003.
[105] Zadeh, J. The role of fuzzy logic in the management of uncertainty in expert systems.
Fuzzy Sets and Systems 11 (1983), 199228.
[106] Zienkiewicz, O. C., and Taylor, R. L. The Finite Element Method Volume 1,
the Basis, fifth ed. Butterwort-Heinemann, Oxford, 2000.
256
Appendix A
A.1
Derivation of Akl (Eq. (5.26))

Chain rule of differentiation for yield function f = f (I1 , J2 ) is used to develop
equation
f I1
f J2
f
Dpqkl =
Dpqkl +
Dpqkl
pq
I1 pq
J2 pq
(A.1)
The first term in the R.H.S of Eq. (A.1) can be written as:

f I1
2
f I1
Dpqkl =
2Gpk ql + K G pq kl
I1 pq
I1 pq
3

f I1
2
f I1
pk ql + K G
pq kl
= 2G
I1 pq
3
I1 pq
(A.2)
where now one can write the first term on the R.H.S of Eq. (A.2) as,

I1
I1
I1
1k +
2k +
3k ql
1q
2q
3q

I1
I1
I1
1l +
2l +
3l 1k
=
11
12
13

I1
I1
I1
1l +
2l +
3l 2k
+
21
22
23

I1
I1
I1
1l +
2l +
3l 3k
+
31
32
33
I1
I1
I1
1l 1k +
2l 2k +
3l 3k
=
11
22
33
I1
pk ql =
pq
(A.3)
257
It is important to note that the differentiation is with respect to radndom stress so that
terms in last line of previous equation cannot be simplified further. Similarly, from Eq. (A.2)
it follows that

f I1
f I1
I1
I1
2G
pk ql = 2G
1l 1k +
2l 2k +
3l 3k
I1 pq
I1 11
22
33
(A.4)
where detailed derivations yield

I1
I1
I1
11 kl +
22 kl +
33 kl
11
22
33

I1
I1
I1
+2
12 kl +
23 kl +
31 kl
12
23
31
I1
I1
I1
kl +
kl +
kl
=
11
22
33
I1
pq kl =
pq
(A.5)
so that the second part of the R.H.S. of Eq. (A.2)) can be written as

f I1
f I1
2
2
I1
I1
K G
pqkl = K G
kl +
kl +
kl
3
I1 pq
3
I1 11
22
33
(A.6)
First part of the R.H.S. of the Eq. (A.1)) can be written as

I1
f
I1
I1
f I1
Dpqkl =
2G
1l 1k +
2l 2k +
3l 3k
I1 pq
I1
11
22
33

f
2
I1
I1
I1
+
K G
kl
+
+
I1
3
11 22 33
(A.7)
while the second part of the same Eq. (A.1)) can be written as

f J2
f J2
2
Dpqkl =
2Gpk ql + K G pqkl
3
J2 pq
J2 pq

2
f J2
f J2
pk ql + K G
pq kl
= 2G
3
J2 pq
J2 pq
(A.8)
258
The second term of the previous equation (Eq. (A.8)) then becomes
J2
pk ql =
pq

J2
J2
J2
1k +
2k +
3k ql
1q
2q
3q
J2
J2
J2
ql 1k +
ql 2k +
ql 3k
=
1q
2q
3q

J2
J2
J2
=
1l +
2l +
3l 1k
11
12
13

J2
J2
J2
1l +
2l +
3l 2k
+
21
22
23

J2
J2
J2
+
1l +
2l +
3l 3k
31
32
33
(A.9)
which can be used to write the first term on the R.H.S. of Eq. (A.2)) as

f
J2
J2
J2
f J2
pk ql = 2G
1l 1k +
2l 1k +
3l 1k
2G
12
13
J2 pq
J2 11

J2
f
J2
J2
+ 2G
1l 2k +
2l 2k +
3l 2k
22
23
J2 21

J2
f
J2
J2
+ 2G
1l 3k +
2l 3k +
3l 3k
32
33
J2 31
(A.10)
and since
J2
J2
J2
J2
pq kl =
11 kl +
22 kl +
33 kl
pq
11
22
33

