Stocastic Soil

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/323627299
STOCHASTIC VARIABILITY OF SOIL PROPERTIES, DATA ANALYSIS, DIGITAL

SIMULATION, EFFECTS ON SYSTEM BEHAVIOR
Thesis · November 1995
CITATIONS READS
45 251
1 author:
Radu Popescu
California Polytechnic State University, San Luis Obispo
64 PUBLICATIONS 1,037 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
DOE Applications View project
All content following this page was uploaded by Radu Popescu on 08 March 2018.
The user has requested enhancement of the downloaded file.

STOCHASTIC VARIABILITY OF SOIL PROPERTIES:
DATA ANALYSIS, DIGITAL SIMULATION,
EFFECTS ON SYSTEM BEHAVIOR
RADU POPESCU
A DISSERTATION
PRESENTED TO THE FACULTY
OF PRINCETON UNIVERSITY
IN CANDIDACY FOR THE DEGREE
OF DOCTOR OF PHILOSOPHY
RECOMMENDED FOR ACCEPTANCE

BY THE DEPARTMENT OF
CIVIL ENGINEERING AND OPERATIONS RESEARCH
NOVEMBER 1995
c Copyright by Radu Popescu, 1995. All rights reserved.
Stochastic Variability of Soil Properties Radu Popescu, 1995
0.1 Abstract
Aside from the measurement induced bias and even within assumed homogeneous soil
strata, natural soils are variable in their properties. The scarcity in eld test results,
characteristic to geotechnical engineering, and a suciently large degree of disorder
exhibited by soil properties leads to the use of statistical methods in describing the
distribution of those properties within a \statistically homogeneous" soil zone.
In this context, it is proposed to model relevant soil parameters as the components
of a multi{dimensional, multi{variate (nD{mV) stochastic eld. The characteristics
of the stochastic eld (cross{spectral density matrix, probability distribution of each
component) are evaluated by nonlinear regression analysis based on sample functions
derived from statistically signicant sets of \in{situ" measurement results. Sample
functions of an nD{mV non{Gaussian stochastic eld are then simulated using an
extension of the spectral representation method. A methodology for evaluating tra-
ditional soil parameters from the results of conventional eld tests is proposed and
employed to derive the spatial distribution of soil parameters which are relevant for
dynamic analyses. Monte Carlo simulations of the behavior of horizontally layered
soil deposits, as well as of structures founded on liqueable soil and subjected to
seismic loads are performed by combining vector eld simulations with nonlinear
dynamic nite element analyses.
The numerical simulation results (predicting excess pore{pressure build{up, lique-
faction index, and liquefaction induced deformations) obtained using stochastic input
constitutive parameters are compared to deterministic input simulation results. It is
concluded that, for the case of dynamically induced pore water pressure build{up, the
results of classical deterministic analyses are on the non{conservative side. Selection
of the optimum discretization mesh size for stochastic analyses, as well as in
uence
on the computational results of assumed spatial correlation distances and probability
distribution functions are also discussed.
The theoretical developments are illustrated with numerical examples conducted
iii
in terms of traditional soil parameters, and based on data obtained from extensive
eld measurement programs. Therefore, the study will provide a basis for character-
ization of spatial variability of soil properties and for liquefaction risk assessment.
iv
0.2 Acknowledgements
First, I would like to gratefully acknowledge the help and guidance of my advisor,
Professor Jean H. Prevost. I thank him for believing in me, oering me the chance
to start this work, and then, during all these years, patiently sharing with me his
knowledge and expertise.
I am also indebted to Professor Georgios Deodatis for his suggestions and rec-
ommendations so much needed in some areas of my research, for his support and
thoughtfulness. To Professor Erik H. Vanmarcke I am grateful for introducing me to
the eld of stochastic mechanics, and for his always kind but challenging recommen-
dations. I would also like to thank the other members of the examination committee,
Professors Ahmet S. C akmak and Peter C.Y. Lee, for their time and consideration.
I would like to acknowledge the nancial support provided by Kajima Corporation
which made possible this project, and to express my sincere appreciation to Dr.
Yoichi Nojiri and Dr. Naoto Ohbo for a wonderful joint research experience. I am
also grateful to Dr. Michael G. Jeeries from Golder Associates for providing the
eld data used in numerical examples.
A special thank you to Professor Dan Stematiu from the Civil Engineering Insti-
tute in Bucharest for opening for me the world of nite elements, guiding my research
for almost 15 years, and supporting me in starting this project.
For their love, patience and devotion in helping me and my family during these
busy years, I thank my mother Eugenia and my mother{in{law Ioana. Away from
their own families, they provided the so much demanded care and love for my children,
trying to substitute for a very busy (and sometimes tired) father.
Life in Princeton and in this country would not have been so enjoyable without
the warm presence of John and Aruna Desai, and all our friends from International
Students Inc. Thank you all for your prayers and spiritual support in times of trial
and in times of joy. For their loyal friendship and helpful introduction to the American
way of life, I thank Carey and Harry Compton, and for the great vacation days which
v
helped \recharging my batteries" I thank Pat and Seth Carpenter. To our friends
Agnes and Michael Sherman I am grateful for their care and love so freely extended
to me and for their willingness to help with the proofreading of the text.
I want to thank my colleagues Igor Krstelj and Arghia Deb for sharing the stu-
dentship experience with its mostly dicult but in general enjoyable days, and for
providing an intellectually stimulating environment. A big thank you to Engineering
Library sta members for their kindness, and for their perseverance in dealing with
my requests for materials.
Last but foremost, I am indebted to my wife Adriana, who during these three
years has endured the sacrices involved by my education, cared for and supported
the family, encouraged me in my work, and been of a great help with editing and
proofreading the text.
vi
Contents
0.1 Abstract : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : iii

0.2 Acknowledgements : : : : : : : : : : : : : : : : : : : : : : : : : : : : v
1 Introduction 1
1.1 Objective : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 1
1.2 Literature review : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 2
1.2.1 Parameter estimation for stochastic elds : : : : : : : : : : : : 2
1.2.2 Digital simulation of stochastic elds : : : : : : : : : : : : : : 3
1.2.3 Evaluation of soil properties : : : : : : : : : : : : : : : : : : : 4
1.2.4 Eects of spatial variability on soil liquefaction : : : : : : : : 5
1.3 Proposed approach : : : : : : : : : : : : : : : : : : : : : : : : : : : : 6
2 Field Data Analysis 11
2.1 Raw data transformations : : : : : : : : : : : : : : : : : : : : : : : : 11
2.1.1 Field data : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 12
2.1.2 Filtering and shortening : : : : : : : : : : : : : : : : : : : : : 14
2.1.3 Standardization : : : : : : : : : : : : : : : : : : : : : : : : : : 15
2.1.4 Test for stationarity : : : : : : : : : : : : : : : : : : : : : : : 18
2.2 Analytical models : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 20
2.2.1 Probability distribution functions : : : : : : : : : : : : : : : : 20
2.2.2 1D correlation structures and SDF : : : : : : : : : : : : : : : 22
2.3 Estimation of spatial variability statistics : : : : : : : : : : : : : : : : 27
2.3.1 Proposed curve tting procedure : : : : : : : : : : : : : : : : 27
2.3.2 Probability distribution function : : : : : : : : : : : : : : : : : 29
2.3.3 Correlation structure in vertical direction : : : : : : : : : : : : 32
2.3.4 Correlation structure in horizontal direction : : : : : : : : : : 39
vii
2.3.5 Spatial cross{correlation between various soil properties : : : : 45

2.4 Conclusions : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 47
3 Simulation of nD{mV, non{Gaussian stochastic elds 50
3.1 Assumptions : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 50
3.2 Simulation of nD{mV Gaussian stochastic elds : : : : : : : : : : : : 51
3.2.1 General formulation : : : : : : : : : : : : : : : : : : : : : : : 51
3.2.2 Use of Fast Fourier Transform : : : : : : : : : : : : : : : : : : 54
3.2.3 Particularities for simulation of random media representing soil
properties : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 56
3.2.4 Computation of power spectral density from sample elds : : : 58
3.2.5 Application | Simulation of a 2D{3V Gaussian stochastic eld 66
3.3 Simulation of non{Gaussian vector elds : : : : : : : : : : : : : : : : 70
3.3.1 Basic algorithm : : : : : : : : : : : : : : : : : : : : : : : : : : 70
3.3.2 Procedure to update the generating spectral density functions
SGrr (k) : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 74
3.3.3 Recovery of the o{diagonal terms in the target spectrum : : 79
3.3.4 Applications { Simulation of non{Gaussian vector elds : : : : 80
3.4 SINOGA { Program for simulation of nD{mV non{Gaussian stochastic
elds : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 93
3.4.1 General presentation : : : : : : : : : : : : : : : : : : : : : : : 93
3.4.2 Computational eort : : : : : : : : : : : : : : : : : : : : : : : 95
3.5 Conclusions : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 96
Appendix 3.A : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 97
Appendix 3.B : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 98
Appendix 3.C : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 98
Appendix 3.D : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 100
4 Evaluation of soil constitutive parameters 103
4.1 The multi{yield plasticity model : : : : : : : : : : : : : : : : : : : : : 103
4.1.1 Basic relations : : : : : : : : : : : : : : : : : : : : : : : : : : : 103
4.1.2 Constitutive model parameters : : : : : : : : : : : : : : : : : 106
4.2 Constitutive parameter evaluation : : : : : : : : : : : : : : : : : : : : 108
viii
4.2.1 From laboratory test results and liquefaction strength : : : : : 108

4.2.2 Estimation from eld data : : : : : : : : : : : : : : : : : : : : 112
4.3 Application { Constitutive parameter evaluation for Erksak sand : : : 120
4.3.1 Empirical correlations : : : : : : : : : : : : : : : : : : : : : : 120
4.3.2 Dilation parameter : : : : : : : : : : : : : : : : : : : : : : : : 123
4.4 Sensitivity analyses : : : : : : : : : : : : : : : : : : : : : : : : : : : : 126
4.4.1 Constitutive parameter evaluation and estimated variability : 126
4.4.2 Numerical simulation results : : : : : : : : : : : : : : : : : : : 129
4.4.3 Remarks : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 132
4.5 Conclusions : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 134
Appendix 4.A : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 136
5 E ects of spatial variability on dynamic behavior 139
5.1 Representation of random data : : : : : : : : : : : : : : : : : : : : : 139
5.1.1 Transfer of random data : : : : : : : : : : : : : : : : : : : : : 139
5.1.2 Selection of mesh size for stochastic analysis : : : : : : : : : : 143
5.2 Eects on dynamic behavior of soil deposits : : : : : : : : : : : : : : 148
5.2.1 Finite element analysis set{up : : : : : : : : : : : : : : : : : : 148
5.2.2 Stochastic vs. deterministic input parameters : : : : : : : : : 149
5.2.3 In
uence of assumed correlation distances : : : : : : : : : : : 151
5.2.4 In
uence of probability distribution of soil properties : : : : : 152
5.3 Application { Structure founded on liqueable soil deposit : : : : : : 152
5.4 Conclusions : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 154
Bibliography : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 157
List of Figures : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 170
List of Tables : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 175
Color Plates : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 176
ix
Chapter 1
Introduction
1.1 Objective
A more rational approach to geotechnical design is made possible by use of statis-
tically based techniques of data analysis. These methods allow designers to rely
more on analytical approaches when dealing with various uncertainties related to soil
properties and induced by: (1) spatial variability of soil properties in natural deposits,
(2) limited availability of information about subsurface conditions, and (3) sampling
and testing errors (Lumb, 1974, Vanmarcke, 1977, De Groot and Baecher, 1993).
In this context, a rst objective of the study is to assess the in
uence of spatial
variability of various soil parameters on system behavior, and to provide the tools for
analyzing the stochastic distribution of soil properties. A methodology is developed
to estimate the statistics of spatial variability based on results of common in{situ soil
tests, and an algorithm is implemented for digital simulation of multi{dimensional,
multi{variate, non{Gaussian stochastic elds representing the spatial distribution
of various soil properties. Several issues related to the importance of accounting for
spatial variability of geomechanical properties, as well as to some details of stochastic
input analysis set{up are also addressed.
The limited availability of eld data is usually dealt with by using various interpo-
lation methods. To obtain reliable estimates of soil properties at untested locations, it
is essential to account for the degree of spatial correlation exhibited by soil properties.
1
A data base containing the general characteristics of spatial variability (probability

distributions, cross{correlation structure) for various types of soil deposits is much
needed for that reason. The present research, based on real results of in{situ soil
tests, provides statistical data on two sites: a loose to medium dense hydraulic ll
sand deposit, and a dense natural sand deposit.
Parameter calibration methods for most soil constitutive models are based on
results of (sometimes expensive) laboratory test programs. On the other hand, the
eld measurements have obvious advantages over laboratory soil tests (Campanella,
1994): the disturbances due to sampling are minimized, the eects of fabric, aging
and natural (in{situ) stress state are preserved, the spatial variability is properly
captured. In this context, a methodology to evaluate traditional soil parameters
based on results of inexpensive in{situ soil tests is developed.
1.2 Literature review

1.2.1 Parameter estimation for stochastic elds
Estimation of statistics of spatial variability of soil properties refers to: average values
(or spatial trends), standard deviation, (cross{) correlation structure, and probability
distribution functions.
Data homogeneity can be achieved by either considering thin horizontal layers (e.g.
Jeeries, 1989), or by removing the spatial trends and normalizing the
uctuations
about mean values (Vanmarcke, 1977 DeGroot and Baecher, 1993 Popescu et.al.,
1995, among others).
Three basic methods for correlation structure estimation are reported in the lit-
erature: method of moments, inverse estimation, and maximum likelihood. Moment
estimators (e.g. Vanmarcke, 1983 Baecher, 1984 Bendat and Piersol, 1986, 1993)
are evaluated either via direct computations (in the spatial domain), or indirect com-
putations (spectral analysis and use of Fourier Transform). The inverse estimation
procedure (e.g. Journel and Huijbregts, 1978 Soulie et.al., 1990), mainly referred
2
in the mining geostatic literature, is also known as \best linear unbiased estimator".
The maximum likelihood method, which simultaneously provides estimates of spa-
tial trends, measurement noise, and correlation structure of residuals about spatial
trends, is referred among others by Magnus (1978), Mardia and Marshall (1984), and
DeGroot and Baecher (1993). Apart from these three methods, there are a series of
practical procedures (Vanmarcke, 1983), based on estimation of the variance reduc-
tion function, and which are appropriate for analyzing data provided by in{situ soil
tests (e.g. Popescu et.al., 1995). Excepting for the moment estimators, all the other
methods require an assumption on the shape of the autocovariance function.
Results of numerous studies on the probability distribution of soil properties are
reported in the literature. Based on eld measurements and empirical correlations,
both Gaussian and non{Gaussian distributions are tted for various soil properties
by various researchers (e.g. Schultze, 1971). A comprehensive review on such studies
on permeability, porosity, and relative density is presented by Ural (1995). It is con-
cluded here that: (1) most soil properties exhibit skewed, non{Gaussian distributions,
and (2) each soil property can follow dierent probability distributions for various
materials and sites, and therefore the statistics and the shape of the distribution
function have to be estimated for each case.
1.2.2 Digital simulation of stochastic elds

Four main groups of methods for digital simulation of random elds are referred in
the literature: (1) Auto{Regressive Moving{Average (ARMA) models (e.g. Gersch
and Yonemoto, 1977 Polhemus and Cakmak, 1981 Spanos, 1983 Elishako, 1983
Samaras et.al., 1985 Kozin, 1988), (2) covariance matrix decomposition method, de-
scribed by Vanmarcke (1983), Yamazaki and Shinozuka (1990), etc., and used, among
others, by Popescu et.al. (1995), (3) spectral representation method introduced by
Rice (1954), but rst applied (Shinozuka and Jan, 1972 Shinozuka, 1972) and sub-
sequently developed by Shinozuka and co{workers, and (4) conditional probability
3
function method (Kameda and Morikawa, 1991, 1992, 1993 Shinozuka and Zhang,
1995).
The spectral representation method is one of the most widely used, among the
above mentioned. Yang (1972, 1973) showed that use of the Fast Fourier Transform
(FFT) can substantially improve the computational eciency. Shinozuka (1974)
extended the use of FFT technique to multi{dimensional cases. Extensions to multi{
variate and non{stationary processes are reported by Deodatis (1995, 1996), and
Deodatis et.al. (1995). Yamazaki and Shinozuka (1988) developed an iterative pro-
cedure to simulate non{Gaussian scalar elds, using a memoryless transformation
(Soong and Grigoriu, 1993 Grigoriu, 1995). The procedure was applied by Ural
(1995) for two{dimensional, univariate elds. An extension of the iterative method
to multi{dimensional, multi{variate, non{Gaussian elds was reported by Popescu
and Deodatis (1995) and Deodatis et.al. (1995). Three review papers on the sub-
ject of simulation using the spectral representation method are written by Shinozuka
(1987), and Shinozuka and Deodatis (1988a, 1991).
1.2.3 Evaluation of soil properties

Estimation of soil geomechanical properties from in{situ test results is advocated in
this work, as compared to evaluation from laboratory soil test results (see e.g. Peck,
1979 Campanella, 1994). A rich literature has been published regarding the estima-
tion of soil parameters based on the results of standard penetration test (Meyerhof,
1956 Schmertman, 1975 Holtz and Gibbs, 1979 Skempton, 1986, etc.), and of cone
penetration test (Robertson and Campanella, 1983 Baldi et.al., 1982, 1989 Been
et.al., 1986, 1987b Robertson, 1990 Jeeries et.al., 1988b, among others). A com-
prehensive reference on estimating engineering soil properties from eld or laboratory
soil test results is represented by the work of Kulhawy and Maine (1990).
Regarding the behavior of saturated sand deposits under seismic excitation, two
basic (and in some respects divergent) approaches describe the liquefaction phe-
4
nomenon (for a detailed discussion, the reader is referred to McRoberts and Sladen,
1992):
1. The cyclic mobility, or Berkeley, approach was initiated by Seed and Lee (1966),
and developed and updated over the subsequent decades (e.g. Seed, 1979,
1987 Seed et.al., 1983, 1984 Tokimatsu and Yoshimi, 1983 Seed and Harder,
1990 Fear and McRoberts, 1995, etc., for correlations with NSPT blow{counts,
and Robertson and Campanella, 1985 Seed and De Alba, 1986 Shibata and
Teparaksa, 1988, etc., using correlations with CPT results).
2. The steady state approach is described in a series of publications by Castro and
co{workers (e.g. Castro, 1969, 1975 Castro and Poulos, 1977 Poulos, 1981).
The concept of a \state parameter" for sands has been introduced by Been and
Jeeries (e.g. Been and Jeeries, 1985 Jeeries and Been, 1987 Been et.al.,
1991).
1.2.4 Eects of spatial variability on soil liquefaction

The eects of spatial variability of soil properties on the extent and pattern of seis-
mically induced pore{water pressure build{up was studied in the past. One{ and
two{dimensional nite element analyses of horizontally layered soil subjected to earth-
quake loads, using stochastic input constitutive parameters, were reported by Fenton
(1990), Ohtomo and Shinozuka (1990), Popescu (1993), Ural (1995) and Popescu
et.al. (1995), among others. Some limitations of those approaches are mentioned
here: (1) the statistics of material property
uctuations (e.g. correlation structure)
were either assumed or evaluated based on relatively scarce eld data (2) not all
soil constitutive parameters which are signicant for liquefaction assessment where
accounted for in the stochastic model (3) either total or no cross{correlation be-
tween the
uctuations of various soil properties was assumed (4) in most cases,
normal assumption was adopted for the soil material property distributions. Due to
those limitations, contradictory results related to the extent of predicted pore pres-
5
sure build{up were obtained: some analysts predicted larger overall pore pressure
build{up by using stochastic than deterministic input parameters, others reported
lower values. However, all studies agreed with respect to the predicted pattern of soil
liquefaction.
1.3 Proposed approach

Direct Monte Carlo simulation and stochastic nite elements are the most widely used
methods for analyzing stochastic structural systems. The rst has the main drawback
of demanding large computational eort, due to sample size dependent accuracy
requirements imposed by statistical methods. On the other hand, the applicability of
stochastic nite elements is limited to linear, or almost linear, materials. Therefore,
the second method is not appropriate for materials which exhibit nonlinear behavior
under the normal loading range: metals (e.g. high performance alloys), composites
(e.g. reinforced concrete), geotechnical materials (e.g. brittle rocks, soils exhibiting
localization type behavior or dynamically induced pore pressure build{up).
The proposed approach is aimed at evaluating the eects of stochastic distri-
bution of soil properties on the system performance under dynamic loads. It uses
the Monte Carlo simulation method in the sense that digital simulations of stochas-
tic elds are combined with deterministic nite element analyses. With reference
to stochastic eld generation, the principal disadvantage of Monte Carlo simulation
methods (namely { large CPU time) is overcome by using an algorithm based on the
spectral representation method which takes advantage of the Fast Fourier Transform
technique for great computational eciency. The dierent soil properties over the
analysis domain are modeled as a multi-dimensional, multi-variate (nD-mV) stochas-
tic eld, in the sense that each component of the multi-variate eld will represent one
of the dierent soil properties (Popescu and Deodatis, 1995, Deodatis et.al., 1995).
If the spatial trends (mean values) of the various geomechanical soil properties mea-
sured at dierent locations over the analysis region are removed, and if the resulting
6
random
uctuations around these spatial trends are normalized, then these random

uctuations can be modeled as a zero-mean, unit-standard-deviation, homogeneous
vector eld. The resulted sample functions of the simulated eld are then properly
transformed to represent stochastic input parameters for the nite element analyses.
Four steps are basically required for the analysis:
{ estimation of the statistics of spatial variability (spatial trends, spatially dependent
variance, probability distribution functions, correlation structure, spatial cross{
correlations between various soil properties) based on results of extensive eld
measurement programs
{ digital generation of an nD{mV, non{Gaussian stochastic eld, each simulated
sample function representing a possible realization of soil property values over
the analysis domain
{ evaluation of soil constitutive model parameters using correlations with in{situ soil
test results
{ deterministic nite element analyses using stochastic parameter input derived from
each sample eld of soil properties a sucient number of nite element simu-
lations has to be performed to derive the statistics of the response.
Whether the vector eld is expressed in terms of soil constitutive parameters
required by the nite element model, or in terms of relevant eld measurement re-
sults makes the dierence between two methods, called here \direct simulation" and
\indirect simulation" method, respectively.
In the \direct" approach, illustrated in Figure 1.1, the required soil constitutive
parameters are rst evaluated from correlations with in{situ test results at each eld
test location, and then used for the random eld analysis. The sample functions of
the vector eld are expressed directly in terms of soil constitutive model parameters
(e.g. friction angle at failure, low strain moduli, permeability, etc.).
In the \indirect" approach, schematically represented in the
ow chart in Fig-
ure 1.2, the random eld analysis is performed in terms of in{situ measurement
7
Figure 1.1: Flowchart for estimating the characteristics of spatial distribution of soil
properties based on field measurement results – “direct simulation” approach
(exemplification for two-dimensional analysis).
8
results, and the sample functions of the vector eld are expressed in terms of relevant
eld test parameters (e.g. NSPT blow{count, cone tip resistance, etc.). The consti-
tutive model parameters required for nite element analyses are evaluated for each
realization of the stochastic eld at the nite element centroids.
Though the \direct simulation" method apparently involves less computational
eort, the current lack of data describing the cross{correlation structure of soil prop-
erties might lead to estimations which would depend too much on the selection of
empirical correlations with eld test results. Therefore, the \indirect simulation"
method, deemed as more robust, is used for all the numerical examples presented in
this study.
According to the
ow chart in Figure 1.2, the subsequent text is organized into four
chapters, each chapter dealing with one of the main steps of the proposed stochastic
analysis. The theoretical developments are illustrated by numerical examples based
on real eld data (cone penetration test results from an extensive soil testing program
performed for the articial islands in the Canadian Beaufort Sea { Gulf Canada
Resources Inc., 1984 { and standard penetration test results from a series of loggings
performed at Akita Harbour { Iai, 1989). The eects of spatial variability of soil
properties on system behavior and the details of the stochastic input nite element
analysis set{up are addressed in the last chapter in terms of behavior of saturated
soil deposits under seismic loads. The nite element analyses are performed using the
multi{yield plasticity soil model (Prevost, 1985, 1989), implemented in the computer
code DYNAFLOW (Prevost, 1981).
The analysis methods and numerical examples presented throughout this work
are specic to dynamic behavior of cohesionless soils. However, it is mentioned here
that the procedures for data analysis and for digital simulation of stochastic elds
developed in Chapters 2 and 3, respectively, are general and applicable to any type
of materials which exhibit stochastic variability of their properties.
9
Figure 1.2: Flowchart for Monte Carlo analysis – “indirect simulation” approach.
10
Chapter 2
Field Data Analysis
2.1 Raw data transformations
A general characteristic of common soil test results is that they are in general inad-
equate for statistical analyses. This is because the distances between measurements
are of the same order, or even larger than the expected correlation distance , while
the optimal sampling distance is x = =2 (Vanmarcke, 1983). This is particularly
manifested in horizontal direction, where the distances between boreholes are usually
of the order of tens of meters, but also in vertical direction in case of SPT data, where
typical sampling distances are z = 1 : : : 1:5m. To obtain a fair estimate of statistical
properties, methods which are considered appropriate for this situation are selected
from the parameter estimation methods presented in literature (e.g. Vanmarcke,
1983, Bendat and Piersol, 1986, 1993).
The proposed methodology for eld data analysis, leading to estimation of the
spatial variability statistics for various material properties, comprises:
1. eld data transformations (data standardization to ensure homogeneity)
2. evaluation of the probability distributions
3. estimation of the correlation structure
4. estimation of cross{correlations between various soil properties.
The respective procedures are described hereafter, and are illustrated with numerical
examples using real eld measurement results from two sites where appropriate data
11
(CPT and SPT results) are available: (1) an articial island made of undensied
hydraulically placed sand, and (2) a quaternary deposit of medium to dense sand.
2.1.1 Field data
The rst location, called Tarsiut P{45, is one of thirty six articial islands con-
structed in the Canadian Beaufort Shelf during the period 1972{1987 and used as
drilling platforms for oil explorations in shallow waters (Jeeries et.al., 1988a). The
island at Tarsiut P{45 is of caisson retained type and consists of a monolithic steel
caisson structure (Molikpaq) founded on an underwater sand berm and lled with
hydraulically placed sand (core) { Jeeries et.al. (1985). A total of 32 piezocone
tests were carried out for core verication (Gulf Canada Resources Inc., 1984) and
a proportion of these were placed on a straight line, at 9 m distance between cen-
ters (proles MAC04, : : : MAC08 { Figure 2.1). Three other CPT proles (MAC31,
: : : MAC33) were performed after a few months to assess the eects of time on cone
tip resistance. There are two pairs of proles (MAC05{MAC32 and MAC08{MAC33)
placed at 1m center{to{center relative distance. The layout of the selected CPT pro-
les, and recorded cone tip resistances are presented in Figure 2.1. Two \statistically
homogeneous" soil layers are analyzed: the core ll, from 0 to {21m, and the sand
berm from {21 to {30m (the elevation is measured from the ground level, and the
ground{water table is at {1.6 : : : {1.7m). Due to uncertainty in locating the bound-
ary between hydraulic ll and sand berm, which is reected in larger variation of the
cone resistance values, as shown in Figure 2.3, the stochastic analysis of CPT results
is performed as follows: layer 1 between {1.6, : : : {19.6m, and layer 2 between {22
: : : {30m. Data for the prole MAC{04, which was not available in electronic format,
was obtained by digitization from Jeeries et.al., 1988b. This data was only used to
estimate the correlation structure in horizotal direction (x2.3.4).
The second site is located at Akita Harbour (Japan), which was hit by the
Nihonkai{Chubu Earthquake (May, 1983, Magnitude 7.7), and extensive damage due
12
9m 9m 9m 9m 9m
0
MAC31 MAC04 MAC05 MAC06 MAC07 MAC08
& MAC32 & MAC33
-2
-4
-6
-8
elevation (m)
-10
-12
-14
-16
-18
-20
-22
0 10 20 0 10 20 0 10 20 0 10 20 0 10 20 0 10 20
qc (MPa)
Figure 2.1: CPT data from the hydraulic ll at Tarsiut P{45 (Gulf Canada Resources
Inc., 1984).
5
borehole 44-2 borehole 44-5 borehole 58-22

0
0
elevation (m)
-5
-5
-5
-10
-10
-10
-15
-15
-15
(x = 0.0m) (x = 35.0m) (x = 70.0m)

-20
-20
-20
0 25 50 0 25 50 0 25 50
NSPT (blows/foot)
Figure 2.2: SPT data from Akita Harbour { Ohama No. 1 Wharf (Iai, 1991).
13
to soil liquefaction was reported. The specic location selected for the present anal-
ysis is Ohama No. 1 Wharf, where three borings with consistent SPT measurements
are available (Figure 2.2, data from Iai, 1991). The horizontal distance between
borings is about 35m. The soil consists of quaternary deposits of medium to dense
sands and did not liquefy during the earthquake. Two layers with slightly dierent
geomechanical properties are identied: a medium grain size sand (D50 0:6mm),
underlain by a ne sand layer, with D50 0:15mm.
2.1.2 Filtering and shortening
The common parameters resulted from cone penetration tests are: (1) directly mea-
sured cone resistance { qc, (2) friction ratio { Fr = fs=qc, where fs is the recorded
sleeve friction, and (3) pore pressure ratio { Bq = u=(qT ; v0) { in the case of
0
piezocone test. The CPT readings usually show a fairly wide scatter. Correct iden-
tication of variations due to small inhomogeneities in the ground from those due
to soil sublayers is essential for a correct interpretation of test results. Continuous
CPT proles (data recorded at 5 : : : 15mm) provide sucient amount of data to allow
evaluation of representative values. However, not all readings should be regarded as
\independent" results, due to the nature of the bearing capacity problem represented
by cone penetration.
An eight{step procedure for ltering and smoothing raw CPT data (qc and Fr)
is described by Harder and von Bloh (1988). Though the procedure is designed for
identication of stratication and soil type, some steps are useful for transformation
of cone resistance results used to estimate the soil properties: (1) data ltering, which
consists in eliminating values falling outside a certain interval dened by mean and
standard deviation, and (2) data shortening, basically an averaging of CPT data over
intervals of the order of 5 : : : 10cm in depth.
Data shortening over intervals z = 10cm is used here for the cone tip resistance
14
recordings. For each CPT prole:

1 XN j
qc(zm j ) = N qc(zi) (2.1)
j i=1
X
where: { Nj = 1(z ; z <z ), 1(A) = 1 if A is true and 1(A) = 0 otherwise]
(j 1)
i
(j )
all data
{ z(j) = j z, j = 1 : : : number of intervals z in the CPT prole
{ zm j = (z(j 1) + z(j))=2
;
The size of averaging interval z is selected to account for the nature of cone tip
resistance measurements: independent results are obtained at distances greater than
two cone diameters (i.e. greater than about 7cm).
In the case of standard penetration test results, there are no such issues as ltering
or shortening since data are already obtained as averages of soil penetration resistance
over a 30.5cm (1ft) distance.
2.1.3 Standardization
Necessary analysis procedures for evaluating statistics of non{stationary data are
substantially more dicult than those for stationary data. Consequently, data non{
stationarity has to be handled within each \statistically homogeneous" soil layer.
Soil properties are markedly inuenced by eective conning stress, therefore trends
(changes in average values) are expected in vertical direction. The vertical non{
stationarity of soil properties can be solved in dierent ways:
1. By subdividing each \homogeneous" soil stratum into horizontal sublayers within
which the properties can be considered approximately stationary (e.g. Jeeries,
1989, subdivides soil strata into 1m thick horizontal sublayers).
2. By expressing the measured values u(x) as sums of spatial trends uav (x) and
residuals (or uctuations about spatial trends) u(x) = uav (x)+ uf (x) (Vanmar-
cke, 1977, DeGroot and Baecher, 1993), where x represents the spatial location.
The residuals uf (x) are further normalized by the sample standard deviation
15
of eld data su(x). Zero{mean, unit{variance, homogeneous elds are obtained

in this manner. For the 2D case, and assuming stationarity of eld test re-
sults in horizontal direction x i.e. uav (x z) = uav (z) and su(x z) = su (z)], the
normalized uctuations are expressed as:
us(x z) = u(x z) ; uav (z)]=su(z) (2.2)
where the trend in vertical direction uav (z) is obtained by regression analysis.
In the rst approach, working with total values is more advantageous when study-
ing the probability distribution of soil properties. The advantages of the second ap-
proach are: (1) the analysis can be performed over a signicantly larger domain,
which leads to: (1a) larger sample size, and (1b) simplications in processing the
analysis results and (2) stochastic analyses of standardized data are usually easier
to perform. The second approach, which is selected in this work, is illustrated in
Figure 2.5 (after Popescu et.al., 1995), based on standard penetration test results
from three borings at Akita Harbour (Figure 2.2).
In the case of CPT data (Figure 2.1), the uctuations about spatial trends for
each CPT prole are:
qf i(z) = qc i(z) ; qav (z) (2.3)
where i = 1 : : : Np with Np = 7 { the number of CPT proles, and the expressions
for qav(z) in core ll and sand berm, evaluated by linear regression, are shown in
Figure 2.3. The sample standard deviation sq (z) is estimated using information from
all proles. A pointwise evaluation:
1 X
N p
sq (zk ) = N ; 1 qf2 i(zk ) zk ; zk 1 = z

2
; (2.4)
p i=1
(where z = 0:1m for the CPT data, x2.1.2) is deemed as inappropriate, since uses
samples which are too small and results in too much variability (dotted lines in
Figure 2.4). An averaging procedure involving larger samples is proposed:
16
0
-1.6m dry
saturated
MAC05
-5
MAC06
MAC07
MAC08
-10
elevation - z(m)
av
q (z) = 2.35 - 0.37 z MAC31
MAC32
-15
MAC33
-20
LAYER 1
LAYER 2
-25
av
q (z) = 6.16 - 0.44 z
-30
0 5 10 15 20 25 30 35 40
qc (MPa)
Figure 2.3: Evaluation of spatial trends { qcav(z) { for the cone tip resistance values
recorded at Tarsiut P45.
0
a. fixed intervals b. moving intervals

-5
-5
∆h = 0.1m ∆h = 0.1m
elevation - z(m)
∆h = 1.0m ∆h = 1.0m
∆h = 2.0m ∆h = 2.0m
-10
-10
-15
-15
-20
-20
0 2 4 6 8 0 2 4 6 8
sq(z) sq(z)
Figure 2.4: Evaluation of depth dependent standard deviation of uctuations about

spatial trends for the hydraulic ll (layer 1), computed over: a. xed intervals h,
and b. moving intervals h.
17
1. Over xed horizontal sublayers, h (Figure 2.4,a):

X
N NX=2
s2q (zm) = N N1 ; 1
p q
qf i(zm + j z)]2 zm ; zm 1 = h (2.5)
;
p q i=1 j = Nq =2
;
where:
{ Nq is the number of measurements in each prole, within the averaging in-
terval h: Nq = h=z, with z = 0:1m (x2.1.2)
{ zm is the elevation at the center of sublayer m.
2. Over \moving" sublayers h (Figure 2.4,b), using the same expression (2.5),
but computing sq (zk ) at all locations, i.e. using zk from Eqn. (2.4) rather than
zm from Eqn. (2.5).
Sample standard deviation of cone resistance uctuations is computed using both

methods, for various values of h (h = z = 0:1m, h = 1m and h = 2m)
and presented in Figure 2.4. The second method, with h = 1m, is selected for the
analysis of CPT data to obtain standardized values qs(x z), and was also employed
for standardization of the SPT results (Figure 2.5 { Popescu et.al., 1995), using a
value h = 7m.
2.1.4 Test for stationarity

