Vous êtes sur la page 1sur 19

Color profile: Generic CMYK printer profile

Composite Default screen

665

Probabilistic slope stability analysis for practice


H. El-Ramly, N.R. Morgenstern, and D.M. Cruden

Abstract: The impact of uncertainty on the reliability of slope design and performance assessment is often significant.
Conventional slope practice based on the factor of safety cannot explicitly address uncertainty, thus compromising the
adequacy of projections. Probabilistic techniques are rational means to quantify and incorporate uncertainty into slope
analysis and design. A spreadsheet approach for probabilistic slope stability analysis is developed. The methodology is
based on Monte Carlo simulation using the familiar and readily available software, Microsoft Excel 97 and @Risk.
The analysis accounts for the spatial variability of the input variables, the statistical uncertainty due to limited data,
and biases in the empirical factors and correlations used. The approach is simple and can be applied in practice with
little effort beyond that needed in a conventional analysis. The methodology is illustrated by a probabilistic slope
analysis of the dykes of the James Bay hydroelectric project. The results are compared with those obtained using the
first-order second-moment method, and the practical insights gained through the analysis are highlighted. The
deficiencies of a simpler probabilistic analysis are illustrated.
Key words: probabilistic analysis, slope stability, Monte Carlo simulation, spatial variability.
Rsum : L'impact de l'incertitude sur la fiabilit de la conception et de l'valuation de la performance des talus est
souvent significative. La pratique conventionnelle des talus base sur le coefficient de scurit ne peut pas traiter de faon explicite l'incertitude, compromettant ainsi la prcision des projections. Les techniques probabilistes constituent des
moyens rationnels pour quantifier et incorporer l'incertitude dans la conception et l'analyse des talus. On dveloppe une
approche de tableur pour l'analyse de la stabilit des talus. La mthodologie est base sur la simulation de Monte Carlo
au moyen de logiciels familiers et facilement accessibles, Microsoft Excel 97 et @Risk. L'analyse tient compte de la
variabilit spatiale des variables entres, de l'incertitude statistique due aux donnes limites et aux distorsions des facteurs empiriques et des corrlations utiliss. L'approche est simple et peut tre mise ne pratique avec peu d'effort en
plus de celui requis par une analyse conventionnelle. La mthodologie est illustre par une analyse probabiliste des talus des digues du projet hydrolectrique de la Baie James. Les rsultats sont compars avec ceux obtenus en utilisant
la mthode du second moment de premier ordre, et les claircissements pratiques obtenus par cette analyse sont mis en
lumire. On illustre les dficiences d'une analyse probabiliste plus simple.
Mots cls : analyse probabiliste, stabilit de talus, simulation de Monte Carlo, variabilit spatiale.
[Traduit par la Rdaction]

El-Ramly et al.

683

Introduction
Slope engineering is perhaps the geotechnical subject most
dominated by uncertainty. Geological anomalies, inherent
spatial variability of soil properties, scarcity of representative
data, changing environmental conditions, unexpected failure
mechanisms, simplifications and approximations adopted in
geotechnical models, and human mistakes in design and construction are all factors contributing to uncertainty. Conventional deterministic slope analysis does not account for
quantified uncertainty in an explicit manner and relies on conservative parameters and designs to deal with uncertain conditions. The impact of such subjective conservatism cannot be
evaluated, and past experience shows that apparently conservative designs are not always safe against failure.
The evaluation of the role of uncertainty necessarily requires the implementation of probability concepts and meth-

ods. Probabilistic analyses allow uncertainty to be quantified


and incorporated rationally into the design process. Probabilistic slope stability analysis (PSSA) was first introduced into
slope engineering in the 1970s (Alonso 1976; Tang et al.
1976; Harr 1977). Over the last three decades, the concepts
and principles of PSSA have developed and are now well established in the literature. In addition to accounting for uncertainty, PSSA is also a useful approach for estimating
hazard frequency for site-specific quantitative risk analyses,
particularly in the absence of representative empirical data.
The merits of probabilistic analyses have long been noted
(Chowdhury 1984; Whitman 1984; Wolff 1996; Christian
1996). Despite the uncertainties involved in slope problems
and notwithstanding the benefits gained from a PSSA, the
profession has been slow in adopting such techniques. The
reluctance of practicing engineers to apply probabilistic
methods is attributed to four factors. First, engineers train-

Received 22 May 2001. Accepted 7 December 2001. Published on the NRC Research Press Web site at http://cgj.nrc.ca on 16 May 2002.
H. El-Ramly.1 AMEC Earth and Environmental, 4810-93 Street, Edmonton, AB T6E 5M4, Canada.
N.R. Morgenstern and D.M. Cruden. Geotechnical and Geoenvironmental Group, Department of Civil and Environmental
Engineering, University of Alberta, Edmonton, AB T6G 2G7, Canada.
1

Corresponding author (e-mail: hassan.el-ramly@ualberta.net).

Can. Geotech. J. 39: 665683 (2002)

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:12 AM

DOI: 10.1139/T02-034

2002 NRC Canada

Color profile: Generic CMYK printer profile


Composite Default screen

666

Can. Geotech. J. Vol. 39, 2002

ing in statistics and probability theory is often limited to basic information during their early years of education. Hence,
they are less comfortable dealing with probabilities than
they are with deterministic factors of safety. Second, there is
a common misconception that probabilistic analyses require
significantly more data, time, and effort than deterministic
analyses. Third, few published studies illustrate the implementation and benefits of probabilistic analyses. Lastly, acceptable probabilities of unsatisfactory performance (or
failure probability) are ill-defined, and the link between a
probabilistic assessment and a conventional deterministic assessment is absent. This creates difficulties in comprehending the results of a probabilistic analysis. All of these issues
are addressed in detail in El-Ramly (2001).
This paper addresses the integration of probabilistic methods into geotechnical practice, focusing on the first three obstacles of the factors listed in the previous paragraph. First,
an overview of some of the concepts involved in quantifying
uncertainty of soil properties is presented. The paper then
describes the development of a practical methodology for
probabilistic slope analysis based on Monte Carlo simulation. The methodology makes use of familiar and readily
available software, namely Microsoft Excel 97 and @Risk.
The implementation of the methodology is illustrated
through a probabilistic analysis of the stability of the James
Bay dykes. Lastly, the results are compared with those of the
first-order second-moment (FOSM) method and the practical
insights gained through the analysis are highlighted.

Statistical analysis of soil data


Components of statistical analysis
Statistical analysis of soil data comprises two main stages:
data review and statistical inference. Data review is largely
judgmental and encompasses a broad range of issues. First,
the consistency of the data, known as the condition of
stationarity, should be ensured. Inconsistency can arise from
pooling data belonging to different soil types, stress conditions, testing methods, stress history, or patterns of sample
disturbance (Lacasse and Nadim 1996). The histogram is a
convenient tool in this regard. A multimodal histogram is an
indication of inconsistent data, or a nonstationary condition.
Data review also aims at identifying trends in the data, biases in measurements, errors and discrepancies in the results, and outlier data values. Decisions should be made
whether to reject outliers or to accept them as extreme values. Baecher (1987) warned that care should be exercised in
this process to avoid rejecting true, important information.
Another concern is any clustering of data in a zone within
the spatial domain of interest which may render the data unrepresentative.
Statistical inferences are commonly based on a simplified
model (Vanmarcke 1977a; DeGroot and Baecher 1993) which
divides the quantity, xi, at any location, i, into a deterministic
trend component, ti, and a residual component, i , where
[1]

xi = ti + i

The trend is evaluated deterministically using regression


techniques and is usually a function of location. The residual
component is a random variable whose probability density
function (PDF) has a zero mean and a constant variance,