J2
J2
J2
12 kl +
23 kl +
31 kl
+2
12
23
31
J2
J2
J2
=
kl +
kl +
kl
11
22
33
(A.11)
with

J2
2
f J2
f
2
J2
J2
K G
pqkl = K G
kl +
kl +
kl
3
3
22
33
J2 pq
J2 11
(A.12)
259
and with

f J2
J2
f
J2
J2
Dpqkl = 2G
1l 1k +
2l 1k +
3l 1k
12
13
J2 pq
J2 11

J2
J2
J2
f
1l 2k +
2l 2k +
3l 2k
+ 2G
22
23
J2 21

f
J2
J2
J2
+ 2G
1l 3k +
2l 3k +
3l 3k
32
33
J2 31

J2
2
J2
J2
f
+ K G
kl +
kl +
kl
3
22
33
J2 11
(A.13)
will finally yield Eq. (A.1)),

I1
f
f
I1
I1
I1
2
Dpqkl =
1l 1k +
2l 2k +
3l 3k + K G
cd kl
2G
pq
I1
11
22
33
3
cd

2
J2
J1
f
(A.14)
ik jl + K G
ab kl
2G
+
ij
3
ab
J2
A.2
Derivation of B (Eq. (5.27))

Using Eq. (A.14), one can write:

f
f
I1
I1
f
I1
Drstu
=
2G
1l 1k +
2l 2k +
3l 3k
rs
tu
I1
11
22
33

f
2
I1
I1
I1
+
K G
kl +
kl +
kl
I1
3
11
22
33

f
J2
J2
J2
+ 2G
1l 1k +
2l 1k +
3l 1k
12
13
J2 11
J2
J2
J2
+
1l 2k +
2l 2k +
3l 2k
21
22
23

J2
J2
J2
1l 3k +
2l 3k +
3l 3k
+
31
32
33

f
J2
J2
J2
f
2
+
kl +
kl +
kl
K G
3
J2
11
22
33
kl
(A.15)
260
where, by using
f
f
=
kl
11
f
f
=
2l 1k
kl
12
f
f
=
3l 2k
kl
23
1l 1k
;
;
;
f
f
=
kl
22
f
f
3l 1k
=
kl
13
f
f
1l 3k
=
kl
31
2l 2k
;
;
;
f
f
=
kl
33
f
f
1l 2k
=
kl
21
f
f
2l 3k
=
kl
32
3l 3k
(A.16)
one can write the first part of the R.H.S. of Eq. (A.15) as

I1
f
I1
I1
f
2G
1l 1k +
2l 2k +
3l 3k
=
I1
11
22
33
kl
#
"

f 2
f 2
f 2
+
+
= 2G
11
22
33

2 "
2

#

f
I1
I1 2
I1 2
= 2G
+
+
I1
11
22
33
(A.17)
while the second part of the same equation (Eq. (A.15)) becomes

J2
J2
J2
1l 1k +
2l 1k +
3l 1k
11
12
13
J2
J2
J2
1l 2k +
2l 2k +
3l 2k
+
21
22
23

J2
J2
f
J2
1l 3k +
2l 3k +
3l 3k
+
31
32
33
kl

2 " 2 2 2
f
J2
J2
J2
= 2G
+
+
13
J2
11
12
2 2 2
J2
J2
J2
+
+
+
21
22
23
2 2 2 #
J2
J2
J2
+
+
+
31
32
33
f
2G
J2
(A.18)
and since
f
f
f
f
+
+
kl =
kl
11 22 33
(A.19)
261
it follows that

I1
I1
I1
f
f
2
kl +
kl +
kl
K G
I1
3
11
22
33
kl

2
f 2 I1
I1
I1 2
= K G
+
+
3
I1
11 22 33
(A.20)
and

f
J2
J2
f
J2
2
kl +
kl +
kl
K G
3
11
22
33
kl
J2
2

2
2
f
J2 J2 J2
= K G
+
+
3
11
22
33
J2
(A.21)
Therefore, adding Eqs. (A.16), (A.17), (A.19), and (A.20) one can write:

)
I1 2
I1 2
I1 2
2G
+
+
11
22
33
#
2

I1
2
ij
+ K G
3
ij
"

2 #
2

J2
2
J2 J2
f
+ K G
ij
2G
+
ij ij
3
ij
J2
f
f
B=
Drstu
=
rs
tu
A.3
f
I1
2 "
(
(A.22)
Derivation of KP (Eq. (5.28))

From the consistency condition (f = 0), one can derive
f
rn
qn
(A.23)
qn =< L > rn
(A.24)
KP =
with
where, < > is the McCaulay bracket and L is the loading index given by
L=
1
Lij ij
KP
(A.25)
262
where, Lij is a vector normal to the loading surface in stress space and dot represents
time derivative. For Drucker-Prager yield criteria (Eq. (5.24)) one can simplify the above
equation (Eq. (A.23)) as
f
(A.26)
=< L >
(A.27)
KP =
where,
By assuming that the friction angle like internal variable is a function of equivalent plastic
strain (epeq ) i.e.,
= (epeq )
(A.28)
one can write, (using the relationship: e peq = 1/ 3 < L > f / J2 )

=
1 d
f
d p
=< L >
p e eq =
p <L>
deeq
J2
3 deeq
and therefore,
KP =
A.4
f
1 f d f
f
rn =
=
p
qn
3 deeq J2
(A.29)
Derivation of Eq. (5.11): Ensemble average form of stochastic continuity equation (Eq. (5.8))
Note: The ensemble average form of the stochastic continuity equation
was derived by Kavvas and Karakas [48] for applications to hydrologic processes.
In the following the same has been re-derived for constitutive equation.
263
The stochastic continuity equation (Eq. (5.8)) corresponding to the general form of constitutive rate equation (Eq. (5.6)) can be written as:
((x, t), t)
= [(x, t), D(x), q(x), r(x), (x, t)] [(x, t), t]

t
(A.30)
The above equation (Eq. (A.30)) in the operator form can be written as:
= ( . .)
t
(A.31)
Now, having defined a time-space varying sure operator, A0 ((x, t), t):
A0 ((x, t), t) = hi . . hi
(A.32)
where, hi denotes the ensemble average operator, one can define a time-space non-stationary
stochastic operator A1 (x, t) from Eqs. (A.31) and (A.32) as:
A1 ((x, t), t) = ( + hi). + .( + hi)
(A.33)
where, is the root mean square of the fluctuations of the random operator on the r.h.s of
Eq. (A.31) around its mean. Now substituting Eqs. (A.32) and (A.33) into Eq. (A.31):
= [A0 ((x, t), t) + A1 ((x, t), t)]

t
(A.34)
which is the operator stochastic differential equation for the continuity equation corresponding to the constitutive rate equation.
264
In the following, Van Kampens approach [95] will be followed in obtaining the
deterministic operator differential equation for the mean of the phase density. To this end,
making interaction substitution:
((x, t), t) =
exp
Z
A0 ((x, ), )d
1 ((x, t))
(A.35)
denotes chronologically ordered exponential, where within the exponential series

where
exp
of the integrals, the arguments within each integral are ordered in time. For example, for a
dummy variable, B( ),
exp
Z
B( )d
0
=1+
Z
X
1
d1
d2 . . .
0
m1
dm B(1 )B(2 ) . . . B(m )
(A.36)
Now, combining Eqs. (A.34) and (A.35):
d1 ((x, t), t)
=
dt

Z t
1
Z t

A0 ((x, ), )d 1 ((x, t), t)

A1 ((x, t), t) exp
A0 ((x, ), )d
exp
0
(A.37)
where,

Z t
1
Z t

A0 ((x, ), )d
= exp
A0 ((x, ), )d
exp
0
Substituting Eq. (A.38) into Eq. (A.37):
(A.38)
265
d1 ((x, t), t)
=
dt

