Probabilistic Assessment of Slope Stability That Considers

the Spatial Variability of Soil Properties

Sung Eun Cho, Ph.D.1
Downloaded from ascelibrary.org by UFPA - Universidade Federal Do Para on 10/03/19. Copyright ASCE. For personal use only; all rights reserved.

Abstract: In this paper, a numerical procedure for probabilistic slope stability analysis is presented. This procedure extends the tradi-
tional limit equilibrium method of slices to a probabilistic approach that accounts for the uncertainties and spatial variation of the soil
strength parameters. In this study, two-dimensional random fields were generated based on a Karhunen-Loève expansion in a fashion
consistent with a specified marginal distribution function and an autocorrelation function. A Monte Carlo simulation was then used to
determine the statistical response based on the generated random fields. This approach makes no assumption about the critical failure
surface. Rather, the critical failure surface corresponding to the input random fields of soil properties is searched during the process of
analysis. A series of analyses was performed to verify the application potential of the proposed method and to study the effects of
uncertainty due to the spatial heterogeneity on the stability of slope. The results show that the proposed method can efficiently consider
the various failure mechanisms caused by the spatial variability of soil property in the probabilistic slope stability assessment.
DOI: 10.1061/共ASCE兲GT.1943-5606.0000309
CE Database subject headings: Slope stability; Monte Carlo method; Probability; Failures; Limit equilibrium; Soil properties.
Author keywords: Slope stability; Monte Carlo method; Probability; Failures; Limit equilibrium.

Introduction sidered the spatial fluctuations of a parameter by using random

field theory 共El-Ramly et al. 2002; Griffiths and Fenton 2004;
The slope stability problem can be considered as a system that has Cho 2007兲.
many potential failure surfaces. However, the determination of Griffiths and Fenton 共2004兲 studied slope stabilities by using
the exact system reliability of the slope stability problem is not the random finite-element method 共RFEM兲, which combines the
possible. Therefore, the probability of failure for the most critical strength reduction method with random fields that are generated
slip surface is commonly used as the estimate for the system by using the local average subdivision method. In the context of
failure probability. This approach assumes that the probability the RFEM, a random field of shear strength is generated and
of failure along different slip surfaces is highly correlated mapped onto the finite-element mesh before carrying out the
共Chowdhury and Xu 1995兲. For the case of highly correlated strength reduction analysis. The critical failure mechanism is then
modes, the contribution to the system failure probability from determined in the process of analysis since the algorithm seeks
failure surfaces other than that associated with the maximum fail- out the most critical path through the generated heterogeneous
ure probability may be small, even though they are infinitely soil mass without assuming the shape and location of the failure
many 共Cornell 1967兲. surface a priori 共Griffiths et al. 2009兲. Although the RFEM has an
The common approaches to a probabilistic analysis are to lo- advantage in assuming no failure mechanism, the method suffers
cate the critical deterministic surface and then calculate the prob- from excessive computational efforts since the strength reduction
ability of failure that corresponds to this surface. However, the method calculates the factor of safety by progressively reducing
surface of the minimum factor of safety may not be the surface of or increasing the shear strength of the material in order to bring
the maximum probability of failure 共Hassan and Wolff 1999兲. As the slope to a state of limiting equilibrium.
an alternative, the critical probabilistic surface that is associated The majority of probabilistic slope stability analyses usually
with the highest probability of failure or the lowest reliability employ the traditional limit equilibrium method 共LEM兲. At the
could be considered 共Li and Lumb 1987; Chowdhury and Xu early stages of the studies, each soil parameter was modeled by
1995; Bhattacharya et al. 2003; Li and Cheung 2001兲. only a single random variable 共Low et al. 1998; Chowdhury and
Soil properties vary spatially, even within homogeneous lay- Xu 1995; Hassan and Wolff 1999兲. Some investigators have cal-
ers, as a result of the depositional and postdepositional processes culated the probability of failure by using the LEM to consider
共Lacasse and Nadim 1996兲. Therefore, recent studies have con- the spatial variability of the soil properties 共Li and Lumb 1987;
El-Ramly et al. 2002; Cho 2007兲. However, when the traditional
1
Senior Researcher, K-Water Institute, Korea Water Resources Corpo- LEM is combined with the random field theory, the influence of
ration, 462-1, Jeonmin-Dong, Yusung-Gu, Daejon 305-730, South Korea. the random field is only taken into account along the predeter-
E-mail: drsecho@hanmail.net mined critical surface since the location where the shear strength
Note. This manuscript was submitted on July 29, 2009; approved on
is calculated continuously changes in the process of searching the
December 20, 2009; published online on December 28, 2009. Discussion
period open until December 1, 2010; separate discussions must be sub- critical failure surface. The LEM is simple and is therefore suit-
mitted for individual papers. This paper is part of the Journal of Geo- able for the Monte Carlo simulation. However, the method re-
technical and Geoenvironmental Engineering, Vol. 136, No. 7, July 1, quires a random field simulation method that is independent of
2010. ©ASCE, ISSN 1090-0241/2010/7-975–984/$25.00. the division of slices in the sliding mass in order to be able to

JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JULY 2010 / 975

J. Geotech. Geoenviron. Eng., 2010, 136(7): 975-984

 calculate the shear strength at any location along the trial slip of X. In n-dimensional hyperspace of the basic variables, g共X兲
surface. = 0 is the boundary between the region in which the target factor
In this study, the effect of spatial variability of shear strength of safety is not exceeded and the region in which it is exceeded.
on the slope stability was studied by using the random LEM. Cornell 共1967兲 pointed out that a slope is actually a series
Random fields of shear strength parameters were generated by the system with an infinite number of components 共or potential
Karhunen-Loève expansion method that is independent of the di- modes of failure兲. Hong and Roh 共2008兲 dealt with the system
vision of slices. Then, statistical responses were obtained without aspect of the slope in the reliability analysis by defining a limit
assuming a failure mechanism by carrying out a series of limit state of the system as a function of the minimum of the factor of
equilibrium analysis based on a search algorithm that can find the safety for all potential slip surfaces.
critical failure surface for the generated random fields. The appli- If ge共X 兩 s兲 is defined as the limit state function of a single
cation of the procedure is illustrated below through example prob- potential slip surface, s, in terms of the factor of safety, ge共X 兩 s兲
Downloaded from ascelibrary.org by UFPA - Universidade Federal Do Para on 10/03/19. Copyright ASCE. For personal use only; all rights reserved.

lems. can be defined as Eq. 共2兲. Then the limit state function g共X兲 that
considers all potential slip surfaces is defined as Eq. 共3兲. Here

Deterministic Slope Stability Analysis ge共X兩s兲 = Fs − 1.0 共2兲

LEM g共X兲 = min关ge共X兩s兲兴 共3兲

all s
Slope stability problems are commonly analyzed by using LEMs
of slices. The failing soil mass is divided into a number of vertical
slices in order to calculate the factor of safety, which is defined as The probability of failure of the slope is then given by the fol-
the ratio of the resisting shear strength to the mobilized shear lowing integral 共Baecher and Christian 2003兲:
stress to maintain the static equilibrium. The static equilibrium of
the slices and the mass as a whole are used to solve the problem.
However, all methods of slices are statically indeterminate and, as P f = P关g共X ⱕ 0兲兴 = 冕
g共X兲ⱕ0
f X共X兲dX 共4兲
a result, require assumptions in order to solve the problem. The
approach presented here adopts Bishop’s simplified method,
which is widely accepted as reasonably accurate and is applicable where f X共X兲 represents the joint probability density function and
to the failure surface of a circular shape. the integral is carried out over the failure domain.
For slope stability problems, direct evaluation of the n-fold
integral is virtually impossible. The difficulty lies in that complete
Search for the Critical Failure Surface probabilistic information on the soil properties is not available
LEMs require that the critical failure surface be determined as and the domain of integration is a complicated function. There-
part of the analysis. The problem of locating the critical circular fore, approximate techniques have been developed to evaluate
surface can be formulated as an optimization problem as follows: this integral.

min Fs共xc,y c,R兲 共1兲 FORM

surface
The first-order reliability evaluation of Eq. 共4兲 is accomplished by
where Fs = objective function 共the factor of safety兲; xc and y c
transforming the uncertain variables, X, into uncorrelated stan-
= coordinates of the center of the circle; and R = radius of the
dard normal variables, Y. The primary contribution to the prob-
circle.
ability integral in Eq. 共4兲 comes from the part of the failure region
Although a circular trial slip surface can be described as a
关G共Y兲 ⱕ 0, for which G共Y兲 is the limit state function in the trans-
function of three shape variables, it could be defined as a series of
formed normal space兴 closest to the origin. The design point is
straight segments. The location of the segment vertices is deter-
defined as point Yⴱ in the standard normal space located on the
mined by shape variables and by the location of a hard stratum
limit state function 关G共Y兲 = 0兴 with the maximum probability den-
into which the slip surface cannot penetrate. The treatment for
sity attached to it. Therefore, the design point, which is the near-
circular slip surface is very simple because there are only three
est point to the origin in the failure region, is an optimum point at
location parameters. Eq. 共1兲 is a type of unconstrained optimiza-
which to approximate the limit state surface. The probability ap-
tion problem that has no equality or inequality conditions. The
proximated at the design point is
BFGS method, which is widely acknowledged to be efficient, was
applied in this study in order to search for the critical circular slip
surface. Kim and Lee 共1997兲 presented a detailed description of P关g共X兲 ⱕ 0兴 ⬇ ⌽共− ␤兲 共5兲
this procedure.
where ␤ = reliability index defined by the distance from the origin
to the design point and ⌽ = standard normal cumulative density
Probabilistic Slope Stability Analysis function.
In the first-order reliability method 共FORM兲, a tangent hyper-
plane is fitted to the limit state surface at the design point. There-
Limit State Function fore, the most important and demanding step in the method is
The problem of the probabilistic slope stability analysis is formu- finding the design point. The design point is the solution of the
lated by a vector, X = 关X1 , X2 , X3 , . . . , Xn兴, which represents a set following nonlinear constrained optimization problem:
of random variables. From the uncertain variables, a limit state
function g共X兲 is formulated to describe the limit state in the space min储Y储 subject to G共Y兲 = 0 共6兲

976 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JULY 2010

J. Geotech. Geoenviron. Eng., 2010, 136(7): 975-984

 Monte Carlo Simulation shape function method. These early methods are relatively ineffi-
cient in the sense that a large number of random variables are
An alternative means of evaluating the multidimensional integral
required to achieve a good approximation of the field. More effi-
of Eq. 共4兲 is to use a Monte Carlo simulation. In a Monte Carlo
cient approaches that use series expansion methods such as the
simulation, discrete values of the component random variables are
Karhunen-Loève expansion and the expansion optimal linear es-
generated in a fashion that is consistent with their probability
timation method have been introduced in previous studies 共Sudret
distribution, and the limit state function is calculated for each
and Der Kiureghian 2002兲.
generated set. The process is repeated many times to evaluate the
All series expansion methods result in a Gaussian field that is
probability of failure by determining whether the limit state func-
represented exactly as a series that involves random variables and
tions are exceeded. However, the Monte Carlo simulation method
deterministic spatial functions that depend on the correlation
is not limited to calculating the probability of failure. Various
structure of the field. The accuracy of the representation depends
statistical properties that are evaluated after the process of simu-
Downloaded from ascelibrary.org by UFPA - Universidade Federal Do Para on 10/03/19. Copyright ASCE. For personal use only; all rights reserved.

on the number of terms used in the series expansion and the kind
lation, such as mean, variance, probability density functions, and
of particular expansion method that is used. In this study, the
cumulative probability distribution functions, can provide a
Karhunen-Loève expansion was adopted to simulate anisotropic
broader perspective and a more comprehensive description of a
random fields of soil properties in two-dimensional space since
given slope. Despite the fact that this approach requires excessive
the method requires the minimum number of terms for a specified
computational effort, the mathematical formulation of the Monte
level of accuracy compared to other mathematical representations
Carlo simulation is relatively simple and the method has the ca-
共Ghiocel and Ghanem 2002兲 and is independent of the division of
pability of handling practically every possible case regardless of
slices.
its complexity.

Karhunen-Loève Expansion
Random Field Model A random field H共x , ␪兲 is a collection of random variables asso-
ciated with a continuous index x 苸 ⍀, where ⍀ is an open set of
Spatial Variability of Soil Rn describing the system geometry and ␪ 苸 ⌰ is the coordinate in
the outcome space. The Karhunen-Loève expansion of a random
One of the main sources of heterogeneity is inherent spatial soil
field H共x , ␪兲 is based on the spectral decomposition of its auto-
variability due to different depositional conditions and different
correlation function ␳共x , x⬘兲, which is bounded, symmetric, and
loading histories 共Elkateb et al. 2003兲. Statistical parameters, such
positive definite. The set of deterministic functions over which
as the mean and variance, are one-point statistical parameters and
any realization of the field H共x , ␪0兲 is expanded is defined by the
cannot capture the features of the spatial structure of the soil
eigenvalue problem as
共El-Ramly et al. 2002兲. Spatial variations of soil properties can be
effectively described by their correlation structure within the
framework of random fields 共Vanmarcke 1983兲.
Autocorrelation distance, defined as the distance to which the
冕⍀
␳共x,x⬘兲␸i共x⬘兲d⍀x⬘ = ␭i␸i共x兲 共8兲

autocorrelation function decays to 1 / e 共e is the base of natural where ␸i and ␭i denote the eigenfunctions and eigenvalues of the
logarithms兲, has been used to describe the spatial extent within autocorrelation function, respectively.
which soil properties show a strong correlation. A large autocor- The series of the deterministic set forms the expansion of
relation distance value implies that the soil property is highly H共x , ␪兲
correlated over a large spatial extent, resulting in a smooth varia-
⬁
tion within the soil profile. On the other hand, a small value
indicates that the fluctuation of the soil property is large. H共x,␪兲 = ␮ + 兺 ␴冑␭i␸i共x兲␹i共␪兲, x苸⍀ 共9兲
i=1
Although an isotropic correlation structure was often assumed
in previous studies, correlations in the vertical direction tend to where ␹i共␪兲 = set of orthogonal random coefficients 共uncorrelated
have much shorter distances than those in the horizontal direction random variables with zero mean and unit variance兲. The approxi-
due to the geological soil formation process for most natural soil mate random field is defined by truncating the ordered series
deposits. A ratio of about 1–10 for these autocorrelation distances given in Eq. 共9兲
is considered common 共Baecher and Christian 2003兲. M

兺 ␴冑␭i␸i共x兲␹i共␪兲,
A Gaussian random field is completely defined by its mean
Ĥ共x,␪兲 = ␮ + x苸⍀ 共10兲
␮共x兲, variance ␴2共x兲, and autocorrelation function ␳共x , x⬘兲. In this i=1
study, an exponential autocorrelation function was used and dif-
ferent autocorrelation distances in the vertical and horizontal di- The number M to be chosen strongly depends on the desired
rections were used as follows: accuracy and on the autocorrelation function of the random field.
In the case of an exponential autocorrelation function 关Eq. 共7兲兴

冉
␳共x,y兲 = exp −
兩x − x⬘兩 兩y − y ⬘兩
lh
−
lv
冊 共7兲
for a one-dimensional case, the eigenvalue problem 关Eq. 共8兲兴 can
be solved analytically. Extension to two-dimensional fields de-
fined for the correlation function on a rectangular domain can be
where lh and lv = autocorrelation distances in the horizontal and obtained as well. Detailed closed-form solutions can be found in
vertical directions, respectively. Ghanem and Spanos 共1991兲.

Simulation of Random Fields Cross-Correlated Gaussian Random Fields

Several methods have been developed to carry out this task, such Typically, more than one random property is involved in geotech-
as the spatial average method, the midpoint method, and the nical problems. For example, the cohesion and the friction angle

JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JULY 2010 / 977

J. Geotech. Geoenviron. Eng., 2010, 136(7): 975-984

 can be considered as random properties in the problem of slope
stability. The present study dealt with cases where all the fields
simulated over a region ⍀ shared an identical autocorrelation
function over ⍀, and the cross correlation structure between each
pair of simulated fields was simply defined by a cross correlation
coefficient. This is reasonable since the spatial correlation struc-
ture is generally caused by changes in the constitutive nature of
the soil over space 共Fenton and Griffiths 2003兲. Therefore, the
modal decomposition of the given autocorrelation function is
done only once. The cross-correlated fields are then expanded by Fig. 1. Example 1: cross section 共Fs = 1.356兲
using the same spectrum of eigenfunctions and eigenvalues, but
Downloaded from ascelibrary.org by UFPA - Universidade Federal Do Para on 10/03/19. Copyright ASCE. For personal use only; all rights reserved.

the sets of random variables used for the expansion of each field
are cross correlated. study, the autocorrelation distance l in the Gaussian field was
In this study, each field of cohesion and friction angle was considered in order to maximally exploit the analytical solution of
expanded by using a set of independent random variables. These Eq. 共8兲.
sets are then correlated with respect to the cross correlation ma- All random variables were assumed to be characterized statis-
trix between two expanded random fields based on the framework tically by a lognormal distribution defined by a mean ␮X and a
presented by Vořechovský 共2008兲. standard deviation ␴X. Once the mean and standard deviation are
Each approximate Gaussian random field Ĥi is then expanded expressed in terms of the dimensionless coefficient of variation,
as follows: defined as COVX = ␴X / ␮X, the mean and standard deviation of the
underlying normal distribution of ln X are then given by
M

Ĥi共x,␪兲 = ␮i + 兺 ␴i冑␭ j␸ j共x兲␹i,j共␪兲 共for i = c,␾兲 共11兲 ␴ln X = 冑ln兵1 + COV2X其 共13兲
j=1

To generate a random field, it is necessary to simulate the corre- ␮ln X = ln ␮X − 0.5␴2ln X 共14兲
lated random vector ␹i,j. In this study, the Latin hypercube sam-
pling technique was used. This technique can be viewed as a
stratified sampling scheme designed to ensure that the upper or Example 1: Application to a Saturated Clay Slope
lower ends of the distributions are well represented. under Undrained Conditions „␾u = 0…
In this example, a probabilistic study is performed on an un-
Transformation to Non-Gaussian Random Fields drained slope with a cross section, as shown in Fig. 1. The un-
drained shear strength cu is modeled as a random field. Table 1
Although a Gaussian random field is often used to model uncer- summarizes the statistical properties of soil parameters for the
tainties with spatial variability for reasons of convenience and a slope. The minimum factor of safety that is associated with the
lack of available data, the Gaussian model is not applicable in critical failure surface based on the mean value of the undrained
many situations where the random variable is always positive. For shear strength is 1.356. The critical deterministic surface is deep
convenience, we find an underlying Gaussian random field that and passes through the foundation soil, as shown in Fig. 1. The
can be easily transformed into the target field. If the random vari- critical probabilistic surface associated with the maximum prob-
ables are considered to be log-normally distributed, then appro- ability of failure, as determined by a search of FORM that as-
priate lognormal random fields can be obtained by exponentiating sumes a perfect spatial correlation, is identical to the critical
the approximate Gaussian field from Eq. 共11兲 as follows: deterministic surface. The reliability index calculated by FORM

冋 册
M
is 0.892 and the probability of failure is 0.186.

兺 ␴i冑␭ j␸ j共x兲␹i,j共␪兲
In the current study, two types of Monte Carlo simulations that
H̃i共x,␪兲 = exp ␮i + 共for i = c,␾兲
j=1
considered the spatial variability of soil property were performed.
In the first approach, random LEM analysis that determined the
共12兲 critical failure surface based on the search algorithm for each
generated random field was carried out in a Monte Carlo simula-
tion. In the second approach, the spatial variability of soil param-
Example Analysis eters was considered only for the previously located critical
surface without considering a search procedure that identifies the
In this section, the application of the presented procedure is illus- critical failure surface in a Monte Carlo simulation.
trated through example problems. For simulations of random In order to obtain accurate statistical responses, 100,000 sets
field, the solution of eigenvalue problem 关Eq. 共8兲兴 and the Latin of random fields were generated for each case based on the sta-
hypercube sampling technique were implemented in Matlab. tistical information. A series of analyses was then performed
Then, the simulated random field was used through direct cou- based on the generated random fields.
pling with the slope stability analysis routine written in FOR-
TRAN.
According to the results of a literature review by El-Ramly et Table 1. Example 1: Statistical Properties of Soil Parameters
al. 共2003兲, the autocorrelation distance is within a range of 10–40 Parameter ␮X COVX
m in the horizontal direction, while in the vertical direction it
Unit weight ␥sat 共kN/ m 兲 3
20 —
ranges from 1 to 3 m. As an illustration, an exponential autocor-
Undrained shear strength cu 共kPa兲 23 0.3
relation structure with lh = 20 m and lv = 2 m is used. In this

978 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JULY 2010

J. Geotech. Geoenviron. Eng., 2010, 136(7): 975-984

 
Table 2. Example 1: Influence of Number of Slices on the Statistical

 
 Response 共lh = 20 m, lv = 2 m, and COVcu = 0.3兲


 


 
Number of slices Pf ␮ Fs ␴ Fs COVFs

45 0.0767 1.2663 0.1993 0.1574
50 0.0763 1.2665 0.1992 0.1573

60 0.0760 1.2669 0.1991 0.1572








  
 slices, the simulated random fields can be better reflected into the
limit equilibrium analyses. As indicated in the table, if a sufficient
Downloaded from ascelibrary.org by UFPA - Universidade Federal Do Para on 10/03/19. Copyright ASCE. For personal use only; all rights reserved.

Fig. 2. Eigenvalues of the autocorrelation function 共lh = 20 m and number of slices are used to represent the random field, increasing
lv = 2 m兲 the number of slices does not significantly alter the responses.
Fig. 3 shows four typical realizations of random field at failure
with the searched critical failure surface. In the figures, the lighter
Random fields were generated based on the solutions of the regions denote a larger strength parameter value while the darker
eigenvalue problem. The number of eigenmodes to be retained regions indicate a smaller strength parameter value. These figures
while simulating a random field depends on the magnitudes of the show that the failure region developed through the weak strength
corresponding eigenvalues. Fig. 2 represents the decaying trends path in the slope. Various failure surface geometries, caused by
of the eigenvalues obtained by solving the Karhunen-Loève inte- the spatial heterogeneity, are not manifested in the deterministic
gral eigenvalue problem. In this example, 100 terms of eigen- analysis or the probabilistic analysis with a single random vari-
mode were used to represent the random field of undrained shear able due to the representation of a homogeneous soil medium.
strength. The results of the simulation are summarized in Fig. 4. Fig.
As previously explained, a continuous random field can be 4共a兲 shows the convergence of the simulation. As indicated in the
obtained for an exponential autocorrelation function on a rectan- figure, the overall probability of failure calculated from the ran-
gular domain based on the analytical solution of the eigenvalue dom LEM 共0.079兲 is much greater compared to that obtained for
problem by the Karhunen-Loève expansion method. In other the fixed critical surface 共0.032兲. Fig. 4共b兲 shows the probability
words, the random field simulation is independent of the division density functions of the factor of safety determined from the two
of slices. The accuracy of the random field generated by the approaches. Fig. 4共c兲 shows the probability distributions of the
Karhunen-Loève expansion method depends on the number of factor of safety with the probability of failure obtained from
terms used in the series expansion, not on the slice width. The FORM. They indicate that the statistical responses calculated
slice width in the sliding mass only controls the accuracy of spa- from the Monte Carlo simulations for the two approaches are
tial representation of the generated random field. significantly different. The reason for this is that the random LEM
The influence of the slice division in the sliding mass on the does not assume a particular failure mechanism but searches for
statistical response for the case where lh = 20 m, lv = 2 m, and the most critical path through the heterogeneous soil mass repre-
COVcu = 0.3 is illustrated in Table 2. With a larger number of sented by random field. There are various possible failure mecha-

   

 

   


 

Fig. 3. Example 1: typical realizations of random field and corresponding analysis result 共lh = 20 m and lv = 2 m兲: 共a兲 Fs = 0.905; 共b兲 Fs
= 0.927; 共c兲 Fs = 0.834; and 共d兲 Fs = 0.983

JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JULY 2010 / 979

J. Geotech. Geoenviron. Eng., 2010, 136(7): 975-984

 0.3
Fixed critical surface

Probability of failure
Random LEM
0.2

0.1

0
1 10 100 1000 10000 100000
Number of simulation
(a)
Downloaded from ascelibrary.org by UFPA - Universidade Federal Do Para on 10/03/19. Copyright ASCE. For personal use only; all rights reserved.

3
Fixed critical Fig. 5. Example 2: cross section 共Fs = 1.204兲
2.5 surface
Deterministic value Random LEM
Fs=1.356
2

mated by the fixed critical surface is accurate only in cases where

Pdf

1.5
the correlation between the safety factors of different surfaces is
1 very high 共Cornell 1967兲.
Table 3 presents summaries of simulation results for cases that
0.5
have three different sets of autocorrelation distances. The relative
0
difference of the two values of probability of failure was calcu-
0 1 2 3
lated based on the results of the two approaches. As can be noted
Factor of safety Fs from the table, the relative difference decreased when the auto-
(b) correlation distance increased, which means that the correlation
between the limit state functions of different surfaces increased
1
and the failure surfaces by random LEM approached the fixed
critical failure surface.
0.8
Cumulative probability

0.6 Example 2: Application to a c-␾ Slope

Fs=1.0 Deterministic value This example concerns the stability for a c-␾ slope as presented
0.4 Fs=1.356
in Fig. 5. As the strength of the soil is spatially distributed, the
Fixed friction angle and cohesion are considered as random fields. Sta-
0.2 FORM
critical surface tistical moments for these parameters are shown in Table 4.
Random LEM
In order to incorporate the dependence between the strength
0 parameters, the cross correlation coefficient r共c , ␾兲 is needed.
0 1 2 3 Wolff 共1985兲 reported the correlation between c and ␾ for
Factor of safety Fs
consolidated-undrained 共CU兲 tests as r = 0.25 and for conso-
(c)
lidated-drained 共CD兲 tests as r = −0.47. Yucemen et al. 共1973兲
Fig. 4. Example 1: results of the Monte Carlo simulation 共COVcu
reported values in a range of −0.49ⱕ r ⱕ −0.24, while Lumb
= 0.3, lh = 20 m, and lv = 2 m兲: 共a兲 convergence of the probability of 共1970兲 noted values of −0.7ⱕ r ⱕ −0.37. A negative correlation
failure; 共b兲 probability density function; and 共c兲 probability distribu- implies that low values of cohesion are associated with high val-
tion of the factor of safety ues of friction angle and vice versa. In other words, a negative
correlation between the cohesion and friction angle means that the
uncertainty in the calculated shear strength is smaller than the
nisms for undrained slopes. The failure surface may go through combined uncertainty in the two-parameter values used to model
the toe or pass through the deeper foundation soil, as shown in the shear strength. This observation arises from the fact that the
Fig. 3, which leads to a low correlation between the limit state variance of the shear strength is reduced if there is a negative
functions of different surfaces. The probability of failure approxi- correlation between the cohesion and friction angle 共Fenton and

Table 3. Example 1: Summaries of Simulation Results

lh lv
共m兲 共m兲 COVcu Pf ␮ Fs ␴ Fs COVFs Relative difference of P f
20 2
0.0316 0.3 1.3510 0.2140 0.1584 0.5842 F
0.0760 1.2669 0.1991 0.1572 S
20 4 0.3 0.0621 1.3528 0.2541 0.1879 0.4308 F
0.1091 1.2830 0.2430 0.1894 S
40 2 0.3 0.0422 1.3526 0.2301 0.1701 0.5280 F
0.0894 1.2705 0.2138 0.1683 S
Note: F = Monte Carlo simulation is performed for the fixed critical surface and S = Monte Carlo simulation is performed based on the search algorithm for
the critical failure surface.

980 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JULY 2010

J. Geotech. Geoenviron. Eng., 2010, 136(7): 975-984

 Table 4. Example 2: Statistical Properties of Soil Parameters The critical failure surface of the deterministic analysis passes
Parameter ␮X COVX Correlation coefficient the toe of slope, as shown in Fig. 5, while the factor of safety was
estimated to be 1.204. The critical probabilistic surface that was
Cohesion c 共kPa兲 10 0.3 −0.7ⱕ r ⱕ 0.5
determined by a search of FORM passed through the same sur-
Friction angle ␾ 共degrees兲 30 0.2
face as the critical deterministic surface. The reliability index was
Unit weight ␥t 共kN/ m3兲 20 — —
1.236 and the probability of failure was 0.108.
Fig. 6 shows three typical realizations of random fields at fail-
Griffiths 2003兲. In this study, a value of ⫺0.5 is considered as a ure with the searched critical failure surface. In the figures, it can
base set, while the range of −0.7ⱕ r ⱕ 0.5 around the base value be observed that the cohesion and friction angle show a negative
is considered. correlation while the failure surfaces tend to pass through near the
toe.
Downloaded from ascelibrary.org by UFPA - Universidade Federal Do Para on 10/03/19. Copyright ASCE. For personal use only; all rights reserved.

In order to obtain statistical responses, a series of analyses was

performed based on the 50,000 sets of random fields generated for The results of the Monte Carlo simulation for the case where
each case. In this study 150 terms of eigenmode were used to r共c , ␾兲 = −0.5, COVc = 0.3, COV␾ = 0.2, lh = 20 m, and lv = 2 m are
represent the random fields of cohesion and friction angle 共see summarized in Fig. 7. Fig. 7共a兲 shows the overall convergence of
Fig. 2兲. the simulations. As indicated in the figure, the overall probability


   
  




   
  




   
  




Fig. 6. Example 2: typical realizations of random fields and corresponding analysis result 关r共c , ␾兲 = −0.5, lh = 20 m, and lv = 2 m兴: 共a兲 Fs
= 0.986; 共b兲 Fs = 0.913; and 共c兲 Fs = 0.985

JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JULY 2010 / 981

J. Geotech. Geoenviron. Eng., 2010, 136(7): 975-984

 0.1 0.14

Coefficient of variation
Fixed critical surface

Probability of failure
0.08
Random LEM 0.12
0.06
0.1
0.04

0.02 0.08 Fixed critical surface

Random LEM
0
0.06
1 10 100 1000 10000 100000
Number of simulation -0.8 -0.4 0 0.4 0.8
(a)
Cross-correlation coefficient r(c,φ)
(a)
Downloaded from ascelibrary.org by UFPA - Universidade Federal Do Para on 10/03/19. Copyright ASCE. For personal use only; all rights reserved.

4
Fixed 5
3.5 critical surface r(c,φ)= −0.7
4.5
3 Random LEM
4 r(c,φ)= −0.5
r(c,φ)= −0.25
2.5 3.5
r(c,φ)= 0
Pdf

2 3
r(c,φ)= 0.25

Pdf
2.5 r(c,φ)= 0.5
1.5
Deterministc value 2
1
Fs=1.204 1.5
0.5 1
0 0.5
0 1 2 3 0
Factor of safety Fs
0 1 2 3
(b) Factor of safety Fs
(b)
1

1
0.8
Cumulative probability

0.8
Cumulative probability

0.6
Fs=1.0
0.6
0.4 Deterministic value r(c,φ)= −0.7
Fs=1.204 Fs=1.0
r(c,φ)= −0.5
Fixed 0.4
0.2 r(c,φ)= −0.25
critical surface
r(c,φ)= 0
FORM
Random LEM
0.2 r(c,φ)= 0.25
0
r(c,φ)= 0.5
0 1 2 3 0
Factor of safety Fs
(c) 0 1 2 3
Factor of safety Fs

Fig. 7. Example 2: results of the Monte Carlo simulation 关r共c , ␾兲 (c)

= −0.5, COVc = 0.3, COV␾ = 0.2, lh = 20 m, and lv = 2 m兴: 共a兲 conver-

Fig. 8. Example 2: influence of the cross correlation coefficient on
gence of the probability of failure; 共b兲 probability density function;
the statistical response obtained from simulation 共COVc = 0.3,
and 共c兲 probability distribution of the factor of safety
COV␾ = 0.2, lh = 20 m, and lv = 2 m兲: 共a兲 COV of factor of safety; 共b兲
probability density function 共random LEM兲; and 共c兲 probability dis-
tribution of the factor of safety 共random LEM兲
of failure calculated from the random LEM 共0.0171兲 is greater
compared to that obtained for the fixed critical surface 共0.0138兲,
but the difference is not significant. Figs. 7共b and c兲 show that the density functions of the factor of safety that was estimated from
probability density functions and the probability distribution of the random LEM. The shape of the probability density function
the factor of safety determined from the two approaches are al- becomes narrower and the uncertainty in the factor of safety de-
most identical. The reason for this is that the correlation between creases when the negative value of the cross correlation coeffi-
the limit state functions of various failure surfaces is very high cient increased. The probability distribution estimated from the
since, for c-␾ slopes, the failure mechanism is nearly always random LEM also shows the same trend, as shown in Fig. 8共c兲.
shallow and passes through the toe, as shown in Fig. 6. Fig. 9 shows the estimated probability of failure against the
In this study, the effect of varying the cross correlation be- cross correlation. The probability of failure decreases when the
tween c and ␾ was investigated. Fig. 8共a兲 shows the effect of negative correlation coefficient increases. Therefore, the assump-
cross correlation on the COV of factor of safety. Since the in- tion of independence between the cohesion and friction angle
crease of one parameter value decreased the other value, the gives conservative results if the actual correlation is negative, but
variation of the total shear strength was reduced, and conse- slightly unconservative results are obtained if the actual correla-
quently, the variation of the factor of safety also decreased. An tion is positive.
opposite effect for the case of a positive value of the correlation The effects of the autocorrelation distance on simulation re-
coefficient was also observed. Fig. 8共b兲 shows the probability sults are summarized in Table 5. As indicated in the table, when

982 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JULY 2010

J. Geotech. Geoenviron. Eng., 2010, 136(7): 975-984

 












 

 !
"
Fig. 10. Example 1: result of the stability analysis by the strength
# "
 reduction method based on the mean value of the undrained shear



strength 共Fs = 1.322兲

 




Downloaded from ascelibrary.org by UFPA - Universidade Federal Do Para on 10/03/19. Copyright ASCE. For personal use only; all rights reserved.









Fig. 9. Example 2: influence of the cross correlation on the probabil- conventional LEMs, the calculations are performed within the
ity of failure 共FORM assumes a single random variable兲 framework of the elastic-perfectly plastic Mohr-Coulomb model
with no hardening or softening. As the strength of the soil is a
spatially distributed random variable, soil strength should be
the autocorrelation distance increased, the relative difference be- modeled as random field considering spatial correlation. A com-
tween the overall probability of failure and the probability of bination of random field and finite difference analysis method was
failure associated with the fixed critical surface decreases. This adopted to seek out the more realistic critical path through the soil
means that the failure mechanisms by random LEM approach the 共Cho and Park 2009兲.
fixed critical surface. The results also show that the vertical auto- The minimum factor of safety calculated from a deterministic
correlation distance has greater influence on the statistical re- analysis based on the mean value of the shear strength is 1.322,
sponse. which is slightly lower than 1.356 from Bishop’s simplified
method. The plot of the shear strain rate shows that the failure
surface is similar to the critical circular surface, as presented in
Discussions Fig. 10. The arrows in Fig. 10 are velocity vectors, which indicate
the pattern of motion at the initiation of failure and show a well-
Finding the critical slip surface that gives the lowest factor of defined circular failure surface. Fig. 11 shows results of the sta-
safety is one of the key issues in any limit equilibrium stability bility analysis by the strength reduction method for the same
analysis. Stratigraphic conditions have a major influence on po- realizations of random field previously shown in Fig. 3. Fig. 11
tential slip surfaces. A simple circular failure surface method is shows that the failure surfaces determined by the strength reduc-
sufficient for a slope in a homogenous soil layer, while for a tion method are quite similar to the critical circular surfaces by
heterogeneous multisoil layer slope, a noncircular failure surface Bishop’s simplified method although some localized noncircular
method should be considered as circular methods can overpredict shape can be observed due to the existence of stratified weak
the factor of safety 共Zolfaghari et al. 2005兲. regions. The differences between the factor of safety for the
In this section, stability analysis based on the strength reduc- strength reduction method and the factor of safety for LEM are
tion method is conducted by FLAC to show that the use of circu- small to be of practical consequence. This suggests that it can be
lar surfaces is a reasonable approximation for the studied problem assumed with reasonable accuracy that the critical slip surface is
共Example 1兲. Strength reduction method can provide an alterna- circular in a single-layered slope. However, it is evident that once
tive to the LEM in providing an assessment of the stability of the anisotropic spatial correlation parameters are incorporated
slopes. The main advantage of the strength reduction method is slopes become stratified. Hence, a search for the critical noncir-
that it requires no assumption on the shape or location of the cular surface is required. Further refinement of the shape for a
failure surface. The critical failure surface is determined auto- circular slip surface can lead to a noncircular slip surface that is
matically from the shear strain arising from the reduction of shear attracted to the lower strength horizontal layers, which will result
strength. FLAC uses an explicit time marching method to solve in lower factors of safety and higher probabilities of failure.
the governing field equations in which every derivative is re- In addition, due to the variety of soil types and multilayered
placed by an algebraic expression written in terms of the field soil profiles that exist in real field cases, future studies should also
variables at discrete points in space. To be consistent with the consider multilayer soil effects on probabilistic slope stability

Table 5. Example 2: Summaries of Simulation Results

COV
lh lv
共m兲 共m兲 c ␾ Pf ␮ Fs ␴ Fs COVFs Relative difference of P f
20 2 0.3 0.2 0.0138 1.2072 0.1077 0.0892 0.1930 F
0.0171 1.1991 0.1057 0.0881 S
20 4 0.3 0.2 0.0323 1.2100 0.1292 0.1068 0.1270 F
0.0370 1.2015 0.1262 0.1051 S
40 2 0.3 0.2 0.0155 1.2086 0.1108 0.0917 0.1885 F
0.0191 1.2003 0.1086 0.0905 S
Note: F = Monte Carlo simulation is performed for the fixed critical surface and S = Monte Carlo simulation is performed based on the search algorithm for
the critical failure surface.

JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JULY 2010 / 983

J. Geotech. Geoenviron. Eng., 2010, 136(7): 975-984

 Cho, S. E., and Park, H. C. 共2009兲. “Effect of spatial variability of cross-
correlated soil properties on bearing capacity of strip footing.” Int. J.
Numer. Analyt. Meth. Geomech., 34共1兲, 1–26.
Chowdhury, R. N., and Xu, D. W. 共1995兲. “Geotechnical system reliabil-
ity of slopes.” Reliab. Eng. Syst. Saf., 47共3兲, 141–151.
  Cornell, C. A. 共1967兲. “Bounds on the reliability of structural systems.” J.
Struct. Div., 93共1兲, 171–200.
El-Ramly, H., Morgenstern, N. R., and Cruden, D. M. 共2002兲. “Probabi-
listic slope stability analysis for practice.” Can. Geotech. J., 39共3兲,
665–683.
El-Ramly, H., Morgenstern, N. R., and Cruden, D. M. 共2003兲. “Probabi-


listic stability analysis of a tailings dyke on presheared clay-shale.”

Downloaded from ascelibrary.org by UFPA - Universidade Federal Do Para on 10/03/19. Copyright ASCE. For personal use only; all rights reserved.

Can. Geotech. J., 40共1兲, 192–208.

Fig. 11. Example 1: results of the stability analysis by the strength Elkateb, T., Chalaturnyk, R., and Robertson, P. K. 共2003兲. “An overview
of soil heterogeneity: Quantification and implications on geotechnical
reduction method for the same realizations of random field shown in
field problems.” Can. Geotech. J., 40共1兲, 1–15.
Fig. 3: 共a兲 Fs = 0.881; 共b兲 Fs = 0.889; 共c兲 Fs = 0.811; and 共d兲 Fs
Fenton, G. A., and Griffiths, D. V. 共2003兲. “Bearing capacity prediction of
= 0.975 spatially random c-␾ soils.” Can. Geotech. J., 40共1兲, 54–65.
Ghanem, R. G., and Spanos, P. D. 共1991兲. Stochastic finite element—A
analysis based on multiscale random fields that varies spatially at spectral approach, Springer, New York.
a large scale as different zonations due to stratification and also at Ghiocel, D. M., and Ghanem, R. G. 共2002兲. “Stochastic finite-element
a smaller scale within each individual zone. analysis of seismic soil-structure interaction.” J. Eng. Mech., 128共1兲,
66–77.
Griffiths, D. V., and Fenton, G. A. 共2004兲. “Probabilistic slope stability
analysis by finite elements.” J. Geotech. Geoenviron. Eng., 130共5兲,
Conclusions 507–518.
Griffiths, D. V., Huang, J., and Fenton, G. A. 共2009兲. “Influence of spatial
This paper proposed a numerical procedure that considers the variability on slope reliability using 2-d random fields.” J. Geotech.
slope stability problems with uncertain quantities. The procedure Geoenviron. Eng., 135共10兲, 1367–1378.
extends the deterministic analysis based on the LEM of slices to a Hassan, A. M., and Wolff, T. F. 共1999兲. “Search algorithm for minimum
probabilistic approach that accounts for the uncertainties and spa- reliability index of earth slopes.” J. Geotech. Geoenviron. Eng.,
tial variation of the soil parameters. 125共4兲, 301–308.
From the example problems of one-layered slope, the follow- Hong, H. P., and Roh, G. 共2008兲. “Reliability evaluation of earth slopes.”
ing observations can be made: J. Geotech. Geoenviron. Eng., 134共12兲, 1700–1705.
Kim, J. Y., and Lee, S. R. 共1997兲. “An improved search strategy for the
1. A combination of random field and LEM enables us to effi-
critical slip surface using finite element stress fields.” Comput. Geo-
ciently consider the various failure mechanisms caused by tech., 21共4兲, 295–313.
the spatial variability of soil property in the probabilistic Lacasse, S., and Nadim, F. 共1996兲. “Uncertainties in characterizing soil
slope stability assessment by searching for the critical path properties.” GSP 58 uncertainty in the geologic environment, C. D.
through the soil. Shackleford, P. P. Nelson, and M. J. S. Roth, eds., ASCE, Reston, Va.,
2. The probability of failure associated with a fixed critical sur- 49–75.
face was smaller than the overall probability of failure that Li, K. S., and Cheung, R. W. M. 共2001兲. “Search algorithm for minimum
comprises all potential failure surfaces since the critical fail- reliability index of earth slopes 共Discussion兲.” J. Geotech. Geoenvi-
ure surface identified by search algorithm always gives a ron. Eng., 127共2兲, 197–198.
smaller or an equal factor of safety compared to that obtained Li, K. S., and Lumb, P. 共1987兲. “Probabilistic design of slopes.” Can.
from the fixed critical surface for each random field. Geotech. J., 24共4兲, 520–535.
3. The results also indicate that the overall probability of failure Low, B. K., Gilbert, R. B., and Wright, S. G. 共1998兲. “Slope reliability
can be significantly higher than the probability of failure as- analysis using generalized method of slices.” J. Geotech. Geoenviron.
sociated with a fixed critical slip surface in a saturated clay Eng., 124共4兲, 350–362.
Lumb, P. 共1970兲. “Safety factors and the probability distribution of soil
slope under an undrained condition due to the spatial hetero-
strength.” Can. Geotech. J., 7共3兲, 225–242.
geneity of shear strength. On the contrary, for c-␾ slopes the Sudret, B., and Der Kiureghian, A. 共2002兲. “Comparison of finite element
limit state functions of slip surfaces are highly correlated. reliability methods.” Probab. Eng. Mech., 17共4兲, 337–348.
Hence, the difference between the overall probability of fail- Vanmarcke, E. H. 共1983兲. Random fields: Analysis and synthesis, MIT
ure and that for the fixed critical surface is not significant. Press, Cambridge, Mass.
Vořechovský, M. 共2008兲. “Simulation of simply cross correlated random
fields by series expansion methods.” Struct. Safety, 30共4兲, 337–363.
References Wolff, T. H. 共1985兲. “Analysis and design of embankment dam slopes: A
probabilistic approach.” Ph.D. thesis, Purdue University, Lafayette,
Baecher, G. B., and Christian, J. T. 共2003兲. Reliability and statistics in Ind.
geotechnical engineering, Wiley, New York. Yucemen, M. S., Tang, W. H., and Ang, A. H. S. 共1973兲. A probabilistic
Bhattacharya, G., Jana, D., Ojha, S., and Chakraborty, S. 共2003兲. “Direct study of safety and design of earth slopes, Structural Research Series
search for minimum reliability index of earth slopes.” Comput. Geo- Vol. 402, University of Illinois, Urbana, Ill.
tech., 30共6兲, 455–462. Zolfaghari, A. R., Heath, A. C., and McCombie, P. F. 共2005兲. “Simple
Cho, S. E. 共2007兲. “Effects of spatial variability of soil properties on genetic algorithm search for critical non-circular failure surface in
slope stability.” Eng. Geol. (Amsterdam), 92共3–4兲, 97–109. slope stability analysis.” Comput. Geotech., 32共3兲, 139–152.

984 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JULY 2010

J. Geotech. Geoenviron. Eng., 2010, 136(7): 975-984