2 [], independent of location and estimated from the scatter


of observations around the trend. The residual component at
location i is spatially correlated with the residual components at surrounding locations.
Uncertainty in trend
Site investigation programs are usually controlled by budget, which limits the number of tests performed. Since the
trend function is estimated from the available sample of observations, it is uncertain. The uncertainty due to the sample
size (number of observations) is known as statistical uncertainty and is typically quantified by considering the regression coefficients of the trend function as variables with
means and non-zero variances, which are inversely proportional to the number of observations, n.
The method of least squares is the most common technique for estimating the means and variances of trendfunction coefficients. Neter et al. (1990) discussed estimating the uncertainties of regression coefficients for linear
trends. Li (1991) pointed out some limitations of the method
of least squares. Baecher (1987) recommended that trend
equations be linear, as the higher the order of the trend function, the more uncertain the regression coefficients. Where
the soil properties do not exhibit a clear trend, a constant
equal to the mean of the observations is assumed.
Bias is another source of uncertainty that could affect the
estimated trend. The measured soil property can be consistently overestimated or underestimated at all locations. Factors such as the testing device used, boundary conditions,
soil disturbance, or the models and correlations used to interpret the measurements may contribute to biases. As an example, Bjerrum (1972) noted that the field vane consistently
overestimated the undrained shear strength of highly plastic
clays. An empirical correction factor is used to adjust the biased measurements. The additional uncertainty introduced
by the use of Bjerrums vane correction factor is, however,
significant. This could also be the case with other empirical
factors.
Spatial variability
Soil composition and properties vary from one location to
another, even within homogeneous layers. The variability is
attributed to factors such as variations in mineralogical composition, conditions during deposition, stress history, and
physical and mechanical decomposition processes (Lacasse
and Nadim 1996). The spatial variability of soil properties is
a major source of uncertainty.
Spatial variability is not a random process, rather it is
controlled by location in space. Statistical parameters such
as the mean and variance are one-point statistical parameters
and cannot capture the features of the spatial structure of the
soil (Fig. 1). The plot in Fig. 1 compares the spatial variability of two artificial sets of data generated using the geostatistics software GSLIB (Deutsch and Journel 1998). Both
sets have similar histograms. The upper plot is highly erratic
and the data are almost uncorrelated, whereas the lower plot
is characterized by high spatial continuity. Additional statistical tools, such as the autocovariance and semivariogram,
are essential to describe spatial variability. The pattern of
soil variability is characterized by the autocorrelation distance (or scale of fluctuation; Vanmarcke 1977a). A large
2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:13 AM

Color profile: Generic CMYK printer profile


Composite Default screen

El-Ramly et al.

667

Fig. 1. A highly erratic spatial structure (upper right) and a highly continuous structure (lower right), both with similar histograms.

autocorrelation distance reflects the spatial continuity shown


in the lower plot of Fig. 1, whereas a short distance reflects
erratic variability (Fig. 1, upper plot). DeGroot (1996) and
Lacasse and Nadim (1996) illustrated the estimation of autocorrelation distance.
Spatial averaging
The performance of a structure is often controlled by the
average soil properties within a zone of influence, rather than
soil properties at discrete locations. Slope failure is more
likely to occur when the average shear strength along the failure surface is insufficient rather than due to the presence of
some local weak pockets. The uncertainty of the average
shear strength along the slip surface, not the point strength, is
therefore a more accurate measure of uncertainty. Baecher
(1987) warned, however, that depending on performance
mode, average properties are not always the controlling factor.
Internal erosion in dams and progressive failure are examples
of cases where extreme values control performance.
The variance of the strength spatially averaged over some
surface is less than the variance of point measurements. As

the extent of the domain over which the soil property is


being averaged increases, the variance decreases. For quantities averaging linearly, the amount of variance reduction depends on the autocorrelation distance. The mean, however,
remains almost unchanged. The numerical data in Fig. 2 (upper right plot) are discretized based on grids of sizes 1 1,
5 5, and 10 10. For each scheme, the averages within
grid squares are calculated and the histogram of the local averages is plotted as shown in Fig. 2. As the averaging area is
increased, the coefficient of variation of the local averages
drops from 1.55 to 0.28 and the minimum and maximum
values become closer to the mean. If the averaging area is
constant, the amount of variance reduction increases as the
autocorrelation distance decreases, i.e., soil variability becomes more erratic.
Theory of random fields
The theory of random fields (Vanmarcke 1977a, 1977b,
1983) is a common approach for modeling spatial variability
of soil properties. It is also the basis of our probabilistic
slope analysis methodology; to illustrate its rationale, some
2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:13 AM

Color profile: Generic CMYK printer profile


Composite Default screen

668

Can. Geotech. J. Vol. 39, 2002

Fig. 2. Variance reduction due to spatial averaging over blocks of sizes 1 1, 5 5, and 10 10.

of the basic concepts of the theory of random fields are presented here.
At any location within a soil layer, a soil parameter is an
uncertain quantity, or a random variable, unless it is measured at that particular location. Each random variable is
characterized by a probability distribution and is correlated
with the random variables at adjacent locations. The set of
random variables at all locations within the layer is referred
to as a random field and is characterized by the joint probability distribution of all random variables. A random field is
called stationary (or homogeneous) if the joint probability
distribution that governs the field is invariant when translated over the parameter space. This implies that the cumulative probability distribution function, the mean, and the
variance are constant for any location within the random
field and that the covariance of two random variables de-

pends on their relative locations, regardless of their absolute


locations within the random field.
The point to point variation of a random field is very difficult to obtain in practice and is often of no real interest
(Vanmarcke 1983). Local averages over a spatial or temporal
local domain are of much greater value. For example, the
hourly or daily rainfall is of interest to hydrologists rather
than the instantaneous rate of fall. Similarly, the average
shear strength of a soil over the area of the slip surface is of
more interest to geotechnical engineers than the variation of
strength from point to point within the layer.
Figure 3 shows a one-dimensional (1D) stationary random
field of a variable, x, with a mean, E[x], a variance, 2 , and
a cumulative probability distribution function, F(x); the
graph could also portray the variation along a line in the parameter space of a homogeneous k-dimensional random
2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:13 AM

Color profile: Generic CMYK printer profile


Composite Default screen

El-Ramly et al.

669

Fig. 3. A realization of a 1D random field of a variable x with a mean E[x], variance 2, and cumulative probability distribution
function F(x), showing local averages over intervals z and z.

field. For this type of random field, Vanmarcke (1977a) defined the dimensionless variance function (z) as
[2]

( z) =

2
2 z

where 2 z is the variance of the local average, X(z), of the


variable, x, over an interval of length z. The variance function is a measure of the reduction in the point variance, 2 ,
under local averaging. For most commonly used correlation
functions, Vanmarcke (1983) showed that (z) can be approximated by
[3]

( z) =
z

z
z

where is the scale of fluctuation. In concept, the scale of


fluctuation has the same meaning as the autocorrelation distance but differs in numeric value. For the common exponential and Gaussian autocorrelation models, is 2.0 and
1.77 times the autocorrelation distance, ro, respectively.
The model indicates no variance reduction, (z) = 1, due
to local averaging up to an averaging interval z equal to .
The rationale of this approximation is that modeling a random phenomenon at a level of aggregation more detailed
than the way the information about the phenomenon is acquired or processed is impractical and unnecessary
(Vanmarcke 1983). In site investigation programs, the spacing or the time interval between observations is often large.
Characterizing the correlation structure at small intervals becomes unreliable and justifies the approximation by a perfect
autocorrelation.
Local averages of the variable x over intervals z and z
(Fig. 3), X(z) and X(z), are spatially correlated. The correlation coefficient, (Xz, X z ), between X(z) and X(z)
is given by eq. [4] (Vanmarcke 1983). It is a function of the
lengths of the two intervals, the separation, Zo, between
them (Fig. 3), and the variance function of the variable being
averaged:
[4]

(X z, X z) =
Z 2o (Z o) Z 12 (Z1) + Z 22 (Z 2 ) Z 23 (Z 3)
2 z z [( z)( z )]0.5

where Z1 is the distance from the beginning of the first interval to the beginning of the second interval, Z2 is the distance
from the beginning of the first interval to the end of the second interval, and Z3 is the distance from the end of the first
interval to the end of the second interval.
Random testing error
Random testing errors arise from factors related to the
measuring process (other than bias), such as operator error
or a faulty device. They can be directly measured only
through repeated testing of the same specimen by different
operators and devices. This is not practical, as most geotechnical field and laboratory tests disturb the sample, so
random measurement errors cannot be known. In data analysis, measurement error is regarded as a random variable with
a zero mean and a constant variance, referred to as random
error variance. The uncertainty due to testing errors is not a
true variation in the soil property and should be discarded in
evaluating parameter variability. Approximate analytical procedures (Baecher 1987) are used to estimate the random error variance. Jaksa et al. (1997), however, pointed to
important inaccuracies associated with such procedures.
Reliable estimation of random error variance requires data
at very small separation distances (r 0), not available in
practice. As a result, the distinction between noise, or random error, and real, short-scale variability (over distances
less than data spacing) is not possible. Hence, the analytically computed variance is composed of random error variance and short-scale variability. Jaksa et al. (1997) also
showed that the computed value of the random error variance is sensitive to data spacing. Kulhawy and Trautmann
(1996) commented that quantitative assessment of testing errors of in situ measurements is rare because of the difficulties in separating natural spatial variability from test
uncertainty. So, the reliability of the analytical estimation of
random error variance is in question. Random testing errors
are best addressed through standardization of testing equipment and procedures, proper personnel training, and continuous inspection.