Z t
Z t

A0 ((x, ), )d
1 ((x, t), t)
A0 ((x, ), )d A1 ((x, t), t) exp
exp
0
(A.39)
and representing the non-commutative operator inside the bracket in Eq. (A.39) by , one
can write Eq. (A.39) as:
d1 ((x, t), t)
= ((x, t), t) 1 ((x, t), t)
dt
(A.40)
The ensemble average form of the above equation (Eq. (A.40)) was shown by Van Kampen
[95] as:
d h1 ((x, t), t)i

=
dt

Z t
d hhh((x, t), t) ((x, t ), t )iii h1 ((x, t), t)i
h((x, t), t)i + 2
0
(A.41)
where, hhhiii represents time-ordered second cumulant. The above equation (Eq. (A.41))
is accurate to the order of 2 c , where c is the correlation time of the stochastic operator
A1 ((x, t), t). In Eq. (A.41), expressing explicitly in terms of Eq. (A.39) and substituting
Eq. (A.35) into Eq. (A.41), one can write:
266

Z t

d
A0 ((x, ), )d h((x, t), t)i =
exp
dt

Z t

0 Z t
A0 ((x, ), )d
A0 ((x, ), )d hA1 ((x, t), t)i exp
exp
0
0

Z t
Z t
2
ds
A0 ((x, ), )d A1 ((x, t), t)
+
exp
0 Z ts
Z 0t

A0 ((x, ), )d exp
exp
A0 ((x, ), )d
0
0

Z ts
A0 ((x, ), )d
h1 ((x, t), t)i (A.42)
A1 ((x, t s), t s)exp
0
By performing various mathematical operations, by exploiting the characteristics of the

time-ordered exponential, one can write:
h((x, t), t)i

t

ds
A1 ((x, t), t)
= A0 ((x, t), t) h((x, t), t)i +
0

Z t
A0 ((x, ), )d A1 ((x, t s), t s)

exp
ts

Z t
A0 ((x, ), )d h((x, t), t)i

(A.43)
exp
2
ts
To obtain an explicit deterministic PDE for the ensemble phase density (h((x, t), t)i),
it is necessary to perform the operations in Eq. (A.43). To this end, the right multiplicative
term of the second additive term on the r.h.s of Eq. (A.43) can be particularized to the
problem in hand (substituting Eq. (A.32) for A0 ) as:

Z t
A0 ((x, ), ) h((x, t), t)i =

exp
ts

Z t
h((x, ), )i
h((x, t), t)i (A.44)

h((x, ), )i
d
exp
ts
267
Now, noting that for any time-dependent dummy vectors/tensors Z( ) and Y ( ), the following operation is valid:
exp
Z
d (Z( ) + Y ( )) =
Z
Z t

Z t
dsZ(s) Y ( )exp
exp
d exp
0
Z t

d Z( )
dsZ(s) exp
0
(A.45)
one can write Eq. (A.44) as:

Z t
A0 ((x, ), ) h((x, t), t)i =

exp
ts
Z t
Z t

h((x, ), )i
d h((x, ), )i
exp
d
exp
h((x, t), t)i
ts
ts
(A.46)
Further, from Eq. (A.33), the stochastic operator A1 ((x, t), t) can be written as:

h((x, t), t)i ((x, t), t)
A1 ((x, t), t) =
+ h((x, t), t)i ((x, t), t)
(A.47)
Utilizing the commutation and product properties of Lie operator in Eq. (A.43) one can
write:
268
exp
Z
Z
A1 ((x, t s), t s) exp
A0 ((x, ), )d h((x, t), t)i

A0 ((x, ), )d
ts
ts

exp[(x,
t), t] x; t s)
exp[(x,
t), t] x; t s)i (
h (
=
h((x, t), t), t)i

+ h (exp[(x, t), t] x; t s)i (exp[(x, t), t] x; t s)