The standardized cone penetration results qs(x z) { Eqn. (2.2) { are tested for ho-
mogeneity using a method proposed by Bendat and Piersol (1986). Data is divided
into N equal intervals in vertical direction, and mean square values for each interval
k and each CPT prole j are computed as:
XN
qj2 k = N1 qs2 xj (k + i ; 1)z] j = 1 : : :Np k = 1 : : :N
1
(2.6)
1 i=1
where N1 = Nq =N is the number of data values in each interval, and Nq is the total
number of qs values in each CPT prole. The resulted values are tested in vertical
direction (for each CPT prole) and in horizontal direction (for N elevations) using
18
Average values of Standardized

field data - Nav(z) Standard deviation - sN(z) data - Ns(x,z)
5
average fluctuations about sN b. 44-2
av b. 44-5
values (N ) average values
b. 58-22
0
0
elevation - z (m)
-5
-5
-5
-10
-10
-10
-15
-15
-15
b. 44-2 b. 44-2
b. 44-5 b. 44-5
b. 58-22 b. 58-22
-20
-20
-20
0 10 20 30 40 50 -25 -15 -5 5 15 25 5 10 15 -3 -2 -1 0 1 2 3
av av
NSPT (blows/foot) NSPT - N sN(z) Ns = (NSPT - N ) / sN
Figure 2.5: Field data standardization for NSPT values recorded at Akita Harbour
(after Popescu et.al., 1995).
the reverse arrangements test (Bendat and Piersol, 1986). The results for the CPT
data, tested at the 5% level of signicance, are as follows:
Layer 1 (hydraulic ll core) { N = 9 N1 = 20:
{ in vertical direction (Np = 8) { all data passed the test

{ in horizontal direction { only the proles situated at 9m center{to{center
distance (Figure 2.1, proles MAC31, MAC04, : : : MAC08) are considered,
and therefore Np = 6 78% of data passed the stationarity test (there is too
much variation in mean square values between elevations {5.6 : : : {9.6m).
Layer 2 (sand berm) { N = 4 N1 = 10 Np = 4: all data passed the test in both
vertical and horizontal direction.
19
2.2 Analytical models

2.2.1 Probability distribution functions
A theoretical model for simulating the probability distribution of soil properties has
to be:
1. exible, i.e. to have a wide diversity of distribution shapes

2. bounded
3. easy to use in automate data analysis.
The theoretical models selected for tting empirical distributions of experimental

data are divided in: (1) common distributions (Gaussian, Lognormal, Uniform, etc.),
and (2) a two parameter distribution which is more exible and, therefore, more
appropriate for curve tting. Beta distribution function is selected for that purpose.
It permits the representation of a wide variety of distributional shapes, for random
variables whose values are limited to a nite interval. Its distribution function over
the interval (0,1):
8 0
>> x<0
< ;(p + q) Z x p 1
F (x p q) = > ;(p);(q) 0 t (1 ; t)q 1dt 0 x 1
; ;
(2.7)
:> 1 x>1
known as the incomplete beta function, is tabulated, and algorithms and computer
programs are available for numerical solutions (IMSL, 1987b).
As shown in x2.1.3, the eld data is standardized, i.e. zero{mean unit{variance
values are used in subsequent analyzes. The expression (2.7) of Beta distribution
function becomes in this case:
8 0 x<a
>>< Z ;;
F (x p q) = > ;( p + q ) x a
tp 1(1 ; t)q 1dt a x b

b a ; ;
(2.8)
>: ;( p);(q ) 0
1 x>b
20
where (a b) is the new interval of variation. The values a and b can be derived from
the condition of zero{mean, unit{variance:
8> p
< X = a + p + q (b ; a) = 0
>: X2 = (b ; a)2 pq (2.9)
(p + q) (p + q + 1)
2
= 1
The admissible solutions for the interval of variation are:
s s
a=; p (p + q + 1) b = q(p + q + 1) (2.10)
q p
The proposed parameter estimation procedure, for evaluating p and q which will
best t a particular empirical distribution, is described in x2.3.1. However, as ex-
plained in x2.3.1, a \pre{estimation" is usually needed for the initial parameter values
to be used in the nonlinear regression algorithm. In the case of Beta distribution,
due to complicated analytical form, maximum likelihood estimates of parameters are
dicult to obtain (Hahn and Shapiro, 1994). Even the method of moments can-
not be used for the case at hand, since the rst two moments are used to evaluate
the variation interval { Eqn. (2.9) { and estimates using higher order moments are
inaccurate.
For suciently large samples, a variation of the method of moments, based on the
extreme values found in the data, is proposed. Given a sample of zero{mean unit{
variance (standardized) experimental data fXig i = 1 : : : N , let X(1) X(2) : : : X(N )
be the sample ordered in ascending order (sample statistics). The bounds a and b of
the variation interval are estimated as:
a^ = 2X(1) ; X(2) and ^b = 2X(N ) ; X(N ; 1) (2.11)
Then, using Eqn. (2.10), the initial estimates p^ and q^ of the distribution parameters
can be computed as:
^ ^
q^ = b(1^ ; a^b) and p^ = ; a^^ q^ (2.12)
b ; a^ b
The initial estimates of p^ and q^ for the cone resistance data in the hydraulic ll
layer are marked with a crossed circle in Figure 2.7,c.
21
2.2.2 1D correlation structures and SDF

The random spatial variability of a homogeneous, scalar eld u(x) is characterized
by the covariance function:
Cuu ( ) = E u(x)u(x + )] ; u(x)]2 (2.13)
or the auto{correlation function:
Ruu ( ) = E u(x)u(x + )] (2.14)
where is the separation distance (space lag vector) and u(x) is the average value
of u(x), which for homogeneous elds is the same for all locations: u(x) = u(x + ).
The corresponding non{dimensional quantity is the correlation function (referred in
some textbooks as correlation coe cient function):
uu ( ) = Cuu2( ) (2.15)
u
In case of zero{mean unit{variance elds, Eqns. (2.13), (2.14) and (2.15) simplify
to:
uu ( ) = Cuu( ) = Ruu ( ) = E u(x)u(x + )] (2.16)
Accounting for the mechanisms of soil deposit formation, leading to dierent spa-
tial variability characteristics in vertical direction (normal to soil strata) as compared
to those in horizontal direction, separable correlation structure models (Vanmarcke,
1983) seem appropriate to simulate spatial variability of soil properties. Moreover,
these models are easier to handle in multi{dimensional analyses. Under the separable
correlation structure assumption, the correlation function is expressed as:
Y
n
uu ( ) = i(i ) (2.17)
i=1
where the subscript u has been dropped, n is the number of spatial dimensions, and
i represents the correlation function in spatial direction i, which is function of the
separation distance i .
22
Table 2.1: Common one{dimensional correlation structures

Correlation Correlation Correlation Normalized
structure function () distance SDF s( )
e() ap
pa
e ( )
2 a 2
Squared ;
a ;
2
exponential (
1 ; a jj a j j
a a sin2 a2
Triangular 0 otherwise 2
2
a2
Exponential e a
; j j
(a > 0) 2 a
a
(a + 2)
2
2b h 1 i
Cosine e b cos a
;
b + a2
2
b
(+a) +b
2 2+ 1
( a) + b
; 2 2
decaying
In the case of separable correlation structure, the n{dimensional spectral den-

sity function1 and the variance function can also be expressed as products of one{
dimensional functions (Vanmarcke, 1983):
Yn Y
n
suu () = si( i) (A) = i(Ai) (2.18)
i=1 i=1
where is the wave{number vector, and A is the n{dimensional averaging domain.
The same stands for the correlation domain, which is expressed as a product of
correlation distances in each spatial direction:
Yn
= i (2.19)
i=1
The proposed methodology for correlation structure parameter estimation is based
on a nonlinear regression procedure. As described in x2.3.1, a theoretical model for
the correlation structure has to be specied rst. Several common correlation models,
listed in Table 2.1, are used for that purpose (Vanmarcke, 1983).
The Triangular correlation structure is easyer to t to sample functions obtained
from scarce data (e.g. Popescu et.al., 1995). The Squared Exponential correlation
1This is actually the normalized spectral density function: suu ( ) = Suu ( )=u2 , which will be
used thereafter.
23
structure has the advantage that, when expressed in more than one dimension, it
becomes isotropic as well as separable (providing that the same parameter values are
used for all spatial dimensions). However, correlation functions depending on more
than one parameter can cover a larger palette of shapes and, hence, can be better
tted to sample functions obtained from eld data. Such a model is derived from
the Exponential Decaying (ED) spectral density function, discussed among others by
Shinozuka and Deodatis (1988b):
sED ( ) = 2;(b 1 + 1) bb1 +1 b e b b1 b2 > 0
2 2 ; 1j j
(2.20)
2
The expression for the corresponding correlation function is derived using the
Wiener{Khinchine relations in the form of Fourier cosine transform (Vanmarcke,
1983):
Z h i
cos (b2 + 1) tan 1 b ;
ED () = 2 s( ) cos d = (2.21)

1
1
b2 +1
0 2
1 + b
2
The Exponential Decaying SDF has the advantages that: (1) has a very exible shape
(Figure 2.6,a,b) and hence is easy to t to real data, and (2) s(0) = 0 and therefore
is more appropriate for the digital simulation procedure presented in Ch. 3.
The sucient condition for existence of correlation distance (Vanmarcke, 1983):
b +1
b
g_ (0) = ;(b + 1) e (b2 ; b1 ) = 0
1
2
b 1 b
2; ; 1
(2.22)
2 =0
is veried for b2 > 1. However, the necessary condition for to exist (Vanmarcke,
1983):
lim 1 Z ()d = 0 (2.23)
!1 " 0
has been veried numerically for values of the parameter b2 which are less than one
(0:3 < b2 1). Therefore it is concluded that the exponential decaying SDF model
has nite correlation distance for b1 > 0 and b2 > 0:3.
None of the classical expressions for evaluation of correlation distance (Vanmar-
cke, 1983) are appropriate for this model. Shinozuka and Deodatis (1988b) use an
24
1
a. Correlation
functions b1 = 0.5 b1 = 2.0
0.5
0.5
ρ(ξ)
0
0
-0.5
-0.5
0 2 0 2 4
1
b1 = 5.0 b2 = 0.5
b2 = 2.0
0.5
b2 = 5.0
ρ(ξ)
0
-0.5
0 2 4 6 8 10
space lag ξ
b. Spectral density functions
0.2
b1 = 0.5
s(κ)
0
0 5 10 15 20
1.2
0.5
b1 = 5.0 b1 = 2.0 b2 = 0.5

b2 = 2.0
b2 = 5.0
1
0.25
s(κ)
0.8
0 5 10
wave number κ
0.6
s(κ)
c. Correlation distance
2.5
values from numerical experiment

0.4
θ/b1 = 1.04 - 1.3 log b2 (0.1 < b2 < 2)

2
1.5
θ / b1
0.2
1
0.5
parameter b2
0
0 5
2 4 6 8 2 4 6 8 2 4 6 8
wave number κ 10-1 100 101 102
Figure 2.6: Exponential decaying SDF model: a. correlation functions, and b. spectral
density functions plotted for various values of the parameters b1 and b2 c. correlation
distance values.
25
expression based on the root{mean{square frequency #2 (Harada and Shinozuka,

1986):
p Z p
=# = 2 s( )d = q
2
1 ;
1
2b 1 2
(2.24)
2 0 (b2 + 2)(b2 + 3)
This leads to values of the correlation distance equal to the space lag corresponding
to a value of the correlation function of about 0.3 : : : 0.5:
j()=0:3:::0:5 (2.25)
which is close to the value of correlation distance as dened by DeGroot and Baecher
(1993), but smaller than the one dened by Vanmarcke (1983) and assumed through-
out this work. Another expression providing results which are more consistent with
Vanmarcke's denition is proposed as:
Z
= ()d (2.26)
1
where 1 = b1 tan 2(b

+ 1) is the point where the correlation function rst crosses
2
the space lag axis ((1) = 0). The results of a parametric study performed for values
0:1 b1 10 0:1 b2 10 indicated that the expression (2.26) gives values which
are 1.5 : : : 2 times larger than those obtained using Eqn. (2.24). However, for the
considered domain of parameter values, the results of Eqn. (2.26) are in a 0.8 : : : 1.1
range of the expression derived by Vanmarcke (1977) for the squared exponential
model:
s
2d (2.27)

u
where du is the average distance between mean crossings. du was numerically eval-
uated from 1D Gaussian stochastic elds simulated as shown in x3.2 for each corre-
sponding set of parameters b1 and b2.
The variation of the correlation distance with parameters b1 and b2 is presented
in Figure 2.6,c. For a certain range of b2 values, the expression (2.26) can be approx-
imated by:
b1 1:04 ; 1:3 log b2 0:1 b2 2 (2.28)
26
Other useful expressions for the correlation structure model can be obtained by
combining simpler models. For example, a Two{Peak (TP) SDF can be formulated
as a sum of two Exponential Decaying spectral density functions:
sTP ( b1 b2 b3 b4 b5) = b5sED ( b1 b2) + (1 ; b5)sED( b3 b4) (2.29)
where sED( bi bj ) is the expression of Exponential Decaying SDF { Eqn. (2.20) {
and bi bj are b1 and b2, respectively (with i j = 1 2 or i j = 3 4). The parameter
b5 (0 < b5 < 1) is a weighting quantity. The expression of the correlation function is
obtained using the Wiener{Khinchine relations { Eqn. (2.21), as:
TP ( b1 b2 b3 b4 b5) = b5ED ( b1 b2) + (1 ; b5)ED ( b3 b4) (2.30)
2.3 Estimation of spatial variability statistics

2.3.1 Proposed curve tting procedure
Given a set of N experimental data fxi yig i = 1 : : : N , the problem is to nd a
theoretical expression g(x b1 b2 : : : bm), depending on m parameters bj , which best
ts the experimental values. The parameters bj are estimated by minimizing the
sum of squared deviations of theoretical values g(xi) from the experimental values
yi. Selection between various analytical models is also based on the minimum sum of
square errors. First, an expression g(x b), b = (b1 b2 : : :bm), is selected. For exam-
ple, for the data shown with dotted line in Figure 2.7,b, and representing an empirical
distribution function, the selected theoretical expression is given in Eqn. (2.8), with
a and b computed according to (2.10), and parameters b1 = p and b2 = q. This
corresponds to the Beta distribution function for zero{mean, unit{variance random
variables (x2.2.1).
The sum of squared deviations of the theoretical values g(xi b) from the experi-
mental data yi (in short { residuals) is expressed as:
X
N
f (b) = yi ; g(xi b)]2 (2.31)
i=1
27
The optimal set of parameter values b^ correspond to the minimum residuals or

least sum{square{error (least{square{method):
X N
" = f (b^ ) = min yi ; g(xi b)]2
b i=1
(2.32)
or
@f (b) = 0 j = 1 : : : m (2.33)
@bj
Eqn. 2.32 represents a nonlinear regression problem which, excepting for polyno-
mial expressions of g(x b) and a few other simple analytical models, does not have
analytical solutions. Therefore, a numerical procedure has to be used to solve 2.33.
The selected procedure for nding the minimum of a function f (b), depending on
m variables b = (b1 b2 : : :bm) (Fletcher, 1972), is implemented in the IMSL library
as subroutine ZXMIN (IMSL, 1987a). The algorithm is based on classical quasi{
Newton methods for solving minimization problems (Gill and Murray, 1972, Sen and
Srivastava, 1990), and assumes the existence of the gradient vector rf (b) and the
Hessian matrix @ 2f (b)=(@bi @bj ). Providing that f (b) is a combination of analytical
functions g(xi b), the above assumptions are true. However, since the functional ex-
pression of the gradient is in general dicult to evaluate, the derivatives @f (b)=@bj
are approximated by nite dierences.
A computer program using the subroutine ZXMIN has been implemented to solve
the nonlinear regression problem (2.32). Various analytical expressions g(x b) for cu-
mulative distribution functions, correlation functions, and spectral density functions
are available. New models can be added to the existing library with minimum changes
in the program. The proposed parameter estimation procedure has the following
steps:
1. Select an expression for g(x b) from the theoretical model library.
2. Input the initial values (initial guess) of the parameters: b(0) = b(0)
1 : : :bm .
(0)
28
3. Solve the nonlinear regression problem (2.32), and nd the best estimates b^
and the residuals:
XhN i2
" = f (b^ ) = yi ; g(xi b^ ) (2.34)
i=1
4. Go to step 1 and select another analytical model g(x b). The optimum solution,
corresponding to the available analytical model library, is achieved for the model
which results in the minimum residuals in Eqn. (2.34).
Remark: When the objective function f (b) is very at, as illustrated by the
contour plot in Figure 2.7,c, or has multiple local minima, the initial guess of the
parameter values (step 2) can be very important in achieving the global minimum.
In such cases, the initial values b(0) can be found by either: (1) performing an m{
dimensional search in the parameter space on a coarse grid, or (2) pre{estimating the
parameters using another method.
2.3.2 Probability distribution function

Given a sample of size N , fxk gk=1 :::N , the empirical distribution function FX (x) is
expressed for values N 1+ 1 FX (x) N N+ 1 (Yamazaki and Shinozuka, 1988), as:
FX (x(k)) = N k+ 1 k = 1 : : :N (2.35)
where x(k) is the kth term in the order statistics of fxk g, and linear interpolation is
employed for values x(k) < x < x(k+1) k = 1 : : :N ; 1. Empirical distributions of
standardized eld test results from the hydraulic ll (Tarsiut P{45 { layer 1), sand
berm (Tarsiut P{45 { layer 2), and Akita Harbour are shown with dotted lines in
Figures 2.7, 2.8 and 2.9, respectively.
The data from Tarsiut P{45 follow skewed, non{Gaussian distributions (the sam-
ple skewness coecients are ^1 = 0:65 for layer 1, and ^1 = ;0:56 for layer 2). The
method described in x2.3.1 is employed to t a Beta distribution to the empirical
29
Figure 2.7: Probability distribution of standardized cone tip resistance values recorded in
the hydraulic fill (layer 1). Comparison of empirical distribution with: a. common
probability models, and b. fitted Beta distribution; c. illustration of the solution search in
the parameter space.
distributions evaluated from field data. First, the initial parameter values are estimated
using the variation of the Method of Moments described in §2.2.1, Eqns. (2.11) and
(2.12). Next, the optimum values of parameters p and q are evaluated by the least square
method using the quasi-Newton minimization procedure. The resulted values are shown
in Table 2.2. The main steps of the iterative search for the optimum solution performed
for data from layer 1 are illustrated in Figure 2.7,c.
30
1
a. b.
0.5 Gaussian CDF
0.5
Uniform CDF
Lognormal CDF Beta CDF
Empirical CDF Empirical CDF
0
0
-4 -2 0 2 4 -4 -2 0 2 4
standardized cone resistance qn standardized cone resistance qn
Figure 2.8: Probability distribution of standardized cone resistance values recorded

in the sand berm (layer 2). Comparison of empirical distribution with: a. common
probability models, and b. tted Beta distribution.
The minimum sums of squared deviations are computed as:
N h
X i2
"B = min
pq yi ; FB (xi)
p q
(2.36)
i=1
where xi yi are the values describing the empirical distribution: xi is the standardized
i , and N is the sample size. F
eld data, yi = N +1 B is the cumulative distribution
p q
function of Beta distribution, Eqn. (2.8). The residuals "B obtained for Beta distribu-
tion are compared to similar values computed assuming Gaussian ("G ), Uniform ("U )
and Lognormal ("LN ) distributions. The comparison results are shown in Table 2.2
in terms of values which are normalized with respect to "G :
"An = ""A A = U LN or B (2.37)
G
Comparisons between empirical distributions of standardized CPT results and
theoretical models are presented in Figures 2.7 and 2.8.
The data from Akita Harbour is tested for normality using the nw2 test (Anderson
and Darling, 1952). This method is used instead of the more common Chi{square test
since only a small sample is available (N = 58), while Chi{square test requires larger
samples. The value nw2 = 0:035 shows that the test is passed for the 95% criterion,
i.e. there is less than 5% discrepancy (in the sense of the nw2 test) between the
empirical distribution and the Gaussian distribution function. Since the test shows
31
Table 2.2: Comparison between results using various probability models

Distribution functions
Field data Uniform Lognormal Beta
"Un "LNn "Bn p q
Standardized CPT results from 1.80 25.9 0.07 3.5 18.4
hydraulic ll (layer 1)
Standardized CPT results from 5.65 107.5 0.42 290.0 21.0
sand berm (layer 2)
Standardized SPT results from 2.45 66.1 0.51 5.6 3.3
Akita Harbour
very little discrepancy (usually a distribution assumption is not rejected if passing

the 10% criterion), the normal assumption is accepted for the standardized SPT
data from Akita Harbour (Popescu et.al., 1995). Comparisons between empirical
distribution and various theoretical models, including the tted Beta distribution,
are presented in Table 2.2 and in Figure 2.9.
1
a. b.
Gaussian CDF
0.5
0.5
Uniform CDF
Lognormal CDF Beta CDF
Empirical CDF Empirical CDF
0
-3 -1.5 0 1.5 3 -3 -1.5 0 1.5 3

standardized SPT resistance Nn standardized SPT resistance Nn
Figure 2.9: Probability distribution of standardized SPT blow{counts recorded at

Akita Harbour. Comparison of empirical distribution with: a. common probability
models, and b. tted Beta distribution.
2.3.3 Correlation structure in vertical direction

The sample correlation function of a series of N data values fuig i = 1 2 : : : N
sampled at equal intervals z from a stationary, zero{mean, unit{variance stochastic
32
process is computed as (Bendat and Piersol, 1986):

;r
NX
uu (rz) = N 1; r u(iz)u(i + r)z] r = 1 : : :m (2.38)
i=1
where rz = z is the lag, r is the lag number and m is the maximum lag number.
This estimate is unbiased and consistent (Bendat and Piersol, 1986).
Sample correlation functions of the standardized CPT records are evaluated using
Eqn. (2.38) for each of the CPT proles. The results are shown in Figure 2.10 for all
data, and Figure 2.11,a (dotted lines) for prole MAC07.
a. Hydraulic fill (layer 1) b. Sand berm (layer 2)
1
1
profiles MAC05, 06, 07, 08, 32, 33 profiles:
profile MAC31 (not included) MAC05, 06, 07, 08
0.5
0.5
0
0
-0.5
-0.5
0 0.5 1 1.5 2 2.5 3 0 0.4 0.8 1.2
space lag ξz space lag ξz
Figure 2.10: Sample correlation functions in vertical direction evaluated for standard-
ized CPT records at Tarsiut P{45.
The next step is tting the sample correlation functions to a theoretical model.
This is performed as shown in x2.3.1, by selecting various expressions from the the-
oretical model library and solving the nonlinear regression problem (2.32). The 1D
correlation models currently available in the theoretical model library are described
in x2.2.2, and are as follows:
1. Squared Exponential correlation structure (SE) { Table 2.1

2. Triangular correlation structure (TR) { Table 2.1
3. Exponential correlation structure (EX) { Table 2.1
4. Cosine Decaying correlation structure (CD) { Table 2.1
5. Correlation structure derived from Exponential Decaying SDF (ED) {
33
a. Sample correlation function and fitted analytical models

1
1
"noise" Exp. Decaying model
Cos. Decaying model
sample correlation fct.
0.5
0.5
soil variability
0
0
Two-Peak model
-0.5
sample correlation function -0.5
0 1 2 3 0 1 2 3
space lag - ξz(m) space lag - ξz(m)
b. Sample spectral density function and fitted analytical models
0.3
0.3
Two-Peak model Exp. Decaying model

sample SDF Cos. Decaying model
sample SDF
0.2
0.2
0.1
0.1
0
0 4 8 12 0 4 8 12
-1 -1
wave number - κ(m ) wave number - κ(m )
Figure 2.11: Sample and theoretical correlation functions (a.), and spectral density
functions (b.) in vertical direction, evaluated using data from CPT prole MAC07,
in the hydraulic ll layer.
34
Eqn. (2.21)
6. Correlation structure derived from Two{Peak SDF (TP) { Eqn. (2.30)
The rst three models, depending on one parameter, are in general dicult to t
to experimental data. They are used to verify the outcomes of the other models in
terms of correlation distances and measure of noise. Models 4. and 5. (CD and ED)
are more exible two{parameter models. The most exible is the Two{Peak SDF
model, depending on 5 parameters, and which can accommodate both small scale
and large scale variability. An example of experimental data tting is illustrated in
Figure 2.11,a, based on the sample correlation function evaluated for CPT prole
MAC07, layer 1.
The sample spectral density function evaluated from experimental data as shown
in x3.2.4 is compared in Figure 2.11,b with the corresponding theoretical SDF's. It
is mentioned that the sample SDF is not used here to evaluate correlation structure
parameters, since it is a biased estimate (Bendat and Piersol, 1986).
The random measurement error, or \noise", (Baecher, 1984 DeGroot and Baecher,
1993) can be identied by using one more parameter in the curve tting proce-
dure (x2.3.1), as shown hereafter. The standardized values us(z) are represented as
(Baecher, 1984):
us(z) = us s(z) + "(z) (2.39)
where us s(z) represents the real spatial distribution of the soil property (here { cone
resistance), and "(z) is a \noise" component resulted from random measurement
error2. The covariance function can be expressed as:
(s)( ) + C (")( )
Cuu () = Cuu (2.40)
uu
Since "(z) is random, its covariance function is a spike:

( 2
Cuu () = 0noise ifotherwise
( " ) =0 (2.41)
2In a study on spatial variability of undrained shear strength in a clay deposit, Soulie et.al.,
(1990) concluded that this \noise" is actually due to very small scale variations of soil properties.
35
2
where noise is the variance due to the random measurement error, and the variance
due to real spatial variability is s2 = 1 ; noise
2 . The variance 2 due to spatial
s
variability can be identied as follows (see also Figure 2.11,a): For any 6= 0, the
covariance function of the standardized random eld us s is written as:
(s)( b 2 ) = 2 ( b)
Cuu (2.42)
s s uu
where uu ( b) corresponds to one of the expressions in the theoretical model library
(x2.2.2). Consequently, by replacing uu ( b) by Cuu (s)( b 2 ) and solving only for
s
values > 0, s2 becomes one more parameter in the nonlinear regression analysis {
Eqn. (2.32).
The correlation structure of the standardized cone resistance recorded at Tarsiut
P{45 is evaluated separately for each soil layer (i.e. hydraulic ll { layer 1, and sand
berm { layer 2). A good t can be obtained for individual proles (Figure 2.11,a).
However, to enforce the homogeneity assumption, a unique correlation function is
evaluated for each soil layer, using information from all CPT proles to estimate
the parameters bi in Eqn. (2.33). The sample correlation function evaluated using
data from prole MAC31, and deemed as being signicantly dierent from the sam-
ple functions obtained from other proles (Figure 2.11,a), was not included in the
analysis.
The curve tting results, in terms of: correlation function parameters bi, variance
due to spatial variability s2, estimated correlation distances z , and residuals ", are
presented in Tables 2.3 and 2.4 for soil layers 1 and 2, respectively. The minimum
residuals resulted for each theoretical model are normalized with respect to "SE ob-
tained for the Squared Exponential correlation structure. For both soil layers, the
Two{Peak SDF correlation structure provided the best t with experimental data.
The resulted optimum solutions are plotted in Figure 2.12 (continuous lines), along
with sample correlation functions and SDF's (dotted lines).
Consistent values of correlation distance z and real spatial variability variance
s2 are obtained for layer 1 from all the theoretical models. The estimated random
36
Table 2.3: Results of correlation structure analysis in vertical direction { layer 1

Correlation Parameters Spatial Correlation Normalized
structure bi variab. s2 dist. z residuals "n
Squared 0.49 0.83 0.87 1.000
Exponential (SE)
Triangular (TR) 0.82 0.88 0.82 1.069
Exponential (EX) 2.22 1.00 0.90 0.911
Cosine 2.08, 0.98 1.00 0.79 0.875
Decaying (CD)
Exponential 0.49, 0.27 0.88 0.88 0.826
Decaying SDF (ED)
Two{Peak 1.28, 4.80 0.90 0.95 0.520
SDF (TP) 2.78, 1.98
0.46
Table 2.4: Results of correlation structure analysis in vertical direction { layer 2

Correlation Parameters Spatial Correlation Normalized
structure bi variab. s2 dist. z residuals "n
Squared 0.29 0.54 0.52 1.000
Exponential (SE)
Triangular (TR) 0.53 0.55 0.53 1.011
Exponential (EX) 5.24 0.86 0.38 0.962
Cosine 5.24, 0.05 0.86 0.38 0.962
Decaying (CD)
Exponential 0.16, 0.02 0.74 0.47 0.959
Decaying SDF (ED)
Two{Peak 5.96, 8.42 0.72 0.48 0.922
SDF (TP) 0.30, 1.45
0.35
37
a. Fitted correlation function and spectral density function - data from layer 1
1
0.3
Two-Peak model Two-Peak model
sample correlation fcts. sample SDF’s
0.5
0.2
0.1
0
-0.5
0
0 1 2 3 0 4 8 12
-1
space lag - ξz(m) wave number - κz(m )
b. Fitted 2 parameter (ED & CD), and 1 parameter models (SE, TR & EX) - layer 1
1
1
ED model SE model
CD model TR model
sample correlation fcts. EX model
0.5
0.5 sample correlation fcts.

0
0
-0.5
-0.5
0 1 2 3 0 1 2 3
space lag - ξz(m) space lag - ξz(m)
c. Fitted correlation function and spectral density function - data from layer 2
1
0.3
Two-Peak Two-Peak model

sample sample SDF’s
0.5
0.2
0.1
0
-0.5
0 0.4 0.8 1.2 0 5 10 15 20 25

-1
space lag - ξz(m) wave number - κz(m )
Figure 2.12: Correlation structure assessment in vertical direction for standardized

CPT data from Tarsiut P{45: a. best t for data from the hydraulic ll layer b. com-
parison between sample correlation functions and some theoretical models c. best t
for data from the sand berm.
38
measurement error noise2 10% is somehow higher than the values reported by
Kulhawy and Mayne, 1990 (5 : : : 10% standard deviation), and Jeeries and Davies,
1993 (6% error at the 50% condence level), estimated from observed repeatability of
CPT results. Larger and more inconsistent values estimated for random measurement
errors in soil layer 2 may be due to availability of a considerable smaller sample size.
The nal values, which will be used in further computations, correspond to the Two{
Peak SDF correlation structure, and are as follows:
{ layer 1: z = 0:95m s2 = 0:90

{ layer 2: z = 0:48m s2 = 0:72
The evaluation of correlation structure characteristics for the SPT data from
Akita Harbour is based on both the curve tting procedure (a triangular correlation
function is tted to the sample correlation function using the least square method),
and estimation of correlation distance using the sample variance reduction function
(Vanmarcke, 1983). An average value z = 1:3m resulted for the correlation distance
in vertical direction. The procedure is described by Popescu et.al. (1995).
2.3.4 Correlation structure in horizontal direction

The assemble of standardized eld measurement results is represented as a 2D stochas-
tic eld u(x z), with values sampled at abscissas xj j = 1 : : : Np (locations of CPT
proles in horizontal direction), and depths zk k = 1 : : : Nq , such that zk ; zk;1 =
z. Np is the number of CPT proles, and Nq is the number of measurements in
each prole. Under the assumption of separable correlation structure, the sample
correlation structure in horizontal direction, for a given space lag x, is computed as:
X N
N1 X
^uu (x) = N 1N
q
u(x1i zk )u(x2i zk ) (2.43)

1 q i=1 k=1
where jx2i ; x1ij = x, and N1 is the total number of pairs of CPT proles which are
situated at a relative horizontal distance equal to x. Eqn. (2.43) can be rearranged
39
as:
2 N
3
1 XN1 1 X
4 1 X N1
1i zk )u(x2i zk )5 =
q
^uu (x) = N u ( x N1 i=1 ^12i (2.44)

1 i=1 Nq k=1
where ^12i is the sample correlation coecient between records obtained from CPT
proles in pair i, situated at locations x1i and x2i. Since various numbers of pairs
N1 are available for various horizontal space lags x, to avoid uneven weighting the
nal averaging in Eqn. (2.44) is omitted, and all values ^12i are used for tting the
theoretical model, as shown in Figure 2.16.
The hydraulic ll at Tarsiut P{45 was discharged into the \Molikpaq" core from a
single spigot, located in a near central position (Jeeries et.al., 1985). This resulted
in a deposit with slightly inclined sand layers, forming a central angle{of{repose
cone with the apex at the spigot location (Jeeries, 1995). This fact can be veried
by computing the sample cross{correlation function for any pair of adjacent CPT
proles, with data from each prole qs(xi z) i = 1 2, considered as an 1D stochastic
eld in vertical direction:
NX;r
^12(rz) = N 1; r
q
qs(x1 kz)qsx2 (k + r)z] (2.45)

q k=1
where x1 and x2 are the locations of the CPT proles in horizontal direction. Such
sample cross{correlation functions, shown in Figure 2.13, are computed for the proles
MAC32 and MAC06, situated at a distance jx2 ; x1j = 8m, using both total cone
resistance values qc and standardized values qs. It can be inferred from the location
h = rz 6= 0 of cross{correlation function peak that the soil layers are slightly
inclined toward prole MAC06, with an average slope h=jx2 ; x1j 8 : : : 9%. The
slight inclination of the soil layers can make a considerable dierence in the evaluation
of correlation distances and assessment of horizontal continuity of loose sand pockets.
For the case illustrated in Figure 2.13, a negative correlation would be estimated
between standardized CPT measurements (dashed line) if perfectly horizontal layers
are considered.
40
1
cross-correlation between total values - qc
0.75
cross-correlation between standardized values - qs
qc
ρ32-06 (0.7) = 0.46
0.5
qs
ρ32-06 (0.8) = 0.34
ρ32-06(ξz)
0.25
0
-0.5 -0.25
-3 -2 -1 0 1 2 3
space lag - ξz(m)
Figure 2.13: Cross{correlation between CPT measurements from two adjacent proles
(MAC32 and MAC08) in the hydraulic ll layer.
To get a correct idea of the correlation distance in horizontal direction in the
context of separable correlation structure, the eect of slightly inclined soil layers
is removed when evaluating the correlation coecients ^12i in Eqn. (2.44). The hy-
draulic ll layer is divided into four horizontal sublayers, of 4.5m each. Within each
sublayer, sample cross{correlation functions are computed for all pairs of CPT proles
located within a reasonable distance (jx2 ; x1j 10m). By simultaneously analyzing
the resulted peak locations (see e.g. Figure 2.14, for sublayer 1), a \most probable"
scenario of sand layer deposition is obtained, as sown in Figure 2.15. The sample
correlation coecients in Eqn. (2.44) are evaluated from the sample cross{correlation
functions in Eqn. (2.45), as:
^12i = ^12i(hi) i = 1 : : :N1 (2.46)
where hi is the estimated shift in vertical direction between CPT proles in pair
i. The analysis is performed separately for each sublayer. The sample correlation
coecients computed between adjacent proles are shown in Figure 2.15, and the
values computed for all pairs of proles situated at relative distances within the
range x i = jx2i ; x1ij = 1 : : : 19m are plotted with markers in Figure 2.16,a.
Because of the variability of the correlation estimates, it is desirable to verify
that a nonzero value of the sample correlation coecient indeed reects the existence
41
Figure 2.14: Cross-correlation functions and estimated sample correlation coefficients

between standardized cone tip resistance values, qs, from adjacent profiles in sublayer 1.
Figure 2.15: Estimated soil layer disposition in the Molikpaq core (the median lines of
the four sublayers are represented by shaded areas).
42
of a statistically signicant correlation. This is accomplished using a test proposed

by Bendat and Piersol (1986). The test is based on the normality assumption of
h 1+^12 i
the random variable w = 21 ln 1;^ 12 (Guttman et.al., 1982). For a given level of
signicance and a number N of pairs of observed values (here N = 45 is the number
of measurements in a CPT prole within each sublayer), a minimum acceptable value
can be assessed:
( )
exp p 2 z = 2 ;1
min
N = ( 2z ) N ; 3 (2.47)
exp p = 2 +1
N ;3
where z is the standardized normal variable. Statistically signicant correlation be-
tween two proles can be expected if:
^12i min
N (2.48)
The value of min N , computed at = 5% level of signicance is 5% 45 = 0:293,
min
and is shown in Figure 2.16,a with a dashed line. Since about 60% of the values ^12i
estimated for proles at relative distances x = 8 : : : 10m are above the minimum
value min
N , there is reason to believe that signicant correlation exists between soil
properties at these distances.
The correlation structure for the hydraulic ll layer is estimated by nonlinear re-
gression, as described in x2.3.1, using the sample correlation coecient values from all
sublayers at the same time. The results, in terms of parameters bi of the correlation
models, correlation distances x, and normalized residuals "n are shown in Table 2.5.
The resulted best t model, corresponding to the Two{Peak SDF correlation struc-
ture, is shown in Figure 2.16,a in terms of correlation function and spectral density
function.
Very scarce data is available to estimate the correlation structure in horizontal
direction for the sand berm (only four proles { MAC05, 06, 07 and 08 { located
in a line at 9m center{to{center distance). The sample correlation coecients ^12i
computed for lag distances x = jx2i ; x1ij = 9m and 18m (shown with markers
43
Table 2.5: Results of correlation structure analysis in horizontal direction { layer 1