Probabilistic slope analysis methods


Probabilistic procedures for slope stability analysis vary
in assumptions, limitations, capability to handle complex
2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:14 AM

Color profile: Generic CMYK printer profile


Composite Default screen

670

problems, and mathematical complexity. Most of them,


however, fall into one of two categories: approximate methods (the FOSM method, the point estimate method), and
Monte Carlo simulation.
Approximate methods make simplifying assumptions that
often limit their application to specific classes of problems.
For example, some used very simple slope models such as
the ordinary method of slices (Tang et al. 1976; Vanmarcke
1980; Honjo and Kuroda 1991; Bergado et al. 1994). Others
dealt only with frictionless soils (Matsuo and Kuroda 1974;
Vanmarcke 1977b; Bergado et al. 1994). A restriction to a
circular (or cylindrical) slip surface is common in many
studies (Alonso 1976; Vanmarcke 1977b; Yucemen and AlHomoud 1990; Bergado et al. 1994). The spatial variability
of soil properties and pore-water pressure is often ignored,
assuming perfect autocorrelation (Nguyen and Chowdhury
1984; Wolff and Harr 1987; Duncan 2000).
Approximate methods allow the estimation of the mean
and variance of the factor of safety but do not provide any
information about the shape of the probability density function, so the failure probability can only be obtained by assuming a parametric probability distribution of the factor of
safety (typically normal or log-normal). Estimates of low
probabilities, required for safe structures, are sensitive to the
assumed distribution.
Few studies used Monte Carlo simulation in slope analysis
when compared to approximate methods. The extensive computational efforts involved in the simulations required researchers to develop their own software to solve slope stability
problems. Computer times needed for the simulations were
also significant. As a result, a Monte Carlo simulation was considered, until recently, to be uneconomic (Tobutt 1982; Priest
and Brown 1983). Studies applying Monte Carlo simulation
also rarely addressed the spatial variability of soil properties
(Major et al. 1978; Tobutt 1982; Nguyen and Chowdhury
1985) because of difficulties in generating random values in
ways that preserved their correlations. Today, rapid developments in software are significantly changing this situation.
The limitations, and sometimes the complexities, of
probabilistic methods combined with the poor training of
most engineers in statistics and probability theory have substantially inhibited the adoption of probabilistic slope stability analysis in practice. To facilitate the integration of
probabilistic methods into practice, the methodologies and
techniques, while being consistent with principles of logic
and mechanics, should be simple, capable of solving real
slope problems, and in formats familiar to engineers.

An integrated probabilistic slope analysis


methodology
Outline of the methodology
The probabilistic slope analysis methodology based on
Monte Carlo simulation developed here has the advantages
of being simple and not requiring a comprehensive statistical
and mathematical background. The methodology is spreadsheet based and makes use of the familiar and readily available Microsoft Excel 97 (Microsoft Corporation 1997) and
@Risk (Palisade Corporation 1996) software. Other similar
software, e.g., Lotus 1-2-3 and Crystal Ball (Decisioneering
Inc. 1996), is likely capable of undertaking the task.

Can. Geotech. J. Vol. 39, 2002

The slope problem (geometry, stratigraphy, soil properties,


and slip surface) and the selected method of analysis are first
modeled in a Microsoft Excel 97 spreadsheet. Available
data are examined and uncertainties in input parameters are
identified and described statistically by representative probability distributions. Only those parameters whose uncertainties are deemed significant to the analysis need to be treated
as variables. A Monte Carlo simulation is then performed as
illustrated schematically in Fig. 4. @Risk draws at random a
value for each input variable from within the defined probability distributions. These values (input set 1, Fig. 4) are
used to solve the spreadsheet and calculate the corresponding factor of safety. The process is repeated sufficient times,
m, to estimate the statistical distribution of the factor of
safety. Statistical analysis of this distribution allows estimating the mean and variance of the factor of safety and the
probability of the factor of safety being less than one. The
procedure is applicable to any method of limit equilibrium
analysis that can be represented in a spreadsheet.
Statistical characterization of input variables
Probability distributions
Different approaches can be adopted to estimate the probability distribution of each input variable. Where there are
adequate amounts of data, the cumulative distribution function (CDF) of the measurements can be used directly in the
simulation process. In geotechnical practice, the observed
CDF (or histogram) may show spikes that would not appear
were more observations available. The CDF is better obtained by resetting the probability associated with each observation to the average of its cumulative probability and
that associated with the next lower observation (Deutsch and
Journel 1998). This procedure smooths unwanted spikes and
allows assigning finite probabilities for values of the input
parameter less than the minimum value measured and more
than the maximum value.
Where observations are scarce or absent, parametric distributions can be assumed from the literature. Studies have
estimated coefficients of variation and probability density
functions of soil properties (Lumb 1966; Chowdhury 1984;
Harr 1987; Kulhawy et al. 1991; Lacasse and Nadim 1996).
Care should be exercised, however, to ensure that the minimum and maximum values of the selected distribution are
consistent with the physical limits of the parameter being
modeled. For example, shear strength parameters should not
take negative values. If the selected distribution implies negative values, then the distribution is truncated at a practical
minimum threshold. Alternatively, variability can be assumed
to follow a triangular distribution with estimates of minimum,
maximum, and most likely values based on expert opinion.
@Risk built-in functions allow great flexibility in modeling input variables. Parametric and nonparametric distributions (e.g., experimental CDF) can be modeled using @Risk
functions. Desired truncations can be easily imposed on any
distribution either through @Risk or using Microsoft Excel 97 functions.
Modeling parameter uncertainty
The probabilistic slope analysis methodology accounts for
the uncertainty of the input parameters based on the concepts
discussed in previous sections and the following equation:
2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:14 AM

Color profile: Generic CMYK printer profile


Composite Default screen

El-Ramly et al.

671

Fig. 4. Monte Carlo simulation procedure using Microsoft Excel 97 and @Risk software.

[5]

V = B(t + )

where V is the input variable adjusted for statistical errors


and bias, B is a bias correction factor, t is the trend component, and is the residual component. The statistical uncertainty in the trend is estimated by standard statistical
methods (Ang and Tang 1975; Neter et al. 1990). The variance of the residual component is equal to the variance of
the measured data, with no explicit consideration of the uncertainty due to random testing errors for two reasons. Estimates of random error variance are unreliable, as discussed
in an earlier section, so any reduction of the observed variance could be unsafe. In addition, random errors fluctuate in
magnitude above and below zero and are uncorrelated spatially. By taking the spatial average of the input variable over
the whole area of interest, such as the slip surface, positive
and negative random errors at the different locations within
the averaging domain tend to cancel out. As a result, the random error variance associated with the averaged quantity is
largely reduced. It is emphasized, however, that a critical
review of the observations should ensure that appropriate
testing standards and procedures are followed. The bias correction factor is also considered a variable with a mean and
variance evaluated by experience and comparison with field
performance or by observations from a more accurate testing
procedure. It does not have to be a multiplier as in eq. [5]; it
could be an addition [V = B + (t + )].
The spatial variability of an input variable is represented
by the correlation structure of the residual component, , and
modeled using the theory of random fields. The spatial variability of along the slip surface is approximated by a 1D
stationary random field. Vanmarckes (1983) approximate
model (eq. [3]) is adopted for the analysis of the random
field. The point to point variability of the residual component along the slip surface is resembled by the variability of
its local averages over segments of the surface. The portion
of the slip surface within the subject layer is divided into
segments of length l not exceeding the scale of fluctuation .
The local average of the residual component over the length
of any of these segments, (l), is considered a variable. The
CDF of the variable (l) is the same distribution function of
the residual component, F(), with no variance reduction.