Z t

Z t
((x, ), )

h((x, ), )i
d
exp
h((x, t), t)i
d
exp
ts
ts

h((x, t), t)i
+ h (exp[(x, t), t] x; t s)i (exp[(x, t), t] x; t s)

(A.48)
where,
Z
exp[(x, t), t] x = exp
d h((x, t), t)i

x
ts
(A.49)
Further, utilizing Eq. (A.48) and performing mathematical operations as dictated by A1 ((x, t), t),
the second additive term on the r.h.s of Eq. (A.43) can be written as:
269

Z t

ds
A1 ((x, t), t)exp
A0 ((x, ), )d
ts

Z t
A1 ((x, t s), t s)
A0 ((x, ), )d h((x, t), t)i
exp
ts
t
0

Z t
A0 ((x, ), )d
ds
A1 ((x, t), t)exp
ts

Z t

A0 ((x, ), )d h((x, t), t)i

exp
A1 ((x, t s), t s)
(A.50)
ts
t
0
+
+

((x, t), t) (
exp[(x,
t), t)] x; t s)
ds Cov0
;

2
(exp[(x, t), t]x; t s)
h((x, t), t)i
+ Cov0 ((x, t), t);
2

Z t
((x, t), t)
ds Cov0
; (exp[(x, t), t] x; t s)
0

(
exp[(x,
t), t]x; t s)
+ Cov0 ((x, t), t);

Z t
Z t

h((x, ), )i
h((x, ), )i
h((x, t), t)i

exp
d
d
exp
ts
ts
Z t
Cov0 [((x, t), t); (

exp[(x,
t), t] x; t s)]
0
Z t
Z t

h((x, ), )i
h((x, ), )i

d
d
exp
exp
ts
ts
2

Z t
Z t
h((x, ), )i
h((x, ), )i

+ exp
d
h((x, t), t)i
exp
d
ts
ts

Z t
((x, t), t)
ds Cov0
; (exp [(x, t), t)] x; t s)
0

h((x, t), t)i

(
exp[(x,
t), t] x; t s)
+ Cov0 ((x, t), t);

Z t
(exp[(x, t), t] x; t s) h((x, t), t)i

dsCov0 ((x, t), t);
Z0 t
dsCov0 [((x, t), t); (

exp[(x,
t), t] x, t s)]
0

Z t
Z t
h((x, ), )i h((x, t), t)i
h((x, ), )i
exp
d
d
exp
ts
ts

2
Z t
h((x, t), t)i
exp[(x,
t), t] x; t s)
dsCov0 ((x, t), t); (
(A.51)
270
Analyzing the orders of magnitudes of the terms on the r.h.s of Eq. (A.51), noting that the
second integral, the terms in the third integral and the argument of the sixth integral have
orders higher than 2 c , and hence dropping those terms:

Z t
A0 ((x, ), )d
A1 ((x, t), t)
exp
ts
0

Z t

A0 ((x, ), )d h((x, t), t)i

exp
A1 ((x, t s), t s)
ts

Z t
((x, t), t) (
exp[(x,
t), t)] x; t s)
ds Cov0
=
;
0

2
(exp[(x, t), t]x; t s)
+ Cov0 ((x, t), t);
h((x, t), t)i
2

Z t
((x, t), t)
; (exp [(x, t), t)] x; t s)

ds Cov0
+
0

h((x, t), t)i

(
exp[(x,
t), t] x; t s)
+ Cov0 ((x, t), t);

Z t
(
exp[(x,
t), t] x; t s) h((x, t), t)i
+
dsCov0 ((x, t), t);
0
Z t
2
h((x, t), t)i
+
dsCov0 [((x, t), t); (
exp[(x,
t), t] x; t s)]
(A.52)
0
2
ds
Further, the first additive term on the r.h.s of Eq. (A.43) can be written as:
A0 ((x, t), t) h(x, t), t)i =
h((x, t), t)i h((x, t), t)i
(A.53)
Combining Eqs. (A.52) and (A.53) and substituting in Eq. (A.43), one obtains the ensemble
average form of the stochastic continuity equation corresponding to the constitutive rate
equation as:
h((xt , t), t)i
=
t