Correlation Parameters Correlation Normalized
structure bi dist. x residuals "n
Squared 8.24 14.6 1.000
Exponential (SE)
Triangular (TR) 12.7 12.7 0.976
Exponential (EX) 0.14 14.3 0.925
Cosine Decaying (CD) 0.15, 0.04 12.4 0.945
Exponential 6.80, 0.10 15.9 0.970
Decaying SDF (ED)
Two{Peak 19.8, 0.60 12.1 0.896
SDF (TP) 1.25, 0.43
0.47
a. Fitted correlation function and spectral density function - data from layer 1
1
2
fitted correlation function (Two-Peak) SDF (TP)
0.5
5% level of
significance
1
sublayer 1
0
sublayer 2
sublayer 3
sublayer 4
-0.5
0 5 10 15 20 0 1 2 3
-1
space lag - ξx(m) wave number - κx(m )
b. Fitted correlation function and spectral density function - data from layer 2
1
fitted correlation function (Two-Peak) SDF (TP)

0.5
5% level of
1
significance
0
sample correlation function

-0.5
0 5 10 15 20 0 1 2 3
-1
space lag - ξx(m) wave number - κx(m )
Figure 2.16: Correlation structure assessment in horizontal direction for standardized

CPT data from Tarsiut P{45: a. best t for data from the hydraulic ll layer b. best
t for data from the sand berm.
44
in Figure 2.16,b) are all below the minimum value min N required at the 5% level of
signicance. Therefore it can be inferred that the correlation distance in horizontal
direction in the sand berm is x < 9m. In this situation (sampling distance greater
than assumed correlation distance) no consistent estimation of the correlation struc-
ture can be performed. The correlation function shown in Figure 2.16,b is estimated
by similarity with the Two{Peak SDF correlation structure obtained for layer 1, and
using the available values of sample correlation coecients. The resulted parameters
are: b1 = 9:1 b2 = 0:38 b3 = 1:9 b4 = 1:1 b5 = 0:36, and x = 5:4m.
Remarks:
1. The parameters of the correlation structures in horizontal direction are estimated
considering the same level of random measurement error as evaluated from the anal-
ysis in vertical direction (x2.3.3).
2. The results obtained from analysis of eld data are checked by direct (visual)
comparison with in{situ spatial distribution. Contours of normalized uctuations of
cone tip resistance measured in the eld are compared in Plate 1 with similar values
resulted from digital simulation. The numerical simulation results are obtained using
the algorithm presented in Chapter 3. Two realizations of a sample eld representing
possible distribution of standardized cone resistance values, and having the same
statistics as derived from analysis of eld data (x2.3.2, 2.3.3, 2.3.4) are used for
comparison.
2.3.5 Spatial cross{correlation between various soil prop-

erties
The eld data analysis is performed for all the relevant eld test results. For the
example on piezocone data, the analysis is also performed for the soil classication
index, Ic, which is related to grain size and soil type. The expression of soil classi-
cation index is (Jeeries and Davies, 1993):
q
Ic = 3 ; log Q(1 ; Bq )]2 + 1:5 + 1:3 log Fs]2 (2.49)
45
where Q = q ;0 0 is the stress normalized cone resistance, Bq is the piezocone pore
T
0
v
v
pressure ratio, and Fs is the stress normalized friction ratio (the expressions for qT Bq
and Fs, and all the notations involved are described in Appendix 4A).
A Beta distribution with parameters p = 5:2 and q = 120 resulted for the soil
classication index in the hydraulic ll layer. The correlation structure estimated
for Ic resulted very similar with that of the cone resistance. A comparison between
sample correlation functions in vertical directions for the two types of eld test results
is presented in Figure 2.17.
1
MAC05 MAC06 MAC07 MAC08

0.5
correlation function ρ(ξz)
0
-0.5
0 1 2 3
1
MAC31 MAC32 MAC33

0.5
0
ρI I (ξz)
c c
ρq q (ξz)
-0.5
c c
0 1 2 3 0 1 2 3 0 1 2 3
space lag - ξz (m)
Figure 2.17: Comparison between sample correlation functions in vertical direction

computed using uctuations about spatial trends of soil classication index Ic (contin-
uous lines) and cone tip resistance qc (dotted lines) data from Tarsiut P{45, layer 1.
The spatial cross{correlation between qc and Ic is analyzed for each soil prole,
as a function of the vertical component z = rz of the space lag. The discretized
form of the direct approach relation (Bendat and Piersol, 1986) is:
NX;r
^qI (rz) = N 1; r qc(iz)Ic(i + r)z] (2.50)
i=1
A typical result, obtained for one of the CPTU proles, is presented in Figure 2.18.
A strong negative correlation can be observed between the two soil parameters. The
shift in the maximum value, z0 = 0:15 : : : 0:20m, coresponds to the distance between
46
0
CPT profile
MAC05 -1 0 1 2
space lag - ξz (m)
-5
elevation - z (m)
cross-correlation between
-0.3
the fluctuations of soil
-10
classification index (Ic) and
-0.6
cone tip resistance (qc)
-15
2 2 0.5
qc (MPa) Ic = { [3 - log10Q(1-Bq)] + [1.5 + 1.3 log10 F] }
20 10 0 1.5 2 2.5 3
Figure 2.18: Study on cross{correlation between soil classication index Ic and cone
tip resistance qc data from CPT prole MAC05, layer 1.
recording devices (for cone resistance, on the one hand, and for sleeve friction and
dynamic pore pressure on the other hand). A rapid drop in correlation with increasing
space lag is observed, which indicates that between the two soil parameters there is
mainly a \point" cross{correlation. A value ^qI ;0:58 of the point cross{correlation
coecient has resulted as an average from all piezocone proles.
2.4 Conclusions
1. A procedure to estimate the statistics of spatial variability of material properties
based on eld measurement results is described. The proposed procedure comprises
the following steps: (1) standardization of the eld data, (2) evaluation of the prob-
ability distribution, and (3) estimation of the correlation structure. Each step is
illustrated with numerical examples based on analysis of real in{situ measurements.
2. Data homogeneity is enforced by a series of transformations: (1) computation

of the uctuations about spatial trends (or spatially variable average values), and
(2) normalization of the uctuations using the spatially varying standard deviation.
A test for checking the homogeneity of the standardized eld data is described.
47
3. The statistics of spatial variability are estimated by curve tting various ana-
lytical models to the sample functions (empirical distribution functions and sample
correlation functions) computed from standardized eld data. The proposed curve
tting procedure is based on nonlinear regression analysis and uses a quasi{Newton
optimization procedure.
4. The sample functions are obtained from eld data as follows:

{ Empirical distribution functions are evaluated based on the order statistics of eld
values.
{ Under the assumption of separable correlation structure, sample correlation func-
tions are separately evaluated, for vertical and horizontal directions, via di-
rect computations. The sample correlation function, which is an unbiased and
consistent estimator, is used: (1) to estimate the random measurement error
(or\noise"), and (2) to t the analytical model for the correlation structure and
to derive the correlation distances in each spatial direction.
5. The results of numerical computations for the analyzed eld data are as follows:
a. CPT data from Tarsiut P{45:
{ Skewed Beta distributions are tted to standardized eld measurement re-
sults obtained from both soil layers.
{ \Two peak" correlation structures, which can accommodate both small scale
and large scale spatial variability, are tted for both soil layers in both
vertical and horizontal directions. The resulted correlation distances are:
layer1 = 0:95m, layer1
vertical layer2 layer2
horizontal = 12:1m, vertical = 0:50m, horizontal = 5:4m.
Due to insucient eld data (sampling distances larger than the esti-
mated correlation distance), the results in terms of correlation structure
in the sand berm (layer 2) are uncertain. Spatial distribution of simulated
48
stochastic sample elds having the resulted theoretical correlation struc-

ture is veried by direct (visual) comparison with the spatial distribution
of the eld data.
b. SPT data from Akita Harbour:

{ An almost symmetrical Beta distribution model best ts the empirical dis-
tribution however, since a test for normality is passed, the Gaussian as-
sumption is accepted for the standardized eld data.
{ A triangular correlation structure with vertical = 1:3m is tted in vertical
direction. The correlation structure in horizontal direction cannot be esti-
mated from the available eld data, and hence it is assumed, using results
of similar analyses.
49
Chapter 3
Simulation of nD{mV,
non{Gaussian stochastic elds
A methodology to simulate multi-dimensional, multi-variate (nD{mV), non-Gaussian,
homogeneous stochastic elds is presented in this chapter. The proposed simula-
tion algorithm combines the spectral-representation-method-based formula for multi-
dimensional, multi-variate, Gaussian, stochastic elds (Shinozuka and Deodatis, 1988a)
with the iterative mapping technique for non-Gaussian elds proposed by Yamazaki
and Shinozuka (1988), and takes advantage of the Fast Fourier Transform technique
for great computational e
ciency. In the following, the proposed algorithm is formu-
lated for stochastic elds representing dierent soil properties, within stochastically
homogeneous soil layers.
3.1 Assumptions
The probability distribution of each soil property and the cross{correlation matrix
describing their spatial variability represent the input for the vector eld simulation
algorithm. The cross{correlation matrix is established based on the results of eld
data analysis, and making the following assumptions:
1. The stochastic vector eld is assumed to be quadrant (Vanmarcke, 1983). De-
noting by R0jk the auto-/cross-correlation function between components j and
k of the vector eld and by the separation distance, the property of quadrant
50
symmetry is expressed as:
R0jk ( 1 2 : : : n ) = R0jk (I1 1 I2 2 : : : In n )

I1 I2 : : : In = 1 j k = 1 2 : : : m (3.1)
where n is the number of spatial dimensions, and m is the number of scalar

components.
2. The spatial variability of soil properties is induced by dierences in soil compo-
sition/density from one location to another. Since these dierences are likely
to inuence all soil properties in similar ways, it is assumed that the auto-
correlation functions of the soil properties R0jj ( ) j = 1 2 : : : m are identical:
R0jj ( ) = R0( ) j = 1 2 : : : m (3.2)
where ( ) = ( 1 2 : : : n).
3. The cross-correlation functions are assumed to be expressed as:
R0jk ( ) = jk R0( ) j k = 1 2 : : : m j 6= k (3.3)
where jk is the correlation coe

cient between components j and k of the vector
eld.
The last two assumptions are conrmed by experimental evidence obtained from
analysis of eld test results from Tarsiut P{45, as described in x2.3.5.
3.2 Simulation of nD{mV Gaussian stochastic elds

3.2.1 General formulation
The algorithm to simulate 1D{mV stationary stochastic processes by spectral rep-
resentation method, proposed by Deodatis (1995) makes possible the generation of
51
ergodic sample functions and takes advantage of the FFT technique for computa-
tional e
ciency. The algorithm is formulated by Deodatis (1995) for the general case
of a cross{spectral density matrix:
0 S 0 (!) S 0 (!) 1
11 1m
0 B
S (!) = @ . .
. . . . ... C A (3.4)
Sm1(!) Smm(!)
0 0
where Sjk0 (!) is the complex cross{spectral density function, such that Sjk0 (!) =
Skj0 (!), and hence S0(!) is hermitian. It can also be shown (e.g. Shinozuka, 1987)
that the cross{spectral density matrix is non{negative denite.
In this section, the method proposed by Deodatis (1995) is generalized to n di-
mensions, i.e. S0(!) in Eqn. (3.4) becomes S0(k), with k = (k1 k2 : : : kn ) { the
n{dimensional wave number vector. The elements Sjk0 (k), of the cross{spectral den-
sity matrix are evaluated from the elements R0jk ( ) of the cross{correlation matrix
using the n{dimensional Wiener{Khintchine relations (e.g. Vanmarcke, 1983):
Z1 Z1
Sjk0 (k) = (21 )n R0jk ( )e;ik d (3.5)
;1
| {z } ;1
n times
where k is the dot product between the wave{number vector and the space{lag
vector, and d = d 1 d 2 d n .
The diagonal elements, Sjj0 (k) j = 1 2 : : : m, are the power spectral density
functions of the m components of the stochastic eld. They are real and non{negative.
The o{diagonal elements, Sjk0 (k) j k = 1 2 : : : m j 6= k, are the cross{spectral
density functions, which are in general complex functions of the vector wave number
k. However, when dealing with random media describing soil property distribution,
the cross{spectral density functions are real. Consequently, the cross{spectral density
matrix is real, symmetric and non{negative denite, and, under certain conditions,
admits Cholesky decomposition (e.g. Burden and Faires, 1993):
S0(k) = H(k)HT (k) (3.6)
where H(k) is lower triangular. Taking advantage of the quadrant symmetry ex-
hibited by elds associated with material properties, and combining the simulation
52
algorithm for nD{1V stochastic elds (Shinozuka and Deodatis, 1988a) with the re-
lationships proposed by Deodatis (1995) for simulation of 1D{mV ergodic processes,
the simulation formula for the rth component of an nD{mV Gaussian eld fG(x),
having zero mean and unit standard deviation, becomes:
Xr X N1 X N2 X
N X
fGr (x) = 2
n
Hrs (k1sl1 k2sl2 : : : knsl )

s=1 l1 =1 l2 =1 l =1 I1 =1 I =1
n
n i
i=2 3 :::n
q h i
k1k2 kn cos I1k1sl1 x1 + I2k2sl2 x2 + Inknsl xn + Is1lI12l:::I
n 2 :::l
n
n
(3.7)
where:
{ the subscript G stands for \Gaussian" eld

{ x = (x1 x2 : : : xn) is the n{dimensional spatial coordinate
{ ki = kNiu is the wave number increment in direction i, expressed in terms of kiu {
i
the upper cut{o frequency, beyond which the elements of the cross{spectral
density matrix can be assumed to be zero { and the number of points Ni to
describe the spectral density function in spatial direction i
{ Is lI l:::I:::l represent m 2n;1 sets of random phase angles uniformly distributed
1 2
1 2
n
n
between 0 and 2.
The functions Hrs (k), which are the elements of the matrix H(k) in Eqn. (3.6),
are computed at wave numbers:
kisl = (li ; 1)ki + ms ki

i
s = 1 2 : : : m li = 1 2 : : : Ni i = 1 2 : : : n (3.8)
i.e., for each s = 1 2 : : : m, there is a frequency shifting:
ks0 = ms ki (3.9)
as proposed by Deodatis (1995).

53
The simulated scalar elds fGr (x) r = 1 2 : : : m are asymptotically Gaussian

as the limits N1 N2 : : :Nn of the summations in Eqn. (3.7) become very large (e.g.
Soong and Grigoriu, 1993) due to the Central Limit Theorem, and are periodic (De-
odatis, 1995) in each spatial direction, with periods given by:
L0 i = m 2k i = 1 2 : : : n (3.10)
i
Sample functions fG(j)(x) of the simulated stochastic eld fG(x) can be obtained
by replacing the sets of random phase angles Is1lI12l:::I n
2 :::l
n
in Eqn. (3.7) with their respec-
tive realizations Is1lI12l:::I
n
2 :::l
n
(j )
(Shinozuka, 1987 Shinozuka and Deodatis 1988a, 1991
Deodatis, 1995).
It can be proven (Deodatis, 1995) that the ensemble expected values E fGr (j ) (x)] r =
1 2 : : : m and the ensemble auto/cross{correlation functions Rrs (x) r s = 1 2 : : : m

of the simulated stochastic elds fGr (j ) (x) are identical to the corresponding targets,
E fGr (x)] and R0rs(x), respectively. Starting from the work of Deodatis (1995), it
can also be generalized that every sample function given by Eqn. (3.7) is ergodic in
mean value and in correlation, when the length of the sample function in each spatial
direction i is equal to the period L0 i given in Eqn. (3.10).
3.2.2 Use of Fast Fourier Transform

To take advantage of the computational e
ciency of the FFT technique, Eqn. (3.7)
used to compute sample functions fG(j)(x) is rewritten as:
(j )
fGr (p1x1 p2(x2 : : :pn xn) "X #)
Xr
( j )
n s k k
= Re grs (q1x1 q2x2 : : :qn xn) exp ipk xk m (3.11)
s=1 k=1
In the above equation, the spatial discretization in each direction k, expressed by
xk , is selected such that:
Mk xk = 2k k = 1 2 : : : n (3.12)
k
with Mk 2Nk , from the condition to avoid aliasing (see e.g. Shinozuka and Deo-
datis, 1988a, 1991 for more details). The nD{mV stochastic eld fG(j)(x) is simulated
54
in each spatial direction k on a domain equal in length to the respective period, L0 k ,

i.e. pk = 0 1 2 : : : m Mk ; 1. However, the auxiliary functions grs (x) in Eqn. (3.11)
are computed on a reduced domain (qk = 0 1 2 : : : Mk ; 1), and the relation between
the two sets of indices in Eqn. (3.11) is:
qk = modfpk Mk g k = 1 2 : : : m (3.13)
i.e. qk is the remainder when dividing pk by Mk . The auxiliary functions grs(x) have
the following expression:
MX 1 ;1 M2 ;1
X MX;1 X
BrsI1Il21:::I
n
grs(j)(q1x1 q2x2 : : :qnxn) = l2:::l

n
l =0 I1 =1 I =1
n
l1 =0 l2 =0 n i
i=2 3 :::n
(X n )
exp iqk xk lk kk Ii r s = 1 2 : : : m (3.14)
k=1
with:
q n (j )o
BrsI1Il21:::I
n
l2 :::l
n
= 2 H (k k
rs 1sl1 2sl2 : : :k nsl
n
) k1 k2 : : : kn exp iIs1lI12l:::I
2 :::l
n
n
r s = 1 2 : : : m (3.15)
Here, the functions Hrs (k) are computed at the same locations (in the frequency
domain) as in Eqn. (3.7), but the wave numbers are expressed as:
kisl = liki + ms ki
i
s = 1 2 : : : m li = 0 1 : : : Mi ; 1 i = 1 2 : : : n (3.16)
to account for the fact that the summations over li i = 1 2 : : : n, start from zero in
Eqn. (3.14).
The summations in Eqn. (3.14) can be e
ciently performed using the Fast Fourier
Transform technique. This is achieved by transforming the expressions in the expo-
nential according to Eqn. (3.12) (for more details the reader is referred to Shinozuka
and Deodatis, 1988a, 1991):
2
xk kk = M k = 1 2 : : : n (3.17)
k
55
and rearranging the terms inside the summations to perform the n{dimensional FFT
as a succession of n one{dimensional transforms:
grs(j)q1 q2:::q =
n
2 1;1 0M ;1 ( )1 ( )3
X 4MX X 2 iq l 2 iq l
@ BrsI1Il21:::I n n
Ii A exp M Ii 5
1 1
n
l2:::l exp
n
I1 =1 I =1 l1 =0 l =0
n
M n 1
i=2 3 :::n | {z }
i n
n;dimensional FFT
r s = 1 2 : : : m (3.18)
such that each function grs(x) is computed as a sum of 2n;1 n{dimensional Fast
Fourier Transforms.
Remark. As shown by Deodatis (1995), in order to simulate ergodic elds, a
frequency shifting { Eqns. (3.8),(3.16) { is performed in Eqns. (3.7) and (3.15). The
exponential in Eqn. (3.11) arises from the frequency shifting theorem of the Fourier
integral (see e.g. Papoulis, 1962).
3.2.3 Particularities for simulation of random media repre-

senting soil properties
At this point it is recalled that the m{variate stochastic eld to be simulated rep-
resents the spatial distribution of various soil properties, and each component (vari-
able) of the eld stands for one specic soil property. To improve the computational
e
ciency of the simulation algorithm for the case of random elds representing ma-
terial properties, the assumptions stated in x3.1 are enforced. The diagonal elements
Sjj0 (k) j = 1 2 : : : m of the cross{spectral density matrix are evaluated using the
Wiener{Khintchine relations (3.5) and Eqn. (3.2):
Z1 Z1
Sjj0 (k) = (21 )n ;1 R0( )e;ik d = S 0(k) j = 1 2 : : : m (3.19)
;1
| {z }
n times
i.e. all the components of the stochastic eld have the same power spectral density
function.
To evaluate the cross{spectral density functions, Sjk0 (k) j 6= k, the third assump-
tion in x3.1 is used. From the Wiener{Khintchine relations (3.5) and making use of
56
Eqns. (3.19) and (3.3), the o{diagonal terms of the cross{spectral density matrix
simplify to:
Z1 Z1
jk
Sjk (k) = (2)n
0
;1
;1 R0( )e;ik d = jk S 0(k) j k = 1 2 : : : m (3.20)
| {z }
n times
and the cross{spectral density matrix can be expressed as:
0 1
1 12 1m
B 21 1 2m C
S (k) = S (k) B@ ... ... . . . ... CCCA = S 0(k) s
0 0 B
B (3.21)
m1 m2 1
where jk = kj jjk j 1 j k = 1 2 : : : m and, therefore, s is a real, symmetric and
non{negative denite matrix, not dependent on the vector of wave numbers k. In
this case, the computational e
ciency is considerably improved, since the Cholesky
decomposition (3.6) has to be performed only once for the constant matrix s:
s = hhT (3.22)
The particular form (3.21) of the cross{spectral density matrix leads to further
simplications of the algorithm described in x3.2.2, as shown hereafter.
The matrix H(k) in Eqns. (3.6) and (3.15) becomes:
q
H(k) = S 0(k) h (3.23)
Eqn. (3.15) is written as:
n I1I2 :::I
l2 :::l = hrs Bs l1 l2 :::l
BrsI1Il21:::I n
r s = 1 2 : : : m (3.24)
n n
where hrs are the elements of the lower triangular matrix h in Eqn. (3.22), and:
q n o
Bs lI11l2I:::l
2 :::I
n
n
= 2 S 0(k1sl1 k2sl2 : : : knsl ) k1 k2 : : : kn exp iIs1lI12l2:::I:::l (j)
n
n
n
s = 1 2 : : : m (3.25)
The functions grs(j)(x) in Eqn. (3.14) simplify to:
grs(j)(q1x1 q2x2 : : :qnxn) = hrs gs(q1x1 q2x2 : : : qnxn)
r s = 1 2 : : : m (3.26)
57
where the functions gs(x) have similar expressions with the functions grs(j)(x), but are
not dependent on the index r:
MX 1 ;1 M2 ;1
X MX;1 X
gs(q1x1 q2x2 : : :qnxn) = Bs lI11l2I:::l
n
2 :::I
n
l =0 I1 =1 I =1
n
l1 =0 l2 =0 n i
i=2 3 :::n
(X n )
exp iqkxk lk kk Ii s = 1 2 : : : m (3.27)
k=1
The m functions gs(x) in Eqn. (3.27), generated with dierent sets of random
phase angles Is1lI12l:::I
n
2 :::l
n
(j )
, are used to simulate m non{correlated nD{1V stochastic
elds:
fs(p1x(1 p2x2 : : :pn xn) "X #)
n s k

= Re gs (q1x1 q2x2 : : :qnxn) exp ipk xk m k
(3.28)
k=1
s = 1 2 : : : m
where ranges and expressions for pk qk k = 1 2 : : : n are given in Eqn. (3.13).
Finally, using Eqns. (3.24), (3.27) and (3.29), Eqn. (3.7) becomes:
X
r
(j )
fGr (x) = hrs fs(x) r = 1 2 : : : m (3.29)
s=1
Consequently, the simulation of the m{variate Gaussian stochastic eld fG(x) is
performed by generating m non{correlated univariate elds fs(x) which are then
combined for each component r of the vector eld fG(x) as shown in Eqn. (3.29),
according to the elements of the cross{spectral density matrix represented here by
the coe
cients hrs.
3.2.4 Computation of power spectral density from sample

elds
As proven by Deodatis (1995), the m scalar components fr(j)(x) of the vector (m{
variate) sample eld f (j)(x) are ergodic. The n{dimensional power spectral density
functions Srr (k) r = 1 2 : : : m, of these components have the general form (see e.g.
Yamazaki and Shinozuka, 1988, for a 2D ergodic scalar eld):
Srr (k) = (2)n L 1L L jFr (k)j2 r = 1 2 : : : m (3.30)
01 02 0n
58
where Fr (k) is the Fourier Transform of fr(j)(x), dened in the discretized form as:
!
k1 k n
Fr t1 m : : :tn m = x1 xn
mMX1;1 mMX;1 (j) ( X n )
:::
n

fr (p1x1 : : : pn xn) exp ;i pk xk tk m k k
p1 =0 p =0n k=1
tk = 0 1 : : : mMk ; 1 k = 1 2 : : : n r = 1 2 : : : m (3.31)
To explain Eqn. (3.31), it is recalled that the components fr(j) (x) r = 1 2 : : : m
of the m{variate eld f (j)(x) are generated over a period L0 k = 2m=kk and at
mMk locations in each spatial direction k. Consequently, by performing the Fourier
Transform in Eqn. (3.31), the functions Fr(k) are computed at mMk locations in the
wave number space, and the wave number increments in each direction are mk k = k
1 2 : : : n. This can be easily understood if Eqn. (3.10) is written as:

L0 k = (k2=m) (3.32)
k
For the sake of clarity, and without loss of generality, the following derivation
of the power spectral density function is explained for the case of an 1D{mV eld
f (j)(x). In this case, Eqns. (3.30) and (3.31) become:
1 jF (k)j2 r = 1 2 : : : m
Srr (k) = 2L (3.33)
r
0
and
! mMX;1 (j) ( )
k
Fr t m = x fr (px)exp ;ipx t m k
p=0
t = 0 1 : : : mM ; 1 r = 1 2 : : : m (3.34)
respectively. Using Eqn. (3.32), the discretized form of Eqn. (3.33) is (Yamazaki and
Shinozuka, 1988):
! ( )2
k 1 1 mMX;1 (j) k
Srr t m = (k=m) mM fr (px)exp ;ipx t m
p=0
t = 0 1 : : : mM ; 1 r = 1 2 : : : m (3.35)
As shown in Eqn. (3.29), each scalar eld fr(j)(x) is a weighted sum of r non{
correlated elds fs(x) s = 1 2 : : : r. Moreover, each of the elds fs(x) are generated
59
using values of the common spectral density function S 0(k) in Eqns. (3.19) and (3.21)]
computed at dierent sets of wave number values for each component s, i.e.:
{ the eld f1(x) is generated using the values S o(k1l), with k1l = lk + m1 k l =
1 2 : : : M ; 1, (this is the 1D form of Eqn. (3.16) for s = 1)
{ the eld f2(x) is generated using the values S o(k2l), with k2l = lk + m2 k, etc.
and the sets fksl jl = 0 1 : : : M ; 1g s = 1 2 : : : r are disjoint. Consequently, when
transforming back to the frequency domain Eqn. (3.34)], each of the terms fs(x) will
give nonzero values for Fr (k) only at the locations in the wave number space which
have been used for simulation, i.e.:
! Xr X;1
mM ( ) X r !
k
Fr t m = x hrs fs (px)exp ;ipxt m = hrsFs t mk
k
s=1 p=0 s=1
t = 0 1 : : : mM ; 1 r = 1 2 : : : m (3.36)
where Fs(k), the Fourier Transform of fs(x), is such that:
! 0 if t = ml + s l = 0 1 : : : M ; 1
Fs t mk is 6=

=0 otherwise (3.37)
Using Eqn. (3.36), the expression (3.33) of the power spectral density becomes:
X 2
1 r
Srr (k) = 2L hrsFs (k) = 2L
1 X r
h2rs jFs(k)j2 r = 1 2 : : : m (3.38)
0 s=1 0 s=1
The result in Eqn. (3.38) is evident for r = 1. The decoupling is possible for r > 1
due to the fact that all the cross terms in the squared summation are zero (as shown
in Eqn. (3.37), the functions Fs(k) have non{zero values only at sets of wave numbers
which are disjoint with each other). More details are presented in Appendix 3.A, for
the case m = 2.
For reasons which will become evident, Eqn. (3.38) is written as:
Xr
Srr (k) = h2rs Sss (k) r = 1 2 : : : m (3.39)
s=1
where:
1 jF (k)j2
Sss (k) = 2L s = 1 2 : : : r (3.40)
s
0
60
and hrs = 0 for s > r. To illustrate how Eqn. (3.39) works, the non{zero values of
the power spectral density functions of an 1D{3V eld are shown in Figure 3.1 by
vertical bars. The original (target) spectral density function S 0(k) is represented by
continuous curves.
For r = 1 (Figure 3.1,a) the only non{zero values of the resulted power spectral
density function S11(k) are concentrated at wave numbers k = t 3k with t =
3l + 1 l = 0 1 : : : M ; 1. This is due to the fact that only those wave numbers
have been used to simulate the component fG(j1)(x), since for r = 1 Eqn. (3.29)
becomes:
fG(j1)(x) = h11 f1(x) (3.41)
If one computes the average of power spectral values over each interval k, i.e.:
" ! #
1 k
S11(k~1 l+1) = 3 h11 S11 lk + 3 + 0 + 0 l = 0 1 : : : M ; 1 (3.42)
2

one obtains the values of the resulted power spectral density function S11 k~1 l+1
at wave numbers k~1 l+1 2 (lk (l + 1)k) { for the case r = 1 m = 3 )

k~1 l+1 = lk + 3k l = 0 1 : : : M ; 1. The values of S11 k~1 l+1 are represented
by open dots in Figure 3.1,a, and they are equal to those of the original (target)

spectral density function: S11 k~1 l+1 = S 0 k~1 l+1 l = 0 1 : : : M ; 1.
For r = 2 (Figure 3.1,b) s takes values s = 1 and s = 2, and the power spectral
density function S22(k) has non{zero values at wave numbers k = t 3k , with
t = 3l + 1 and t = 3l + 2. The average of spectral density values over each
interval k is computed as:
" ! ! #
1 k 2 k
S22(k~2 l+1) = 3 h221 S11 lk + 3 + h222 S22 lk + 3 + 0
= S 0(k~2 l+1) l = 0 1 : : : M ; 1 (3.43)
(vertical dashed lines in Figure 3.1,b).
61
a. |r=1 u
e
h2
S values of average
SDF | S (k~ )
1s
6
ss
averaging
2
=1 for = 1 over l
ss r l +1
h11
h
2 =0
z
these}|values{
12
h
2 =0 u
13
u
=0
l l =1 l =2 l =3
averaging
- - - -
f
for = 0 over
`
l
these}|values{
`
! a
! a
f
" b
" b
z f
" b
" b
a
"
"
u
"
" a
k
! a
! a
f
!
u u u u u u u u
!
0 k z }| { z }| {
-
2 + 3 3 + 3
k 2k k k
|r=2
3 3 k k
b.
u
h22 Ss ss
u
target SDF
6
21 = 0 4
h
2
: u
22 = 0 6
e
S (k )
2
h : `
0
2 =0
`
u e
! a
! a
h 23 u u
;
" b
" b
" b
;
" b
e
a
u
"
"
"
" a
u e k
! a
! a
u u u u
~ ~ ~ ~
!
k21 k22 k23 k24

c. | =3 r
h23 S
6
s ss
averaging
2
=03 for = 3 l
h31
=04
:
u
u over these
2
h32 :
e u
`
u e values
=03
`
2 z }| {
! a
! a
h33 :
u
" b
" b
u
" b
u
" b
ue
a
a
"
u
"
ue
"
u
" a
k
! a
! a
u
!
0
!
k 2 k 3 k 4 k
!
Figure 3.1: Computed power spectral density functions Sss (lk + s 3k ) { continuous
bars { and Sss(k~r l+1 ) { dashed bars { for an 1D{3V sample eld (n = 1 m = 3 r =
1 2 3).
62
for r = 3 = m (Figure 3.1,c) there is spectral power at all values k = lk +

3 k s = 1 2 3 l = 0 1 : : : M ; 1. The power spectral density over each
s
interval k is again computed as:

X !
1 3
s k
S33(k3 l+1) = 3 h3sSss lk + 3 = S 0(k~3 j+1)
~ 2
s=1
l = 0 1 : : : M ; 1 (3.44)
The generalization to the n{dimensional case is straightforward. By omitting zero

terms, Eqns. (3.42),: : : (3.44) can be written as:
Xr !
1 s k
Srr (k~1rl1+1 : : : k~nrl +1) = mn h2rsSss l1k1 + m : : :lnkn + m
1 s k n
n
s=1
lk = 0 1 : : : Mk ; 1 k = 1 2 : : : n r = 1 2 : : : m (3.45)
The averaging in Eqn. (3.45) is over mn locations in the wave number volume
k1 k2 kn, at which the values of the spectral density function are evaluated,
according to Eqn. (3.35). As shown in Figure 3.2 for the 2D{2V case, according to
the stochastic eld simulation algorithm, nonzero values may appear only at locations
situated on diagonals parallel with the rst diagonal of the wave{number space.
The terms Sss (k) in Eqn. (3.45) have the discretized expression:
!
k 1 k n
Sss t1 m : : :tn m = k k M 1
1 Mn2 mn
1 n 2
mM ;1 0mM ;1 2ipn tn 1 2ip1t1 2
X1 X
@ fs(p1x1 : : :pn xn) exp ; mM A exp ; mM
n
n 1
| p1=0 p =0
n
{z }
n;dimensional FFT
tk = 0 1 : : : mMk ; 1 k = 1 2 : : : n s = 1 2 : : : m (3.46)
The wave numbers k~krl +1 k = 1 2 : : : n, at which the values Srr (k~ r ) are com-
k
puted in Eqn. (3.45) can be evaluated using Eqns. (3.42),: : : (3.44) and assuming that
Srr (k) const. over the n{dimensional intervals k, as shown in Figure 3.3:
Xr
k~krl +1 = lk kk + smkk h2rs
k
s=1
lk = 0 1 : : : Mk ; 1 k = 1 2 : : : n r = 1 2 : : : m (3.47)
63
a. |r=1
6
h21s Sss k2
;
;
u
k111~ u
~
k112 k =1
u u ;

3

3
r =1 l1 +1=2
2
u
;
=1

h11

2 f f 3k2

=0 f
;
u u u u u u
h12
p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p

u
p p ;
;
p p p p p p p p p p ; p p p p p p p p p ; p p p p p p p p p p ; p p p p p p p p p p ; p p p p p p p p p p ; p p p p p p p p p ;
~
k213
;
u u ;
u ; ; ; ; ;
-
f
p p p p p
;
p p p p p p ;
p p p p p p p p p p ;
p p p p p p p p p ;
f
2k2
f u
;
u u u u ;
u u ; ; ; ; ;
u
= k~212
; ; ; ; ; ; ;
k2
k2 +
u
u u u
2
; ; ; ; ; ; ;
; ; ; ; ; ; ;
k2
f f
f
;
u u u;
u u u ; ; ; ; ;
= k~211
; ; ; ; ; ; ;
k2
u u u
2
; ; ; ; ; ; ;
k1
p ; p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p
; ; ; ; ; ;
z }| {
; ; ; ; ; ; ; -
0 k1
k1 k1 ~213
2k1 + =k
|r=2
2
2
b.
~
h 2
2s Sss k2
~121 u
k122 k =1
6
;
;
k r =2 l1 +1=2
;
u u
uu f u
; ;
2
; ;
f
;
= 0:4
uu f
h21
u
; ;
2
= 0:6 3k2 u
;
u
;
~
;
h22
u
k223
; ;
-
f
; ; ; ; ; ; ; ;
u
f
p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p ; ; ;
u u ; ; ; ; ;
u
;
f
u
; ; ;
p p p p p
p p p ;
p p p ;
p p p pp p p p p p p p p ;
p p p p p p p ;
p p p p
p p p p p p p p p p p p p p p ;
p ;
p p p p p p p p p ;
p p p p p p
p p p p p p p p p pp p p p p p p p p p p p p p p p p p p p
p p p p p p p p p p p p p p p p p p p p p p p p
2k2 u
p p p p p p p
p p p p p p p p p p p p p p p p p p p p p p p
p p p p p p p p
p p p p p p p p p p p p p p p p p p p p p p
~
; ; ;
u ;
u ; ;
u ; ; ;
u =2
k222 k
u
; ; ;
-
; ; ; ; ; ; ;
p p p p
p p p p p p p p p p p p p p p p p p p p p
p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p
p p p p p
p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p
=2 +1=2
p p p p p p
p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p
u u u
; ; ;
r l2
; ; ; ; ; ; ;
f
u f
f u
; ; ;
p p ;
p p p p p p p p p ;
p p p p p p p p p ;
k2 u
u u u
; ; ;
~
; ; ; ; ; ; ;
k221
; ; ; -
; ; ; ; ; ; ;
p
p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p
p p
p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p
p p p
p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p
;
u ; ;
u ; ;
u ; ;
f
k1
p ; p p p p p p p p p p p p p p p p p p p p p p p p p p
; ; p p p p p p p p p p p p p p p p p p p p
; ; p p p p p p p p p p p p p ; ;
values of
; ; ; ; ; ; ; -
0 k1 2k1 3k1 ~krl +1 )

Sss (k k

l1 =0 -
l1 =1 -
l1 =2 -
Figure 3.2: Computed power spectral density functions Sss (l1k1 + s 2k1 l2k2 +
s 2k2 ) { continuous bars { and Sss(k~krl +1) { dashed bars { for a 2D{2V sample eld
(n = 2 m = 2 r = 1 2).
k
64
a. 1D3V eld = 3 r
2 r
X
hrs
f h2rs =1
xk h2rs = s mkk h2rs
6
r
X r
X s=1
s=1|
s=1 {z }
k=3 k=3 k=3
=1
- - -
uh232
k~krl +1 = lk k + xk
k uh231
uh233
-
~ ( + 1)
-
l k k3 l k
l +1
b. 2D4V eld = 3 = 1 2 r k
k2
k1=4 k1=4 k1=4 k1=4 ( )=0
6
v Srr k
( = 0)
- - - -
pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp
C
pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp hrs
(l2 + 1)k2 pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp
6
C
pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp k2
( )=0

pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp vpp pp pp pp pp pp pp pp pp pp pp pp
6
4

Srr k
( = 0)

pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp

~
?
hrs
pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp hpp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp k2

6

k23l2 4
pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp vpp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp

hlocation of

6
X
?
k2
~ +1

pp pp pp pp pp pp xpp 2pp pp pp pp pp pp vpp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp

6
4

pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp xpp 1pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp

kkrlk

pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp pp k2

4

-

l2 k2
p p p p p pppp p p p pppppp p p p p ppp p ppp p p p ppp p p p p p p p p p p
? ?
k1
~ ( 1 + 1)
-
l1 k1 k13l1 l k1
Figure 3.3: Computation of the wave numbers k~krl +1 at which the resulted spectral
density functions are evaluated, under the assumption that Srr (k) const. over each
k
n{dimensional interval k.
65
For r = 1, h11 = 1 and from Eqn. (3.47): k~k1l +1 = lkkk + mk , as it is also shown
k
k
in Figure 3.1,a. An illustration of the case n = 2 m = 2 is presented in Figure 3.2,

showing how the values of Srr(k~ r ) and Sss (t1 2k1 t2 2k2 ) are distributed in the wave
number space.
3.2.5 Application | Simulation of a 2D{3V Gaussian stochas-

tic eld
The relations presented so far are used to simulate a two{dimensional, three{variate
(2D{3V) Gaussian stochastic eld. A separable correlation structure is assumed for
the common autocorrelation function, R0( ) in Eqn. (3.2), and hence, the expression
of the (common) spectral density function is of the form:
S 0(k) = S110 (k1) S220 (k2) (3.48)
where Sii0 (ki) is the two sided 1D spectrum in the spatial direction i. An exponen-
tially decaying expression is selected for the 1D spectrum in both spatial directions
(Shinozuka and Deodatis, 1988b):
Sii0 (ki ) = 1 bai +1 ka exp f;b jk jg

i
i = 1 2 (3.49)
i
i i
2 ;(ai + 1) i
where the parameters ai and bi control the shape of the spectrum and the correlation
distances, and are usually evaluated from eld measurement results. For this example,
the shape parameters ai are selected a1 = a2 = 2 and the correlation distances:
1 = 10m, in horizontal direction, and 2 = 2m, in vertical direction. For ai = 2 the
p
parameters bi are computed as bi = 6i (Shinozuka and Deodatis, 1988b).
The selected correlation structure is quadrant symmetric, and the rst quadrant in
the wave{number domain is delimited as follows: 0 k1 ku1 and 0 k2 ku2, with
ku1 = 0:5s;1 and ku2 = 2:0s;1 { the cut{o wave numbers beyond which the elements
of the cross{spectral density matrix are assumed to be zero. The discretization in
the frequency domain consists of 64
64 points (N1 = N2 = 26).
66
In the spatial domain, the mesh consists of 384

384 points (i.e. mM1 = mM2 =
384 m = 3 variables, M1 = M2 = 27). The dimensions of the simulated eld are
L01
L02 = 2400
600m.
Some of the numerical simulation results are presented in Figures 3.4, : : : 3.8. The
original, nonsymmetric spectral density function (target SDF) is shown in Figure 3.4
along with 2D views of the simulated scalar components.
2.5
Target spectral density function Simulated Field #2

2
4
target spectrum
1.5
f2 ( X1, X2 )
2
1
0
tal 0.5
0.5
-2
zon .4
ori 3 0
-4
0
10
- h 0. 0
k1.1 0.2
2 80
1.5
1 60 100
k2 - vertica
0
l direction 0.5 80
0
0 X2 40 60
20 40 X1
20
Simulated Field #1 Simulated Field #3
4
4
f1 ( X1, X2 )
f3 ( X1, X2 )
2
-2 0 2
-2 0
-4
-4
10 10
0
X2 8 X20 8
(co0 (co 0
rre 60 100 rre 60 100
lat 80 ) lat 80 m)
ion 40
ion 4
60 ce=10
m 60 ce=10
dis0 dis an
ta 0
4 n dis t a n tan 20 40 n dist
nc 20 ce 20 rrelatio
e= 20 orrelatio =2
2m 0 (c m) 0 0 1 (co
) 0 X1 X
Figure 3.4: Simulation of a 2D{3V stochastic eld { target SDF and rst 100
100m
of the scalar components of the simulated eld (total dimensions: 2400
600m).
In Figures 3.5 and 3.6 the resulted spectral density functions of the three simulated
scalar components are compared with the target SDF. The errors "r { Eqn. (3.56)]
of the resulted spectra relative to the target spectrum are: (1) very small for the rst
scalar component { "1 = 3
10;4 { and (2) larger for the other two components
("2 = "3 = 4
10;3 ), due to the approximation in evaluating the locations k~krl +1 k
{ Eqn. (3.47) { for r > 1.

To illustrate the inuence of the cross{covariance terms on the simulated eld,
1D cuts in both spatial dimensions are plotted in Figure 3.7:
67
2.5
2.5
Target spectral density function SDF - Field #2
2
2
target spectrum
1.5
1.5
SDF #2
1
1
tal 0.5
tal 0.5
0.5
0.5
zon .4
zon .4
ori 3 0
ori 3 0
0
0
- h 0.
- h 0.
k1.1 0.2
k1.1 0.2
2 2
1.5 1.5
1
k2 - vertica 1
k2 - vertica
0
0
l direction 0.5 l direction 0.5
0
0
0 0
2.5
2.5
SDF - Field #1 SDF - Field #3
2
2
1.5
1.5
SDF #1
SDF #3
1
1
tal 0.5
tal 0.5
0.5
0.5
zon .4
zon .4
ori 3 0
ori 3 0
0
0
- h 0.
- h 0.
k1.1 0.2
k1.1 0.2
2 2
1.5 1.5
k2 - verti 1 k2 - verti 1
0
0
cal direct 0.5 cal direct 0.5
ion ion
0
0
0 0
Figure 3.5: Simulation of a 2D{3V stochastic eld { comparison between target and
resulted spectral density functions
2.5
2
Target SDF
-4
Resulted SDF - Field #1 - ε1 = 3 x 10
Target and resulted SDF
-3
1.5
-3
1
0.5
0
0 0.4 0.8 1.2 1.6 2

2 2 0.5
wave number: (k1 + k2 )
Figure 3.6: Target and resulted spectral density functions represented along the wave{
number space diagonal
68
{ comparisons between strongly correlated scalar components #1 and #2 in Fig-

ures 3.7,a and b (12 = 0:8)
{ comparisons between weakly correlated scalar components #1 and #3 in Fig-

ures 3.7,c and d (13 = 0:2).
First 150m in direction X1 (X2 = 0) First 150m in direction X2 (X1 = 0)

Total simulated length: L01 = 2400m Total simulated length: L02 = 600m
Correlation distance: θ1 = 10m Correlation distance: θ1 = 2m
3
3
a. b.
2
2
1
1
f ( X1, 0 )
f ( 0, X2 )
0
0
-1
-1
field #1
-2
-2
field #2
-3
-3
0 50 100 150 0 50 100 150

3
c. d.
2
2
1
1
f ( X 1, 0 )
f ( 0, X2 )
0
0
-1
-1
field #1
-2
-2
field #3
-3
-3
0 50 100 150 0 50 100 150

X1 X2
Figure 3.7: Simulation of a 2D{3V stochastic eld { 1D sections of the resulted scalar
components: a,b. elds #1 and #2 { strongly correlated { 12 = 0:8 c,d. elds #1
and #3 { weakly correlated { 13 = 0:2.
The eect of input correlation distances is made evident by the comparison be-
tween the 1D cuts in horizontal direction (1 = 10m { Figures a and c) vs. those in
vertical direction (2 = 2m { Figures b and d).
The resulted distributions of the simulated scalar components are very close to
the normal distribution, as shown in Figure 3.8,a.
69
1
cummulative distribution a. b.
resulted
0.5
0.5
stnd. normal
CDF field #1 simulation #1
0
0
-3 -2 -1 0 1 2 3 0 0.5 1
stochastic field variable input
Figure 3.8: Simulation of a 2D{3V stochastic eld: a. resulted cumulative distribu-
tions vs. standard normal distribution b. input and resulted correlation coe
cients.
To chech how well the o{diagonal terms of the resulted cross{correlation matrix
match the target spectrum, three independent simulations are performed, employing
dierent values for the seed used to generate the sets of random phase angles Is1lI12l2
in Eqn. (3.25). Sample correlation coe
cients between the zero mean, unit standard
deviation scalar components of the simulated elds are computed as follows:
1 X1 mM
mM X2
ij = m2M M fi(x1k1 x2k2 )fj (x1k1 x2k2 ) i j = 1 2 3 i 6= j (3.50)
1 2 k1 =1 k2 =1
and compared with the input values in Figure 3.8,b. The input correlation coe
cients
are: 012 = 0:8 013 = 0:2 and 023 = 0:5. Excepting for one value (13 { simulation #3),
all the resulted sample correlation coe
cients are within 1% of the input values.
The solution time on an IBM minicomputer was 60 sec. and the total memory
required for an in{core solution { about 8 Mbytes.
3.3 Simulation of non{Gaussian vector elds

3.3.1 Basic algorithm
An ergodic, multi{variate stochastic eld fG(j)(x) can be simulated as shown in x3.2,
based on a cross spectral density matrix S0(k) { which will be refered to as tar-
get spectrum. In the following, the subscript (j) indicating a sample eld will be
70
dropped, for clarity. The simulated eld is Gaussian by virtue of the Central Limit
Theorem, and its cross spectral density matrix is theoretically identical with the tar-
get spectrum: SG (k) S0(k) (Deodatis, 1995). Deviations from this identity might
be due to: (1) round{o errors, and (2) details of resulted spectrum computations,
as shown in x3.2.4.
A non{Gaussian eld fB (x) can be obtained starting from the simulated Gaussian
eld fG (x) by a mapping technique, proposed by Yamazaki and Shinozuka, (1988):
fB (x) = FB;1 fFG fG (x)]g (3.51)
where FG and FB are the Gaussian and non{Gaussian eld distribution functions,
respectively. Eqn. (3.51) is a memoryless nonlinear transformation (Soong and Grig-
oriu, 1993 Grigoriu, 1995), which yields stationary non{Gaussian elds when the
original Gaussian elds are stationary (Grigoriu, 1984).
However, because the transformation (3.51) is nonlinear, the cross spectral density
SB (k) of the non{Gaussian eld does not coincide with the target spectrum. There-
fore, an iterative process is necessary to modify the cross spectral density matrix
SG (k) of the original Gaussian eld fG (x) in order to obtain in the end SB (k) S0(k).
The algorithm is described by Yamazaki and Shinozuka (1988) for scalar eld gener-
ation, and is modied here to accommodate multi{variate elds.
Starting from S(1)
G (k) = S (k), at iteration (1), the following steps are performed
0
at each iteration (i):

Step 1. Simulate the Gaussian eld fG(i)(x), based on the cross spectral density
matrix S(Gi)(k), according to Eqns. (3.25), (3.27): : : (3.29). As shown in x3.2.3,
the simulation of the m correlated scalar components fGr (x) is performed by
rst generating m non{correlated scalar elds fs(x), which are then combined
according to Eqn. (3.29). Each of the elds fs(x) are computed starting from
sets of locations in the wave number space:
!
s k 1 s k n
k = l1k1 + m : : : lnkn + m lk = 0 1 : : : Mk ; 1 (3.52)
s
71
which are disjoint for various s values, as shown in x3.2.4. The spectral density
functions of the elds fs(x) are Sss (k) in Eqns. (3.45) and (3.46), which at the
rst iteration are theoretically identical with the (common) diagonal term of
the target spectrum, computed at the respective locations:
Sss(1)(ks) = S 0(ks) (3.53)
Consequently, for multi{variate eld simulation the generating spectral density

function SG(i)(k) is expressed as a collection of spectra Sss(i)(ks) s = 1 2 : : : m
computed at disjoint sets of wave numbers ks.
Step 2. Map the Gaussian eld fG(i)(x) into a non{Gaussian eld fB(i)(x), as shown
in Eqn. (3.51). The function FG can be evaluated as the empirical cumulative
distribution of the generated eld fG(i)(x). Since a relatively large computational
eort is required to evaluate the empirical distribution function, and accounting
for the fact that fG(i)(x) is, with a good approximation, a Gaussian eld with
zero mean and unit standard deviation, FG in Eqn. (3.51) is replaced by the
standard normal distribution function, :
n h io
fB(i)(x) = FB;1 fG(i)(x) (3.54)
(i)
Step 3. Evaluate the resulted spectral density functions SBrr (k) r = 1 2 : : : m of
(i)
each scalar component fBr (x) of the simulated non{Gaussian eld fB(i)(x), by
performing inverse n{dimensional Fast Fourier Transform. The resulted spectra
have non{zero values at the locations in the wave{number domain which have
been used to generate the Gaussian eld fG(i)(x) { ks in Eqn. (3.52). By averaging
the values resulted in each n{dimensional interval k = k1 kn , one gets
(i) ~r
the spectral density functions SBrr (k ) r = 1 2 : : : m, with
k~r = (k~1rl +1 : : : k~nrl +1)

1 n
lk = 0 1 : : : Mk ; 1 (3.55)
where k~krl +1 are computed according to Eqn. (3.47).

k
72
Step 4. Check convergence by comparing SBrr

(i)
(k) with the target SDF S 0(k), for
all the scalar components (r = 1 2 : : : m). The proposed convergence check is
based on the relative error of the resulted spectra with respect to the target
SDF, computed for each scalar component r as:
n times
Zz k 1 }| Z k { (i)
u

un
S ( k) ; S 0(k) dk
Brr
"(ri) = 0 Z k 01 Z k (3.56)
0 S 0(k) dk
u un
| {z }
0
n times
The iterative process ends when the relative errors computed for all scalar
components of the simulated eld are within a prescribed tolerance:
"(ri) TOL r = 1 2 : : : m (3.57)
As long as convergence is not achieved, the spectral density functions Sss (k)
used to generate the non{correlated Gaussian scalar elds fs(x) are updated,
as shown in the next section.
Remark: The convergence check is only performed for the diagonal terms SBrr(k)
of the cross spectral density matrix SB (k). Considerations on recovering the o{
diagonal terms (i.e. correlation coe
cients) are presented in x3.3.3.
The ow chart of the proposed algorithm and several of its computational phases
are illustrated in Figure 3.9, where target and computed spectral density functions
(SDF) of the rst component of a 2D-3V vector eld are displayed. The target SDF
S 0(k) of this rst component is shown in Figure 3.9,a and is used to generate a sample
function of the Gaussian eld fG(1)1 (x) superscript (1) stands for the rst iteration].
The Gaussian sample function fG(1)1 (x) is then transformed into fB(1)1 (x) that follows a
lognormal distribution this is done using Eqn. (3.51)]. The computed SDF SB(1)11(k) of
fB(1)1 (x) is shown in Figure 3.9,b and is obviously dierent from the target SDF shown
in Figure 3.9,a because of the nonlinearity of the transformation in Eqn. (3.51).
73
During the iteration process, the SDF SG(i11) (k) used to generate the Gaussian sample
function fG(i1)(x) at the i-th iteration is updated according to the formula shown in the
ow chart. The SDF SG(30) 11 (k) at the 30-th (last) iteration is shown in Figure 3.9,c.
This is the spectral density function used to generate the Gaussian sample function
1 (x), which is then transformed into the lognormal sample function fB 1 (x) that
fG(30) (30)
has the SDF SB(30) 11 (k) shown in Figure 3.9,d. Note that SB 11 (k) (Figure 3.9,d) is
(30)
practically identical to the target SDF S 0(k) (Figure 3.9,a).

It has been therefore demonstrated in Figure 3.9 how the iterative scheme is
working to generate sample functions of a multi-dimensional, multi-variate stochastic
eld, that are compatible both with a prescribed cross-spectral density matrix and
with prescribed non-Gaussian probability distribution functions (see also Deodatis
et.al., 1995 Popescu et.al., 1996).
3.3.2 Procedure to update the generating spectral density

functions SGrr(k)
The spectral density functions SGrr (k) are updated at every iteration i by modifying
the spectra Sss(i)(k) used to simulate the non{correlated scalar elds fs(x). The
correction is performed as:
Sss(i+1)(ks) = Fs(i+1)(ks) Sss(i)(ks) s = 1 2 : : : m (3.58)
where Fs(i+1)(ks) is the updating factor for the spectrum Sss (ks). The updating
(i)
factors are evaluated on the basis of the resulted spectra SBrr (k) corresponding to
(i) (x), simulated at iteration i. Consequently, the proposed
the scalar components fBr
correction procedure has the shortcoming that the updating factors are evaluated
(i) (x) originates
based on coupled eld spectra { because each scalar component fBr
in a summation of r non{correlated univariate elds fs(x) { and they are used to
modify the spectral density functions Sss (ks) of the non{correlated elds fs(x). Two
methods to evaluate the updating factors Fs(i+1)(ks) are proposed and discussed in
the following:
74
Input data:
3
Target Cross{Spectral Density Matrix: S0 (k) a. Target Spectral Density Function
Prescribed Probability Distribution Functions:
SDF
2
FBr , r = 1 2 : : :m
0.5
Assume initial SDF to generate Gaussian
?
0
2
nD{mV eld: 1.5
1
0.5
G (k) = S (k)
S(1)
0
0 0
and set iteration counter i = 1
3
b. Computed SDF of non-Gaussian component #1
i := i + 1 SDF
(first iteration)
2
- -
Generate Gaussian sample eld using FFT:

1
?
0.5
(i) ( i)
SGrr (k) fGr (x) r = 1 2 : : :m
0
2
1.5
1
0.5
0
0
Transform Gaussian eld into non{Gaussian,

?
for each component

n h
r of theionD{mV eld: c. SDF used to generate Gaussian component #1
(i) ;1 FG f (i) (x)
fBr (x) = FBr
4
Gr (last iteration)
3
SDF
Compute SDF of each component of

?
2
transformed non{Gaussian vector eld:

(i) (i)
fBr (x) SBrr (k) r = 1 2 : : :m
1
0.5
Check convergence
0
for r = 1 2 : : :m
?
P
2
1.5
1
YES
0.5
P
0
(i)
SBrr (k) S 0 (k)
P
0
STOP
P
-
P
P
3
d. Computed SDF of non-Gaussian component #1

P
(last iteration)
NO
SDF
2
Update SDF used to generate Gaussian

?
1
nD{mV eld:
0.5
(k) S(i)(k)
0
(i+1) (i)
SGrr (k) = SGrr
0
SBrr (k)
1
)
(m -
2
1.5
r = 1 2 : : :m 1
κ2 (m -1) 0.5
1
κ
0
Figure 3.9: Flowchart for numerical simulation of nD{mV non{Gaussian stochastic

elds, and illustration of dierent computational phases of proposed iterative process.
75
1. Global correction:
Xm S 0 (k)
Fs(i+1)(ks) = m1 s = 1 2 : : : m (3.59)
r=1 SBrr (k)
(i)
where k1 : : : km and k are in the same n{dimensional interval k.

2. Individual correction:
Fs(i+1)(ks) = S(i)(k )s s = 1 2 : : : m
0 s
(3.60)
SBss(k )
Accounting for the way each component fs(x) of the multi{variate stochastic
eld f (x) is simulated { Eqn. (3.29) { the individual correction is \exact" (in
the sense discussed by Yamazaki and Shinozuka, (1988) for univariate elds)
for the scalar component f1(x).
In the case of global correction, a unique updating factor is used for all scalar
components (F1(i+1)(k1) = = Fm(i+1)(km)), while in the second case a specic
updating factor is evaluated for each scalar component s.
An example, based on the results of the numerical simulation of an 1D{2V non{
Gaussian stochastic eld, is used to analyze the eectiveness of the proposed updating
procedures. Various situations, involving resulted eld distribution and correlation
between scalar components are selected for exemplication. Two extreme situations
related to the skewness coe
cient
1 of the resulted cumulative distributions are
analyzed:
{ Case A (symmetric distributions) { eld #1 { Beta distribution with p = 4 q = 4
and
1 = 0, and eld #2 Beta distribution with p = 3 q = 5 and
1 0:3.
{ Case B (skewed distributions) { eld #1 { Lognormal distribution with
1 3:0,
and eld #2 Beta distribution with p = 1 q = 7 and
1 1:21.
For each of these cases, two situations are analyzed, involving the input correlation
coe
cient between the scalar components:
1 The skewness coe cients
1 are evaluated for the actual resulted distributions
76
{ Case 1 (low correlation) { 12 = 0:2.

{ Case 2 (high correlation) { 12 = 0:8.
a. Case A1: b. Case A2:

- symmetric distributions (γ1 = 0 . . . 0.3) - symmetric distributions (γ1 = 0 . . . 0.3)
0
0
10
10
- low correlation (ρij = 0.2) - high correlation (ρij = 0.8)
6 8
6 8
4
4
individual corr.
error sum: ε1 + ε2
global correction
2
2
-1
-1
10
10
6 8
6 8
individual corr.
global correction
4
4
2
2
-2
-2
10
10
0 5 10 15 20 0 5 10 15 20
c. Case B1: d. Case B2:

- skewed distributions (γ1 = 1.2 . . . 3.0) - skewed distributions (γ1 = 1.2 . . . 3.0)
0
0
10
10
- low correlation (ρij = 0.2) - high correlation (ρij = 0.8)

6 8
6 8
4
4
error sum: ε1 + ε2
2
-1
-1
10
10
6 8
6 8
individual corr. individual corr.

global correction global correction
4
4
2
2
-2
-2
10
10
0 5 10 15 20 0 5 10 15 20
iteration iteration
Figure 3.10: Comparison between updating procedures { simulation of an 1D{2V
stochastic eld.
The results of the study, presented in Figure 3.10 in terms of sum of relative errors
" = "1 + "2 resulted from the two scalar components plotted vs iteration number,
show that the individual correction procedure is more eective in all cases. The
convergence speed is also dependent on the resulted skewness coe
cient and on the
degree of correlation between the scalar components: both the minimum (residual)
relative error and the number of iterations necessary to achieve it are increasing with
increasing skewness and correlation coe
cients. The explanation is that:
77
{ higher the skewness, more nonlinear is the mapping relation (3.54), and hence a
larger number of iterations is necessary (Figures 3.10,a and b)
{ higher the correlation between scalar components, larger is the interference be-
tween errors induced by each scalar component, while the individual correction
procedure accounts separately for these errors, and hence both the number of
iterations and the residual error is higher (Figures 3.10,a and b on the one hand,
and c and d on the other hand).
More details on how the individual correction procedure handles the relative errors
between resulted and target SDF are presented in Figure 3.11:
b. correlated fields - ρij = 0.7
100
100
a. non-correlated fields
5
field #1 field #1 - Lognormal

relative error - εr (r=1,2,3,4)
field #2 field #2 - Beta (1,7)

10-1
10-1

5
5
10-2
10-2
5
5
10-3
10-3
0 5 10 15 20 0 10 20 30 40 50
iteration iteration
Figure 3.11: Convergence analysis using the individual correction procedure { simula-
tion of an 1D{4V stochastic eld: a. non{correlated scalar components b. correlated
scalar components.
{ in the case of non{correlated scalar components (Figure 3.11,a) the convergence is

relatively rapid for all components, and a relative error "r 10;2 r = 1 2 3 4
is achieved in 8 iterations
{ for correlated scalar components (Figure 3.11,b) both the necessary number of
iterations (30 iterations to achieve "r 10;2 ) and the residual errors are higher
than in the rst case it is noteworthy that the convergence behavior of the rst
78
scalar component (r = 1), which is not aected by correlation with the other
components, is the same as in the case of non{correlated elds.
3.3.3 Recovery of the o

{diagonal terms in the target spec-
trum
Based on the assumptions stated in x3.1, the o{diagonal terms of the target spectrum
S0(k) are expressed in Eqn. (3.20) as: Sik0 (k) = 0ik S 0(k), i.e. as a product between the
target SDF, common to all scalar components, and the prescribed sample correlation
coe
cients between the scalar components fi(x) and fk (x). The correlation between
scalar components is determined by the way univariate elds fs(x) are combined
in Eqn. (3.29). Consequently, in case of Gaussian eld simulation, the prescribed
correlation between scalar components is preserved (as shown in Appendix 3.B, for
the case of a 2V eld, and in Figure 3.8,b for a 2V{3D eld), and the o{diagonal
terms of the target spectrum are recovered with good accuracy.
However, in the case of non{Gaussian eld simulation, some of the initial corre-
lation between scalar components is lost due to the nonlinear transformation (3.51),
as shown in Appendix 3.C and Figure 3.27 for a 2V Lognormal stochastic eld.
Therefore, to obtain in the end the desired correlation coe
cient 0ik between scalar
components, one should initially prescribe larger values 0ik . For example, for the case
described in Appendix 3.C, in order to recover the value 012, the prescribed value has
to be:
h i
012 = ln 012(e ; 1) + 1 (3.61)
Unfortunately, for most distribution functions there are no such analytical solu-
tions. On the other hand, some uncertainties in recovering the correlation coe
cient
between generating Gaussian elds fGr (x) are induced by possible correlation be-
tween the univariate elds fs(x). An attempt to an empirical solution is presented
in Figure 3.12: the results obtained from various numerical simulations are plotted
79
1.2
1
0.8
/ ρik
0
0.6
computed
data from application in Section 3.2.5

data from application in Section 3.3.2
data from applications in Section 3.3.4
ρik
0.4
data resulted from other simulations

linear regression: y = 1.0 - 0.086x
2
quadratic regression: y = 0.99 - 0.049x
0.2
+/- standard deviation for linear regression (σ = 0.061)

+/- standard deviation for quadratic regression (σ = 0.061)
0
0 0.5 1 1.5 2 2.5 3

med
average skewness coefficient: γ1,ik = (γ1,i + γ1,k)/2
Figure 3.12: Ratio between resulted and prescribed correlation coe

cients vs. average
skewness coe
cients of the simulated scalar components.
computed
as ratio ik 0 vs. the average
1med ik =

1 i +
1 k of the skewness coe
cients of
2
ik
resulted scalar components fBi(x) and fBk (x). (Here, the skewness coe
cient is con-
sidered a measure of the nonlinearity in Eqn. (3.51) ). One can use either the linear
or the quadratic regression curves shown in Figure 3.12 to compute the prescribed
value as a function of the desired correlation coe
cient, as:
0ik = 1 ; 0:086
0
ik (3.62)

1med
ik
for the linear regression, or:
0ik = 0:99 ; 0:049(
0
ik (3.63)

1med
ik )
2
if using the quadratic regression.
3.3.4 Applications { Simulation of non{Gaussian vector elds

The results obtained from numerical simulation of several non{Gaussian stochastic
elds (1D{4V, 2D{3V and 3D{2V elds) are discussed in the following. Various types
80
of spectral density functions, having separable correlation structure, are adopted as

target SDF. The scalar components of the simulated elds have dierent prescribed
cumulative distributions, ranging from symmetric distributions (Gaussian, Uniform,
Beta with p = q), to skewed distributions like Beta with p = 2 and q = 6, and
Lognormal.
0.6
Target SDF
-3
Resulted SDF - Field #1 - ε1 = 1.8 x 10
-3
0.4
-3
-3
0.2
0
0 0.5 1 1.5 2 2.5

-1
wave number - k(m )
Figure 3.13: Simulation of an 1D{4V non{Gaussian eld { comparison between target
and resulted spectral density functions (the relative errors shown in the legend are
computed after 20 iterations).
Simulation of an 1D{4V stochastic eld

1. Input data:
{ Target SDF: derived from Gaussian correlation structure, with correlation distance
= 2m (Figure 3.13):
8 !29
< =
S 0(k) = 2p exp :; k2 (3.64)
{ Cross{spectral density terms: various correlation coe

cients 0ij = 0:2 : : : 0:7 be-
tween scalar components are prescribed, and the matrix s in Eqn. (3.21) is:
0 1:0 0:5 0:4 0:2 1
B C
s = BB@ 00::54 10::07 01::70 00::36 CCA (3.65)
0:2 0:3 0:6 1:0
81
{ Prescribed cumulative distributions:

{ component #1 { Lognormal distribution
{ component #2 { Beta distribution with p = 2 and q = 6
{ component #4 { Gaussian distribution.
{ Discretization:
{ in wave{number domain: ku = 2:5s;1 N = 29 and k = 0:0049s;1
{ in space: L0 = 5144m, M = 211 and x = 0:63m.
2. Simulation results (Figures 3.13 : : : 3.16):

{ Comparison between target SDF and the spectral density functions of the non{
Gaussian eld, resulted after 20 iterations { in Figure 3.13.
{ 100m long portions of the simulated scalar components, plotted along with the
components of the original Gaussian eld simulated at the rst iteration (i.e.
(20) (1)
fBr (x) vs. fGr (x) r = 1 2 3 4) are shown in Figure 3.14.
{ Comparisons between resulted cumulative distributions of each scalar component

vs. prescribed distributions are presented in Figure 3.15.
{ The relative error evolution during the iterative process is shown in Figure 3.16,a.
For all the components, the relative error is less than 1% ("r 10;2 r =
1 2 3 4) after 10 iterations.
{ Prescribed and resulted correlation coe

cients between scalar components are
shown in Figure 3.16,b. Excepting for 14, all the resulted sample correlation
coe
cients are within the one standard deviation ranges shown in Figure 3.12.
82
a. Scalar component #1 - Lognormal distribution
6
non-Gaussian field
Gaussian field at the first
3 iteration (for comparison)
f1(x)
0
-3
100 120 140 160 180 200
b. Scalar component #2 - Beta distribution with p=2 and q=6

3
non-Gaussian field
Gaussian field
0
f2(x)
-3
100 120 140 160 180 200
c. Scalar component #3 - Beta distribution with p=3 and q=5

3
non-Gaussian field
Gaussian field
0
f3(x)
-3
100 120 140 160 180 200
d. Scalar component #4 - Gaussian distribution

3
final result
initial Gaussian field
0
f4(x)
-3
100 120 140 160 180 200

x(m)
Figure 3.14: Simulation of an 1D{4V eld { part of the resulted eld (total simu-
lated length L0 = 5144m) comparison between non{Gaussian components fBr(x)
(continuous lines) and the corresponding Gaussian components generated at the rst
(1)
iteration fGr (x) (dotted lines).
83
a. Field #1 - Lognormal distribution b. Field #2 - Beta distribution (p=2, q=6)
1
cumulative distribution
0.5
0.5
prescribed
resulted prescribed
stand. normal resulted
(for comparison)
0
0
-3 0 3 6 -3 0 3
c. Field #3 - Beta distribution (p=3, q=5) d. Field #4 - Gaussian distribution

1
1
0.5
0.5
prescribed prescribed
resulted resulted
0
-3 0 3 -3 0 3
stochastic field variable stochastic field variable
Figure 3.15: Simulation of an 1D{4V eld { comparison between prescribed and
resulted cumulative distributions of the scalar components.
a. Solution convergence b. Correlation coefficients

0
10
field #1 - Lognormal ρ12

5
relative error - εr (r=1,2,3,4)
field #2 - Beta (2,6) ρ13

-1
field #3 - Beta (3,5)

10
ρ23
field #4 - Gaussian
5
ρ14
-2
resulted
ρ24
0.5
10
5
ρ34
-3
10
5
-4
10
0 5 10 15 20 0 0.5 1
iteration input
Figure 3.16: Simulation of an 1D{4V eld { recovery of the target spectrum: a. rela-
tive errors of resulted SDF with respect to the target SDF during the iterative process
b. input and resulted correlation coe
cients.
84
Simulation of a 2D{3V stochastic eld

1. Input data:
{ The target SDF and the discretization in both wave{number and spatial domain
are identical to those used for the application in x3.2.5.
{ Cross{spectral density terms: 012 = 0:5 013 = 0:4 and 023 = 0:7.
{ component #1 { Uniform distribution
{ component #3 { Gaussian distribution.

{ The resulted spectral density functions of the three simulated scalar components
are compared with the target SDF in Figures 3.17 (for the whole two{dimensional
wave{number domain) and 3.18 { along the wave{number space diagonal. The
resulted SDF are shown after 25 iterations.
{ Prescribed and resulted cumulative distributions of the simulated scalar compo-

nents are compared in Figure 3.19.
{ The relative errors between resulted and target SDF are less than 1% for all com-
ponents ("r 10;2 r = 1 2 3) after 13 iterations (Figure 3.20,a).
{ Given the relatively low skewness of the target cumulative distributions, the re-
sulted sample correlation coe
cients between scalar components are in a close
range of the prescribed values (computed
ik =0ik = 0:95 : : : 1:01), as shown in Fig-
ure 3.20,b.
85
2.5
2.5
Target spectral density function SDF - Field #2 - Beta distribution (p=2, q=6)
2
2
target spectrum
1.5
1.5
SDF #2
1
1
tal 0.5
tal 0.5
0.5
0.5
zon .4
zon .4
ori 3 0
ori 3 0
0
0
- h 0.
- h 0.
k1.1 0.2
k1.1 0.2
2 2
1.5 1.5
1
k2 - vertica 1
k2 - vertica
0
0
l direction 0.5 l direction 0.5
0
0
0 0
2.5
2.5
SDF - Field #1 - Uniform distribution SDF - Field #3 - Gaussian distribution
2
2
1.5
1.5
SDF #1
SDF #3
1
1
tal 0.5
tal 0.5
0.5
0.5
zon .4
zon .4
ori 3 0
ori 3 0
0
0
- h 0.
- h 0.
k1.1 0.2
k1.1 0.2
2 2
1.5 1.5
k2 - verti 1 k2 - verti 1
0
0
cal direct 0.5 cal direct 0.5
ion ion
0
0
0 0
Figure 3.17: Simulation of a 2D{3V non{Gaussian eld { comparison between target

and resulted spectral density functions
2.4
Target SDF
-3
2
-3
-3
1.6
1.2
0.8
0.4
0
0 0.3 0.6 0.9 1.2 1.5

-1
wave number - k(m )
Figure 3.18: Target and resulted spectral density functions represented along the
wave{number space diagonal (the relative errors shown in the legend are computed
after 25 iterations).
86
a. Field #1 - Uniform distribution b. Field #2 - Beta distribution (p=2, q=6)
1
0.5
0.5
resulted resulted
0
0
-2 0 2 -3 0 3
stochastic field variable
c. Field #3 - Gaussian distribution
1
0.5
prescribed
resulted
0
-3 0 3
stochastic field variable
Figure 3.19: Simulation of a 2D{3V eld { comparison between prescribed and re-
sulted cumulative distributions of the scalar components.
a. Solution convergence b. Correlation coefficients

0
10
field #1 - Uniform ρ12

relative error - εr (r=1,2,3)
field #3 - Gaussian ρ23

-1
10
resulted
0.5
-2
10
-3
10
0 5 10 15 20 25 0 0.5 1
iteration input
Figure 3.20: Simulation of a 2D{3V eld: a. evolution of relative errors during the
iterative process b. input and resulted correlation coe
cients.
87
Simulation of a 3D{2V stochastic eld

1. Input data:
{ A separable target SDF is adopted: S o(k) = S110 (k1) S220 (k2) S330 (k3), with Sii0(ki)
given in Eqn. (3.49), and: ai = 2 i = 1 2 3 1 = 2 = 10m 3 = 2m.
{ The prescribed correlation coe
cient between simulated scalar components is 012 =
0:5.
{ Discretization:
{ in wave{number domain: ku1 = ku2 = 0:5s;1 ku3 = 2s;1 N1 = N2 = N3 =
25 k1 = k2 = 0:0156s;1 k3 = 0:0625s;1 .
{ in space: L01 = L02 = 792m L03 = 198m M1 = M2 = M3 = 26 x1 =
x2 = 6:28m x3 = 1:57m.
{ The comparison between target and resulted spectral density functions is presented
in Figure 3.21 on three 2D sections of the 3D wave{number domain, at various
values of the wave{number k3 (in vertical direction). In Figure 3.22, the spectra
are also plotted along the wave{number space diagonal. Values resulted after
5 iterations are shown in these gures.
{ Prescribed and resulted cumulative distributions of the simulated scalar compo-
nents are compared in Figure 3.23.
{ The relative errors between resulted and target SDF are less than 1% ("1 "2
10;2 ) after 4 iterations (Figure 3.24,a).
88
{ The resulted correlation coe

cient { computed
12 = 0:495 { is very close to the pre-
scribed value, as illustrated in Figure 3.24,b.
Simulation of a 2D{2V eld representing the spatial distribution of in{situ

test results from Tarsiut P{45
Another example, involving a two{dimensional, two{variate (2D{2V), non-Gaussian
stochastic eld is presented in the following. The sample functions of the simulated
vector eld represent possible realizations of standardized cone tip resistance, qc, and
soil classication index, Ic, in the hydraulic ll layer at Tarsiut P{45. In accordance
with the separable correlation structure assumption (x2.2.2), the expression for the
(common) spectral density function S 0(1 2) is of the form Eqn. (2.18)]:
S 0(1 2) = S110 (1) S220 (2) (3.66)
where Sii0 (i) is the one-dimensional spectral density function in spatial direction i.
The expressions for the one-dimensional spectral density functions in each spatial
direction i become:
Sii0 (i) = b5i 2;(b 1 + 1) bb12i +1b2 e;b1 jj + (1 ; b5i) 2;(b 1 + 1) bb34i +1 b4 e;b3 jj(3.67)
i i i i i i
2i 4i
where the parameters bji j = 1 : : : 5 i = 1 2 are estimated using eld test results
from the hydraulic ll layer, as shown in x2.3.4 for i = 1 and in x2.3.3 for i = 2.
Upper cut-o wave numbers, beyond which the elements of the cross-spectral
density matrix are assumed to be zero (e.g. Deodatis, 1995), are specied as: u1 =
1:33 m;1 and u2 = 4:0 m;1. The prescribed correlation coe
cient between the
scalar components is 12 = ;0:58 (x2.3.5). Finally, the prescribed probability dis-
tribution functions (zero-mean and unit-standard-deviation) for the two components
of the vector eld are Beta distributions with parameters derived from eld data
analysis, as shown in x2.3.2.
The target SDF is compared in Figure 3.25,a with the spectral density functions
computed from the components of a simulated sample function (obtained at the end
89
Stochastic Variability of Soil Properties
90
Radu Popescu, 1995
Figure 3.21: Simulation of a 3D2V non-Gaussian field – comparison between target and resulting spectral density functions.
8
Target SDF
-3
-3
6
4
2
0
0 0.3 0.6 0.9 1.2 1.5

2 2 2 0.5
wave number - (k1 + k2 + k3 )
Figure 3.22: Target and resulted spectral density functions represented along the
wave{number space diagonal (the relative errors shown in the legend are computed
after 5 iterations).
a. Field #1 - Beta distribution (p=2, q=6) b. Field #2 - Beta distribution (p=4, q=4)
1
1
0.5
0.5
resulted resulted
0
-3 0 3 -3 0 3
stochastic field variable stochastic field variable
Figure 3.23: Simulation of a 3D{2V eld { comparison between prescribed and re-
sulted cumulative distributions of the scalar components.
91
a. Solution convergence b. Correlation coefficient
10-1
1
6 8
relative error - εr (r=1,2)
4
2 field #2 - Beta (4,4)
10-2
resulted
0.5
6 8
4
2
10-3
0
1 2 3 4 5 0 0.5 1
iteration input
Figure 3.24: Simulation of a 3D{2V eld: a. evolution of relative errors during the
iterative process b. input and resulted correlation coecients.
a. Target and resulted spectral density functions b. Prescribed and resulted

0.15
cross-correlation coeff.
Target SDF
0
Resulted SDF - qc - εr = 0.27%
0.1
Resulted SDF - Ic - εr = 2.21%

SDF
prescribed
ρ12 = -0.58
ρ12 = -0.565
0.05
-1
0
0 2 4 6 -1 resulted 0
2 2 0.5 -1
wave number - [κ1 + κ2 ] (m )
c. Prescribed and resulted probability density functions

1
1
0.5
0.5
resulted resulted
0
-2 0 2 4 -2 0 2 4
standardized cone tip resistance - qc standardized soil classification index - Ic
Figure 3.25: Digital simulation of a 2D{2V non{Gaussian stochastic eld representing

the eld conditions at Tarsiut P{45: comparison between prescribed and resulted
statistics of a sample eld.
92
Table 3.1: Solution time and required memory

Simulated Number of points in Memory Solution time for "r 2%
eld spatial discretization (Kbytes) (r = 1 2 : : : nr. of variables)
1D{4V 4 211 = 8192 740 10 it. 5s./it. = 50s.
2D{3V (3 27)2 = 147 456 7,940 13 it. 60s./it. = 13min.
3D{2V (2 26)3 = 2 097 152 80,750 4 it. 22min./it. = 1h 28min.
2D{2V (2 29)2 = 1 048 576 48,270 5 it. 9min./it. = 45min.
of the iteration process). Relative errors between prescribed and resulted SDF's of
the order of 2% are obtained after 5 iterations. The comparison is presented in two
dimensions, along the wave number space diagonal. The prescribed and resulted
values of the cross{correlation coecient are shown in Figure 3.25,b. The prescribed
probability distribution functions for the components of the vector eld are compared
with the corresponding probability distribution functions computed from the sample
function in Figure 3.25,c.
Computational e ort
The computational eort is presented in Table 3.1 in terms of solution time on an IBM
minicomputer and necessary memory for an in{core solution, using double precision
computations:
3.4 SINOGA { Program for simulation of nD{

mV non{Gaussian stochastic elds
3.4.1 General presentation
The algorithm described in the previous sections is implemented in the computer pro-
gram SINOGA (SImulation of NOn{GAussian elds), written in FORTRAN. In its
actual version, SINOGA is an independent code, with its own input/output sections.
However, with minimal changes, the program can be attached to DYNAFLOW, to
generate correlated stochastic elds for soil material properties to be used for lique-
93
faction analyses.
The main features of the program SINOGA are as follows:
An iterative process is used to generate multi{dimensional, multi{variate (nD{
mV) stochastic elds with zero mean and unit standard deviation. The m scalar
components have a given common n{dimensional spectral density function {
target SDF (more precisely: the iterative process ends when the spectral density
function of each component is close enough to the target SDF). There are also
prescribed cumulative distribution functions for each scalar component of the
simulated eld.
The required correlation between the scalar components of the simulated eld is
achieved with very good precision for variables having symmetric distributions.
As the skewness coecient of the probability distribution increases, some of
the initial correlation is lost in the process of mapping { Eqn. (3.51) { and a
correction is necessary, as discussed in x3.3.3.
The target SDF can be (1) computed, starting from given correlation structures
(which can be dierent for dierent spatial directions), or (2) read from a le
(e.g. to use a sample SDF evaluated on the basis of eld measurements).
Similarly, the cumulative distribution of each scalar component can be either
specied as an existent probability model, or input as a sample cumulative dis-
tribution. Dierent distributions can be assigned to various scalar components.
The input data contains : (1) control variables (number of spatial dimensions,
number of variables, i/o options) (2) information on the target SDF and cross{
spectral terms (3) information on the cumulative distributions of the scalar
components and (4) convergence check options.
The output consists of the simulated vector eld and resulted spectral density
function of each scalar component.
94
3.4.2 Computational eort

The computation time is reduced to a minimum (1) by using the Fast Fourier Trans-
form technique to generate the stochastic eld { Eqn. (3.18) { and to compute the
resulted spectra { Eqn. (3.46) { and (2) by using a number of points equal to a power
of 2 in both wave{number and spatial domains, which enhances the eectiveness of
the FFT algorithm, especially for multiple dimensions.
The time necessary for one iteration vary as a function of (1) the number of
spatial dimensions and variables, (2) mesh size, and (3) type of selected cumulative
distributions. Some values are presented in x3.3.4. The number of iterations necessary
for achieving a certain result accuracy increases with the skewness coecient of the
target distributions and with the required correlation between scalar components, as
shown in x3.3.2.
An in{core solution is selected for the present version of the program. This speeds
up the computations and is eective for relatively high accuracy simulations, involving
a larger number of iterations, but requires a large amount of memory. The memory
requirements for double precision computations, as a function of problem dimensions
are presented in Figure 3.26.
It is mentioned that, at this stage, no special eort has been made to reduce
memory requirements at the expense of computational time and/or result accuracy.
However, for real problems involving simulation of spatial distribution of soil proper-
ties and starting from a sample target SDF, which is inherently evaluated with some
(usually large) approximation, the required computational accuracy can be reduced,
which leads to a signicant drop in the number of iterations. The required memory
can be also substantially reduced:
{ by 50% if single precision calculations are performed
{ by another 50 { 70% if using external storage and/or recomputing some arrays at
every iteration.
95
5
10
necessary memory (Kbytes)
4
10
1D-4V
3
10
1D-5V
1D-6V
2D-2V
2D-3V
2
2D-4V
10
3D-1V
3D-2V
3D-3V
1
10
24 26 28 2 10 2 12 2 14 2 16
number of points in one spatial direction
Figure 3.26: Memory requirements (upper bounds) for in{core double precision
computations vs. the number of points in one spatial direction Mk (for the case
M1 = : : : = Mn).
3.5 Conclusions
1. An algorithm for digital simulation of multi{dimensional, multi{variate, non{
Gaussian stochastic elds is presented. Sample functions are generated accord-
ing to a prescribed cross-spectral density matrix and prescribed (non-Gaussian)
probability distribution functions, using an extension of the spectral represen-
tation method.
2. Multi{dimensional, Gaussian vector elds are rst generated using the spectral
representation method, and taking advantage of the Fast Fourier Transform for
great computational eciency. The performance of the proposed algorithm is
further enhanced by making a series of assumptions that are appropriate for
random media representing material properties. A procedure to compute the
power spectral density function from sample vector elds is also presented.
96
3. Sample functions of a non{Gaussian vector eld that are compatible with (1) a
prescribed cross{spectral density matrix, and (2) prescribed (non{Gaussian)
probability distribution functions, are obtained using memoryless nonlinear
transformations and an iterative scheme.
4. The eectiveness of proposed algorithm in reproducing (1) the target spectral

density function, (2) the prescribed cross{correlations between scalar compo-
nents, and (3) the prescribed probability distribution function of each compo-
nent, is veried by a series of numerical examples.
5. The algorithm for simulation of non{Gaussian vector elds is implemented in

a computer program. Input/output options, numerical results accuracy, and
computational eort (CPU time and memory requirements) are discussed.
Appendix 3.A Explanation of Eqn. (3.38)

Computation of the spectral density function according to Eqn. (3.38) for the
second scalar component (r = 2) of an 1D{2V stochastic eld (n = 1 m = 2):
S22(k) = 2L 1 jh F (k) + h F (k)j2

21 1 22 2
0
= 2L 1 h(h ReF + h ReF )2 + (h ImF + h ImF )2i
21 1 22 2 21 1 22 2
0
= 2L 1 hh2 (ReF )2 + h2 (ImF )2 + h2 (ReF )2 + h2 (ImF )2
21 1 21 1 22 2 22 2
0
+ 2h21h22 (ReF1 ReF2 + ImF1 ImF2 )] (3.68)
= 2L 1 hh2 jF (k)j2 + h2 jF (k)j2i
21 1 22 2
0
because the cross products are zero (i.e. ReF1 ReF2 = 0, ImF1 ImF2 = 0) since:
! (
k
F1 t 2 is = 0 if t = 2l
6!= 0 (if t = 2l + 1 and
F2 t 2k is 6= 0 if t = 2l
= 0 if t = 2l + 1 (3.69)
l = 0 1 : : : M ; 1
97
Appendix 3.B Correlation between the components of a 2V

Gaussian eld
For a 2V eld (m = 2), the coecients hrs in Eqn. (3.29) are: h11 = 1 h12 =
q
012 h22 = 1 ; (012)2, with 012 { the prescribed correlation coecient between the
scalar components. Consequently, Eqn. (3.29) becomes:
fG1(x) = f1 (x)
q
fG2(x) = 012f1 (x) + 1 ; (012)2f2 (x) (3.70)
The resulted sample correlation coecient 12 between the zero mean unit stan-
dard deviation scalar components of the simulated eld is:
q
12 = E fG1 (x)fG2(x)] = E f1 (x) 012f1 (x) + 1 ; (012)2f2 (x)
q
= 12E f1 (x)f1 (x)] + 1 ; (012)2E f1 (x)f2 (x)]
0 (3.71)
q
= 012 + 1 ; (012)212
where 12 is the correlation coecient between f1 (x) and f2 (x)2.
Consequently, if the univariate elds f1 (x) and f2 (x) are truly non{correlated
(12 = 0), then the prescribed correlation between scalar components is achieved by
the simulated Gaussian eld (12 = 012).
Appendix 3.C Correlation between the components of a 2V
Lognormal eld
Let 012 be the prescribed correlation coecient, and assume that f1 (x) and f2 (x)
are truly non{correlated. Then E fG1(x)fG2(x)] = 012, where fG1(x) and fG2(x) are
the generating Gaussian elds, having zero mean and unit standard deviation. Using
the relations for the statistics of a Lognormal random variable Y which is derived
from a Gaussian random variable X as Y = expfX g:
1
E Y ] = exp E X ] + 2 X 2 (3.72)
2 A similar relation stands for the general case, with arbitrary mean and standard deviation.
98
h n o i n o
Y2 = exp X2 ; 1 exp 2E X ] + X2 (3.73)
one obtains the transformed Lognormal elds with zero mean and unit standard
deviation as:
exp f f (x )g ; pe
Gr
fBr(x) = q r = 1 2 (3.74)
e(e ; 1) 1
resulted ratio: r = ρ12 / ρ12
0
0
ρ 0
0.9
r = ( e 12 - 1 ) / [ ρ12 ( e - 1 ) ]
0.8
0.7
0.6
0.5
0 0.2 0.4 0.6 0.8 1

0
prescribed correlation coefficient: ρ12
Figure 3.27: Ratio between resulted and prescribed correlation coecients vs. pre-
scribed correlation coecient, for a 2V Lognormal eld.
The resulted sample correlation coecient between the Lognormally distributed
scalar components has the expression:
12 = E fB1(x)fB2(x)] = e(e 1; 1) (E exp ffG1(x) + fG2(x)g] ; e) (3.75)
Let G(x) = fG1 (x) + fG2(x) be a Gaussian eld with zero mean and variance:
h i
G2 = E (fG1(x) + fG2(x))2 = 2(1 + 012) (3.76)
By replacing Eqn. (3.76) in (3.75) and using Eqn. (3.72) the correlation coecient
between the scalar components of a Lognormal 2V eld is obtained as:
12 = exp fe;121g ; 1
0
(3.77)
99
The ratio 12=012 computed using Eqn. (3.77) is plotted in Figure 3.27 as a function
of the prescribed correlation coecient 012.
Appendix 3.D List of notations for Chapter 3

ai bi { shape parameters of the target SDF
f (x) { n{dimensional m{variate stochastic eld
fB (x) { non{Gausian eld
fG (x) { Gausian eld
fs (x) { non{correlated univariate Gausian elds
fr (x) { rth scalar component of the vector eld f (x)
FB { non{Gaussian distribution function
FG { Gaussian distribution function
Fs(i+1) { updating factor at iteration (i)
Fs (k) { n{dimensional Fourier Transform of fs (x)
Fr (k) { n{dimensional Fourier Transform of fr (x)
grs (x) { auxiliary functions used to simulate the multi{variate elds
h { lower triangular matrix resulted from Cholesky decomposition of s
H(k) { lower triangular matrix resulted from Cholesky decomposition of S0(k)
i { index for spatial dimensions
(i) { iteration counter
Ik { set of indicators for the quadrants of wave{number space, taking
values 1
k { index for spatial dimensions
k { position in the n dimensional wave{number domain (wave{number
vector)
kku { upper cut{o wave{number in direction k
k~r k~krl +1 { wave numbers at which resulted SDF are evaluated (vector and
k
components)
100
ks kksl k
{ wave{numbers at which generating SDF are evaluated (vector and
components)
lk { counter for the discretization in wave{number domain
L0k { period of the simulated eld in spatial direction k
m { number of variables
Mk { number of points in direction k, in the spatial discretization
n { number of spatial dimensions
Nk { number of points to describe the spectral density functions in
direction k
p q { parameters of Beta distribution
pk qk { counters for the discretization in spatial domain
r { index for spatial dimensions
R0( ) { cross{correlation matrix
R0( ) { (common) auto{correlation function
s { index for spatial dimensions
s { matrix of correlation coecients
SDF { spectral density function
S0(k) { target cross{spectral density matrix (target spectrum)
S 0(k) { (common) diagonal term of the target spectrum (target SDF)
SBrr(k) { resulted SDF of the rth scalar component of fB (x)
SGrr (k) { resulted SDF of the rth scalar component of fG(x)
Sss (k) { SDF of fs (x)
tk { counter for the discretization in wave{number domain
x { position in the n dimensional spatial domain (space vector)
1 { skewness coecient
kk { wave{number increment in direction k
xk { spatial increment in direction k
"r { relative error of the resulted SDF of scalar component r with respect
101
to the target SDF

i { correlation distance in direction i
{ space{lag vector
jk { sample correlation coecient between scalar components j and k of
the stochastic eld
{ standard normal distribution function
Isl1 I12l2:::I:::l
n
n
{ sets of random phase angles
102
Chapter 4
Evaluation of soil constitutive
parameters
According to the ow chart in Figure 1.2, the values of relevant in{situ test results
generated over the analysis domain as shown in Chapter 3 are used to evaluate
the constitutive parameters of the soil constitutive model used for stochastic input
nite element analyses. The constitutive model selected here for that purpose is the
multi{yield plasticity model (Prevost 1985, 1989, 1993) implemented in the computer
program DYNAFLOW (Prevost, 1981). The model performance in predicting the
behavior of soil deposits and soil{structure systems under seismic excitation has been
validated in the past, as shown in x4.2. The constitutive parameters of the multi{
yield plasticity model are expressed in terms of traditional soil properties. A short
description of the model and of its parameters is presented in x4.1.
4.1 The multi{yield plasticity model

4.1.1 Basic relations
The multi{yield constitutive soil model (Prevost 1989, 1993) is a kinematic hardening
model based on a relatively simple plasticity theory (Prevost, 1985) and is applicable
to both cohesive and cohesionless soils. The concept of a \eld of work{hardening
moduli" (Iwan, 1967, Mr"oz, 1967, Prevost, 1977) is used by dening a collection
f0 f1 : : :fn of nested yield surfaces in the stress space (Figure 4.1,a). Drucker{
103
Prager type surfaces are employed for frictional materials (sands).

The yield function is described for each yield surface as (Prevost, 1989):
s
fi(0 i Mi) = js ; p0ij + 23 Mi p0 i = 0 1 : : : n (4.1)
where 0 is the eective stress tensor, s = 0 ; p0 is the deviatoric stress tensor,
p0 = 13 tr0 is the eective mean normal stress and i are kinematic tensors dening
the coordinates of the yield surface centers in the deviatoric stress space. Mi are
material parameters denoting the size of each yield surface. The yield surfaces dene
regions of constant shear moduli in the stress space, and in this manner the model
discretizes the smooth elastic{plastic stress{strain curve into n linear segments. The
outermost surface fn (called failure surface) corresponds to zero shear strength.
The plastic ow rule:
"_ p =<
_ > P (4.2)
where "_ p is the plastic strain rate, P is the plastic potential: P = P0 + P 00, and

_ is the plastic load factor, is associative in its deviatoric component. To account
for experimental evidence from tests on frictional materials (e.g. Lade, 1987), a
non{associative ow rule is used for the dilatational component (Figure 4.1,b):
=$)2 ; 1 X
3P 00 = ((= (4.3)
$)2 + 1 pp
where $ and Xpp are material parameters. Eqn. (4.3) expresses the dependence of
soil dilatational behavior on the mobilized stress ratio = ( 23 s : s) 21 =p0
The material hysteretic behavior and shear stress{induced anisotropic eects are
simulated by a purely kinematic hardening rule. Upon contact, the yield surfaces
are translated in the stress space by the stress point, as illustrated in Figure 4.1,c.
The direction of translation is selected such that the yield surfaces do not overlap,
but remain tangent to each other at the stress point.
The constitutive equations:
_ 0 = E : (_" ; "_ p) (4.4)
104
Figure 4.1: Main features of the multi-yield plasticity soil constitutive model.
105
where "_ is the rate of deformation tensor and E is the fourth order isotropic elastic
tensor, are integrated numerically using a stress relaxation procedure (Figure 4.1,d).
The return mapping algorithm proposed by Simo and Ortiz (1985) is modied for
the multi{yield plasticity case by Prevost (1989, 1993).
4.1.2 Constitutive model parameters

The required constitutive parameters of the multi{yield plasticity model are summa-
rized in Table 4.1 and are as follows:
1. State parameters: mass density of the solid phase s and, for the case of
porous saturated media, porosity nw and permeability k.
2. Low strain elastic parameters: low strain moduli G0 and B0, which are
related through the value of Poisson coecient :
G0 (1 + )
B0 = 23(1 (4.5)
; 2 )
The dependence of moduli on the eective mean normal stress p0, assumed in the
form:
!
0 n
!
0 n
p
G = G0 p0 B = B0 p0 p (4.6)
0 0
is accounted for by introducing two more parameters: power exponent n and reference
eective mean normal stress p00 .
3. Yield and failure parameters: these parameters describe the position i,
size Mi and plastic modulus Hi0, corresponding to each yield surface fi i = 0 1 : : : n.
For a given number n + 1 of yield surfaces, all these parameters can be evaluated
based on experimental stress{strain curves obtained from triaxial or simple shear soil
tests (Prevost, 1989, 1993, Ohbo et.al., 1990). However, stress{strain curves obtained
from laboratory tests may be aected by sample disturbance eects and stochastic
spatial variability of soil properties. Therefore generation of stress{strain curves based
on limited eld information is preferable (Prevost, 1989). A modied hyperbolic
expression proposed by Prevost (1989) and Griths and Prevost (1990) is used to
106
simulate soil stress{strain relations. The necessary parameters are: (1) the initial
gradient, and (2) the stress and strain levels at failure. The initial gradient is given
by G0 { the low strain shear modulus. The maximum strain level "max dev is estimated
from laboratory soil test results, and the stress level at failure is expressed as a
function of the friction angle at failure and the stress path. Recently, Hayashi et.al.
(1992) improved the modied hyperbolic model based on a study of experimental
stress{strain curves obtained for sands with a wide variety of grain size distributions.
They introduced a new parameter { { depending on the maximum grain size Dmax
and uniformity coecient Cu. Consequently, besides the low strain shear modulus
G0, three more parameters (friction angle at failure , maximum deviatoric strain
"max
dev , and stress{strain curve coecient ) are needed to describe the sizes Mi and
plastic moduli Hi0 of the yield surfaces. One more parameter { the coecient of lateral
stress k0 { is necessary to evaluate the initial positions i of the yield surfaces.
4. Dilation parameters: are used to evaluate the volumetric part of the plastic
potential 3P 00 in Eqn. (4.3). They are:
{ dilation (or phase transformation) angle $, related to the parameter $ in Eqn. (4.3)
by:
$
$ = 3 6sinsin$
C
E
(4.7)
where, in the triaxial stress state, C stands for triaxial compression and E for
triaxial extension, and for the general three{dimensional stress state, $ = $C if
tr s3 < 0 and $ = $E if tr s3 > 0 (Prevost, 1989)
{ dilation parameter Xpp, which is a scale parameter for the plastic dilation, and
depends basically on relative density and sand type (fabric, grain size).
Excepting for the dilation parameter, all the required constitutive model parame-
ters are traditional soil properties, and can be derived from the results of conventional
107
Table 4.1: Multi{yield plasticity model constitutive parameters

Constitutive parameter Symbol Type
Mass density - solid s State
Porosity nw parameters
Permeability k
Low strain elastic moduli G0 B0 Low strain
Reference eective mean normal stress p00 elastic
Power exponent n parameters
Friction angle at failure Yield and
Stress{strain curve coecient failure
Maximum deviatoric strain "max
dev parameters
Coecient of lateral stress k0
Dilation angle $ Dilation
Dilation parameter (cyclic) Xpp parameters
laboratory (e.g. \triaxial", \simple shear") and in{situ (e.g. \cone penetration",
\standard penetration", \wave velocity") soil tests (Prevost, 1989, 1993). The dila-
tion parameter Xpp is evaluated based on the results of liquefaction strength analysis,
as shown in x4.2.1.
4.2 Constitutive parameter evaluation

4.2.1 From laboratory test results and liquefaction strength
The methodology to evaluate soil constitutive parameters for the multi{yield plas-
ticity model based on triaxial soil test results is presented by Prevost (1985, 1989)
and Ohbo et.al. (1990). Actual estimation of constitutive parameters from results of
monotonic and cyclic laboratory soil test results is described by Popescu and Prevost
(1993a,b). A method to estimate the dilation parameter Xpp based on liquefaction
strength analysis and parametric studies using element tests is proposed by Popescu
et.al. (1992) and Popescu and Prevost (1993a), and is briey described hereafter.
First, liquefaction strength analysis is performed, which results in the value of
cyclic stress ratio (CSR) necessary to induce initial liquefaction in a given number
108
of stress driven cycles (usually NL = 15 cycles for seismic magnitude M = 7.5 – e.g. Seed
et al., 1983). The CSR value is derived either from the results of cyclic laboratory soil
tests (e.g. Popescu and Prevost 1992, 1993a,b), or from correlations with normalized
penetration resistance (N1 if using SPT results, or qn for CPT results) – e.g. Seed et al.
(1984), Robertson and Campanella (1985), etc. – Figure 4.2,a. If neither laboratory nor
field test results are available, the penetration resistance can be estimated from empirical
correlations with other soil properties (e.g. with relative density).
Figure 4.2: Evaluation of the dilation parameter Xpp from the results of liquefaction
strength analysis: a. relationship between normalized penetration resistance and cyclic
stress ratio which causes liquefaction in NL cycles; b. illustrative example of the
parametric studies using element tests; c. evaluation of the dilation parameter using
“regula falsi”.
The next step consists in numerical simulation of cyclic triaxial or simple shear
109
laboratory tests (element tests). Liquefaction strength curves (CSR vs. NL) are
constructed using various values of the dilation parameter, and the actual Xpp value
is found by interpolation, as illustrated in Figure 4.2,b. A more ecient search for
the dilation parameter value consists in numerically solving the equation:
Ncycles (Xpp CSR) = NL (4.8)
for a given cyclic stress ratio CSR. The procedure, using \regula falsi", is shown in
Figure 4.2,c.
The proposed method of parameter evaluation from laboratory test results and liq-
uefaction strength analysis has been validated in the past by post{analyses (Popescu
et.al., 1992, Popescu and Prevost, 1993a) and class \A" predictions (Popescu and
Prevost, 1993d, Prevost and Popescu, 1994). Some results of a comparative study
involving the performance in predicting excess pore pressure build{up achieved by
the soil constitutive models employed for the class \A" predictions performed within
the framework of the VELACS project are presented in Figure 4.3 (after Popescu
and Prevost, 1995a). The VELACS project (Arulanandan and Scot, 1993) is an NSF
sponsored geotechnical study on the eects of earthquake{like loading on a variety of
soil models (nine groups of centrifuge experiments were performed, using various soil
and soil{structure models). A total of 56 numerical predictions, representing 75% of
the Class \A" predictions submitted by various analysts have been available for the
comparative study presented in Figure 4.3. The markers indicate the magnitude of
normalized root{mean{square errors of predicted excess pore pressures with respect
to the average of centrifuge experimental recordings, for each of the nine centrifuge
models involved in the VELACS project. It can be seen from Figure 4.3 that, for most
centrifuge models, the multi{yield soil constitutive model (root{mean{square errors
represented by black circles) provided closer predictions of excess pore pressures than
the other soil constitutive models involved in the project.
110
Figure 4.3: Results of a study on the performance of various soil constitutive models used
fro Class “A” predictions (VELACS project) in predicting excess pore pressure build-up:
a. averages of the root-mean-square error index over all the pore pressure transducers of
each centrifuge model; b. evaluation of the root-mean-square error εrms between predicted
and recorded excess pore pressure time histories (after Popescu and Prevost, 1995a).
More detailed comparisons for pore pressures, displacements and accelerations are
available on-line at: http://cee.princeton.edu/~radu/soil/velacs/comp.html.
111
4.2.2 Estimation from eld data

All the validations presented before involve comparisons with excess pore pressures,
displacements and accelerations obtained in centrifuge experiments. The soil used in
centrifuge modeling has the same characteristics as the one used in laboratory tests,
i.e.: (1) is disturbed from its natural state, and (2) is homogeneous. However, both
shear strength and liquefaction resistance of soils in natural, undisturbed state are
found to be signicantly larger than laboratory estimated values (e.g. Yoshimi et.al.,
1989, Seed and De Alba, 1986). In this context, the advantages of eld measurements
over laboratory testing are summarized by Campanella (1994): the disturbances are
minimized, the eects of fabric and aging on measured properties are preserved, and
the soil is tested in its natural in{situ stress state. Also, the so called \homogeneous"
soil layers are in fact only \statistically homogeneous", i.e. there might be signicant
variations of soil properties from a spatial location to another. The proposed esti-
mation procedures for spatial variability of soil properties are discussed in details in
Chapter 2. The recommended material property evaluation procedure for full scale
situations, involving soil materials in their natural state, is based on correlations with
eld test results, and is presented hereafter.
Numerous methods are available for in{situ evaluation of geotechnical properties.
Among them, two groups of tests are most popular:
1. Dynamic penetration tests { standard penetration test (SPT), Baecher pene-
tration test (BPT), and dynamic cone penetration test (DCPT).
2. Cone penetration tests { mechanical CPT, electric CPT, piezocone (CPTU),
seismic piezocone (SCPTU), resistivity piezocone (RCPTU), and horizontal
stress cone (HSC).
The advantages and disadvantages of each type of eld test, as well as their
applicability in estimating various soil properties are discused by Campanella (1994).
Mentioned here (Table 4.2) are comparative features of standard penetration test
112
Table 4.2: Comparison between SPT and CPT (after Campanella, 1994)
SPT CPT
empirical database best historical database limited, but growing SPT
database can be used thru
correlations qc ; NSPT
repeatability not repeatable repeatable
accuracy numerous sources of errors accurate
sampling distance infrequent sampling continuous data
soil sample provided not provided
equipment equipment variability standardized
cost (Kulhawy and $50 : : : $150/test $13 : : : $20/m
Maine, 1990)
(SPT) and cone penetration test (electric CPT and piezocone), for which a large
database for empirical correlations with geotechnical properties is available.
Both types of tests have good applicability in estimating various soil properties:
soil type, moduli and strength, and liquefaction resistance. However, given its advan-
tages { precision, repeatability, continuous logging, multiple channel measurements,
as well as ease of use in oshore testing (Jeeries and Davies, 1993) { the cone
penetration test is becoming increasingly more popular for in{situ investigations.
The results of penetration type soil tests represent the response of soil to an
imposed deformation. Geotechnical properties can be then estimated using either
(1) an appropriate constitutive model and an analytical solution, or (2) empirical
correlations. Various theories have been developed to express analytical dependence
between cone tip resistance and soil strength. However, there is currently no satisfac-
tory general solution for interpreting cone test results (Been et.al., 1986). Therefore
empirical correlations are used to derive soil properties.
A procedure for estimating geotechnical properties from standard penetration re-
sistance (NSPT ) or cone penetration resistance (qc) and in-situ stress state is presented
in the following. It is based on a vast amount of information on empirical correlations
113
between SPT and CPT results and various soil properties. Both relative density ap-
proach (Seed et.al., 1984) and state parameter approach (Been and Jeeries, 1985)
are used. A general scheme showing possible correlations between multi{yield plas-
ticity model constitutive parameters and eld test results is presented in Figure 4.4.
Correlations are indicated by arrows, and the numbers placed next to each soil pa-
rameter indicate references, as listed in Table 4.3. Each soil property can be evaluated
as a function of one ore more other properties, as indicated by merging arrows. In
general, more than one empirical correlations are available to derive each particular
soil property. However, the accuracy of estimations is highly dependent on selection
of the appropriate correlations for the respective material.
The proposed procedure can be easily t to any amount of available information.
On the one hand, one can estimate the required soil parameters starting from relative
density only (e.g. Popescu et.al., 1992). On the other hand, the condence in the
estimates can be increased by using available general and laboratory soil test results,
as shown in Figure 4.4. It is also recommended to use collateral information for
selecting the appropriate correlation formula for each soil parameter, as shown in
x4.3.1.
The methodology for soil constitutive parameter evaluation from in{situ standard
penetration resistance has been validated by post{analysis of the behavior (1) of a
building founded on liqueable soil during the 1964 Niigata Earthquake (Popescu
et.al., 1992), and (2) of the soil{seawall interaction at a wharf in Akita Harbour
during the 1983 Nihonkai{Chubu Earthquake (Popescu and Prevost, 1993c).
114
Figure 4.4: Evaluation of soil properties based on field test results.
115
Table 4.3 Correlations between soil parameters and eld test results
Param. Reference in Fig. 4.4 Basic relation Remarks
qc=NSPT 1. Robertson et.al., 1982 chart: q N(bars)
c
55
vs: D50 various sands
2. Seed and De Alba, 1986 chart: q N(tsf)
60
c
vs: D50 D50 1mm
3. Kulhawy and Mayne, q =p = 5:44D (mm) SD = 1:03
50
c a
N60
1990 R2 = 0:7
4. Jeeries and Davies, chart: q (MPa)
N60 vs: D50
c
0:01 D50 10
1993 (mm)
5. Kulhawy and Mayne, q =p = 4:25 ; F
c a
SD = 0:89
N60 41:3
1990 R2 = 0:414
6. Jeeries and Davies, q (MPa) = f(q 0 B ) ve sites
T v0 q
c
N60
1993
qn=(N1)60 7. Olsen and Farr, 1986 chart: qn ; f 0 ; N1 0
v
s
sand : : : clay
(N1)60 1. Seed, 1979 chart: CN = f(v0 0 Dr ) Dr = 50% 70%
(or CN ) 2. Skempton, 1986 CN = 1+ 20=p 0 ne sands
CN = 2+ 30=p coarse sands
v a
CN = 0:7+1:70=p q
a
overcons. sands
v
0
a
3. Liao and Whitman, CN = p 0 all sands

v
a
0
1986
v
qn 1. Shibata and Cq = 0:7+1:70=p 0 data from Baldi

(or Cq ) Teparaksa, 1988 et.al., 1982
v a
q
2. Kulhawy and Mayne, Cq = p 0 0
a
1990
v
3. qc=v0 0 is also used for various correlations with

G0 1. Ohsaki and Iwasaki, G0 = a(NSPT )b Japanese
1973 practice
2. Ohta and Goto, G0 = f(NSPT depth, corr. with shear
1978 age, soil type) wave velocity
3. Richart et.al., 1970 G0 ; p ; e { fabric
0
lab. tests
p
a p a
4. Hardin and Richard, G0 = f(e p0 fabric) 10;4

1963
5. Hardin, 1978 Go = f(e OCR p0)
6. Bellotti et.al., G0 = 400 pa
Ticino sand
0:43
1986 expf1:39Dr g pp 0
(ne)
chart: Gq 0 ; pq
a
7. Baldi et.al., 1989 c

c
0
Ticino sand
v 0
116
Table 4.3 (cont.) { Correlations between soil parameters and eld test results
Dr 1. Marcuson and Dr = f(NSPT v0 0 coarse sands
Bieganousky, 1977 OCRrCu )
2. Tokimatsu and Dr = 21 0=pN +0:7 + 1F:7
0
J
NJ = N78
Yoshimi, 1983 Jap. stnd.
v a
3., 4. Kulhawy and (N1 )60 = 60 + 25 log D NC sands

D2 50
Mayne, 1990 Dr = f(test cond: aging
r
(N1)60 OCR)
q
5. Meyerhof, 1957 Dr = 21 NJ =( pp + 0:7)
0
a
Jap. stnd.
6. Holtz and Gibbs, Dr = f(NSPT v0 0) calibration
1979 chamber
q
7. Skempton, 1986 Dr (N1)60=C with C = 65 (coarse)
60 (medium) 55 (ne) and 40 (NC lls)
8. Jamiolkowski et.al., Dr = 68log qnp=Kq ; 1] calibration
a
1985 chamber
9. Kulhawy and Mayne, studies
1990 Dr = f(qn Qc OCR)
10. Baldi et.al., 1982 chart: Dr = f(qc h0 0) uncem., unaged
quartz sands
11. Huntsman et.al., chart: fs ; h0 ; Dr
0 calibration
1986 chamber
Soil 1. Schmertmann, 1978 various charts showing mech. CPT
classi- 2. Douglas and Olsen, the soil type as electric
cation 1981 a function of friction
and 3. Robertson and qc=pa and Fr CPT
OCR Campanella, 1983
4. Olsen and Farr, 1986 soil type = f(qn f 0 ) 0
s
electric CPT
5. Senneset and Janbu, soil type = f(Bq qp ) various
v
T
1985 correlations
a
6. Jones and Rust, 1982 soil type = f(u qT v0) 0 with

7. Robertson, 1990 soil type = f(qT Fs v0 0) results
8. Jeeries and Davies, chart: soil type = of CPTu
1991 f(qT Bq Fs v0 0)
1. Hayashi et.al., 1992 = 0:037Dmax ; 0:027Cu various types
+0:217 of sand
117
Table 4.3: (cont.) { Correlations between soil parameters and eld test results
2 30:34
1. Schmertmann, 1975 tan;1 4 NSPT 0 5 depth > 1 : : : 2m
12:2 + 20:3 p
0
v
2. Parry, 1977 chart: = f(NSPT =v0 0)

a
3. NAVFAC, 1982 chart: = f(Dr ) ne sands

4. Parry, 1977 chart:" = f(Dr ) #
5. Robertson and ;1
tan 0:1 + 0:38 log 0 q c
v0
Campanella, 1983
6. Kulhawy and = 17:6 + 11:0 log qp n
SD = 2:8
Maine, 1990 R2 = 0:64
a
7., 8. Olsen and Farr, chart: = f(qn fs=v0 0) clean sands

1986 chart: = f(qn soil type) sand : : : clay
9. Jeeries and Been, = 32(1 ; 1:67) SD 2
1987
k0 1. Jaky, 1948 k0 = 1 ; sin NC sands
2. Ochiai, 1977 k0 = f(tan )
3. Marchetti, 1985 chart: k0 ; ; qc=v0 0 low compr. sands
4. Kulhawy et.al., h0 = (qc=pa )
0 1 : 25
calibration
1989 p a 35 exp f D r = 20g chamber data
k 1. Shahabi et.al., k = 1:2Cu0:735D100:89 1+e3e medium and
1984 ne sands
2. Cassagrande and k = f(soil type) all soils
Fadum, 1940
3. Hazen, 1930 k = (1 : : : 1:5)D102 D10(cm)
4. Creager et.al., 1945 k = f(D20)
5. Kenney et.al., 1984 k(cm=s) = D52 (mm2) Dr 80%
6. JSSMFE, 1988 k = f(D10) Jap. practice
n 1. Conventional values for cohesionless soils: n = 0:4 : : : 0:6
2. Popescu and n = 0:7 Nevada sand
Prevost, 1993a (ne)
3. Bellotti et.al., n = 0:65 : : : 0:75 Ticino sand
1986 n 0:80 Po{River sand
118
Table 4.3 (cont.) { Correlations between soil parameters and eld test results

$ 1. Koerner, 1970 $ = f(Dr gr. size, fabric) single mineral soil
2. Been and Jeeries, ; $ = f() $ 30 : : : 32
1985
1., 2. Been et.al., 1987a = f(qc p0, soil type) various sands
= f(qc p
ss )
0
Lique- 1. Robertson and zone of liquef. sands on electric

faction Campanella, 1985 qc ; Fr ; soil type] chart CPT
Resis- 2. Seed et.al., 1984 CSR ; (N1)60 ; F
tance 3. Fear and CSR ; (N1)60 ; F updates
(CSR) McRoberts, 1995 Seed's chart
4., 5. Tokimatsu and CSR ; (N1)J ; F Japanese
(charts) Yoshimi, 1983 standards
d ; D ; ax. strain from laboratory
r
2p0
test results
6. Christian and CSR ; Dr
Swiger, 1975
7. Yoshimi et.al., d ; (N ) undist. samples,
1 78
1989 2p0 clean sands
8. Robertson and CSR ; qn ; D50
Campanella, 1985
9. Seed and De Alba, CSR ; qn ; D50 or F unconservative( )
1986
10. Shibata and CSR ; qn ; D50 unconservative
Teparaksa, 1988 for D50 = 0:05( )
11. Mitchell and CSR ; qn ; D50
Tseng, 1990
12., 13. Seed and K ; v0 0 correction for
Harder, 1990 v0 0 > 1 barr
K ; 0h ; Dr corr. for static
v0 shear stress
(*) { as resulted from a study by Mitchell and Tseng (1990).
119
4.3 Application { Constitutive parameter evalu-

ation for Erksak sand
4.3.1 Empirical correlations
The method for evaluation of soil properties from eld data is employed to estimate
the geotechnical properties of a hydraulic ll deposit made of Erksak sand, and used
for construction of articial islands in the Canadian Beaufort Sea (Jeeries et.al.,
1985, 1988a). Cone penetration test results obtained from hydraulic lls for a large
range of depths and two deposition methods are available (Gulf Canada Res. Inc,
1984). A short description of the analyzed soil deposit, and of the eld measurement
program is given in x2.1.1. The index properties of Erksak sand are reported by Been
et.al. (1987b, 1991), and listed in Table 4.4.
The correlation formulas used to evaluate soil constitutive parameters as functions
of cone tip resistance qc and eective overburden stress v0 0 are shown in Figure 4.5
with continuous lines and boldface numbers. The specic correlations are selected on
the basis of:
{ sand type: uniformly graded, medium to ne grained sand, with subrounded shaped
particles and mainly quartz minerals
{ deposit history: normally consolidated hydraulic ll.
Analytical formulas are only selected , to allow automate data processing for the study
on liquefaction strength presented in x4.3.2, and for evaluation of soil properties at
any spatial location (Ch. 5.). Whenever more than one correlation formulas seem
appropriate, the selection is made based on information from other types of eld
tests (Jeeries et.al., 1987, 1988a), and laboratory test results reported by Been
et.al., 1987b, 1991 (dashed lines in Figure 4.5).
A particularity of the analyzed deposit is the relatively large value of the coecient
of lateral stress k0. Estimations of k0 in the hydraulic ll core (Jeeries et.al., 1985)
resulted in values ranging from k0 = 0:7 near the surface to k0 = 1:6 at the base
120
Figure 4.5: Numerical example: evaluation of constitutive model parameters from cone
penetration test results and general soil test results (continuous lines); available results
from other field tests (self-boring pressuremeter, horizontal stress cone) and laboratory
soil tests are used to select appropriate correlation formulas (dashed lines).
121
Table 4.5: Estimation of geotechnical properties of Erksak sand
Soil property Correlation / Estimated value Reference

Mass density s = 2660Kg/m3 Been et.al., 1991
8 0:17
Overburden stress < 0 +0:07 v0 0:05MPa
> Shibata and
0
0
correction factor Cq = > q Teparaksa, 1988

v
(qn = Cq qc) : pa =v0 0 v0 0 > 0:05MPa Kulhawy and Maine,

1990
relative density 1 q
Dr2 = 305 n
Kulhawy and Maine,
p
void ratio e = emax ; Dr (emax ; emin) 1990
a
porosity nw = e=(1 + e)
permeability k = 1:2Cu0:735D100:89 1+e3e Shahabi et.al., 1984

0:5
low strain shear G0 = 70 (2:171+;ee)2 pp0 Hardin and Richard,
0
modulus 1963
a
power exponent n = 0:5 estimated

Poisson's coe. = 0:35 h i estimated
friction angle = tan;1 0:1 + 0:38 log q 0
0
c
Robertson and
at failure v
Campanella, 1983
stress{strain curve = 0:217 + 0:037Dmax Hayashi et.al., 1992
coecient ;0:027Cu = 0:26
max. dev.strain "$dev = 7% compr., 4% ext.
max Been et.al., 1991
dilation angle $ = 31 Been and Jeeries, 1985
cyclic stress ratio CSR = 0:1 15+ q n
Shibata and
25;q n
Teparaksa, 1988
dilation parameter Xpp = f(qc v0 0) { Eqn. (4.11) Figure 4.6,c
of the caisson. Data obtained from self boring pressuremeter and horizontal stress
cone tests performed in similar hydraulic lls (Jeeries et.al., 1987) indicated a range
k0 = 0:7 : : : 0:9. A value k0 = 0:7 is used by Jeeries et.al. (1988a) to analyze
CPT data in hydraulically placed sands. A value k0 = 0:7, constant with depth, is
considered in this study.
Numerous empirical correlations between relative density Dr , penetration resis-
tance, and in{situ stress are available (Figure 4.4). Most of them, however, are based
on the value of in{situ eective vertical stress, and assume a value k0 = 0:4 : : : 0:5
for the coecient of lateral stress. Comparisons with charts by Baldi et.al. (1982)
and Kulhawy et.al. (1989), showing correlations between cone resistance, eective
122
horizontal stress and relative density, led to selection of the formula proposed by
Kulhawy and Maine, 1990 (Figure 4.5 and Table 4.3). The void ratio e is evaluated
based on relative density, emax and emin. Then, porosity nw , permeability k, and low
strain shear modulus G0 are estimated from correlations with void ratio, grain size
and fabric, as shown in Table 4.5.
As in the case of relative density, the relation for estimating friction angle
at failure as a function of penetration resistance and eective conning stress
is selected by accounting for the unusually large values of the coecient of lateral
stress k0. The range of average friction angle values and the variation with depth
are estimated from average values (50% condence level) of the state parameter
reported by Jeeries et.al. (1985), and also using a correlation proposed by Marchetti
(1985), relating friction angle, cone resistance and the coecient of lateral stress. The
resulted range and the tendency of the friction angle values to diminish with depth in
the hydraulic ll core (as reported by Jeeries et.al., 1985, 1988a) are best matched by
the correlation formula proposed by Robertson and Campanella (1983) { Table 4.5.
The other soil parameters (mass density { s , power exponent { n, stress{strain
curve coecient { , maximum deviatoric strain { "$max $
dev and dilation angle { ) are
assumed invariant with respect to cone tip resistance values, as shown in Table 4.5.
4.3.2 Dilation parameter

The dilation parameter Xpp is evaluated as a function of cone tip resistance qc and
eective vertical stress v0 0. To obtain the desired correlation formula, liquefaction
strength analysis is performed for a range of cone resistance and eective stress val-
ues which includes the values from cone penetration test measurements shown in
Figure 2.3.
First, the liquefaction resistance (CSR) is estimated using the correlation pro-
posed by Shibata and Teparaksa (1988) between cyclic stress ratio necessary to pro-
duce liquefaction in NL = 15 cycles and normalized cone tip resistance qn (Table 4.5).
123
Next, cyclic simple shear soil tests are simulated (element tests { x4.2.1), using the
values of all other soil parameters, estimated as shown in Table 4.5. The resulted val-
ues of dilation parameter computed for various combinations qc ; v0 0 are plotted vs.
the corresponding liquefaction resistances CSR in Figure 4.6,a. A relatively strong
correlation can be observed between Xpp and CSR. The relationship Xpp ; CSR is
linearized by means of transformations of each variable. Linear regression analysis of
data shown in Figure 4.6,b results in:
log Xpp = 7:47 ; 10:93 CSR0:1 (4.9)
which can be expressed in terms of cone resistance qc and eective vertical stress v0 0
as:
2 q 0 3b3
15 + q c pa =v0 5
log Xpp = b1 + b2 40:1 q (4.10)
25 ; qc pa=v0 0
with b1 = 7:47 b2 = ;10:93 and b3 = 0:1.
The correlation represented by Eqn. (4.10) is very good for intermediate values of
the eective vertical stress v0 0 = 0:1MPa. One can notice from Figure 4.6,b that the
markers corresponding to larger values of v0 0 (v0 0 = 0:2 : : : 0:3MPa) systematically
fall below the regression line, while those corresponding to smaller values (v0 0 =
0:02 : : : 0:05MPa) are plotted above the line. Though the dierences seem small
in terms of log Xpp, errors of up to 25% are given by Eqn. (4.10) in terms of Xpp
values. Therefore, a correction coecient for the eective stress value is included in
Eqn. (4.10), as the power exponent b4 of v0 0:
"
15 + q p
c pa =(v0 0 )b4
#b3
log Xpp = b1 + b2 0:1 25 ; q pp =(0 )b4 (4.11)
c a v0
The coecients b1 : : : b4 are evaluated by nonlinear regression analysis, using

the procedure described in x2.3.1, extended to two dimensions. The values resulted
for Erksak 330/0.7 sand are: b1 = 7:07 b2 = ;11:4 b3 = 0:13, and b4 = 0:39.
The regression analysis results (continuous lines) are compared in Figure 4.6,c with
124
a. Results of element tests b. Linearization of Xpp - CSR relation

0.12
-0.5
effective vertical stress initial regression line
0.1
log Xpp = 7.47 - 10.93 CSR
0.1
σ’vo = 0.02 MPa

dilation parameter - Xpp
-1
σ’vo = 0.05 MPa
0.08
σ’vo = 0.10 MPa
log Xpp
0.06
-1.5
σ’vo = 0.20 MPa
σ’vo = 0.30 MPa
0.04
-2
R2 = 99.2%
0.02
-2.5
0
0 0.1 0.2 0.3 0.4 0.75 0.8 0.85 0.9

0.1
cyclic stress ratio - CSR CSR
c. Regression analysis results (element test results are shown with markers)
0.12
Erksak 330/0.7 sand

0.1
Xpp = f(qc, σ’vo)

0.08
dilation parameter - Xpp
element test results

σ’vo = 0.02 MPa
σ’vo = 0.05 MPa
0.06
σ’vo = 0.10 MPa

σ’vo = 0.20 MPa
0.04
σ’vo = 0.30 MPa

0.02
0
0 5 10 15 20 25 30
cone tip resistance - qc (MPa)
Figure 4.6: Evaluation of dilation parameter Xpp for Erksak sand: a. correlation with
cyclic stress ratio CSR b. estimation of initial parameter values for nonlinear regres-
sion analysis c. resulted correlation between dilation parameter, cone tip resistance,
and eective vertical stress.
125
the results of liquefaction strength analysis (markers) for various values of cone tip
resistance and eective vertical stress.
Remark: For values of eective vertical stress v0 0 > 0:05MPa, the normalized
penetration resistance qn is computed using the relation proposed by Kulhawy and
Mayne (1990), instead of that used by Shibata and Teparaksa (1988) to derive their
correlation with cyclic stress ratio (Table 4.5). This resulted in slightly larger qn
values, and consequently in larger estimated liquefaction resistance. However, this
compensates for the fact that the relationship proposed by Shibata and Teparaksa
for cyclic stress ratio is on the conservative side, as compared to similar correlations
reported by Ishihara (1985), Seed and De Alba (1986), and Mitchell and Tseng (1990).
4.4 Sensitivity analyses

4.4.1 Constitutive parameter evaluation and estimated vari-
ability
To perform sensitivity analyses for the parameters of multi{yield soil constitutive
model, a set of parameter values are estimated based on real site measurement results.
The selected soil conditions correspond to the hydraulic ll used for articial islands
in the Beaufort Sea (Jeeries et.al., 1988a). A 22m thick hydraulic ll soil layer made
of Erksak sand is considered in the analysis. The water table is at about 2m below
the ground level (Figure 4.8,a).
The parameter evaluation method, based on correlations with in{situ measure-
ment results, is described in x4.2.2. With reference to Figure 4.4, the particular
correlation formulas used for the present analysis are shown in Figure 4.7. Results
of the general soil test results performed for Erksak sand and reported by Been et.al.
(1987b, 1991) are also used. The soil properties are estimated considering an average
value of the cone resistance, qc = 6:9 MPa, recorded at the mid depth (z = 12m)
of the saturated soil layer. The estimated eective vertical stress at that location
is v0 0 = 0:123MPa. Eleven soil constitutive model parameters are selected for the
126
Figure 4.7: Sensitivity analysis: evaluation of constitutive model parameters from field
data and general soil test results.
To get an idea of the expected range of variation for each soil parameter value
resulting from such an estimation method, other correlation formulas presented in
127
Table 4.6: Soil parameter values estimated for a hydraulic ll layer
Constitutive parameter Estimated Estimated variation coecient
value (Pi ) Vi 100 (%) References(1)
Mass density { solid, s 2660 Kg/m3 0.8 estimated(2)
Porosity(3), nw 0.427 1.2 2,3,4,6, : : : 10
Permeability, k 2:3 10;4 m/s 32.6 1, : : : 6
Low strain shear modulus, G0 80 MPa 7.0 1, : : : 7
Poisson's coecient(4), 0.29 7.0 1, : : : 4
Power exponent, n 0.5 10.0 estimated
Friction angle at failure, 37 5.1 1, : : : 9
Stress{strain curve coe., 0.26 25.7 estimated(5)
Max. deviatoric strain, "max
dev 6% 28.8 estimated(6)
Dilation angle, $ 31 3.2 estimated(7)
Dilation parameter, Xpp 0.065 19.6 2,4,5,8,9,10,11
(1) { as in Figure 4.4 and Table 4.3
(2) { data from Hough, 1969
(3) { evaluated from relative density and void ratio
(4) { evaluated from k0 = 0:41 (Ref. 1] in Figure 4.4)
(5) { data reported by Hayashi et.al. (1992) for various sand types
(6) { data from triaxial soil tests performed on Nevada sand (Arulmoli et.al., 1992)
(7) { suggested range (Been and Jeeries, 1985)
128
Figure 4.4 and Table 4.3, which are considered appropriate for the selected type of
soil, are also used. For each constitutive parameter value Pi i = 1 2 : : : 11 (e.g.
P1 = s P2 = nw etc.), the variation coecient is computed as: Vi = Ps , where
i
h i1
P$i = N1 PNk=1 Pik is the mean value, and si = N 1;1 PNk=1(Pik ; P$i) 2 is the sample
i
i i
i i
standard deviation. Ni represents the number of correlation formulas available for

estimating the parameter Pi (e.g. N7 = 9 for P7 = , etc.), and Pik is the value of
parameter Pi estimated from empirical relation k in Figure 4.4.
The variation coecients Vi computed for each soil parameter Pi are listed in
Table 4.6. This variations originate in dierences due to use of various correla-
tion formulas. The dierences are mainly produced by : (1) variations in soil type
(2) spatial variations of soil properties within statistically homogeneous soil layers
and (3) measurement errors. Therefore they reect to some extent the soil property
variability which the analyst has to face in real life.
4.4.2 Numerical simulation results

The soil behavior is analyzed under the assumption of an innite horizontally layered
soil deposit. The 1D discretization mesh is presented in Figure 4.8,a. The soil deposit
is subjected to a sinusoidal base input acceleration, with a duration T1 = 5:5sec
(Krstelj and Prevost, 1993), scaled at 0.1g { Figure 4.8,b. To capture the pore
water pressure dissipation phenomenon, the total analysis time is selected T = 15sec.
The numerical simulations are performed with the nite element code DYNAFLOW
(Prevost, 1981).
A rst analysis uses the parameters listed in Table 4.6. Some of the computation
results are presented in Figure 4.8, as: relative displacements and horizontal accelera-
tions at the ground level (Figures 4.8,c and d, respectively), and excess pore pressure
ratio r with respect to the initial eective vertical stress (Figure 4.8,e { time histories,
and Figure 4.8,f { values at the end of shaking). The following representative results
are selected for the sensitivity analysis:
129
c. Horizontal displacements (node 12) d. Horizontal acc.
0.02
2
(node 12)
1
a (m/s )
2
δ (m)
0
0
a. Finite element
discretization
-1
-0.02
mesh
-2
node 12 0 time (s) 15 0 time (s) 8
element 11 e. Time histories of excess pore f. E.P.P ratio at the end
(dry) pressure ratio: r(t) = u(t) / σv0 of shaking (t = T1)
1
-2
element 10 element 10
0.5
r
-4
1 0
-6
r - max
r - max
element 8
0.5
r
-8
element 8
1 0
-10
depth (m)
element 6
0.5
r
-12
element 6
1 0
-14
element 4 element 4
0.5
r
-16
1 0
-18
element 2 element 2
0.5
r
-20
0
-22
0 5 10 15
time (s) 0 0.5 1
r(T1) = u(T1) / σv0
b. Base input motion
1
a (m/s )
2
T = 15 s - total analysis time

T1 = 5.5 s - shaking time
-1
0 5 10 15
time (s)
Figure 4.8: Sensitivity analysis set{up, and numerical computation results: a. nite
element mesh b. base input motion c. horizontal relative displacements at the ground
level d. horizontal accelerations at the ground level e. time histories of excess pore
pressure ratio f. excess pore pressure ratios at the end of shaking.
130
1. Liquefaction index (Shinozuka and Ohtomo, 1989) at the end of shaking:

Z H p(T1 z )
R1 = Q = H1 0 (z) dz (4.12)
0 v0
where H = 20m is the dimension of the saturated soil layer, p(t z) is the excess
pore pressure computed at time t and elevation z, and v0 0(z) is the initial
eective vertical stress at elevation z.

2. Maximum excess pore pressure ratio R2 = rmax = p(T01(zz)) z=;7m , computed at
0
v
the end of shaking (t = T1) and elevation z = ;7m (element #8) where the
maximum excess pore pressure build{up occurs (Figure 4.8,e).
3. Pore pressure index (Popescu and Prevost, 1995b):
ZT
R3 = T0 (z)1 p(t z)dt (4.13)
v0 0 z=;7m
computed at z = ;7m.
4. Acceleration amplication factor:
n o
max aground level(t)
0tT
R4 = n base o (4.14)
max
0tT
a (t)
where a(t) is the horizontal acceleration.
5. Maximum relative horizontal displacement:
n o
R5 = 0max
tT
jground level(t) ; base(t)j (4.15)
where (t) is the horizontal displacement.

The constitutive model sensitivity is expressed in terms of normalized variations
Rji=Rj of each representative computational result Rj j = 1 : : : 5, produced by
variations of each soil parameter Pi i = 1 : : : 11. The values Rji are computed as:
Rji = Rj (P1 : : : Pi + 21 Pi : : : P11) ; Rj (P1 : : : Pi ; 21 Pi : : : P11) (4.16)
131
where ∆Pi=0.1Pi. Consequently, each variation in computational results ∆Rj,i corresponds

to a 10% variation of the value Pi of parameter i. The normalized variations ∆Rj,i/Rj are
listed in Table 4.7. The liquefaction index values, R1=Q, are presented in Figure 4.9,a.
Figure 4.9: Results of the sensitivity analysis in terms of normalized variations ∆R1,i / R1
of the resulting liquefaction index: a. corresponding to relative variations ∆Pi / Pi = 10%;
b. corresponding to expected variations ∆Pi = one standard deviation.
4.4.3. Remarks
1. The values ∆Rj,i / Rj presented in Table 4.7 are computed for the set of parameter
values P = P1, P2, … P11 listed in Table 4.6. These results are likely to vary with
respect to the position in the parameter space, i.e.: ∆R j,i ≠

∆R j,i
where P and P’
Rj Rj
P P'
are two different sets of soil parameter values. However, the same trends are
expected, and hence the results in Table 4.7 can give qualitative information on the
132
Table 4.7: Results of the sensitivity analysis in terms of normalized variations

R (%), corresponding to variations P = 10% of the parameter values
j i
R P
i
j i
Soil Liquef. Maximum Pore press. Acceleration Max. horiz.

parameter index e.p.p. ratio index ampl. factor displacement
R1 (%) R2 R3 R4 R5
Pi R1
i
R2 (%)
i
R3 (%) R4 (%)
i i
R5 (%)
i
1. s -5.62 -4.47 -9.36 7.43 4.86

2. nw 2.71 2.58 5.22 -3.38 -2.25
3. k -0.87 -2.09 -5.07 0.33 0.00
4. G0 3.26 -0.49 -2.23 6.95 -1.21
5. 2.58 4.37 6.54 -8.48 -1.15
6. n -2.11 -0.87 0.17 2.38 0.96
7. -6.42 -4.54 -7.42 -0.76 0.66
8. -0.22 -0.69 -0.95 2.62 1.79
9. "max
dev 0.07 -0.50 -1.56 1.67 1.38
10. $ 6.94 8.34 14.52 -5.05 0.18
11. Xpp 3.53 3.84 6.86 -1.81 0.09
relative importance of each parameter.
2. As shown in Table 4.6, the values of soil parameters are estimated with various
degrees of accuracy. Assuming that the expected variability in evaluating the soil
parameters is that shown in Table 4.6, the image of how each constitutive parameter
aects the computational results is considerably changed. The values Rji=Rj shown
in Table 4.8 and Figure 4.9,b are computed for variations of each soil parameter value
equal to one standard deviation (i.e. Pi = si, or Pi =Pi = Vi ).
3. Assuming that the liquefaction resistance is correctly evaluated, the eects of the
errors Pi in estimating some parameter values may be attenuated while deriving
the value of dilation parameter Xpp by means of liquefaction strength analysis. This
is illustrated in Table 4.8 by numbers in brackets, and in Figure 4.9,b by black bars.
Values of parameters which signicantly inuence liquefaction resistance are only
shown (namely: G0 n , and $). As expected, the liquefaction strength analysis
only attenuates the errors of computational results which are in terms of excess pore
pressure response.
133
Table 4.8: Results of the sensitivity analysis in terms of normalized variations

R (%), corresponding to expected variations P = V (1)
j i
R P i
i
j i
Soil Liquef. Maximum Pore press. Acceleration Max. horiz.

param. index e.p.p. ratio index ampl. factor displacement
R1 (%) R2 R3 R4 R5
Pi R1
i
R2 (%)
i
R3 (%)
i
R4 (%)
i
R5 (%)
i
1. s -0.45 -0.36 -0.75 0.59 0.39

2. nw 0.33 0.31 0.63 -0.41 -0.27
3. k -2.83 -6.81 -16.54 1.09 0.00
4. G0 (1.76) 2.28 (-0.91) -0.34 (-2.62) -1.56 (5.10) 4.86 (-0.86) -0.85
5. (-0.49) 1.80 (0.46) 3.06 (-0.05) 4.58 (-5.00) -5.94 (-0.86) 0.80
6. n (-1.54) -2.11 (-0.28) -0.87 (1.24) 0.17 (2.24) 2.23 (0.98) 0.96
7. (-1.17) -3.27 (-0.09) -2.32 (0.63) -3.78 (1.14) -0.39 (0.49) 0.34
8. -0.58 -1.76 -2.44 6.73 4.6
9. "max
dev 0.19 -1.45 -4.49 4.78 3.97
10. $ (1.57) 2.22 (1.89) 2.67 (3.13) 4.65 (-0.93) -1.62 (0.04) 0.06
11. Xpp 6.92 7.53 13.45 -3.55 0.18
(1) { variations in calculated response including the eects of liquefaction strength
analysis are shown in brackets.
4.5 Conclusions
1. The required constitutive parameters of the multi{yield plasticity model are tra-
ditional soil parameters, which can be derived from the results of conventional
laboratory and in{situ soil tests. Liquefaction strength analysis and paramet-
ric computations using element tests are however necessary for estimating the
dilation parameter, which is a scaling coecient for the plastic dilation.
2. The multi{yield plasticity soil model and the methodology for constitutive pa-
rameter evaluation from laboratory soil test results and liquefaction strength
analysis have been validated by post{analyses and class \A" predictions. Close
results in terms of predicted excess pore pressures, accelerations, and displace-
ments, as compared to the corresponding recorded values, have been obtained
by numerical simulation of a large variety of centrifuge geotechnical experiments
involving soil and soil{structure models.
134
3. Both shear strength and liquefaction resistance of the soil samples used for
laboratory tests are dierent from the corresponding in{situ properties. Conse-
quently, the recommended parameter evaluation procedure for full{scale anal-
yses, involving soil materials in their natural, undisturbed state, is based on
correlations with eld test results. A large number of correlations between the
results of eld measurements (SPT and CPT) and soil geotechnical properties
are selected from literature and organized in a suggestive ow chart. The pro-
posed methodology is exible, i.e. it can be adapted to the available amount of
data.
4. The proposed procedure for soil constitutive parameter evaluation includes:

(1) use of appropriate empirical correlation formulas (2) estimation of liquefac-
tion resistance and (3) parametric studies involving numerical element tests.
It is presented on the support of the analysis of Erksak sand.
5. Sensitivity analyses are performed for the parameters of the multi{yield plas-
ticity soil constitutive model. Variations in the numerical predictions (excess
pore pressures, accelerations, and displacements), due to: (1) relative variations
of 10% in each soil parameter value, and (2) expected variations of each param-
eter value, are estimated. Two of the conclusions of the sensitivity analyses are
mentioned here:
{ the variations (errors) in the computed response in terms of predicted excess

pore pressure due to errors in the estimation of some of the constitutive
parameters are attenuated due to use of liquefaction strength analysis
{ accounting for: (1) the inuence on the predicted response, and (2) the ex-
pected estimation error of each soil parameter, some of the constitutive
parameters (namely: dilation parameter, friction angle, permeability, low
strain shear modulus, dilation angle) are considered to need special atten-
tion.
135
Appendix 4.A List of notations for Chapter 4

B0 { low strain bulk modulus
Bq { CPTu pore pressure parameter: Bq = qu;;u0 0
T
0
v
CN { NSPT correction for overburden stress

Cq { qc correction for overburden stress
Cu { uniformity coecient: Cu = D60=D10
CPT (U ) { cone penetration test (piezocone test)
CSR { cyclic stress ratio causing liquefaction in NL = 15 cycles: CSR = l=v0 0
Dmax { maximum particle size
Dn { particle size at "n"% ner
D50 { mean grain size
Dr { relative density: Dr = (emax ; e)=(emax ; emin )
E { fourth order isotropic elastic tensor
F { percent passing No. 200 sieve (percent of nes)
Fr { cone friction ratio: Fr = fs =qc
Fs { stress normalized CPTU friction ratio: Fs = q ;f 0 100%
T
s
G0 { low strain elastic (dynamic) shear modulus

Hi0 { plastic modulus corresponding to yield surface fi
Kq { calibration chamber correction factor: Kq = 1 + (Dr ; 0:3)=3
K { correction for static driving shear stress
K { correction for overburden stress v0 0 > 1 barr
Ic { soil classication index
Mi { material parameter denoting the size of yield surface fi
(N1)60 { N60 corrected to reference stress of one atmosphere: (N1)60 = CN N60
N60 { NSPT corrected for eld procedures
NC { normally consolidated
NJ { NSPT { Japaneze standards
NL { number of cycles to liquefaction
136
NSPT { standard penetration resistance

OC { overconsolidated
OCR { overconsolidation ratio
P { plastic potential
P0 P 00 { deviatoric and volumetric components of the plastic potential P
Q { liquefaction index
Qc { compressibility coecient
R2 { coecient of determination
SD { standard deviation
SPT { standard penetration test
Xpp { dilation parameter
a { unequal end area ratio
e { void ratio
emax(emin) { maximum (minimum) void ratio
fs { CPT sleeve friction
fi { yield surface in the stress space
k { coecient of permeability
k0 { coecient of lateral stress
kG { modulus number
n { power exponent for variation of moduli with conning stress
nw { porosity: nw = e=(1 + e)
p0 { eective mean normal stress
p00 { reference eective mean normal stress
pa { atmospheric pressure
qT { qc corrected for unequal area eects: qT = qc(1 ; a u)
qc { cone tip resistance
qn { qc standardized to reference stress of one atmosphere: qn = Cq qc
s { deviatoric stress tensor
137
u { dynamic pore pressure (CPTU )

u0 { hydrostatic pressure
{ stress{strain curve coecient
i { kinematic tensor dening the center of yield surface fi
{ shear strain
{ unit tensor
(_"p) "_ { (plastic) strain rate tensor
"max
dev { maximum deviatoric strain

1
{ mobilized stress ratio: = 23 s : s 2 =p0
$ { stress ratio at phase transformation

ss { slope of the steady state line
{ Poisson's ratio
s { mass density { solid phase
0 { eective stress tensor
1 3 { principal stresses
d { double amplitude cyclic deviatoric stress (triaxial): d = 1 ; 3
h0 0 { eective horizontal stress
v0 { total vertical stress
v0 0 { eective vertical stress
h { static driving shear stress on a horizontal plane
l { average cyclic shear stress
{ friction angle at failure (for triaxial boundary conditions)
$ { dilation angle
{ state parameter
138
Chapter 5
Eects of spatial variability on
dynamic behavior
5.1 Representation of random data
5.1.1 Transfer of random data
The continuous stochastic eld which describes the spatial distribution of soil prop-
erties is represented by a nite set of random variables. The sample elds generated
as shown in Chapter 3 consist of realizations of those random variables at prede-
ned spatial locations. The distances xk { Eqn. (3.12) { which are referred here
as \stochastic eld mesh size", depend on the upper cut{o frequencies kuk and the
ratio between Nk and Mk { the number of simulation points in wave number domain
and in spatial domain, respectively { 3.2.1:
x
xk = 2k M1 = k2 M Nk k = 1 2 : : : n (5.1)

k k ku k
where the subscript k is the index for spatial dimensions and kk is the mesh size in
the wave number domain.
From the condition to avoid aliasing Mk 2Nk , 3.2.2. A reliable discretization
x
in the wave number domain requires relatively large values of Nk and, consequently,
a ratio Mk =Nk 4 is imposed by computer memory limitations in the case of multi{

dimensional random eld simulations. Since the upper cut{o frequencies kku are
functions of the particular correlation structure, the analyst is quite restricted in
139
selecting the stochastic eld mesh size in the spatial domain. On the other hand, as
shown in the next section, totally di erent criteria are used for selecting the nite
element mesh size. Consequently, when using the spectral representation method for
stochastic eld generation, it is very likely to have two di erent meshes: one for the
random eld discretization, and the other for the nite element structural analysis. A
transfer of data is therefore necessary from one mesh to another. Numerous methods
are proposed for that purpose by various authors, in the context of stochastic nite
element analysis (see e.g. Brenner, 1991, for a detailed literature review). Two such
methods are investigated here:
1. The midpoint method (Shinozuka and Dasgupta, 1986 Yamazaki et.al., 1988
Der Kiureghian and Ke, 1988 Deodatis, 1989, etc.) is a point discretization
type method. The random eld is represented by its values at the centroids of
each nite element.
2. The local averaging (or spatial averaging) method, proposed by Vanmarcke

(1977) and Vanmarcke and Grigoriu (1983), assigns to each nite element a
value obtained as an average of stochastic eld values over the element domain.
A comparison between the two methods for stochastic eld representation is pre-
sented hereafter, based on two sample functions of a stochastic eld representing the
normalized uctuations of cone tip resistance (data from Tarsiut P{45).
The resulted variance of the discretized eld is represented in Figure 5.1,a as a
function of the ratio A= between the nite element size A = xz and the charac-
teristic area = xz . The variance obtained by applying the midpoint discretization
method is basically the same as the point variance (black squares in Figure 5.1,a).
The variance of the spatially averaged process very closely follows the theoretical
variance reduction function: (A) = x(x) z (z). The values of 1D variance
reduction functions x(x) and z (z) are calculated here from the 1D correlation
140
a. Variance reduction
1.2
1
1
0.8 1D variance reduction functions theoretical variance reduction
γx(∆x) or γz(∆z)
function - γ(A)
γx(∆x)
γ(A) = γx(∆x) x γz(∆z) (for ∆x/∆z = 12)
2
σA / σ0
γz(∆z)
0.6
A = ∆x x ∆z - finite element area

2
∆x/θx or ∆z/θz
midpoint method
0
0.4
0 2 4 6
local averaging method
0.2
2
A / α (α = θx x θz = 11.5 m - characteristic area)
0
0 1 2 3 4 5 6 7
b. Reduction of the skewness coefficient - γ1 c. Reduction of the kurtosis coefficient - β2

1.25
1.25
1
1
- 3)
orig
midpoint method
0.75
0.75
orig
- 3) / (β2 local averaging method

/ γ1
mesh
0.5
0.5
γ1
mesh
midpoint method
(β2
0.25
0.25
local averaging method

A/α A/α
0
0 1 2 3 4 5 0 1 2 3 4 5
d. Empirical distribution function

1
midpoint
method
A / α = 0.26 A / α = 1.04 A / α = 4.15

0
-2 0 2 4 -2 0 2 4 -2 0 2 4
1
local prescribed Beta CDF

averaging resulted empirical CDF
Gaussian CDF
method
A / α = 0.26 A / α = 1.04 A / α = 4.15
0
-2 0 2 4 -2 0 2 4 -2 0 2 4
Figure 5.1: Comparison between local (spatial) averaging and midpoint discretization
methods (results based on two sample functions of a stochastic eld describing the
cone tip resistance data at Tarsiut P{45).
141
functions (x ) and (z ), respectively, using the expression (Vanmarcke, 1983):
Z x !
x(x) = x 2 x
1 x (x)dx etc:
; (5.2)
0
and numerical integration. From studies on random eld discretization, using Gaus-
sian random elds with exponential correlation structure, Der Kiureghian and Ke
(1988) concluded that the midpoint method over{represents the variability of the
random eld, while the spatial averaging method tends to under{represent the true
variability, and therefore the two methods bracket the real result. A further study
on this issue is presented in the next section.
For Gaussian random elds, the probability distribution function of the discretized
eld remains Gaussian for both midpoint and spatial averaging methods. On the
other hand, for non{Gaussian random elds the original distribution of the point
eld is dicult or impossible to obtain when using the spatial averaging method
(Der Kiureghian and Ke, 1988). This is illustrated in Figure 5.1,b,c (open squares),
and d, based on a numerical example. The Beta distribution of the original eld
(Table 2.2) is reduced to the normal distribution by the local averaging process:
lim
A=!1 1
= 0 (Figure 5:1 b) (5.3)
and
lim = 3
A=!1 2
(Figure 5:1 c) (5.4)
where 1 and 2 are the coecient of skewness and the coecient of kurtosis, re-
spectively. The resulted empirical distribution function (dotted lines in Figure 5.1,d)
switches from the prescribed Beta distribution (continuous lines) to the normal dis-
tribution (dashed lines), as the averaging domain increases. On the other hand, the
midpoint method preserves the probability distribution of the original point eld, as
resulted from the numerical example presented in Figure 5.1.
For Gaussian elds, the spatial averaging method seems to be a more logical ap-
proach to random eld discretization, and was proven to provide better accuracy
142
than the midpoint method for coarse meshes (Der Kiureghian and Ke, 1988). This
method was used for some of the computations performed with data from Akita Har-
bour (Popescu et.al., 1995), illustrated in Plate 7,b,g,h. For the case of non{Gaussian
elds, the midpoint method, which preserves the original probability distribution, is
deemed as most appropriate and used hereafter.
5.1.2 Selection of mesh size for stochastic analysis

The mesh size used in deterministic nite element computations is usually dictated
by the expected gradients in results. However, when using stochastic input the dis-
tance between two random variables has to be short enough to capture the essential
features of the random eld (Brenner, 1991). Consequently, the upper bounds for
the mesh size xk depend on the correlation distances k . Various values of these
upper bounds have been reported for various correlation structures, e.g.: xk
(0:25 : : : 0:5)k (Der Kiureghian and Ke, 1988) for an exponential correlation struc-
ture, xk (0:7 : : : 0:8)k (Popescu et.al., 1995) for a triangular correlation struc-

ture. It is believed that the upper bounds depend on the type (shape) of correlation
structure as well as on correlation distances. Some results of a study based on eld
data obtained from Tarsiut P{45, which was found to have a Two{Peak correlation
structure, are presented hereafter. The study also addresses the issue of variance
reduction discussed in 5.1.1.
x
Sample 1D correlation functions of the qc component of the generated stochastic

eld are represented in Figure 5.2, along with the target correlation functions in
vertical and horizontal directions. The sample correlation functions are evaluated
from the realization of the sample eld in the following way:
1. In vertical direction (Figure 5.2,a):
NXv ;r
v
qq (rz) = N v r1 qs(x0 iz )qsx0 (i + r)z] r = 1 : : : mv =4 (5.5)
i=1
;
where x0 is a xed horizontal location, and mv z = 128m is the dimension of

143
a. 1D correlation functions in vertical direction at x0 = 4.5m

1
target correlation function in vertical direction

sample correlation function for ∆z = 0.25m
0.8

0.6

0.4
0.2
0
space lag - ξz(m)

-0.2
0 0.5 1 1.5 2
b. 1D correlation functions in horizontal direction at z0 = 1.5m

1
target correlation function in horizontal direction

sample correlation function for ∆x = 0.75m
0.8

0.6

0.4
0.2
space lag - ξx(m)

0
0 5 10 15 20
Figure 5.2: 1D sample correlation functions evaluated using various sampling intervals
(data from the hydraulic ll at Tarsiut P{45).
144
the simulated eld in vertical direction.

2. In horizontal direction (Figure 5.2,b):
h ;r
NX
hqq (rx) = 1 qs(ix z0)qs(i + r)x z0] r = 1 : : :mh =4 (5.6)
Nh ; r i=1
where z0 is a xed location in elevation, and mhx = 384m is the dimension

of the simulated eld in horizontal direction.
Various values for the sampling distances (or mesh dimensions) { x and z are
used. One can notice from Figure 5.2 that:
1. Larger is the sampling distance x or z, poorer is the t between sample and
theoretical correlation functions.
2. There are limit values xmax and zmax below which acceptable t is obtained:
1:5m x=8 < xmax < 3:0m x=4 in horizontal direction and zmax =

0:5m z =2 in vertical direction.

It is obvious that di erent limiting sampling distances are obtained for each spatial
direction because of di erent shapes of the correlation structures.
Some results of a numerical experiment are presented in the following, in terms of
predicted excess pore pressure ratio with respect to the initial e ective vertical stress.
The purpose of this experiment is to study the inuence of the variance reduction
(addressed in 5.1.1) and of the selected nite element mesh size on the computa-
x
tional results, in the context of stochastic input nite element analyses. A saturated
soil deposit is subjected to seismic excitation, and the numerical predictions are per-
formed using various mesh sizes. The nite element analysis set{up (dimensions of
the analysis domain, materials, boundary conditions, input motion) is described in
x5.2.1.
Several realizations of the sample eld describing stochastic variability of soil
properties at Tarsiut P{45 have been generated and are used to evaluate the stochastic
145
12
Sample field #1 (as in Plates 2 and 3)
elevation - z (m)
9
6 1.5m x 0.25m elem. ("exact" solution)
deterministic input
2 2
3.0m x 0.5m elements (σA = γ(A) σ0 )
3
2 2
liquefaction index - Qv 3.0m x 0.5m elements (σA = σ0 )
0
0 0.2 0.4 0.6 0.8

0.6
liq. index - Qh
0.5
0.3 0.4
horizontal distance - x (m)

0 15 30 45 60
12
Sample field #2
elevation - z (m)
9
1.5m x 0.25m elem. ("exact" solution)

6
deterministic input
2 2
3
2 2
0
0 0.2 0.4 0.6 0.8

0.6
liq. index - Qh
0.5
0.3 0.4

0 15 30 45 60
12
Sample field #3
elevation - z (m)
9
1.5m x 0.25m elem. ("exact" solution)

6
deterministic input
2 2
3
2 2
0
0 0.2 0.4 0.6 0.8

0.6
liq. index - Qh
0.5
0.3 0.4

0 15 30 45 60
Figure 5.3: E ects of variance reduction on predicted liquefaction index.
146
input soil parameters for the nite element analyses. The results are presented in
Plates 2 and 3, and Figures 5.3 and 5.4 in terms of excess pore pressure ratio with
respect to the initial e ective vertical stress, and liquefaction index computed (1) in
vertical direction { Eqn. (5.8) { and (2) in horizontal direction { Eqn. (5.9). The
midpoint discretization method is employed for data transfer from the stochastic
eld mesh to the nite element mesh.
Two situations are analyzed for the variance reduction:
Case 1 : variance reduction according to the theoretical variance reduction function:
A2 = (A)02, where A = x z is the nite element area, and (A) is
evaluated from the corresponding 1D components { Eqn. (5.2).
Case 2 : no variance reduction: A2 = 02.
The results obtained for various mesh sizes are compared with the assumed \ex-
act" solution, corresponding to a very ne mesh (1:50m 0:25m). From the compu-

tational results presented in Plates 2 and 3, and in Figure 5.3, it can be concluded
that:
{ variance reduction corresponding to the nite element mesh size (and according to
the spatial averaging method) leads to severe loss of accuracy in computational
results, as the mesh size increases (Plate 2) as expected, smaller is the variance,
closer are the results to those obtained using deterministic input parameters
{ for the case of \no variance reduction", meshes up to 0:25x 0:50z (i.e. 3:0m

0:5m, corresponding to the limit values derived by means of correlation structure

analysis) provide close results (Plate 3) acceptable results are also obtained
using slightly larger meshes { here: 4:5m 0:75m, corresponding to x 0:37x

and z 0:75z ).

A comparison in terms of the average pore water pressure build{up:

1 Z H Z L p(x z )
P = LH dxdz (5.7)
0 0 v0 0 (x z )
147
where L and H are the dimensions of the rectangular analysis domain, is presented
in Figure 5.4. In the case of variance reduction (circles) the amount of predicted
pore pressure build{up drops as soon as the mesh size increases, as compared to the
\exact" solution. In the case of no variance reduction (square markers), the P {index
value only drops for large meshes (6m 1m).
average pore pressure build-up - P-index
0.5
"exact" solution
2 2
no variance reduction: σA = σ0
0.48
2 2
variance reduction: σA = γ(A) σ0
0.46
deterministic input
0.44
0.42
relative mesh size - A/α

0.4
0 0.2 0.4 0.6
Figure 5.4: Inuence of variance reduction and selected mesh size on predicted average
pore pressure build{up.
In the case of data from Akita Harbour, where a Gaussian eld has been simu-
lated (Popescu et.al., 1995), the spatial averaging method and corresponding vari-
ance reduction led to acceptable results for mesh dimensions up to: x 0:25x and
z 0:75z (Plates 7,b,g,h and 8). An explanation may reside in the shape of the

variance reduction functions, which do not decrease so fast with increasing averaging
interval as in the case of CPT data from Tarsiut P{45.
5.2 Eects on dynamic behavior of soil deposits

5.2.1 Finite element analysis set{up
The e ects on the computational results of various factors related to stochastic input
parameters are investigated in the following based on dynamic analyses of a saturated
sandy soil deposit subjected to seismic excitation. The multi{yield plasticity model
(Prevost, 1985, 1989) implemented in the computer code DYNAFLOW (Prevost,
148
1981) is used for this purpose.

A loose to medium dense soil deposit, with geomechanical properties as well as
spatial variability characteristics corresponding to the hydraulic ll at Tarsiut P{45
{ 2.1.1 { is selected for the numerical examples. The analysis domain extends 60m
x
in horizontal direction (Plate 4). A 12m thick \statistically homogeneous" saturated

sand deposit is overlaid by an 1.6m dry sand layer. A rigid, impervious bedrock
is assumed at the base, where the input horizontal acceleration is applied. The
base input motion is selected as the rst 6 sec. of the N{S component of the 1964
Niigata Earthquake accelerogram, recorded at Akita Prefectural Oce. The input
acceleration is scaled at 0.15 g. A semi{in nite horizontal deposit is simulated by
imposing the same displacements in both horizontal and vertical directions, and for
both solid and uid phases, at the lateral nodes situated in the same horizontal
planes (same elevations). In the case of stochastic input, this assumption induces
some boundary e ects, but they are supposed to be con ned in a small region along
lateral boundaries.
A re ned discretization mesh (1920 1:5m 0:25m nite elements for the saturated

material, and 40 1.5m 1:6m nite elements for the dry sand) is used, to eliminate

any e ects induced by larger mesh sizes.

Similar analyses have been performed for a dense sand deposit, with characteristics
corresponding to the soil at Akita Harbour { Ohama No. 1 Wharf. The details of the
nite element analysis set{up are presented by Popescu et.al. (1995).
5.2.2 Stochastic vs. deterministic input parameters

A deterministic analysis is performed for comparison, for the same soil deposit de-
scribed in the previous section. A soil column is discretized into 24 nite elements
for the saturated zone and one element representing the dry sand (Plate 4). The
soil constitutive parameters for the deterministic input analysis are variable with
depth, corresponding to the linear variation exhibited by the spatial trends of the
149
eld measurement results in terms of cone tip resistance qc ( 2.1.3 and Figure 2.3),
x
and soil classi cation index Ic. Consequently, the deterministic input analysis is per-
formed using the same average values of soil parameters as for the stochastic input
computations.
A comparison between results of deterministic and stochastic parameter input
computations is presented in Plate 4 in terms of predicted excess pore pressure ratio
with respect to the initial e ective vertical stress, and liquefaction index. Four real-
izations of the vector eld of material properties are employed to derive four sets of
stochastic input soil parameters. Consequently, four di erent responses in terms of
excess pore pressure ratio are predicted by the stochastic input analysis. There are
however a series of common features of the stochastic input computational results,
as compared to the results of the deterministic input analysis:
{ more pore pressure build{up is predicted by the stochastic input model than by
the deterministic input model (see also Figure 5.4) a possible explanation is
o ered in 5.2.4
x
{ local values of excess pore pressure ratio r = p=v0 0 up to 1.0 (i.e. liquefaction) are
predicted by the stochastic input analysis, while in case of deterministic input
a maximum value r = 0:6 is predicted
{ in the case of stochastic input results, there are patches with large excess pore
pressure, corresponding to the presence of loose pockets in the material this
predicted pattern of excess pore pressure build{up better explains the phe-
nomenon of sand boils observed in areas a ected by soil liquefaction.
Average predicted pore pressures are also compared in Plate 4 in terms of the
liquefaction index (Shinozuka and Ohtomo, 1989), computed by averaging the excess
pore pressure ratio in horizontal direction:
1 Z L p(x z )
Qv (z) = L 0 (x z) dx (5.8)
0 v0
150
or in vertical direction:
1 Z H p(x z )
Qh(x) = H 0 (x z) dz (5.9)
0 v0
where L = 60m and H = 12m are the dimensions of the analysis domain. The
computed values clearly indicate higher excess pore pressure build{up predicted by
the stochastic input than by the deterministic input analysis, for the loose hydraulic
ll deposit.
Similar comparison results obtained for the dense sand deposit at Akita Harbour
are presented in Plates 7,a: : : f and 8. The same features of the results can be observed,
though the di erences in excess pore pressure build{up are not so obvious as in case
of the loose sand deposit. Larger di erences between deterministic and stochastic
analyses in terms of location of the most a ected layer are exhibited.
5.2.3 Inuence of assumed correlation distances

Four realizations of a stochastic eld having correlation distances three times larger
than derived from eld data analysis are generated to investigate what are the inu-
ences of correlation distances on dynamic behavior. To obtain correlation functions
which have similar shapes with the ones used for previous computations, only the
parameters b1 and b3 of the target correlation functions are modi ed (assumed three
times larger).
The predicted excess pore pressure ratios and liquefaction indices are shown in
Plate 5. The liquefaction indices obtained for the input stochastic eld with larger
correlation distances (red lines) are compared with the results from previous com-
putations (blue lines). Excepting for more uniformity in the predicted pattern, the
increase in correlation distances seems to have little inuence on the excess pore pres-
sure magnitude. Similar conclusions have been obtained from the study on the dense
sand layer at Akita Harbour (Popescu et.al., 1995).
151
5.2.4 Inuence of probability distribution of soil properties

Some results of a study on the inuences of probability functions of soil properties
on the amount of pore water pressure build{up are presented in Plate 6.
The stochastic input parameters are obtained starting from three realizations of
the stochastic eld simulated using the same seed number for the generation of the
random phase angles in Eqn. (3.7), but di erent target distribution functions (case 1
{ Beta distribution, as derived from eld data analysis, case 2 { Normal distribution,
and case 3 { Lognormal distribution). Consequently, a similar pattern of pore water
pressure distribution is predicted for all three cases. The average values of input
soil parameters, as well as the variance of their spatial variability are also similar for
all cases. However, the resulted amount of excess pore pressure build{up is clearly
di erent for the three situations analyzed, and is apparently dependent on the left
tail of the distribution function (the cumulative distributions of the standardized eld
test results are shown in Plate 6): larger pore pressure build{up is predicted as the
left tail of the probability distribution is longer.
This important conclusion also indicates an explanation for predicting more pore
pressure build{up in case of stochastic input computations: the increase in pore
pressure is facilitated by the presence of looser zones, and more variability in soil pa-
rameter spatial distribution leads to increased likelihood of loose pockets. Of course,
denser zones of material are correspondingly represented, however their contribution
seems to be negligible in the case of dynamically induced soil liquefaction.
5.3 Application { Structure founded on lique-

able soil deposit
A structure representing a four story building is founded on a relatively loose sand
deposit, as shown in Figure 5.5. The saturate soil layer is discretized into 840 -
nite elements (1.5m 0:5m). Semi{in nite soil deposit conditions are prescribed at

the lateral boundaries of the analysis domain. Impervious soil{structure interface is

152
assumed. The base input acceleration (Figure 5.5) is selected as the strong motion
recorded at Akita Prefectural Oce during the 1964 Niigata Earthquake, and scaled
at 0.15 g.
node 950

FINITE ELEMENT

ANALYSIS SET−UP

10.40m
structure perfect slip condition

impervious surface
H=12m

B=7.5m node 938

1.60m
free field DRY SAND
element 636 element 648

boundary SATURATED SAND
conditions
12.00m
(free field) (below structure)
finite element dimensions:
∆x = 1.50m, ∆z = 0.50m input motion
22.50m 7.50m 22.50m

2
acceleration (m/s2)
BASE INPUT ACCELERATION

1
(N−S component of the 1964 Niigata

0
Earthquake acceleration, recorded at

Akita Prefectural Office, and scaled
-1
at 0.15g) time (sec.)

-2
0 3 6 9
Figure 5.5: Structure founded on lique able sand deposit: nite element analysis
set{up.
The same four realizations of the stochastic eld of soil properties generated based
on results of piezocone tests performed at Tarsiut P{45, and used for the study pre-
sented in 5.2.2 are employed to derive the soil constitutive parameters. Some of
x
the results, in terms of predicted excess pore pressures, are presented in Plate 9.
Computational results obtained using deterministic parameter input, with soil prop-
erty values corresponding to the average values of the stochastic soil parameters, are
shown for comparison.
The same conclusions can be inferred as obtained in 5.2.2: more pore pressure x
is predicted by stochastic than by deterministic analyses, especially in the free eld.

More important however seem to be here the predicted pattern of pore pressure
153
4
relative horizontal displacements (cm)
deterministic input
stochastic input
2
0
time (sec.)
-2
0 3 6 9
Figure 5.6: Predicted horizontal displacements of the top of structure, relative to the
base of structure (node
x
950 node 938).
x ;
build{up which, for some possible realizations, may lead to large di erential settle-
ments of the structure, as shown in Figure 5.6 (prescribed time histories of horizontal
displacements at the top of structure relative to the displacements at the base |
node
x
950 node 938), and Figure 5.7 (deformed mesh at the end of shaking). It can be
x ;
observed that, for some realizations of the stochastic eld, the stochastic input anal-
ysis predicts remanent deformations which are larger than predicted by deterministic
analysis.
5.4 Conclusions
1. The e ects of spatial variability of material properties on system behavior are
investigated based on the results of stochastic input nite element analyses,
obtained for a saturated soil deposit and for a soil{structure system subjected
to seismic excitation.
2. For non{Gaussian elds, the transfer of random data from the discretization
of simulated stochastic eld to the nite element mesh is performed using the
midpoint method, which preserves the probability distribution of the original
154
155
point eld. The spatial averaging method is used for Gaussian elds.
3. In the case of stochastic input analyses there is an upper bound for the nite
element mesh size, which is imposed by the condition to capture the essential
features of spatial variability, and therefore depends on the correlation structure
(both on correlation distances and on shape of correlation function). For the
various types of correlation structures used in this study, the upper bound for
the nite element mesh size ranges between 0.25, : : : 0.75 of the correlation
distances.
4. Based on the example of a real site behavior, it is concluded that both the
pattern and the amount of dynamically induced pore water pressure build{up
are strongly aected by spatial variability of soil properties. More pore pressure
build{up is predicted by the stochastic input model than by the deterministic
input model. An explanation resides in the fact that soil liquefaction is triggered
by the presence of loose pockets in the soil deposit. Since deterministic input
analyses cannot account for such variability in soil properties, they are deemed
as providing non{conservative results.
5. It is found that the amount of predicted excess pore pressure is strongly aected
by the probability distribution of soil parameters, and more speci cally by the
left tail of the distribution, corresponding to presence of looser pockets in the
soil deposit. Therefore, for the case of real distributions skewed to the right, the
Gaussian assumption for soil properties results in too conservative predictions
of pore pressure build{up.
6. The previous conclusions are con rmed by the results of a study involving a
soil{structure system: more pore pressure is predicted by stochastic than by de-
terministic analyses, especially in the free eld. The predicted pattern of lique-
faction occurrence, which depends on each speci c realization of the stochastic
eld, is important for assessing the dierential settlements of structures founded
on lique able soil deposits.
156
Bibliography
1. Anderson, T.W. and D.A. Darling, (1952). Asymptotic theory of certain \good-
ness of t" criteria based on stochastic processes. Ann. Math. Stat., 23:193{
212.
2. Arulanandan, K. and R.F. Scott, editors, (1993). Proc. Int. Conf. on Verif.
Numerical Proc. for the Analysis of Soil Liq. Problems, volume 1. Balkema,
Rotterdam.
3. Arulmoli, K., K.K. Muraleetharan, and M.M. Hossain, (1992). VELACS {
Laboratory testing program soil data report. Technical report, The Earth
Technology Corporation, Project No. 90-0562.
4. Baecher, G.B. (1984). Just a few more tests and we'll be sure! In Probabilistic
Characterization of Soil Properties: Bridge Between Theory and Practice, pages
1{18, Atlanta, Georgia, ASCE.
5. Baldi, G., R. Bellotti, M. Jamiolkowski, and E. Pasqualini, (1982). Design
parameters for sand from CPT. In Proc. 2nd European Symp. on Penetration
Testing, volume 2, pages 425{432, Amsterdam.
6. Baldi, G., R. Bellotti, V. Ghionna, and M. Jamiolkowski, (1989). Stiness
of sands from CPT, SPT and DMT { a critical review. In Proc. Geotechn.
Conf. \Penetration Testing in the UK", pages 299{305, Birmingham, Inst.
Civil Engineers.
7. Been, K., J.H.A. Crooks, D.E. Becker, and M.G. Jeeries, (1986). The cone
penetration test in sands: Part I, state parameter interpretation. Geotechnique,
36(2):239{249.
8. Been, K. and M.G. Jeeries, (1985). A state parameter for sands. Geotech-
nique, 35(2):99{112.
9. Been, K., M.G. Jeeries, J.H.A. Crooks, and L. Rothenburg, (1987a). The cone
penetration test in sands: Part II, general inference of state. Geotechnique,
37(3):285{299.
157
10. Been, K., B.E. Lingnau, L.H.A. Crooks, and B. Leach, (1987b). Cone penetra-
tion test calibration for Erksak (Beaufort Sea) sand. Can. Geotech. Journ.,
24:601{610.
11. Been, K., M.G. Jeeries, and J. Hachey, (1991). The critical state of sands.
Geotechnique, 41(3):365{381.
12. Bellotti, R., V. Ghionna, M. Jamiolkowski, R. Lancellotta, and G. Manfredini,
(1986). Deformation characteristics of cohesionless soils from in{situ tests. In
S.P. Clemence, editor, Use of In{Situ Tests in Geotechnical Engineering, pages
47{73. ASCE, New York.
13. Bendat, J.S. and A.G. Piersol, (1986). Random Data, Analysis and Measure-
ment Procedures. Wiley{Interscience, New York, 2nd edition.
14. Bendat, J.S. and A.G. Piersol, (1993). Engineering Applications of Correlation
and Spectral Analysis. Wiley{Interscience, New York, 2nd edition.
15. Brenner, C.E., (1991). Stochastic nite element methods: Literature review.
Technical Report 35{91, Inst. of Engineering Mechanics, Univ. of Innsbruck,
Austria, November.
16. Burden, R.L. and J.D. Faires, (1993). Numerical Analysis. PWS Publishing
Co.
17. Campanella, R.G., (1994). Field methods for dynamic geotechnical testing: An
overview of capabilities and needs. In Proc. Symp. on Dynamic Geotechnical
Testing II, pages 3{23, San Francisco, CA.
18. Casagrande, A. and R.E. Fadum, (1940). Notes on soil testing for engineering
purposes. Technical report, Harvard University.
19. Castro, G., (1969). Liquefaction of Sands. PhD thesis, Harvard University,
Cambridge, Mass.
20. Castro, G., (1975). Liquefaction and cyclic mobility of saturated sands. J.
Geotech. Eng. Div., ASCE, 101(GT6):551{569.
21. Castro, G. and S.J. Poulos, (1977). Factors aecting liquefaction and cyclic
mobility. ASCE Journ. Geotechn. Engnrg. Div., 103(GT6):501{516.
22. Christian, J.T. and W.F. Swiger, (1975). Statistics of liquefaction and SPT
results. ASCE Journ. Geotechn. Engnrg. Div., 101(GT11):1135{1150.
23. Creager, W.P., J.D. Justin, and J. Hinds, (1945). Engineering for dams. John
Wiley & Sons, New York.
158
24. DeGroot, D.J. and G.B. Baecher, (1993). Estimating autocovariance of in{situ
soil properties. Journ. Geotechnical Engnrg., 119(1):147{166.
25. Deodatis, G., (1989). Stochastic FEM sensitivity analysis of nonlinear dynamic
problems. Probabilistic Engineering Mechanics, 4(3):135{141.
26. Deodatis, G., (1995). Simulation of ergodic multi{variate stochastic processes.
Journ. of Engnrg. Mech., (submitted).
27. Deodatis, G., (1996). Nonstationary stochastic vector processes: Seismic ground
motion applications. Journ. of Engnrg. Mech., (submitted).
28. Deodatis, G., R. Popescu, and J.H. Prevost, (1995). Simulation of stochastic
processes and elds for Monte Carlo simulation applications: Some recent de-
velopments. In Proc. 15th Biennial Conf. on Mech. Vibrations and Noise,
Boston, MA.
29. Der Kiureghian, A. and J.B. Ke, (1988). The stochastic nite element method
in structural reliability. Probabilistic Engineering Mechanics, 3(2):83{91.
30. Douglas, B.J. and R.S. Olsen, (1981). Soil classi cation using electric cone
penetrometer. In G.M. Norris and R.D. Holtz, editors, Cone Penetration
Testing and Experience. ASCE, New York.
31. Elishako, I., (1983). Probabilistic Methods in the Theory of Structures. John
Wiley.
32. Fear, C.E. and E.C. McRoberts, (1995). Reconsideration of initiation of lique-
faction in sandy soils. Journ. Geotechnical Engnrg., 121(3):249{261.
33. Fenton, G.A., (1990). Simulation and Analysis of Random Fields. PhD thesis,
Princeton Univ.
34. Fletcher, R., (1972). Fortran subroutines for minimization by quasi{Newton
methods. Technical report, R7125 AERE, Harwell, England, June.
35. Gersch, W. and J. Yonemoto, (1977). Synthesis of multi{variate random vi-
bration systems: A two{stage least squares ARMA model approach. Journ.
of Sound and Vibr., 30(4):553{565.
36. Gill, P.E. and W. Murray, (1972). Quasi{Newton methods for unconstrained
optimization. Journ. of the Inst. of Mathematics and its Applications, 9:91{
108.
37. Griths, D.V. and J.H. Prevost, (1990). Stress{strain curve generation from
simple triaxial parameters. Num. Analyt. Meth. Geomech., 14:587{594.
159
38. Grigoriu, M., (1984). Crossings of non{Gaussian translation processes. Journ.

of Engnrg. Mech., 110(4):610{620.
39. Grigoriu, M., (1995). Applied Non{Gaussian Processes. Prentice Hall.
40. Gulf Canada Resources Inc, (1984). Frontier development { Molikpaq, Tar-
siut delineation { 1984/85 season. Technical Report 84F012, Gulf Canada
Resources Inc.
41. Guttman, I., S.S. Wilks, and J.S. Hunter, (1982). Introductory Engineering
Statistics. Wiley, New York, 3rd edition.
42. Hahn, G.H. and S.S. Shapiro, (1994). Statistical Models in Engineering. Wiley{
Interscience.
43. Harada, T. and M. Shinozuka, (1986). Scale of correlation for stochastic elds.
Technical report, Dept. of Civil Engrg. and Engrg. Mech., Columbia Univ.
44. Harder, H. and G. von Bloth, (1988). Determination of representative CPT{
parameters. In Proc. Geotechn. Conf. Penetration Testing in the UK, pages
237{240, Birmingham, Inst. of Civil Engineers.
45. Hardin, B.O., (1978). Nature of stress{strain behavior for soils. In Proc. ASCE
Spec. Conf. Earthq. Eng. Soil Dynamics, volume 1, pages 3{90, Pasadena,
CA.
46. Hardin, B.O. and F.E. Richard, (1963). Elastic wave velocities in granular soils.
ASCE Journ. Soil Mech. and Found. Div., 89(SM1):33{65.
47. Hayashi, H., M. Honda, and T. Yamada, (1992). Modeling of nonlinear stress
strain relations of sands for dynamic response analysis. In Proc. 10th World
Conf. on Earthquake Eng., Madrid, Spain. Balkema, Rotterdam, Invited
paper.
48. Hazen, A., (1930). Water supply. In American Civil Engineers Handbook.
Wiley.
49. Holtz, W.G. and H.J. Gibbs, (1979). Discussion of \SPT and relative density
in coarse sand". J. Geotech. Eng. Div., ASCE, 105(GT3):439{441.
50. Hough, B.K., (1969). Basic Soils Engineering. Ronald Press, New York, 2nd
edition.
51. Huntsman, S.R., Mitchell J.K., and Klejbuk L.W, (1986). Lateral stress mea-
surement during cone penetration. In S.P. Clemence, editor, Use of In{Situ
Tests in Geotechnical Engineering, pages 617{634. ASCE, New York.
160
52. Iai, S., (1991). A strain space multiple mechanism model for cyclic behavior
of sand and its application. Technical report, EERG, Port and Harbour Res.
Inst., Japan.
53. IMSL, (1987a). MATH/Library { FORTRAN Subroutines for Mathematical
Applications: User's Manual.
54. IMSL, (1987b). STAT/Library { FORTRAN Subroutines for Statistical Anal-
ysis: User's Manual.
55. Ishihara, K., (1985). Stability of natural deposits during earthquakes. In Proc.
11th Int. Conf. Soil Mech. and Found. Engnrg., volume 1, pages 321{376.
56. Iwan, W.D., (1967). On a class of models for the yielding behavior of continuous
and composite systems. Journ. Appl. Mech., Trans. ASME, 34(E3):612{617.
57. Jaky, J., (1948). Pressure in soils. In Proc. 2nd ICSMFE, volume 1, pages
103{109.
58. Jamiolkowski, M., C.C. Ladd, J.T. Germaine, and R. Lancellotta, (1985). New
developments in eld and laboratory testing of soils. In Proc. 11th Int. Conf.
on Soil Mech. and Foundation Engnrg., volume 1, pages 57{153, San Francisco.
59. Jeeries, M.G., (1989). Observed strength distributions in some arctic clays.
In Proc. 8th Int. Conf. O shore Mechanics and Arctic Engnrg., pages 125{132,
The Hague.
60. Jeeries, M.G., (1995). Personal communication.
61. Jeeries, M.G. and K. Been, (1987). Use of critical state representations of sand
in the method of stress characteristics. Can. Geotech. Journ., 24:441{446.
62. Jeeries, M.G. and M.P. Davies, (1991). Soil classi cation by the cone pene-
tration test: Discussion. Can. Geotech. Journ., 28:173{176.
63. Jeeries, M.G. and M.P. Davies, (1993). Use of CPTu to estimate equivalent
SPT N60. Geotechnical Testing Journal, 16(4):458{468.
64. Jeeries, M.G., L. Jonsson, and K. Been, (1987). Experience with measurement
of horizontal geostatic stress in sand during cone penetration test pro ling.
Geotechnique, 37(4):483{498.
65. Jeeries, M.G., B.T. Rogers, H.R. Stewart, S. Shinde, D. James, and S. Williams-
Fitzpatrick, (1988a). Island construction in the Canadian Beaufort Sea. In
Proc. ASCE Spec. Conf. Hydr. Fills, pages 816{883, Fort Collins, American
Society of Civil Engineers.
161
66. Jeeries, M.G., B.T. Rogers, and K.M. Grin, (1988b). Characterization of
sand lls with the Cone Penetration Test. In Proc. Geotechn. Conf. Penetra-
tion Testing in the UK, pages 199{202, Birmingham, Inst. of Civil Engineers.
67. Jeeries, M.G., H.R. Stewart, R.A.A. Thompson, and B.T. Rogers, (1985).
Molikpaq development at Tarsiut P{45. In Proc. Conf. Arctic '85, pages
1{27, San Francisco, CA, American Society of Civil Engineers.
68. Jones, G. and E. Rust, (1982). Piezometer penetration testing. In Proc. 2nd
European Symp. on Penetration Testing, volume 2, pages 607{613, Amsterdam.
69. Journel, A.G. and C.J. Huijbregts, (1978). Mining Geostatistics. Academic
Press, London, England.
70. JSSMFE, (1988). The Explanations of the Applied Mathematics, the Figures
and the Tables of Soil Mechanics. Japanese Society of Soil Mechanics and
Foundations Engineering editors. Soils and Foundations Engineering Library.
71. Kameda, H. and H. Morikawa, (1991). Simulation of conditional random elds
{ a basis for regional seismic monitoring for urban earthquake hazard mitiga-
tion. In U.S.{Italy{Japan Workshop/Seminar on Intelligent Systems, Univ. of
Perugia, Italy.
72. Kameda, H. and H. Morikawa, (1992). An interpolating stochastic process for
simulation of conditional random elds. Prob. Engnrg. Mech., 7(4):242{254.
73. Kameda, H. and H. Morikawa, (1993). Conditioned stochastic process for con-
ditional random elds. Journ. of Engnrg. Mech., 120(4):855{875.
74. Kayen, R.E., J.K. Mitchell, R.B. Seed, A. Lodge, S. Nishio, and R. Coutinho,
(1992). Evaluation of SPT{, CPT{, and shear wave{based methods for lique-
faction potential assessment using Loma Prieta data. In Proc. 4th U.S.{Japan
Workshop on Liquefaction, Large Ground Deformation and their E ects on
Lifelines, pages 177{204.
75. Kenney, T.C., D. Lau, and G.I. Ofoegbu, (1984). Permeability of compacted
granular materials. Canadian Geotechnical Journal, 21:726{729.
76. Koerner, R.M., (1970). Eect of particle characteristics on soil strength. J.
Soil Mech. Found. Div., ASCE, 96(SM4):1221{1234.
77. Kozin, F., (1988). Auto{regressive moving{average models of earthquake records.
Journ. of Prob. Engnrg. Mech., 3(2):58{63.
78. Krstelj, I. and J.H. Prevost, (1993). VELACS Class A centrifuge model test
12. Technical report, Princeton University.
162
79. Kulhawy, F.H. and P.W. Mayne, (1990). Manual on estimating soil properties
for foundation design. Final Report 1493{6, EL{6800, Electric Power Research
Institute, Palo Alto, CA.
80. Kulhawy, F.H., C.S. Jackson, and P.W. Mayne, (1989). First{order estimation
of k0 in sands and clays. In K.H. Kulhawy, editor, Foundation Engineering:
Current Principles and Practices, pages 121{134. ASCE, New York.
81. Lade, P.V., (1987). Behavior and plasticity theory for metals and frictional
materials. In Proc. 2nd Int. Conf. Const. Laws Engnrg. Mat., pages 327{334,
Tucson, Arizona.
82. Liao, S.C. and R.V. Whitman, (1986). Overburden correction factors for SPT
in sand. Journal of Geotechnical Engineering, 112(3):373{377.
83. Lumb, P., (1974). Application of statistics in soil mechanics. In I.K. Lee,
editor, Soil Mechanics { New Horizons. Newnes{Butterworths, London.
84. Magnus, J.R., (1978). Maximum likelihood estimation of the GSL model with
unknown parameters in the disturbance covariance matrix. J. Econometrics,
7(3):281{312.
85. Marchetti, S., (1985). On the eld determination of k0 in sand. In Proc. 11th
Int. Conf. Soil Mech. Found. Eng., volume 5, pages 2667{2672, San Francisco,
CA.
86. Marcuson, W.F. and W.A. Bieganousky, (1977). SPT and relative density in
coarse sands. Journ. Geotechnical Engnrg. Div., 103(GT11):1295{1309.
87. Mardia, K.V. and R.J. Marshall, (1984). Maximum likelihood estimation of
models for residual covariance in spatial regression. Biometrika, 71(1):135{
146.
88. McRoberts, E.C. and J.A. Sladen, (1992). Observations on static and cyclic
sand{liquefaction methodologies. Canadian Geotechnical Journal, 29:650{665.
89. Meyerhof, G.C., (1956). Penetration tests and bearing capacity of cohesionless
soils. J. Soil Mech. Found. Div.,ASCE, 82(SM1):458{482.
90. Meyerhof, G.G., (1957). Discussion of session 1. In Proc. 4th Int. Conf. on
Soil Mechanics and Foundation Engnrg., volume 3, London.
91. Mitchell, J.K. and D.J. Tseng, (1990). Assessment of liquefaction potential by
cone penetration resistance. In J.M. Duncan, editor, Proc. H. Bolton Seed
Memorial Symp., volume 2, pages 335{350.
163
92. Mroz, A., (1967). On the description of anisotropic workhardening. Journ.

Mech. Phys. Solids, 15:163{175.
93. NAVFAC, (1982). Naval Fac. Eng. Command, Soil Mechanics (DM 7.1),
Alexandria.
94. Ochiai, H., (1977). The estimated equation of the coecient of earth pressure
at rest k0 for sand by using N{value. JSSMFE, 17(3):93{101.
95. Ohbo, N., N. Nojiri, K. Hayashi, K. Ueno, and J.H. Prevost, (1990). Liq-
uefaction of sandy soil deposits { analysis and experiments. In Proc. 3rd
U.S. { Japan Workshop on Liquefaction, Large Ground Deformation and their
E ects on Lifelines, pages 387{402.
96. Ohsaki, Y. and R. Iwasaki, (1973). On dynamic shear moduli and Poisson's
ratio of soil deposits. Soils and Foundations, 13(4):61{73.
97. Ohta, Y. and N. Goto, (1978). Empirical shear wave velocity equations in terms
of characteristic soil indexes. Earthq. Engnrg. Struct. Dyn., 6:167{187.
98. Ohtomo, K. and M. Shinozuka, (1990). Two{dimensional spatial severity of
liquefaction. In Proc. 8th Japan Earthquake Engnrg. Symp., Tokyo.
99. Olsen, R.S. and J.V. Farr, (1986). Site characterization using the cone pen-
etrometer test. In S.P. Clemence, editor, Use of In{Situ Tests in Geotechnical
Engineering, pages 854{868. ASCE, New York.
100. Papoulis, A., (1962). The Fourier Integral and its Applications. McGraw{Hill.
101. Parry, R.H.G., (1977). Estimating bearing capacity in sand from SPT values.
J. Geotech. Eng. Div., ASCE, 103(GT9):1014{1019.
102. Peck, R.B., (1979). Liquefaction potential: Science versus practice. J. Geotech.
Eng. Div., ASCE, 105(GT3):393{398.
103. Polhemus, N.W. and A.S. Cakmak,
(1981). Simulation of earthquake ground
motions using ARMA models. Earthq. Engnrg. Struct. Dyn., 9(4):343{354.
104. Popescu, R., (1993). Numerical simulations of soil liquefaction using stochastic
input parameters. Technical report, Dept. of Civil Engineering and Operations
Res., Princeton Univ., January.
105. Popescu, R. and G. Deodatis, (1995). Simulation of multi{dimensional, multi{
variate, non{Gaussian stochastic elds. Technical report, Dept. of Civil Engi-
neering and Operations Res., Princeton Univ., January.
164
106. Popescu, R. and J.H. Prevost, (1992). Material parameter identi cation for
Fuji River sand. Technical report, Dept. of Civil Engineering and Operations
Res., Princeton Univ., report to Kajima Corporation.
107. Popescu, R. and J.H. Prevost, (1993a). Centrifuge validation of a numerical
model for dynamic soil liquefaction. Soil Dynamics and Earthquake Eng.,
12(2):73{90.
108. Popescu, R. and J.H. Prevost, (1993b). Numerical Class \A" predictions for
Models No. 1, 2, 3, 4a, 4b, 6, 7, 11 and 12. In Proc. Int. Conf. on Verif.
Numerical Proc. for the Analysis of Soil Liq. Problems, volume 1, pages 1105{
1207. Balkema, Rotterdam.
109. Popescu, R. and J.H. Prevost, (1993c). Dynamic analysis of Akita Harbour {
Ohama No. 2 wharf seawall. Technical report, Dept. of Civil Engineering and
Operations Res., Princeton Univ., part of Ph. III of Princeton Univ. { Kajima
Corp. Joint Res. Progr.
110. Popescu, R. and J.H. Prevost, (1993d). Comparison between numerical Class
A predictions and experimental results for VELACS Project centrifuge models.
Technical Report BCS{9120028, Dept. of Civil Engineering and Operations
Res., Princeton Univ., Princeton, NJ.
111. Popescu, R. and J.H. Prevost, (1995a). Comparison between VELACS numer-
ical \class A" predictions and centrifuge experimental soil test results. Soil
Dynamics and Earthquake Eng., 14(2):79{92.
112. Popescu, R. and J.H. Prevost, (1995b). Reliability assessment of centrifuge soil
test results. Soil Dynamics and Earthquake Eng., 14(2):93{101.
113. Popescu, R., J.H. Prevost, and G. Deodatis, (1996). Inuence of spatial vari-
ability of soil properties on seismically induced soil liquefaction. In Proc.
Uncertainty '96, Madison, Wisconsin, (submitted).
114. Popescu, R., J.H. Prevost, N. Ohbo, and K. Hayashi, (1992). Numerical simu-
lations of soil liquefaction. In Proc. 4th U.S.{Japan Workshop on Liquefaction,
Large Ground Deformation and their E ects on Lifelines, pages 269{282.
115. Popescu, R., J.H. Prevost, and E.H. Vanmarcke, (1995). Numerical simulations
of soil liquefaction using stochastic input parameters. In Proc. 3rd Int. Conf.
Rec. Adv. in Geotech. Earthq. Eng. Soil Dyn., St. Louis, MO.
116. Poulos, S.J., (1981). The steady state of deformation. ASCE Journ. Geotechn.
Engnrg. Div., 107(GT5):553{562.
165
117. Prevost, J.H., (1977). Mathematical modeling of monotonic and cyclic undrained
clay behavior. Int. J. Num. Meth. Geom., 1(2):195{216.
118. Prevost, J.H., (1981). DYNAFLOW: A nonlinear transient nite element anal-
ysis program. Dept. of Civil Eng. and Op. Research, Princeton University,
Last update 1995.
119. Prevost, J.H., (1985). A simple plasticity theory for frictional cohesionless soils.
Soil Dynamics and Earthquake Eng., 4:9{17.
120. Prevost, J.H., (1989). DYNA1D a computer program for nonlinear seismic
site response analysis Technical documentation NCEER-89-0025. Technical
report, Dept. of Civil Eng. and Op. Research, Princeton University.
121. Prevost, J.H., (1993). Nonlinear dynamic response analysis of soil and soil{
structure interacting systems. In Proc. Seminar Soil Dyn. Geotech. Earthq.
Eng., Lisboa, Portugal, pages 49{126. Balkema, Rotterdam.
122. Prevost, J.H. and R. Popescu, (1994). Overview presentation of predictions
for Models No. 4a and 4b. In Proc. Int. Conf. on Verif. Numerical Proc.
for the Analysis of Soil Liq. Problems, volume 2, pages 1469{1514. Balkema,
Rotterdam.
123. Rice, S.O., (1954). Mathematical analysis of random noise. In N. Wax, editor,
Selected Papers on Noise and Stochastic Processes. Dover.
124. Richart, F.E., J.R. Hall, and R.D. Woods, (1970). Vibrations of Soils and
Foundations. Prentice Hall.
125. Robertson, P.K., (1990). Soil classi cation by the cone penetration test. Can.
Geotech. Journ., 27:151{158.
126. Robertson, P.K. and R.G. Campanella, (1983). Interpretation of cone penetra-
tion tests { Part I: sand. Can. Geotech. Journ., 20:718{733.
127. Robertson, P.K. and R.G. Campanella, (1985). Liquefaction potential of sands
using the CPT. Journ. Geotechnical Engnrg., 111(3):384{403.
128. Robertson, P.K., R.G. Campanella, and A. Wightman, (1982). SPT{CPT
correlations. Technical report, Soil Mech. Series No. 62, Dept. of Civil Engnrg.,
Univ. of British Columbia.
129. Samaras, E., M. Shinozuka, and A. Tsurui, (1985). ARMA representation of
random processes. Journ. of Engnrg. Mech., 111(3):449{461.
166
130. Schmertmann, J.H., (1975). Measurement of in{situ shear strength. In Proc.

ASCE Spec. Conf. on In{Situ Meas. Soil Prop., volume 2, pages 57{138,
Raleigh.
131. Schmertmann, J.H., (1978). Guidelines for cone penetration test performance
and design Rep. FHWA{TS{78{209. Technical report, U.S. Dept. Trans.,
Washington, DC.
132. Schultze, E., (1971). Frequency distributions and correlations of soil properties.
In 1st Int. Conf. Appl. Stat. Prob. Soil Struct. Eng., pages 371{387.
133. Seed, H.B., (1979). Soil liquefaction and cyclic mobility evaluation for level
ground during earthquakes. Journ. Geotechnical Engnrg. Div., 105(2):201{
255.
134. Seed, H.B., (1987). Design problems in soil liquefaction. Journ. Geotechnical
Engnrg. Div., 113(8):827{845.
135. Seed, H.B. and P. De Alba, (1986). Use of SPT and CPT tests for evaluating
the liquefaction resistance of sands. In S.P. Clemence, editor, Use of In{Situ
Tests in Geotechnical Engineering, pages 281{302. ASCE, New York.
136. Seed, H.B., I.M. Idriss, and I. Arango, (1983). Evaluation of liquefaction po-
tential using eld performance data. J. Geotech. Eng., ASCE, 109(3):458{482.
137. Seed, H.B. and K.L. Lee, (1966). Liquefaction of saturated sands during cyclic
loading. ASCE Journ. Soil Mech. and Found. Div., 92(6):105{134.
138. Seed, H.B., K. Tokimatsu, L.F. Harder, and R.M. Chung, (1984). The inuence
of SPT procedures in soil liquefaction resistance evaluations. Technical report,
EERC 84{15 U.C. Berkeley.
139. Seed, R.B. and L.F. Harder, (1990). SPT{based analysis of cyclic pore pressure
generation and undrained residual strength. In J.M. Duncan, editor, Proc. H.
Bolton Seed Memorial Symp., volume 2, pages 351{376.
140. Sen, A. and M. Srivastava, (1990). Regression Analysis { Theory, Methods and
Applications. Springer{Verlag.
141. Senneset, K. and N. Janbu, (1985). Shear strength parameters obtained from
static CPT. In R.C. Chaney and K.R. Demars, editors, Strength Testing of
Marine Sediments. ASTM, Philadelphia.
142. Shahabi, A.A., B.M. Das, and A.J. Tarquin, (1984). An empirical relation for
coecient of permeability of sands. In Proc. 4th Australia{New Zeeland Conf.
on Geomechanics, volume 1, pages 54{57.
167
143. Shibata, T. and W. Teparaksa, (1988). Evaluation of liquefaction potential of

soils using cone penetration tests. Soils and Foundations, 28(2):49{60.
144. Shinozuka, M., (1972). Monte Carlo solution of structural dynamics. Com-
puters and Structures, 2(5+6):855{874.
145. Shinozuka, M., (1974). Digital simulations of random processes in engineering
mechanics with the aid of FFT technique. In S.T. Ariaratnam and H.H.E.
Leipholz, editors, Stochastic Problems in Mechanics, pages 277{286. Univ. of
Waterloo Press.
146. Shinozuka, M., (1987). Stochastic elds and their digital simulation. Stochastic
Methods in Structural Dynamics, (Eds. G.I. Schueller and M. Shinozuka),
Martinus Nijho Pub., Dordrecht.
147. Shinozuka, M. and G. Dasgupta, (1986). Stochastic nite elamant methods in
dynamics. In Proc. 3rd Conf. of Dynamic Response of structures, pages 44{54,
ASCE, Univ. of California, L.A., CA.
148. Shinozuka, M. and G. Deodatis, (1988a). Stochastic process models for earth-
quake ground motion. Journ. of Probab. Engnrg. Mech., 3(3):114{123.
149. Shinozuka, M. and G. Deodatis, (1988b). Response variability of stochastic
nite element systems. Journ. of Engnrg. Mech., 114(3):499{519.
150. Shinozuka, M. and G. Deodatis, (1991). Simulation of stochastic processes by
spectral representation. Appl. Mech. Reviews, ASME, 44(4):191{204.
151. Shinozuka, M. and C.M. Jan, (1972). Digital simulation of random processes
and its applications. Journ. of Sound and Vibrations, 25(1):111{128.
152. Shinozuka, M. and K. Ohtomo. Spatial severity of liquefaction, (1989). In
Proc. 2nd U.S.{Japan Workshop on Liquefaction, Large Ground Deformation
and their E ects on Lifelines.
153. Shinozuka, M. and R. Zhang, (1995). Equivalence between Kriging and CPDF
methods for conditional simulation. Journ. of Engnrg. Mech., (accepted for
publication).
154. Simo, J.C. and M. Ortiz, (1985). A uni ed approach to nite deformation
elastoplastic analysis based on the use of hyperelastic constitutive equations.
Comp. Meth. Appl. Mech. Engnrg., 49:221{245.
155. Skempton, A.W., (1986). Standard penetration test procedures and the ef-
fects in sands of overburden pressure, relative density, particle size, aging and
overconsolidation. Geotechnique, 36(3):425{447.
168
156. Soong, T.T. and M. Grigoriu, (1993). Random Vibrations of Mechanical and
Structural Systems. Prentice Hall.
157. Soulie, M., P. Montes, and V. Silvestri, (1990). Modelling spatial variability of
soil parameters. Can. Geotech. Journ., 27:617{630.
158. Spanos, P.{T.D., (1983). ARMA algorithms for ocean spectral modeling. Journ.
of Energy Res. Tech., ASME, 105(3):300{309.
159. Tokimatsu, K. and Y. Yoshimi, (1983). Empirical correlation of soil liquefaction
based on SPT N{value and nes content. Soils and Foundations, 23(4):56{74.
160. Ural, D., (1995). The E ects of Spatial Variability of Soil Properties on Lique-
faction. PhD thesis, Princeton Univ.
161. Vanmarcke, E.H., (1977). Probabilistic modeling of soil pro les. Journ.
Geotech. Eng. Div., 103(GT11):1227{1246.
162. Vanmarcke, E.H., (1983). Random Fields: Analysis and Synthesis. The MIT
Press.
163. Vanmarcke, E.H. and M. Grigoriu, (1983). Stochastic nite element analysis
of simple beams. Journ. of Engineering Mechanics, 109(5):1203{1214.
164. Yamazaki, F. and M. Shinozuka, (1988). Digital generation of non{Gaussian
stochastic elds. Journ. of Engnrg. Mech., 114(7):1183{1197.
165. Yamazaki, F. and M. Shinozuka, (1990). Simulation of stochastic elds by
statistical preconditioning. Journ. of Engnrg. Mech., 116(2):268{287.
166. Yamazaki, F., M. Shinozuka, and G. Dasgupta, (1988). Neumann expan-
sion for stochastic nite element analysis. Journ. of Engineering Mechanics,
114(8):1335{1354.
167. Yang, J.{N., (1972). Simulation of random envelope processes. Journ. of
Sound and Vibr., 25(1):73{85.
168. Yang, J.{N., (1973). On the normality and accuracy of simulated random
processes. Journ. of Sound and Vibr., 26(3):417{428.
169. Yoshimi, Y., Tokimatsu K, and Y. Hosaka, (1989). Evaluation of liquefaction
resistance of clean sands based on high{quality undisturbed samples. Soils and
Foundations, 29(1):(3{104.
169
List of Figures
1.1 Flow chart for estimating the characteristics of spatial distribution of
soil properties based on eld measurement results { \direct simulation"
approach (exempli cation for two{dimensional analysis). : : : : : : : 8
1.2 Flow chart for Monte Carlo analysis { \indirect simulation" approach. 10
2.1 CPT data from the hydraulic ll at Tarsiut P{45 (Gulf Canada Re-
sources Inc., 1984). : : : : : : : : : : : : : : : : : : : : : : : : : : : : 13
2.2 SPT data from Akita Harbour { Ohama No. 1 Wharf (Iai, 1991). : : 13
2.3 Evaluation of spatial trends { qcav (z) { for the cone tip resistance values
recorded at Tarsiut P45. : : : : : : : : : : : : : : : : : : : : : : : : : 17
2.4 Evaluation of depth dependent standard deviation of uctuations about
spatial trends for the hydraulic ll (layer 1), computed over: a. xed
intervals h, and b. moving intervals h. : : : : : : : : : : : : : : : 17
2.5 Field data standardization for NSPT values recorded at Akita Harbour
(after Popescu et.al., 1995). : : : : : : : : : : : : : : : : : : : : : : : 19
2.6 Exponential decaying SDF model: a. correlation functions, and b. spec-
tral density functions plotted for various values of the parameters b1
and b2 c. correlation distance values. : : : : : : : : : : : : : : : : : : 25
2.7 Probability distribution of standardized cone resistance values recorded
in the hydraulic ll (layer 1). Comparison of empirical distribution
with: a. common probability models, and b. tted Beta distribution
c. illustration of solution search in the parameter space. : : : : : : : : 30
2.8 Probability distribution of standardized cone resistance values recorded
in the sand berm (layer 2). Comparison of empirical distribution with:
a. common probability models, and b. tted Beta distribution. : : : : 31
2.9 Probability distribution of standardized SPT blow{counts recorded at
Akita Harbour. Comparison of empirical distribution with: a. common
probability models, and b. tted Beta distribution. : : : : : : : : : : 32
2.10 Sample correlation functions in vertical direction evaluated for stan-
dardized CPT records at Tarsiut P{45. : : : : : : : : : : : : : : : : : 33
170
2.11 Sample and theoretical correlation functions (a.), and spectral density
functions (b.) in vertical direction, evaluated using data from CPT
pro le MAC07, in the hydraulic ll layer. : : : : : : : : : : : : : : : : 34
2.12 Correlation structure assessment in vertical direction for standardized
CPT data from Tarsiut P{45: a. best t for data from the hydraulic
ll layer b. comparison between sample correlation functions and some
theoretical models c. best t for data from the sand berm. : : : : : : 38
2.13 Cross{correlation between CPT measurements from two adjacent pro-
les (MAC32 and MAC08) in the hydraulic ll layer. : : : : : : : : : 41
2.14 Cross{correlation functions and estimated sample correlation coe-
cients between standardized cone resistance values { qs { from adjacent
pro les in sublayer 1. : : : : : : : : : : : : : : : : : : : : : : : : : : : 42
2.15 Estimated soil layer disposition in the Molikpaq core (the median lines
of the four sublayers are represented by shaded areas). : : : : : : : : 42
2.16 Correlation structure assessment in horizontal direction for standard-
ized CPT data from Tarsiut P{45: a. best t for data from the hydraulic
ll layer b. best t for data from the sand berm. : : : : : : : : : : : 44
2.17 Comparison between sample correlation functions in vertical direction
computed using uctuations about spatial trends of soil classi cation
index Ic (continuous lines) and cone tip resistance qc (dotted lines)
data from Tarsiut P{45, layer 1. : : : : : : : : : : : : : : : : : : : : : 46
2.18 Study on cross{correlation between soil classi cation index Ic and cone
tip resistance qc data from CPT pro le MAC05, layer 1. : : : : : : : 47
3.1 Computed power spectral density functions Sss (l k + s 3k ) { contin-
uous bars { and Sss(k~r l+1) { dashed bars { for an 1D{3V sample eld
(n = 1 m = 3 r = 1 2 3). : : : : : : : : : : : : : : : : : : : : : : : 62
3.2 Computed power spectral density functions Sss (l1 k1 + s 2k1 l2 k2 +
s 2k2 ) { continuous bars { and Sss (k~krl +1) { dashed bars { for a 2D{2V
k
sample eld (n = 2 m = 2 r = 1 2). : : : : : : : : : : : : : : : : : : 64

3.3 Computation of the wave numbers k~krl +1 at which the resulted spec-
k
tral density functions are evaluated, under the assumption that Srr (k) const.
over each n{dimensional interval k. : : : : : : : : : : : : : : : : : : 65
3.4 Simulation of a 2D{3V stochastic eld { target SDF and rst 100
100m of the scalar components of the simulated eld (total dimensions:
2400 600m). : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 67
3.5 Simulation of a 2D{3V stochastic eld { comparison between target
and resulted spectral density functions : : : : : : : : : : : : : : : : : 68
3.6 Target and resulted spectral density functions represented along the
wave{number space diagonal : : : : : : : : : : : : : : : : : : : : : : : 68
171
3.7 Simulation of a 2D{3V stochastic eld { 1D sections of the resulted

scalar components: a,b. elds #1 and #2 { strongly correlated {
12 = 0:8 c,d. elds #1 and #3 { weakly correlated { 13 = 0:2. : : : 69
3.8 Simulation of a 2D{3V stochastic eld: a. resulted cumulative dis-
tributions vs. standard normal distribution b. input and resulted
correlation coecients. : : : : : : : : : : : : : : : : : : : : : : : : : : 70
3.9 Flowchart for numerical simulation of nD{mV non{Gaussian stochas-
tic elds, and illustration of dierent computational phases of proposed
iterative process. : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 75
3.10 Comparison between updating procedures { simulation of an 1D{2V
stochastic eld. : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 77
3.11 Convergence analysis using the individual correction procedure { sim-
ulation of an 1D{4V stochastic eld: a. non{correlated scalar compo-
nents b. correlated scalar components. : : : : : : : : : : : : : : : : : 78
3.12 Ratio between resulted and prescribed correlation coecients vs. av-
erage skewness coecients of the simulated scalar components. : : : : 80
3.13 Simulation of an 1D{4V non{Gaussian eld { comparison between
target and resulted spectral density functions (the relative errors shown
in the legend are computed after 20 iterations). : : : : : : : : : : : : 81
3.14 Simulation of an 1D{4V eld { part of the resulted eld (total simu-
lated length L0 = 5144m) comparison between non{Gaussian compo-
nents fBr(x) (continuous lines) and the corresponding Gaussian com-
ponents generated at the rst iteration fGr (1)
(x) (dotted lines). : : : : : 83
3.15 Simulation of an 1D{4V eld { comparison between prescribed and
resulted cumulative distributions of the scalar components. : : : : : : 84
3.16 Simulation of an 1D{4V eld { recovery of the target spectrum: a. rel-
ative errors of resulted SDF with respect to the target SDF during the
iterative process b. input and resulted correlation coecients. : : : : 84
3.17 Simulation of a 2D{3V non{Gaussian eld { comparison between tar-
get and resulted spectral density functions : : : : : : : : : : : : : : : 86
wave{number space diagonal (the relative errors shown in the legend
are computed after 25 iterations). : : : : : : : : : : : : : : : : : : : : 86
3.19 Simulation of a 2D{3V eld { comparison between prescribed and re-
sulted cumulative distributions of the scalar components. : : : : : : : 87
3.20 Simulation of a 2D{3V eld: a. evolution of relative errors during the
3.21 Simulation of a 3D{2V non{Gaussian eld { comparison between tar-
get and resulted spectral density functions. : : : : : : : : : : : : : : : 90
172
wave{number space diagonal (the relative errors shown in the legend
are computed after 5 iterations). : : : : : : : : : : : : : : : : : : : : : 91
3.23 Simulation of a 3D{2V eld { comparison between prescribed and re-
sulted cumulative distributions of the scalar components. : : : : : : : 91
3.24 Simulation of a 3D{2V eld: a. evolution of relative errors during the
3.25 Digital simulation of a 2D{2V non{Gaussian stochastic eld repre-
senting the eld conditions at Tarsiut P{45: comparison between pre-
scribed and resulted statistics of a sample eld. : : : : : : : : : : : : 92
3.26 Memory requirements (upper bounds) for in{core double precision
computations vs. the number of points in one spatial direction Mk
(for the case M1 = : : : = Mn). : : : : : : : : : : : : : : : : : : : : : : 96
3.27 Ratio between resulted and prescribed correlation coecients vs. pre-
scribed correlation coecient, for a 2V Lognormal eld. : : : : : : : : 99
4.1 Main features of the multi{yield plasticity soil constitutive model. : : 105
4.2 Evaluation of dilation parameter Xpp from the results of liquefaction
strength analysis: a. relationship between normalized penetration re-
sistance and cyclic stress ratio which causes liquefaction in NL cycles
b. illustrative example of the parametric studies using element tests
c. evaluation of dilation parameter using \regula falsi". : : : : : : : : 109
4.3 Results of a study on the performance of various soil constitutive mod-
els used for Class \A" predictions (VELACS project) in predicting ex-
cess pore pressure build{up: a. averages of the root{mean{square error
index over all the pore pressure transducers of each centrifuge model
b. evaluation of the root{mean{square error "rms between predicted
and recorded excess pore pressure time histories (after Popescu and
Prevost, 1995a). : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 111
4.4 Evaluation of soil properties based on eld test results. : : : : : : : : 115
4.5 Numerical example: evaluation of constitutive model parameters from
cone penetration test and general soil test results (continuous lines) re-
sults from other eld tests (self{boring pressuremeter, horizontal stress
cone) and laboratory soil tests are used to select the appropriate cor-
relation formulas (dashed lines). : : : : : : : : : : : : : : : : : : : : : 121
4.6 Evaluation of dilation parameter Xpp for Erksak sand: a. correlation
with cyclic stress ratio CSR b. estimation of initial parameter val-
ues for nonlinear regression analysis c. resulted correlation between
dilation parameter, cone tip resistance, and eective vertical stress. : 125
4.7 Sensitivity analysis: evaluation of constitutive model parameters from
eld data and general soil test results. : : : : : : : : : : : : : : : : : : 127
173
4.8 Sensitivity analysis set{up, and numerical computation results: a. -

nite element mesh b. base input motion c. horizontal relative displace-
ments at the ground level d. horizontal accelerations at the ground
level e. time histories of excess pore pressure ratio f. excess pore pres-
sure ratios at the end of shaking. : : : : : : : : : : : : : : : : : : : : 130
4.9 Results of the sensitivity analysis in terms of normalized variations
R1
R1 of the resulted liquefaction index: a. corresponding to relative
i
variations PP = 10% b. corresponding to expected variations Pi =

i
one standard deviation. : : : : : : : : : : : : : : : : : : : : : : : : : : 132

i
5.1 Comparison between local (spatial) averaging and midpoint discretiza-

tion methods (results based on two sample functions of a stochastic
eld describing the cone tip resistance data at Tarsiut P{45). : : : : : 141
5.2 1D sample correlation functions evaluated using various sampling in-
tervals (data from the hydraulic ll at Tarsiut P{45). : : : : : : : : : 144
5.3 Eects of variance reduction on predicted liquefaction index. : : : : : 146
5.4 Inuence of variance reduction and selected mesh size on predicted
average pore pressure build{up. : : : : : : : : : : : : : : : : : : : : : 148
5.5 Structure founded on lique able sand deposit: nite element analysis
set{up. : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 153
5.6 Predicted horizontal displacements of the top of structure, relative to
the base of structure (node 950 ; node 938). : : : : : : : : : : : : : : :
x x 154
5.7 Structure founded on lique able sand deposit { comparison between
deformations at the end of shaking, prescribed using deterministic and
stochastic parameter input. : : : : : : : : : : : : : : : : : : : : : : : 155
174
List of Tables
2.1 Common one{dimensional correlation structures : : : : : : : : : : : : 23
2.2 Comparison between results using various probability models : : : : : 32
2.3 Results of correlation structure analysis in vertical direction { layer 1 37
2.4 Results of correlation structure analysis in vertical direction { layer 2 37
2.5 Results of correlation structure analysis in horizontal direction { layer 1 44
3.1 Solution time and required memory : : : : : : : : : : : : : : : : : : : 93
4.1 Multi{yield plasticity model constitutive parameters : : : : : : : : : : 108
4.2 Comparison between SPT and CPT (after Campanella, 1994) : : : : 113
4.3 Correlations between soil parameters and eld test results : : : : : : 116
4.4 Index properties of Erksak 330/0.7 sand (after Been et.al., 1991) : : : 121
4.5 Estimation of geotechnical properties of Erksak sand : : : : : : : : : 122
4.6 Soil parameter values estimated for a hydraulic ll layer : : : : : : : : 128
R P
R (%), corresponding to variations P = 10% of the parameter values133
j i i

j i
R P
R (%), corresponding to expected variations P = Vi : : : : : : : 134
j i (1)
i
j i
175
Color Plates
1. Comparison between in{situ measurements (a.) and realizations of sample elds
(b.) of normalized uctuations of cone resistance ;qs(x z).
2. Selection of the mesh size and inuence of the assumed variance reduction on
computational results in terms of predicted excess pore pressure ratio ( eld data
from Tarsiut P{45) the variance of the stochastic eld is reduced according to:
A2 = (A)02.
3. Selection of the mesh size and inuence of the assumed variance reduction on
computational results in terms of predicted excess pore pressure ratio ( eld
data from Tarsiut P{45) no variance reduction is considered.
4. Comparison between deterministic and stochastic parameter input computa-
tions in terms of predicted excess pore pressure ratio ( eld data from Tar-
siut P{45).
5. Inuence of assumed spatial correlation distances on predicted excess pore pres-
sure ratio ( eld data from Tarsiut P{45).
6. Inuence of assumed probability distribution functions on predicted excess pore
pressure ratio ( eld data from Tarsiut P{45).
7. Computed excess pore pressure ratio with respect to the initial eective verti-
cal stress: comparison between results obtained using deterministic parameter
input (a.), and stochastic parameter input (b, : : : ,f.) various realizations of a
sample eld obtained using data from Akita Harbour are employed g,h. selec-
tion of the nite element mesh size i, : : : ,l. inuence of assumed correlation
distances.
8. Liquefaction index in vertical and horizontal directions, computed for the cases
shown in Plate 7 (data from Akita Harbour).
9. Structure founded on lique able soil ( eld data from Tarsiut P{45).
176
View publication stats

Stocastic Soil

Diunggah oleh

Informasi Dokumen

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Stocastic Soil

Diunggah oleh

Hak Cipta:

Format Tersedia

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

STOCHASTIC VARIABILITY OF SOIL PROPERTIES, DATA ANALYSIS, DIGITAL

Thesis · November 1995

DOE Applications View project

The user has requested enhancement of the downloaded file.

RECOMMENDED FOR ACCEPTANCE

0.1 Abstract : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : iii

2.3.5 Spatial cross{correlation between various soil properties : : : : 45

4.2.1 From laboratory test results and liquefaction strength : : : : : 108

A data base containing the general characteristics of spatial variability (probability

1.2 Literature review

1.2.2 Digital simulation of stochastic elds

1.2.3 Evaluation of soil properties

1.2.4 Eects of spatial variability on soil liquefaction

1.3 Proposed approach

2.1.1 Field data

borehole 44-2 borehole 44-5 borehole 58-22

(x = 0.0m) (x = 35.0m) (x = 70.0m)

2.1.2 Filtering and shortening

recordings. For each CPT prole:

of eld data su(x). Zero{mean, unit{variance, homogeneous elds are obtained

us(x z) = u(x z) ; uav (z)]=su(z) (2.2)

sq (zk ) = N ; 1 qf2 i(zk ) zk ; zk 1 = z

a. fixed intervals b. moving intervals

Figure 2.4: Evaluation of depth dependent standard deviation of uctuations about

1. Over xed horizontal sublayers, h (Figure 2.4,a):

Sample standard deviation of cone resistance uctuations is computed using both

2.1.4 Test for stationarity

Average values of Standardized

Layer 1 (hydraulic ll core) { N = 9 N1 = 20:

{ in vertical direction (Np = 8) { all data passed the test

2.2 Analytical models

1. exible, i.e. to have a wide diversity of distribution shapes

The theoretical models selected for tting empirical distributions of experimental

tp 1(1 ; t)q 1dt a x b

2.2.2 1D correlation structures and SDF

Table 2.1: Common one{dimensional correlation structures

In the case of separable correlation structure, the n{dimensional spectral den-

ED () = 2 s( ) cos d =     (2.21)

b1 = 5.0 b1 = 2.0 b2 = 0.5

values from numerical experiment

θ/b1 = 1.04 - 1.3 log b2 (0.1 < b2 < 2)

expression based on the root{mean{square frequency #2 (Harada and Shinozuka,

where 1 = b1 tan 2(b

b1 1:04 ; 1:3 log b2 0:1 b2 2 (2.28)

2.3 Estimation of spatial variability statistics

The optimal set of parameter values b^ correspond to the minimum residuals or

1. Select an expression for g(x b) from the theoretical model library.

2.3.2 Probability distribution function

0.5 Gaussian CDF

Figure 2.8: Probability distribution of standardized cone resistance values recorded

Table 2.2: Comparison between results using various probability models

very little discrepancy (usually a distribution assumption is not rejected if passing

-3 -1.5 0 1.5 3 -3 -1.5 0 1.5 3

Figure 2.9: Probability distribution of standardized SPT blow{counts recorded at

2.3.3 Correlation structure in vertical direction

process is computed as (Bendat and Piersol, 1986):

1. Squared Exponential correlation structure (SE) { Table 2.1

a. Sample correlation function and fitted analytical models

sample correlation function -0.5

Two-Peak model Exp. Decaying model

Since "(z) is random, its covariance function is a spike:

Table 2.3: Results of correlation structure analysis in vertical direction { layer 1

1.2.4 Eects of spatial variability on soil liquefaction

recordings. For each CPT prole:

of eld data su(x). Zero{mean, unit{variance, homogeneous elds are obtained

us(x z) = u(x z) ; uav (z)]=su(z) (2.2)

sq (zk ) = N ; 1 qf2 i(zk ) zk ; zk 1 = z

Figure 2.4: Evaluation of depth dependent standard deviation of uctuations about

1. Over xed horizontal sublayers, h (Figure 2.4,a):

Sample standard deviation of cone resistance uctuations is computed using both

Layer 1 (hydraulic ll core) { N = 9 N1 = 20:

{ in vertical direction (Np = 8) { all data passed the test

1. exible, i.e. to have a wide diversity of distribution shapes

The theoretical models selected for tting empirical distributions of experimental

ED () = 2 s( ) cos d = (2.21)

where 1 = b1 tan 2(b

b1 1:04 ; 1:3 log b2 0:1 b2 2 (2.28)

{ layer 1: z = 0:95m s2 = 0:90

qs(x1 kz)qsx2 (k + r)z] (2.45)

of a statistically signicant correlation. This is accomplished using a test proposed

4. The sample functions are obtained from eld data as follows:

stochastic sample elds having the resulted theoretical correlation struc-

R0jj ( ) = R0( ) j = 1 2 : : : m (3.2)

R0jk ( ) = jk R0( ) j k = 1 2 : : : m j 6= k (3.3)

3.2 Simulation of nD{mV Gaussian stochastic elds

{ the subscript G stands for \Gaussian" eld

between 0 and 2.

kisl = (li ; 1)ki + ms ki

ks0 = ms ki (3.9)

The simulated scalar elds fGr (x) r = 1 2 : : : m are asymptotically Gaussian

grs(j)(q1x1 q2x2 : : :qnxn) = l2:::l

for r = 3 = m (Figure 3.1,c) there is spectral power at all values k = lk +

interval k is again computed as:

n{dimensional interval k.

Sii0 (ki ) = 1 bai +1 ka exp f;b jk jg

3.3 Simulation of non{Gaussian vector elds