This is because the variance function (l), eq. [3], equals


one for l . The correlation coefficients between the variables representing the local averages over different segments
of the slip surface are estimated using eq. [4]. Choosing the
length of segments equal to eliminates the correlation coefficients between most of the variables and greatly simplifies
the simulation. Figure 5 is a schematic illustration of the
model, where is the correlation coefficient between variables.
Critical slip surface
For any slope, there is an unlimited number of potential
slip surfaces. The slope may fail, or perform unsatisfactorily,
along any of these surfaces. The total probability of unsatisfactory performance of a slope is, thus, the joint probability
of failure occurring along the admissible slip surfaces. These
surfaces, however, are highly correlated, since they are all
analyzed using the same input variables and the same analytical model. Furthermore, the close proximity of many of
them add an element of spatial correlation. Evaluating the
total probability of unsatisfactory performance of the slope
is a mathematically formidable task.
The strong correlation between slip surfaces, however,
significantly reduces the difference between the total probability of unsatisfactory performance and that of the most
critical surface (Mostyn and Li 1993). Hence, the probability
estimate from the most critical slip surface is considered a
reasonable estimate of slope reliability (Vanmarcke 1977b;
Alonso 1976; Yucemen and Al-Homoud 1990). In this study,
the probability of unsatisfactory performance is obtained by
independently analyzing a number of predetermined slip surfaces. The highest estimated probability value is considered
representative of the total probability of unsatisfactory performance of the slope. The question, however, is which surfaces to consider?
Chowdhury and Tang (1987) and Hassan and Wolff (1999)
indicated that the deterministic critical slip surface is not always the most critical surface in a probabilistic analysis.
Hassan and Wolff (1999) proposed a search algorithm for locating the slip circle with the minimum reliability index. It
searches for the slip surface with the longest length within
2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:15 AM

Color profile: Generic CMYK printer profile


Composite Default screen

672
Fig. 5. Modeling the spatial variability of an input parameter
along the failure surface using Vanmarckes (1983) approximation of a 1D random field.

the soil layer whose uncertainty contributes the most to the


uncertainty of the factor of safety. Their algorithm, however,
does not directly account for the spatial variability of soil
properties and pore-water pressure. In locating the probabilistic critical slip surface, variance reduction due to spatial
averaging of soil parameters over the length of the slip surface has to be considered. The reduction depends on the
autocorrelation distance and the length of slip surface, which
is not known beforehand. As a result, an increase in the variance of the factor of safety, due to a longer portion within a
highly uncertain layer, could be offset by a variance reduction due to spatial averaging. Hence, in slopes dominated by
uncertainty due to spatial variability, the slip surface based
on the Hassan and Wolff algorithm may not be the most critical surface in a probabilistic analysis.
An essential part of the probabilistic analysis is, thus, to
consider all slip surfaces that may be hazardous. These may
include the deterministic critical slip surface, the minimum
reliability index surface according to Hassan and Wolff
(1999), noncircular, structurally controlled surfaces, and surfaces through weak or presheared layers.
Spreadsheet modeling
The spreadsheet software Microsoft Excel 97 is used in
modeling the slope geometry, stratigraphy, and slope analysis
method. Using analytical geometry, the equations describing
the ground profile, the boundaries between soil layers, and
the slip surface are first established with reference to a coordinate system. The equations are then modeled in a
Microsoft Excel 97 spreadsheet. The coordinates of the
points of intersection of the different boundaries, for example, where the slip surface meets the borders between layers,
can be obtained within the spreadsheet and used in calculating the slice information (width, coordinates of mid-base
point, total height, and thickness in each soil type) for factor
of safety computations.
Next, the slip surface within each soil layer is divided into
segments of length in addition to a residual segment of
smaller length, similar to the illustration in Fig. 5. The local
averages of a soil parameter over the length of each segment
are considered variables. The spatial variability of the soil
parameter along the slip surface is modeled by the correlations between these variables. The variables are characterized by the CDF of the residual component, F(), with no

Can. Geotech. J. Vol. 39, 2002

variance reduction, since the variance function (z ) is


equal to one (eq. [3]). @Risk functions are used to assign
appropriate probability distributions for the spreadsheet cells
containing the input variables. The correlations between
variables are modeled using @Risk functions IndepC and
DepC or by creating a correlation coefficients matrix using the Correlate command. In the simulation process,
correlated variables are sampled so as to preserve input correlation coefficients.
Slope analysis methods, such as the Bishop, Janbu, and
Spencer methods, can be easily modeled in a spreadsheet.
However, more advanced and complex methods such as the
MorgensternPrice method and three-dimensional (3D)
methods, are, for now, cumbersome to model in a spreadsheet. In this paper, the Bishop method of slices is used. The
spreadsheet model is built by creating a calculation table
similar to that proposed by Bishop (1955) for mechanical
computations. The appropriate soil parameters for each slice
are automatically chosen by a series of nested IF statements that compares the X coordinate of the mid-base point
with the coordinates of the intersections of the slip surface
with soil layers. The Microsoft Excel 97 circular reference feature is used for the iteration process until the difference between the assumed and the calculated factors of safety
is within the user-specified range. Further discussions and an
illustration of the structure of the spreadsheet are presented in
the context of the case study analyzed in a later section.
Issues in simulation
Monte Carlo simulation requires the generation of random
numbers, which are used in sampling the CDFs of the input
variables. @Risk software (Palisade Corporation 1996) allows two sampling techniques, Monte Carlo sampling (or
random sampling) and Latin Hypercube sampling. The latter
allows sampling the entire CDF curve, including the tails of
the distribution, with fewer iterations, and consequently less
computer time. It is adopted in this study.
The output of a Monte Carlo simulation is sensitive to the
number of iterations, m. When m is large, the random samples
drawn for each input variable are also large and the match between the CDF recreated by sampling and the original input
CDF is more accurate. Hence, the level of noise in the simulation diminishes and the output becomes more stable at the
expense of increasing computer time. The optimum number
of iterations depends on the sizes of the uncertainties in the
input parameters, the correlations between input variables,
and the output parameter being estimated. A practical way to
optimize the simulation process is to repeat the simulation using the same seed value and an increasing number of iterations. A plot of the number of iterations, m, against the
probability of unsatisfactory performance indicates the minimum number of iterations at which the probability value stabilizes.
Interpretation of the output
Probability of unsatisfactory performance
The output of the simulation is the probability density function of the factor of safety from which the mean and standard
deviation can be estimated. The probabilistic safety measure
used in this study is the failure probability. It is the probabil 2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:15 AM

Color profile: Generic CMYK printer profile


Composite Default screen

El-Ramly et al.

ity of the factor of safety, FS, being less than one, or simply
the number of iterations with FS 1.0 relative to the total
number of iterations, m. The term failure, however, implies
that the collapse of the slope is the event of concern to the designer, which is not necessarily the case. The serviceability of
the slope is as important as slope collapse and requires thorough evaluation and assessment. Serviceability issues include,
among others, slope movement and cracking (without the
slope collapsing), high water seepage, and surface erosion.
The term failure may also imply to clients, particularly nonprofessional clients, that the slope would fail. We recommend
using a different terminology. The U.S. Army Corps of Engineers (1992, 1995) used probability of unsatisfactory performance, Pu, instead of failure probability. The same term,
unsatisfactory performance, is adopted here to address failure
mechanisms. Another safety measure, perhaps probability of
unsatisfactory serviceability, should be considered to address
serviceability criteria. The evaluation of slope serviceability
is, however, beyond the scope of this study.
Because the simulation process uses random sampling of
the input variables, the calculated probability of unsatisfactory performance is also a variable. The probability value
from a single simulation could differ from the true value.
Law and McComas (1986) described relying on the results
of a single simulation run as one of the most common and
potentially dangerous simulation practices. It is, therefore,
essential to repeat the simulation using different seed values
to assess the consistency of the estimates. By running the
simulation many times, the histogram of the probability of
unsatisfactory performance, the mean probability, and the
95% confidence interval around the mean could also be estimated. Using @Risk macro functions, the process of running a number of simulations can be fully automated. A
simple macro file can be designed to run a number of simulations and save the output of each simulation in a separate file.
In reflecting on the computed probability of unsatisfactory
performance, three points should be noted. First, the computed probability of unsatisfactory performance does not
address temporal changes in input variables. If pore-water
pressures, as an example, vary with time, the impact of such
variations on stability can only be assessed by undertaking
further analyses using the new pore-pressure distributions.
Second, the computed probability of unsatisfactory performance represents the probability of occurrence of a hazardous
event over the lifetime of the conditions considered in the
analysis, such as loading conditions and environmental conditions. The calculated probability is, thus, associated with a
time frame that varies from one problem to another. If the annual probability of unsatisfactory performance is required, as
is the case in quantitative risk analyses, the reference time
needs to be estimated. For example, if the pore-water pressure
used in the analysis is the result of a rainstorm, the annual
probability of unsatisfactory performance is estimated based
on the return period of the storm. In the case that none of the
input variables is time dependent, the annual probability could
be referenced to the lifetime of the slope.
Third, despite efforts to include all sources of uncertainty,
there is the possibility of undetected uncertainties such as
human mistakes affecting slope performance. The contribution of these unknown uncertainties to the probability of unsatisfactory performance is not considered, so the computed

673

probability could be a lower bound to the actual probability


of unsatisfactory performance. That is why comparison with
computed probabilities of different designs is believed to be
of greater value.
Having estimated the probability of unsatisfactory performance, the next step is to assess whether it is acceptable.
Typically, this is achieved through comparing the computed
values with a probabilistic slope design criterion. Commonly,
a criterion based on the observed frequency of slope failures
and judgement is used. A major drawback to this approach
is that the site- and case-specific features such as slope geometry, site conditions, and sources and levels of uncertainty
of the case histories constituting the database are not addressed. Applying such a global criterion to any slope is a
significant generalization.
The authors are not aware of any well-founded probabilistic criteria for the acceptability of a design. A reliable approach to estimate such criteria is to calibrate the computed
probabilities of unsatisfactory performance of slopes with
their observed field performance. This requires probabilistic
slope stability analyses of cases of failed slopes and of
slopes performing satisfactorily. Through the comparison of
the probabilities associated with the two classes, a probabilistic slope design criterion could be established. This approach has been developed by El-Ramly (2001) and will be
published separately.
Reliability index
The reliability index, , another common probabilistic
safety measure, is given by
[6]

E[FS] 1
[FS]

where E[FS] and [FS] are the mean and standard deviation
of the factor of safety, respectively. The above definition is
accurate if the performance function (factor of safety equation) is linear, which is not the case for slope analysis models. Mostyn and Li (1993) suggested, however, that the
performance functions of slopes are reasonably linear and recommended ignoring the nonlinearity when calculating . Ang
and Tang (1984) provided a convenient theoretical background
of the meaning and estimation of the reliability index.
Sensitivity analysis
Through the software @Risk, Spearman rank correlation
coefficients between the factor of safety and the input variables can be calculated for a sensitivity analysis. The
Spearman coefficient is a correlation coefficient based on the
rank of the data values within the minimummaximum
range, not the actual values themselves. It varies between 1
and 1. A value of zero indicates no correlation between the
input variable and the output, a value of 1 indicates a complete positive correlation, and a value of 1 indicates a complete negative correlation. Spearman correlation coefficients
could be used as measures of the relative contribution of
each input variable to the uncertainty in the factor of safety.
This contribution comprises two elements, the degree of uncertainty of the input parameter and the sensitivity of the
factor of safety to changes in that parameter.
The results of sensitivity analyses are of significant practical value, as they quantify the contributions of the various
2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:15 AM

Color profile: Generic CMYK printer profile


Composite Default screen

674

sources of uncertainty to the overall design uncertainty. Resources, whether intellectual or physical, can thus be rationally allocated towards reducing the uncertainty of the
variables with large impacts on design. Also, the relative impacts of systematic uncertainty (statistical uncertainty and
bias) and uncertainty due to inherent spatial variability can
be estimated. This information is of interest because systematic uncertainty has a consistent effect at all locations within
the domain of the problem and can have a major impact on
design. Unlike the uncertainty due to inherent spatial variability, systematic uncertainty can be reduced by increasing
the size of data sample and by avoiding highly uncertain empirical correlations and factors.

Probabilistic analyses of James Bay dykes


Spreadsheet-based probabilistic slope analysis
The application of the probabilistic slope analysis methodology described earlier in the paper is illustrated through the
analysis of the well-documented dykes of the James Bay hydroelectric project in Quebec, Canada. Although the dykes
were not built, the design was the subject of extensive studies
quantifying the sources of uncertainty, analyzing the spatial
variability of soil properties (Ladd et al. 1983; Souli et al.
1990), and evaluating the stability of the dyke probabilistically
using the FOSM method (Christian et al. 1994).
The project called for the construction of 50 km of earth
dykes in the James Bay area. Various design options were
investigated. Only one design is considered in this study, the
single stage construction of a 12 m high embankment with a
slope angle of 18.4 (3h:1v) and a 56 m wide berm at midheight. Figure 6 shows the geometry of the embankment and
the underlying stratigraphy.
The embankment is on a clay crust, about 4.0 m thick,
overlying a sensitive marine clay which in turn is underlain
by a lacustrine clay. The marine clay is about 8.0 m thick
and the lacustrine clay is about 6.5 m thick. The undrained
shear strength of both clays, measured by field vane tests at
1.0 m depth intervals, exhibited a large scatter. The mean
undrained shear strength of the marine clay is about 35 kPa
and that of the lacustrine clay is about 31 kPa. The lacustrine clay is underlain by a stiff till.
The uncertainty in soil parameters was quantified by Ladd
et al. (1983) and Christian et al. (1994). They identified the
input parameters whose uncertainties are deemed important
to the stability of the dykes and estimated their statistical parameters. The following probabilistic stability analyses are
based on the conclusions of these studies. Other case studies
demonstrating the complete analysis starting with field and
laboratory data, through the quantification of parameter uncertainty, the probabilistic assessment, and the estimation of
the probability of unsatisfactory performance will be published separately by the present authors.
Eight input parameters are considered variables: the unit
weight and friction angle of embankment material, the thickness of the clay crust, the undrained shear strength and
Bjerrum vane correction factors for the marine and lacustrine clays, and the depth of till layer. Table 1 summarizes
the means and variances of all input variables.

Can. Geotech. J. Vol. 39, 2002

The variances of the unit weight and friction angle of the


embankment material are evaluated judgementally (Christian
et al. 1994) to reflect potential variability in fill properties.
The bias in vane measurements is adjusted by Bjerrums
vane correction factor, which is also considered uncertain
due to the scatter in Bjerrums database. The uncertainty in
the depth of the till layer, Dtill, is entirely statistical uncertainty due to the limited number of borings. In the absence
of the actual probability distributions of the data, all variables are assumed normal. The spatial variability of all uncertain soil parameters is characterized by an isotropic
autocorrelation distance of about 15 m, which was inferred
from Christian et al. (1994).
As the strength of the lacustrine clay is relatively low, the
depth to the till layer controls the location of the critical slip
circle. The uncertainty in Dtill, thus, introduces uncertainty
in the location of the slip circle. To examine the impact of
this uncertainty, deterministic stability analyses were performed varying Dtill incrementally between 15.5 and 21.5 m
(3 standard deviations). The minimum factor of safety varied between 1.69 and 1.30, implying that the uncertainty in
Dtill could have an important impact on the reliability of the
design. All the critical slip circles were tangent to the top of
the till, daylighted within a short distance at the top of the
embankment, and shared almost the same x coordinate for
the centres.
The equations describing the dyke profile and soil layers
with reference to a coordinate system are estimated and
modeled in a Microsoft Excel 97 spreadsheet. The general
layout and structure of the spreadsheet are illustrated in Appendix A. Since the depth of the till is considered a variable,
a different value is sampled for each simulation iteration and
consequently the critical slip circle varies from one iteration
to another. To minimize the computer time, some restrictions
were imposed on the geometry of the slip circles based on
the results of the previous parametric study. The slip circles
are assumed tangent to the till layer, daylight at a fixed point
at the top of the embankment (X1 = 4.9, Y1 = 36.0), and have
a common X coordinate for the centres (X0 = 85.9). The
equation of the slip circle is then added to the spreadsheet as
a function of the Y coordinate of the centre, Y0, and the radius, R; both depend on the sampled value of Dtill. In each
iteration the sampled value of Dtill thus defines only one critical slip circle. Using the principles of analytical geometry,
the intersections of the slip circle and the boundaries between layers and the breakage points in ground profile are
computed.
The spatial variabilities of soil parameters are modeled as
1D random fields, assuming exponential autocovariance
functions. The spatial variability of the unit weight of the
embankment is, in fact, a two-dimensional (2D) random
field. Vanmarcke (1983) showed that the process of averaging a 2D random field over a rectangular area with one side
of the area smaller than the scale of fluctuation in the same
direction could be approximated by averaging a 1D random
field in the perpendicular direction. If the cross section of
the embankment is regarded as a rectangle, the variability of
unit weight could be approximated by a 1D random field in
the horizontal direction. This approximation is not likely to
have any effect on the analysis, as the impact of the spatial
2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:16 AM

Color profile: Generic CMYK printer profile


Composite Default screen

El-Ramly et al.

675

Fig. 6. Cross section and stratigraphy of the James Bay dykes, showing the approach adopted to account for spatial variability of soil
properties.

Table 1. Input variables and their statistical parameters (based on Christian et al. 1994)
Soil parameters
Variance

Bias factors

Input variable

Mean

Inherent variability

Statistical uncertainty

Mean

Variance

Fill friction angle, fill ()


Fill unit weight, fill (kN/m3)
Clay crust thickness, tcr (m)
Shear strength of marine clay, SuM (kPa)
Bjerrum vane correction factor for SuM, M
Shear strength of lacustrine clay, SuL (kPa)
Bjerrum vane correction factor for SuL, L
Depth of till, Dtill (m)

30.0
20.0
4.0
34.5

31.2

18.5

1.00
1.00
0.19a
66.26

74.82

3.00
1.00
0.04
0.90

3.00

1.00

1.0

1.0

0.006

0.023

Variance is reduced by 80% to account for spatial averaging based on the assessment of Christian et al. (1994).

variability of unit weight on design reliability is minimal


(Alonso 1976; Nguyen and Chowdhury 1984). This is attributed to the small spatial variability of the unit weight (a coefficient of variation of only 5% in the James Bay case) and
the insensitivity of stability calculations to variations in unit
weight.
The slip surface within each soil layer is divided into segments of lengths not exceeding = 30 m ( = 2ro for exponential autocovariance functions), as illustrated in Fig. 6.
Hence, the average undrained shear strength over the length
of each segment has the same variance as the input data (Table 1) with no reduction. The undrained shear strength of the
lacustrine clay, for example, is modeled by three variables
representing the average strength over three segments of the
slip surface, as illustrated in Fig. A1 in Appendix A. Similarly, the embankment fill is divided into five zones and the
average unit weight within each zone is considered a random
variable. The correlation coefficients between the variables
are estimated using eq. [4], and the correlation coefficient
matrices are shown in Appendix A. Additional variables are
added to the spreadsheet to model statistical uncertainties
and the uncertainties in Bjerrum vane correction factors. In

Fig. 7. Estimating the optimum number of iterations for Monte


Carlo simulation.

2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:16 AM

Color profile: Generic CMYK printer profile


Composite Default screen

676

Can. Geotech. J. Vol. 39, 2002

Fig. 8. Histogram and probability distribution function of the factor of safety.

Fig. 9. Histogram of the probability of unsatisfactory performance based on the results of 25 simulations.

total, 19 variables are defined in the spreadsheet (shaded


cells in Fig. A1 in Appendix A) to account for the various
components of parameter uncertainty.
Using @Risk options, truncation limits are imposed on
the probability distributions of the undrained shear strength
to prevent sampling negative values. Truncation limits at the
mean 3 are also imposed on the probability distributions
of the thickness of the clay crust and the depth of the till
layer. The purpose of these limits is to avoid sampling extreme low or high values that may cause discrepancies in the
sequence of layers.
The stability calculations are based on the Bishop method
of slices. The tables used for factor of safety computations are

shown in Fig. A2 in Appendix A. For each set of sampled


input parameters, the iterative process for factor of safety calculations is repeated until the difference between the factors
of safety in two consecutive calculations is less than 0.01.
Trial simulations indicated that the optimum number of iterations is 32 000 (Fig. 7). Using a seed number of 31 069
(an arbitrary value), the mean factor of safety is estimated to
be 1.46, with a standard deviation of 0.20. The probability of
unsatisfactory performance is 4.70 103. Figure 8 shows
the histogram and the probability distribution function of the
factor of safety. The histogram is slightly right skewed, with
a coefficient of skewness of 0.29. After 25 simulations, the
mean factor of safety is 1.46, with a standard deviation of
0.20 and a coefficient of skewness of 0.30. The mean probability of unsatisfactory performance is 4.70 103, with the 95%
confidence interval between 4.50 103 and 4.90 103. Figure 9 shows the histogram of the probability of unsatisfactory
performance. The reliability index is calculated to be 2.32.
A sensitivity analysis (Fig. 10) shows Spearman rank correlation coefficients for all 19 input variables. It is interesting that many of the factors with major contributions to the
uncertainty of the factor of safety are not related to soil
property measurements. For example, Bjerrums correction
factor for the undrained shear strength of the lacustrine clay,
the statistical uncertainty in the depth of the till, and the statistical uncertainty in the unit weight of the fill (which was
evaluated judgementally) are among the main factors affecting the reliability of the design. The results highlight the
significance of the additional uncertainty that could be introduced by the designer through the use of empirical factors
and subjective estimates of uncertainty. Although subjective
estimates based on experience are acceptable in practice,
quantitative estimates of uncertainty, such as variance, are
not yet reliable. Therefore care should be exercised in making such judgements.
First-order second-moment (FOSM) analysis
The FOSM method is an approximate approach based on
Taylors series expansion of the performance function g(x1,
2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:17 AM

Color profile: Generic CMYK printer profile


Composite Default screen

El-Ramly et al.

677

Fig. 10. Sensitivity analysis results. Spearman rank correlation coefficients for all input variables.

x2,,xn) around its mean value. The performance function g


could be Bishops method of slices, which is a function of a
number of input variables xn that represents soil properties,
pore pressure, and slope geometry. For uncorrelated input
variables, the mean and variance of the factor of safety are
given by eqs. [7] and [8], respectively:
[7]

E[FS] g(E[x1], E[x2],,E[xn])

[8]

g 2
[FS]
[x i ]
i =1 x i

where E[] and 2 [] denote the mean and variance, respectively. For most geotechnical models, the analytical evaluation of the derivatives (g/x i ) is cumbersome. A finite
difference approach is commonly used to approximate the
partial derivatives as follows:
[9]

g
FS1 FS 2
FS

=
1
2
xi
xi xi
xi

where FS1 and FS2 are the factor of safety calculated for values of the input variable xi equal to x1i and x i2 , respectively,
with all other input variables assigned their mean values.
Usually, x1i and x i2 are taken equal to one standard deviation
above and below the mean E[xi]. To account for spatial variability and the reduction in the variance of the average parameters, the variance of measured data, 2 [xi], is reduced
by a reduction factor f. Vanmarcke (1977a) suggested that
the variance reduction factor can be approximated by

[10]

2ro
L

where L is the length over which the parameter of interest is


being averaged.
When the stability of James Bay dykes is analyzed
probabilistically using the FOSM method, the mean factor of
safety is estimated to be 1.46. Table 2 summarizes the calculations of the variance of the factor of safety. For computing
the variance reduction factor f, the length L is taken equal to
the length of the slip surface within each layer, except for
the fill unit weight, where L is taken equal to the length of
the embankment in cross section. The standard deviation of
the factor of safety is computed to be 0.19, thus the reliability index is 2.42. To estimate the probability of unsatisfactory performance, a form of the probability density function
of the factor of safety must be assumed. For normal and lognormal probability distributions, the probability of unsatisfactory performance is estimated to be 8.4 103 and 2.5
103, respectively. Table 3 compares the results of the FOSM
method and the spreadsheet-based Monte Carlo simulation.
The FOSM method appears to be a reasonable approach for
estimating the mean and variance of the factor of safety.
However, the uncertainty about the shape of the probability
density function of the factor of safety introduces uncertainties in estimating the probability of unsatisfactory performance, as illustrated in Table 3.
Simplified analysis
The spatial variability of soil properties and pore-water
2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:17 AM

Color profile: Generic CMYK printer profile


Composite Default screen

678

Can. Geotech. J. Vol. 39, 2002

Table 2. FOSM calculations of the variance of the factor of safety.


Variance, 2 [x]
Input variable
fill ()
fill (kN/m3)
tcr (m)
SuM-1 (kPa)
SuM-2 (kPa)
SuL (kPa)
Dtill (m)
Output

2 [FS]
[FS]

Inherent
variability

Systematic
uncertainty

FS/x

f 2ro /L

Inherent variability
(FS/x)2f 2 [x]

Systematic uncertainty
(FS/x)2 2 [x]

1.00
1.00
0.19a
66.26
66.26
74.82
0.00

3.00
1.00
0.04
7.60
7.60
24.90
1.00

0.009
0.061
0.007
0.005
0.005
0.022
0.055

1.00
0.24

1.00
1.00
0.37
0.00

0.0001
0.0009
0.0000
0.0019
0.0019
0.0129
0.0000

0.0003
0.0037
0.0000
0.0002
0.0002
0.0115
0.0030

0.0180
0.0370
0.1920

0.0190

Variance is reduced by f = 0.2 based on the assessment of Christian et al. (1994).

Table 3. Comparing the outputs of different analysis approaches.


Method of analysis

E[FS]

[FS]

Skewness

Pu

Spreadsheet-based probabilistic slope analysis


FOSMa
Simplified analysis

1.46
1.46
1.46

0.20
0.19
0.25

0.30
Not available
0.32

4.70103
8.40103; 2.50103
2.37102

2.32
2.42
1.84

a
The value 8.40 103 is based on the assumption that the probability density function of the factor of safety is normal, and the value 2.50 103
assumes the probability density function of the factor of safety is log-normal.

Table 4. Input variables and statistical parameters for the simplified analysis (based on Christian
et al. 1994).
Soil parameters

Bias factors

Input variable

Mean

Variance

Mean

Variance

Fill friction angle, fill ()


Fill unit weight, fill (kN/m3)
Clay crust thickness, tcr (m)
Shear strength of marine clay, SuM (kPa)
Bjerrum vane correction factor for SuM, M
Shear strength of lacustrine clay, SuL (kPa)
Bjerrum vane correction factor for SuL, L

30
20
4
34.5

31.2

4
2
1
66.26

74.82

0.006

0.023

pressure is a major source of parameter uncertainty. However, probabilistic analyses ignoring the issues of spatial
variability and statistical uncertainty and relying only on the
means, variances, and probability distributions of measured
data are not uncommon (e.g., Nguyen and Chowdhury 1984;
Wolff and Harr 1987; Duncan 2000). Such an approach, referred to hereafter as a simplified analysis, can be erroneous
and misleading.
To illustrate the errors incurred by the simplified approach,
the James Bay dykes are reanalyzed probabilistically, ignoring the spatial variability of soil properties. The analysis is
based directly on the probability distributions of the data
with no variance reduction and no considerations of statistical uncertainties. Table 4 summarizes the input variables and
the statistical parameters used in the analysis, based on
Christian et al. (1994). All variables are assumed to be normally distributed.
The deterministic critical slip surface almost coincided
with the surface of minimum reliability index according to

the algorithm of Hassan and Wolff (1999). Based on trial


simulations, the latter slip surface yielded a slightly higher
probability of unsatisfactory performance and was subsequently used in the probabilistic assessment. A Microsoft
Excel 97 model and a Monte Carlo simulation using 32 000
iterations and a seed number of 31 069 gave a mean factor
of safety of 1.46, with a standard deviation of 0.25 and a coefficient of skewness of 0.32. The probability of unsatisfactory performance is calculated to be 2.39 102. Based on
the results of 25 simulations, the mean probability of unsatisfactory performance is 2.37 102, with the 95% confidence interval between 2.34 102 and 2.41 102. The
reliability index is 1.84. Table 3 compares the results of the
simplified analysis with those of the previous analyses. The
simplified analysis overestimates the probability of unsatisfactory performance by a factor of 5. This is attributed to the
significant reduction in the uncertainty due to soil variability
as a result of spatial averaging, which is taken into account
in all the analyses but the simplified analysis. In slopes dom 2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:18 AM

Color profile: Generic CMYK printer profile


Composite Default screen

El-Ramly et al.

inated by uncertainties due to the spatial variability of soil


properties, the simplified analysis could significantly overestimate the probability of unsatisfactory performance.

Conclusions
Conventional slope design practice addresses uncertainty
only implicitly and in a subjective manner, thus compromising the adequacy of projections. Without proper consideration of uncertainty, the factor of safety alone can give a
misleading sense of safety and is not a sufficient safety indicator. Probabilistic slope stability analysis is a rational
means to incorporate quantified uncertainties into the design
process. An important conclusion of this study is that probabilistic analyses can be applied in practice without an extensive effort beyond that needed in a conventional analysis.
The stated obstacles impeding the adoption of such techniques into geotechnical practice are more apparent than real.
The developed probabilistic spreadsheet approach, based
on Monte Carlo simulation, makes use of Microsoft Excel
97 and @Risk software, which are familiar and readily available to most engineers. The underlying procedures and concepts are simple and transparent, requiring only fundamental
knowledge of statistics and probability theory. At the same
time, the analysis accounts for the spatial variability of the
input variables, the statistical uncertainty due to limited data,
and the bias in the empirical factors and correlations used.
The approach is flexible in handling real slope problems, including various loading conditions, complex stratigraphy, c
soils, and circular and noncircular slip surfaces.
Probabilistic slope stability analysis provides practicing
engineers with valuable insights that cannot be reached otherwise. For example, the level of uncertainty in the factor of
safety is quantified through the variance of the factor of
safety and the probability of unsatisfactory performance.
This could have an important impact on decisions about a
design factor of safety. If the reliability of the computed factor of safety is deemed high, the profession may be willing
to adopt lower design factors of safety than usual, provided
that the serviceability of the slope is not compromised. With
increasingly sparse funds, many agencies and organizations
prioritize slope repair and maintenance expenditures, as is
the case of hydraulic structures, based on safety levels.
Comparing the safety of different structures based on the
factor of safety alone is inadequate because the underlying
sources and levels of uncertainty are not addressed. By combining the most likely value of the factor of safety and the
uncertainty in that value, the probability of unsatisfactory
performance and the reliability index provide a sounder basis for the comparison.
The practical value of quantifying the relative contributions of the various sources of uncertainty to the overall uncertainty of the factor of safety through sensitivity analyses,
using Spearman rank correlation coefficient, cannot be underestimated. Such information allows available resources,
whether intellectual or physical, to be rationally allocated towards reducing the uncertainties of the variables with the
largest impact on design. The sensitivity analyses undertaken
in this study showed that the uncertainty of Bjerrums vane
correction factor is substantial and could have a large impact
on the reliability of the design. This warns that the reliability

679

of a design could be significantly reduced by the use of


empirical factors and correlations without the designer even
realizing this. Understanding the limitations and, more importantly, the reliability of such factors and correlations
prior to using them is essential.
It is our view that combining conventional deterministic
slope analysis and probabilistic analysis will be beneficial to
slope engineering practice and will enhance the decisionmaking process. It is important to note, however, that simplified probabilistic analyses can be erroneous and misleading.
For example, ignoring the spatial variability of soil properties
and assuming perfect autocorrelations as in our simplified
analysis can significantly overestimate the probability of unsatisfactory performance.

Acknowledgements
The authors would like to thank the Natural Sciences and
Engineering Research Council of Canada for providing the
financial support for this research. We are also grateful for
useful discussions with colleagues at the University of Alberta and elsewhere.

References
Alonso, E.E. 1976. Risk analysis of slopes and its application to
slopes in Canadian sensitive clays. Gotechnique, 26: 453472.
Ang, A.H-S., and Tang, W.H. 1975. Probability concepts in engineering planning and design. Vol. I. Basic principles. John Wiley,
New York.
Ang, A.H-S., and Tang, W.H. 1984. Probability concepts in engineering planning and design. Vol. II. Decision, risk and reliability. John Wiley, New York.
Baecher, G.B. 1987. Statistical analysis of geotechnical data. Final
Report GL-87-1, U.S. Army Corps of Engineers, Waterways Experiment Station, Vicksburg, Miss.
Bergado, D.T., Patron, B.C., Youyongwatana, W., Chai, J.C., and
Yudhbir. 1994. Reliability-based analysis of embankment on
soft Bangkok clay. Structural Safety, 13: 247266.
Bishop, A.W. 1955. The use of the slip circle in the stability analysis of slopes. Gotechnique, 5: 717.
Bjerrum, L. 1972. Embankments on soft ground. In Proceedings of
the American Society of Civil Engineers Specialty Conference on
Performance of Earth and Earth-Supported Structures, Purdue
University, Lafayette, Ind., June 1114, Vol. 2, pp. 154.
Chowdhury, R.N. 1984. Recent developments in landslide studies:
probabilistic methods. In Proceedings of the 4th International
Symposium on Landslides, Toronto, Ont., September 1621. Canadian Geotechnical Society, Vol. 1, pp. 209228.
Chowdhury, R.N., and Tang, W.H. 1987. Comparison of risk models for slopes. In Proceedings of the 5th International Conference on Applications of Statistics and Probability in Soil and
Structural Engineering, Vancouver, B.C., May 2529, Vol. 2,
pp. 863869.
Christian, J.T. 1996. Reliability methods for stability of existing
slopes. In Proceedings of Uncertainty 96. Geotechnical Special
Publication 58, Vol. 2, pp. 409419.
Christian, J.T., Ladd, C.C., and Baecher, G.B. 1994. Reliability and
probability in stability analysis. Journal of Geotechnical Engineering, ASCE, 120: 10711111.
Decisioneering Inc. 1996. Crystal ball: Monte Carlo simulation software
for spreadsheet risk analysis. Decisioneering Inc., Denver, Colo.
2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:18 AM