Z t
((x, t ), t )
d Cov0 ((x, t), t);
h((x, t), t))i +
h((xt , t), t)i
0

Z t

h((xt , t), t)i
d Cov0 [((x, t), t); ((x, t ), t )]

(A.54)
+

Probabilistic Elasto Plasticity and Its Applications in Finite Element Simulations of Stochastic Elastic-Plastic Boundary Value Problems

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Probabilistic Elasto Plasticity and Its Applications in Finite Element Simulations of Stochastic Elastic-Plastic Boundary Value Problems

Uploaded by

Copyright:

Available Formats

Probabilistic ElastoPlasticity and Its Application in Finite Element

Simulations of Stochastic ElasticPlastic Boundary Value Problems

Probabilistic ElastoPlasticity and Its Application in Finite Element Simulations of

Professor Boris Jeremic

Motivation and Theoretical Background

2 Probability Theory Background

2.5.1 Classification of Random Processes/Fields . . . .

Uncertain and Spatially Uncertain Material Properties

3 Characterization & Quantification of Uncertainties in Material Properties

4 Random Field Modeling of Uncertain Material Properties

III Material (Constitutive) Level Stochastic Simulation: Probabilistic

6 Probabilistic ElastoPlasticity: Numerical Examples and Verifications 116

ElasticPlastic Shear Constitutive Behavior with Probabilistic Yield Criteria 147

IV Stochastic Simulations of Solids and Structures with Uncertain

8 Stochastic Finite Elements: Numerical Examples and Verifications

Conclusions and Future Research Directions

A.4 Derivation of Eq. (5.11): Ensemble average form of stochastic continuity

Interpreted Effective Stress Strength Parameters at Opelika NGES (after

Anticipated Influence of Material Fluctuations on Stress-Strain Behavior . .

Approximation of Dirac delta initial condition (Eq. (5.52)) with a Gaussian

KL eigenvalues of exponential covariance kernel having variance = 1000 kP a2

Schematic of static 1D soil column (shear beam) example . . . . . . . . . .

8.26 Time evolution of standard deviation of displacement at the top node of

Classification of random processes based on state and parameter space . . .

Representative values of variabilities in consolidation parameters (after Baecher

Maximum likelihood estimated constant mean Gauss-Markov model parameters 76

Comparison of results (at top node) of FPKE-based spectral stochastic finite

Motivation and Theoretical

and speaking of natural geomaterials, they are

material properties as random. Usually the number of random material properties in

material properties are probabilistic. For a 3-D elasto-plastic model

at least 10, 0003 = 1012 runs of computationally expensive deterministic model.

simulations of behaviors of solids and structures made of stochastic elasticplastic materials,

Probabilistic characterization and rational quantification of uncertainties in material

probability theory with theory of elastoplasticity and subsequent development of a stochastic

Probability Theory Background

The definitions are mainly following Montgomery and Runger [72]

Properties of Single Random Variable

analyze the complete probabilistic description of a single random variable.

Probability Distribution Function

defined as A = 2sin/ 3(3 sin), where is the friction angle of a material

can be interpreted as:

P [A = 0.25] = P [0.245 < A < 0.255] =

Cumulative Distribution Function

for - < x <

Moments of Probability Distributions

for < x <

for < x <

for < x <

derived in Subsection 5.2.1

G G2 /(G + cu / 3)), where cu = 0.3 (deterministic)) is the rate of evolution of internal

for < x <

V ar[X] = E[X E(X)]2

for < x <

x2 fX (x)dx (E[X])2 for < x <

and the COV of von Mises elasticplastic shear modulus as:

Characteristic Function and its relation with Moments and Cumulants

characteristic function also completely characterizes the probabilistic behavior of a random

By expanding the characteristic function, x in terms of MacLaurin series:

and evaluating the various derivatives using Eq. (2.10):

xfX (x)dx = i hXi

one can express the j th moment of X in terms of characteristic function of X as: