Professional Documents
Culture Documents
walls
Goudappa Dodagoudar*, Sajna Sayed and K. Rajagopal
Conventional finite element method (FEM) relies on deterministic values to represent structural and
load parameters and it does not have the capability of considering random parameters in response
evaluation of the structures. In the context of stochastic computational mechanics, uncertainty
modeling using random finite element method (RFEM) has by far received the major attention in
recent years. The objective of the paper is to develop a computer program for stochastic analysis of
reinforced retaining walls considering serviceability limit-state. The paper presents the formulation of
the RFEM for random field modeling of the reinforced retaining walls. The developed program is
validated using 1D bar and 2D plane-stress problems that involve random material properties. The
random field is simulated by Cholesky decomposition of the covariance matrix. The Monte Carlo
simulation (MCS) technique has been used in the finite element solution of the deformation response
variability of the reinforced soil retaining wall. Results of the random finite element analysis are useful
for serviceability limit-state design of the reinforced retaining walls.
Keywords: Spatial variability, Finite element method, Monte Carlo simulation, Reinforced wall
mE mean value of elastic modulus using local average subdivision method to study the
md(x) mean value of displacement at any point x stochastic response behavior of various geotechnical
wr friction angle of reinforced fill structures.
wb friction angle of backfill Based on the review of literature, it is noted that a very few
cr unit weight of reinforced fill studies have been attempted to study the probabilistic
cb unit weight of backfill response of reinforced soil retaining walls (Yuan et al.,
y dilation angle 2003; Chalermyanont and Benson, 2004; Chun et al., 2004;
H Lode angle Sayed et al., 2008) and stochastic response of conventional
retaining walls (Fenton et al., 2005). Moreover, no sys-
tematic study has been attempted in the literature on the
Introduction influence of spatial variability in soil properties on the overall
Probability theory provides a rational and efficient means response of the reinforced soil retaining walls. The previous
of characterizing uncertainty which is prevalent in studies on probabilistic response analysis of reinforced soil
computational stochastic mechanics. In recent years, the retaining walls (Yuan et al., 2003; Chalermyanont and
rapid development of probabilistic methods in the field of Benson, 2004; Chun et al., 2004; Sayed et al., 2008) have not
structural design has overshadowed considerably the been taken into account the effect of spatial variation of soil
fundamental role of limit-state design as it takes into properties. However, the conventional reliability analysis
account various uncertainties. Need for professional using explicit limit-state functions tends to over-represent its
judgment and experience in the selection of potential limit variability (Mahadevan and Haldar, 1991). It is to be noted
states is greater in foundation design than in structural that sampling at all locations to obtain the material pro-
design because geologic materials are among the most perty values is usually impractical and uneconomical, and
variable of all engineering materials. Conventional finite measurement and testing errors tend to dilute the value of the
element method (FEM) used in various fields of engineer- information. Therefore, the tasks of prediction, analysis,
ing for stress and deformation analysis without accounting and decision-making must usually proceed on the basis
for uncertainties is not a rational approach. Random finite of incomplete information about the medium, and this
element method (RFEM) or stochastic finite element fact renders its modeling as a random field meaningful
method (SFEM), which combines probability theory with (Vanmarcke et al., 1986).
deterministic FEM procedures, is becoming robust enough A brief description of the theoretical background of the
to allow engineers to estimate the risk of structural RFEM using MCS is given in the first part of the paper.
systems. In geotechnical engineering, Bathurst et al. Illustrative examples including one- and two-dimensional
(2008) applied reliability theory in the determination of (1D and 2D) structural problems are presented to
load and resistance factors in the load and resistance demonstrate the validation of the developed computer
factor design (LRFD) of reinforced soil walls and Fenton program for stochastic finite element analysis. Final part
et al. (2005) used RFEM for settlement analysis of shallow of the paper presents a numerical example of random finite
foundation resting on spatially variable soils. element analysis of reinforced soil retaining wall consider-
Applications of stochastic finite element analysis appear ing deformation of the wall as a response quantity. It is
to have been initiated in the early 1970s (Astill et al., 1972; noted that the results of the random finite element analysis
Der Kiureghian and Ke, 1988; Vanmarcke and Grigoriu, are useful for serviceability limit-state design of the
1983; Mahadevan and Haldar, 1991). The SFEM using reinforced soil retaining walls under the conditions of
first-order and second-order perturbation methods has random material properties.
been implemented by Handa and Anderson (1981),
Shinozuka and Yamazaki (1988), and Chakraborty and
Dey (1995) for stochastic response analysis of various
Random finite element method
structural systems. Another methodology is known as the Traditional deterministic analysis, such as the FEM, uses a
Neumann expansion method which has been found to be single design point, considering it as sufficient to represent
efficient when coupled with the Monte Carlo simulation the response. This simulation of a single design point is
(MCS) technique (Chakraborty and Dey, 1995; Shinozuka inadequate and unrealistic when characterizing systems
and Dasgupta, 1988; Yamazaki et al., 1988; Ren and under varying loads and material properties. These
Elishakoff, 1998; Kaminski, 2007). In geotechnical engi- uncertainties fluctuate over space and time domains, and
neering, although applications such as those described by the responses of the structure are accordingly affected by
Thurner and Schweiger (2000) combined the deterministic these parameters. The mathematical model of the varia-
FEM with reliability methods, they do not take advantage bility, parameterized by the correlation between different
of the FEM’s power to model spatial variation. Baecher locations, can be characterized by means of a random
and Ingra (1981) and Righetti and Harrop-Williams field.
(1988) implemented the SFEM to incorporate spatial Material properties are treated as random field that do
variation of random parameters for stress and settlement not vary randomly in space; rather such variation is gradual
analyses of foundations. Ishii and Suzuki (1987) and Gui and follows a pattern that can be quantified using spatial
et al. (2000) used the SFEM for slope stability analysis. correlation structures, where properties are treated as
Fenton and Griffiths, and their co-workers (Griffiths and random variables. The spatial correlation structure is often
Fenton, 2004; Fenton and Griffiths, 2005; Jaksa et al., expressed in terms of covariance function. If the variability
2005) have carried out random finite element analyses of the random field is entirely random, the covariance
function will decay asymptotically to zero. A convenient et al., 2007). Determining the covariance matrix among all
measure of the variability of a random field is the elements of the fluctuation part of the stiffness matrix
correlation length, also sometimes referred to as the scale requires extremely high computational effort in the case of
of fluctuation, which is the distance within which points of a Neumann expansion method (Matthies et al., 1997). The
random field are significantly correlated. To curve fit the spectral SFEM has been applied to linear problems, and it
covariance functions obtained from analysis of field data, is not applicable to general non-linear problems yet (Sudret
spatial correlation models such as spherical, exponential, and Der Kiureghian, 2000). However, as far as accuracy is
and Gaussian are used. These models help to determine the considered all the methods are equally good. In this study,
spatial correlation between field data at any separation the MCS method is used for carrying out random finite
distance and in different directions. The RFEM takes element analysis of 1D and 2D problems of the computa-
account of spatial correlation, which is the tendency for tional stochastic mechanics.
properties of soil elements ‘‘close together’’ to be correlated
while soil elements ‘‘far apart’’ are uncorrelated. Random field discretization
The stochastic analysis can be carried out using MCS by
generating random fields. For this purpose, the problem In most random finite element applications, it is necessary
domain is divided into appropriate number of finite to represent a continuous-parameter random field in terms
elements and the material property value for each element of a vector of random variables. This process is known
will be the summation of the values at the Gauss points. as discretization of the random field. A random field
So if there are n Gauss points, then there are n property discretization method is characterized by its efficiency, i.e.
values associated with each zero mean homogeneous its ability to accurately represent the random field with as
random field. This assessment is accomplished by calculat- few random variables as possible. The discretization
ing an autocovariance matrix whose value depends on the methods can be divided into three groups, namely point
characteristics of spatial correlation between the material discretization, average discretization, and series expansion
properties. This process is repeated several times and methods. The integration point method coming under
finally the mean and variance of nodal displacements are point discretization method mentioned by Matthies et al.
evaluated. (1997) referring to Brenner and Bucher (1995) is used
in the present study. Assuming that every integration
appearing in the finite element resolution scheme is
Monte Carlo simulation method obtained from integrand evaluation at each Gauss points
Monte Carlo simulation is known as a simple random of each element; the authors’ discretized the random
sampling method or statistical trial method that makes field by associating a single random variable to each of
realizations based on randomly generated sampling sets these Gauss points. This gives accurate results for short
for uncertain variables. Application of the MCS method correlation length. However, the total number of random
to probabilistic structural mechanics problems has been variables involved increases dramatically with the size of
attempted by several researchers (Vanmarcke et al., 1986; the problem.
Takada, 1991; Fenton and Griffiths, 2008; Shinozuka and One- and two-dimensional problems
Yamazaki, 1988). A simulation in RFEM proceeds by the A stochastic field is defined by the expected value and
following steps (Fenton and Griffiths, 2008): variance of the random variable and by the autocorrela-
1. From the raw data, histogram, statistical estimators, tion function. The spatial variation of some property such
and goodness-of-fit tests, a distribution with which to as the Young’s modulus can be written as
model the problem is decided upon. Parameters such
as the mean, variance, and correlation length are aðxÞ~a0 ½1zaðxÞ (1)
determined to characterize the randomness.
2. A random field following the distribution decided where a0 is the mean value of the property, a (x) is zero
upon in the previous step is defined. mean homogenous process, and x indicates the position
3. A realization of the random field is generated using a vector of a point on the domain where the problem is
random field generator. defined. The a (x) has a zero mean value and its
autocovariance function is given by
4. The response of the system to the random input
generated in the previous step is evaluated. Ra ðjÞ~E ½aðxÞ, aðxzjÞ~Caa (2)
5. The above algorithm is repeated from step 3 for as
many times as are feasible, recording the responses. where E[] represents the expected value and j is the
The MCS method has the advantage that it is adaptable to distance separating two points with coordinates x and
all types of problems and the results can be obtained to xzj. The autocovariance function is used to represent the
desired accuracy. However, for problems with many 1D stochastic field, given as
random variables or small failure probabilities this proce- Ra ðjÞ~s2a exp½{ðjjj=bl Þ (3)
dure is usually too expensive, since a large number
of solutions are needed to obtain reliable results. The where b is the autocorrelation length parameter, l is the
perturbation method is also known as the Taylor series length of finite element mesh, and sa is the standard
expansion method and, generally, its effectiveness is deviation of the fluctuation [a (x)] of the random variable.
restricted in that the random fluctuations must be small, The following autocovariance function is used for the 2D
i.e. the coefficient of variation should be less than 0?2 (Choi stochastic field
jj j jj j
Ra ðjÞ~E ½aðxÞ aðxzjÞ~s2a exp { 1 z 2 (4)
b1 l b2 l
where x5(x1, x2) and xzj5(x1zj1, x2zj2) are the
coordinates of two points, b1 and b2 are the autocorrela-
tion length parameters, l is the length of the finite element
mesh, and sa is the standard deviation of the fluctuation [a
(x)] of the random variable. Vector a can be generated by
a~LZ (5)
where L is a lower triangular matrix generated by
Cholesky decomposition of covariance matrix Caa given
by
0 1
Varða1 Þ . . . CVða1 ,an Þ 1 Bar subjected to linear body force distribution: a
B .. .. C Geometry and loads b Mesh discretization
Caa ~B
@ . P .
C
A (6)
CVða1 , an Þ Varðan Þ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ð1{Pf ÞPf =Ns
COVðPf Þ~ (11)
in which Var(a1) is the variance of a (x), CV(a1,a2) is the Pf
covariance of a (x), and Z is a vector of uncorrelated
Gaussian random variables with zero mean value and
The COV(Pf) is used as a criterion to decide when to stop
standard deviation equal to one. A randomly generated
the simulations. Typically a tolerable or target COV(Pf)
number r in the interval [0, 1], with a uniform distribution,
[dtarget] is specified and the simulations are terminated when
can be considered as input and Z can be computed by the
COV(Pf),dtarget. Usually the dtarget lies between 0?01 and
Box–Muller transformation to get normal deviates. Once 0?05 for most of the structural mechanics problems.
the stochastic field is defined, the direct MCS method uses The MCS with samples ranging from 1000–5000 have
the same formulation of the deterministic finite element been used to generate the uncertainty in the response variable
technique. For each simulated sample, it is necessary to 2 the displacement. The random variation of material
assemble the global stiffness matrix. The expected value of properties such as elastic modulus of soil is implemented by
displacement U at a particular node is given by developing new subroutines in FORTRAN. These subrou-
X
Ns tines incorporate the random field simulation using integra-
E ½U %-
U~ Ui =Ns (7) tion point method. Later these subroutines are incorporated
i~1 in the finite element program ‘‘GEOFEM’’ which was
where Ns is the number of simulations and Ui represents developed by Rajagopal (1998). Example problems reported
in the literature have been used to validate the developed
the displacement at any node for each simulation. The
program. As an example application, results of the random
standard deviation of displacement is given by
finite element analysis of the reinforced soil retaining wall are
N
Ps 2 presented in the paper.
Ui {-
U
s½U ~ i~1 (8) Validation examples: results and
Ns {1
where U
- is the mean of the displacement at a particular discussion
node. Given the values of the input variables, the MCS The methodology proposed using the MCS method is
method can easily evaluate performance function g(x) for applied to perform the random finite element analysis of
each run and then compute the probability of failure by
performing several simulations. The performance function
may be written as
g(x)~Uan {Un (9)
where Uan is the allowable displacement at any node n and
Un is the calculated displacement at the node n. The
probability of failure (Pf) is given by
Pf ~P½gðxÞv0~Nf =Ns (10)
where Nf is the number of simulations with g(x),0 (i.e.
number of failures) and Ns is the total number of
simulations. The accuracy of equation (10) can be
estimated by evaluating the coefficient of variation of
probability of failure [COV(Pf)] as 2 Histogram of the generated elastic modulus values
two validation examples and one numerical example of the and autocovariance function given by equation (3). A
reinforced soil retaining wall. mesh discretization involving 16 uniformly spaced nodes is
shown in Fig. 1b.
Example 1: One-dimensional bar The random finite element methodology developed in
A 1D bar example problem given in Rahman and Rao this study is applied to determine the axial displacement of
(2001) is considered for the validation of the developed the bar. Figure 2 shows the histogram of simulated elastic
random finite element program. Authors presented a modulus values. It follows normal distribution with mean
stochastic meshless method for solving boundary-value value (m)51 and standard deviation (s)50?1. Figures 3
problems in linear elasticity that involves random material and 4 show the mean md(x) and standard deviation sd(x)
properties. The material property was modeled as a respectively, of the axial displacement as a function of x. It
homogeneous random field and meshless formulation is can be seen from the figures that the results obtained using
used to predict the stochastic structural response. GEOFEM program are in good agreement with those
The bar AB of length, l51 units, is subjected to a linear obtained by Rahman and Rao (2001). Figure 5 shows the
body force distribution, p(x)5x, in x-direction as shown in variation of probability of failure with sample size and it
Fig. 1a. The point A of the bar is fixed and the point B is supports the fact that the fluctuation in the value of
free. The bar has a constant cross-sectional area, A51 probability of failure reduces as the sample size increases
units. The modulus of elasticity thereby enhancing confidence in the results.
E ðxÞ~mE ½1zaðxÞ (12) Example 2: Two-dimensional problem
is random with mean, mE51 units and a (x) is a Shinozuka and Yamazaki (1988) considered a 2D plane-
homogeneous Gaussian random field with mean zero stress problem (Fig. 6) with elastic modulus as random
13 Variation of mean displacement along the height of 14 Variation of standard deviation of displacement along
retaining wall the height of retaining wall
matrix (D) is first formulated based on the current tangent has been quoted as having a range from 2 to 42% with a
modulus value and Poisson’s ratio. During the plastic recommended value of 30% (Lee et al., 1983). In this study,
flow, correction is applied to this matrix to obtain the the coefficient of variation of elastic modulus [COV(Es)] of
elastoplastic constitutive matrix (Dep). the soil is varied from 10 to 40% and accordingly the
The reinforcement is modeled using three-noded bar displacement of the wall is obtained (Figs. 13 and 14). It is
elements, which are compatible with the eight-noded seen that the standard deviation of displacement increases
quadrilateral elements used for modeling the soil. The compared to the mean value with the increasing value of the
interfaces between the soil and the panel wall are modeled COV(Es). This is partly due to the fact that the standard
using six-noded joint elements. The finite element analysis deviation results only from deviatoric component of the
is carried out in two stages. In the first stage, the self- sample. A similar observation was also made by Shinozuka
weight of the soil is applied. The wall panels are restrained and Yamazaki (1988). Given the probability distribution of
in the lateral direction using external props simulated the limiting tolerable displacement for the reinforced
using nodal link elements. In the second stage, the uniform retaining wall, the tolerable displacement for deterministic
surcharge pressure is applied in increments of 0?2 kPa per design can be obtained, for example, by dividing the mean
load step. The maximum number of iterations per load limiting displacement by safety factor. Wahls (1994)
step is limited to 25. The properties of the elements used in reported a value of factor of safety equal to 1?5. For the
the finite element analysis are given in Tables 2–4. example problem analyzed in this study, the tolerable
lateral displacement at the top of the wall is 22?5 mm. For
Results design purposes, reliability charts can be prepared involving
In the 2D random finite element analysis, the backfill and allowable wall displacement as a criterion, based on which
reinforced soil properties are considered as random fields. the type of reinforcement required for the reinforced
The coefficient of variation of Young’s modulus of the soil retaining wall can be selected when the soil and reinforce-
ment properties are treated as random fields in the
Table 2 Properties of continuum elements
deformation analysis of the wall.