Color profile: Generic CMYK printer profile


Composite Default screen

680
DeGroot, D.J. 1996. Analyzing spatial variability of in situ soil
properties. In Proceedings, Uncertainty 96. Geotechnical Special Publication 58, Vol. 1, pp. 210238.
DeGroot, D.J., and Baecher, G.B. 1993. Estimating autocovariance
of in-situ soil properties. Journal of Geotechnical Engineering,
ASCE, 119(1): 147166.
Deutsch, C.V., and Journel, A.G. 1998. GSLIB: geostatistical software
library and users guide. Oxford University Press, New York,
543 pp.
Duncan, J.M. 2000. Factors of safety and reliability in geotechnical
engineering. Journal of Geotechnical and Geoenvironmental Engineering, ASCE, 126: 307316.
El-Ramly, H. 2001. Probabilistic analyses of landslide hazards and
risks: bridging theory and practice. Ph.D. thesis, University of
Alberta, Edmonton, Alta.
Harr, M.E. 1977. Mechanics of particulate media a probabilistic
approach. McGraw-Hill, New York, 543 pp.
Harr, M.E. 1987. Reliability-based design in civil engineering.
McGraw-Hill Inc., New York, 290 pp.
Hassan, A., and Wolff, T. 1999. Search algorithm for minimum reliability index of earth slopes. Journal of Geotechnical and Geoenvironmental Engineering, ASCE, 125: 301308.
Honjo, Y., and Kuroda, K. 1991. A new look at fluctuating geotechnical data for reliability design. Soils and Foundations, 31:
110120.
Jaksa, M.B., Brooker, P.I., and Kaggwa, W.S. 1997. Inaccuracies
associated with estimating random measurement errors. Journal
of Geotechnical and Geoenvironmental Engineering, ASCE,
123: 393401.
Kulhawy, F.H., and Trautmann, C.H. 1996. Estimation of in-situ test
uncertainty. In Proceedings of Uncertainty 96. Geotechnical Special Publication 58, Vol. 1, pp. 269286.
Kulhawy, F.H., Roth, M.S., and Grigoriu, M.D. 1991. Some statistical
evaluation of geotechnical properties. In Proceedings of the 6th International Conference on Applications of Statistics and Probability
in Civil Engineering, Mexico City, Vol. 2, pp. 705712.
Lacasse, S., and Nadim, F. 1996. Uncertainties in characterizing
soil properties. In Proceedings of Uncertainty 96. Geotechnical
Special Publication 58, Vol. 1, pp. 4975.
Ladd, C.C., Dascal, O., Law, K.T., Lefebrve, G., Lessard, G., Mesri,
G., and Tavenas, F. 1983. Report of the subcommittee on embankment stability annex II. Committee of Specialists on Sensitive Clays on the NBR Complex, Socit dEnergie de la Baie
James, Montral, Que.
Law, A.M., and McComas, M.G. 1986. Pitfalls in the simulation of
manufacturing systems. In Proceedings of the Winter Simulation
Conference, Washington, D.C., pp. 539542.
Li, K.S. 1991. Discussion: probabilistic potentiometric surface mapping. Journal of Geotechnical Engineering, ASCE, 117: 14571458.
Lumb, P. 1966. The variability of natural soils. Canadian Geotechnical
Journal, 3: 7497.
Major, G., Ross-Brown, D.M., and Kim, H-S. 1978. A general
probabilistic analysis for three-dimensional wedge failures. In
Proceedings of the 19th U.S. Symposium on Rock Mechanics,
Stateline, Nev., 13 May, pp. 4556.
Matsuo, M., and Kuroda, K. 1974. Probabilistic approach to design
of embankments. Soils and Foundations, 14: 117.
Microsoft Corporation. 1997. Microsoft Excel 97. Microsoft Corporation, Redmond, Wash.
Mostyn, G.R., and Li, K.S. 1993. Probabilistic slope stability analysis state-of-play. In Probabilistic methods in geotechnical
engineering. Edited by K.S. Li and S.-C.R. Lo. A.A. Balkema,
Rotterdam. pp. 89109.

Can. Geotech. J. Vol. 39, 2002


Neter, J., Wasserman, W., and Kutner, M.H. 1990. Applied linear
statistical models. Richard D. Irwin Inc., Boston.
Nguyen, V.U., and Chowdhury, R.N. 1984. Probabilistic study of
spoil pile stability in strip coal mines two techniques compared. International Journal of Rock Mechanics, Mining Science
and Geomechanics, 21: 303312.
Nguyen, V.U., and Chowdhury, R.N. 1985. Simulation for risk
analysis with correlated variables. Gotechnique, 35: 4758.
Palisade Corporation. 1996. @Risk: risk analysis and simulation
add-in for Microsoft Excel or Lotus 1-2-3. Palisade Corporation,
Newfield, N.Y.
Priest, D.S., and Brown, E.T. 1983. Probabilistic stability analysis
of variable rock slopes. Transactions of the Institute of Mining
and Metallurgy, 92: A1A12.
Souli, M., Montes, P., and Silvestri, V. 1990. Modelling spatial
variability of soil parameters. Canadian Geotechnical Journal,
27: 617630.
Tang, W.H., Yucemen, M.S., and Ang, A.H.S. 1976. Probabilitybased short-term design of slopes. Canadian Geotechnical Journal, 13: 201215.
Tobutt, D.C. 1982. Monte Carlo simulation methods for slope stability. Computers and Geosciences, 8: 199208.
U.S. Army Corps of Engineers. 1992. Reliability assessment of
navigation structures. Engineering Technical Letter 1110-2-532,
U.S. Army Corps of Engineers, Washington, D.C.
U.S. Army Corps of Engineers. 1995. Introduction to probability
and reliability methods for use in geotechnical engineering. Engineering Technical Letter 1110-2-547, U.S. Army Corps of Engineers, Washington, D.C.
Vanmarcke, E.H. 1977a. Probabilistic modeling of soil profiles.
Journal of the Geotechnical Engineering Division, ASCE, 103:
12271246.
Vanmarcke, E.H. 1977b. Reliability of earth slopes. Journal of the
Geotechnical Engineering Division, ASCE, 103: 12471265.
Vanmarcke, E.H. 1980. Probabilistic stability analysis of earth slopes.
Engineering Geology, 16: 2950.
Vanmarcke, E.H. 1983. Random fields: analysis and synthesis. MIT
Press, Cambridge, Mass.
Whitman, V.W. 1984. Evaluating calculated risk in geotechnical
engineering. Journal of Geotechnical Engineering, ASCE, 110:
145188.
Wolff, T.F. 1996. Probabilistic slope stability in theory and practice. In Proceedings of Uncertainty 96. Geotechnical Special
Publication 58, Vol. 2, pp. 419433.
Wolff, T.F., and Harr, M.E. 1987. Slope design for earth dams. In
Proceedings of the 5th International Conference on Applications
of Statistics and Probability in Soil and Structural Engineering,
Vancouver, B.C., May 2529, Vol. 2, pp. 725732.
Yucemen, M.S., and Al-Homoud, A.S. 1990. Probabilistic threedimensional stability analysis of slopes. Structural Safety, 9: 120.

List of symbols
b slice width
B bias correction factor
C cohesion
Dtill depth of till
E[] mean
f variance reduction factor
F(x) cumulative probability distribution function of variable x
FS factor of safety
2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:18 AM

Color profile: Generic CMYK printer profile


Composite Default screen

El-Ramly et al.
g(x1, x2,
, xn)
h
hcr
hfill
hL
hM
l1, l2, l3
L
m
n
Pu
r
ro
ru
Su
SuL
SuM
t
tcr
ti
u
V
W
x
xi
X(z)
z
Z0
Z1

performance function
total height of slice
height in clay crust
height in embankment fill
height in lacustrine clay
height in marine clay
segments of slip surface of lengths l1, l2, and l3
length over which soil parameters are averaged
number of iterations in Monte Carlo simulation
number of observations
probability of unsatisfactory performance
separation distance between two variables in the space
of a random field
autocorrelation distance
pore-pressure ratio
undrained shear strength
undrained shear strength of lacustrine clay
undrained shear strength of marine clay
trend component for variable x
clay crust thickness
trend component at location i
pore-water pressure
input variable adjusted for statistical uncertainty and bias
total height of slice
random variable
value of random variable x at location i
local average of variable x over a length z
distance
separation distance between two intervals z, z
distance from the beginning of the first interval to the
beginning of the second interval

681
Z2 distance from the beginning of the first interval to the
end of the second interval
Z3 distance from the end of the first interval to the end of
the second interval
angle to the horizontal of slice base
reliability index
scale of fluctuation
residual component, equal to difference between x and t
i residual component at location i
friction angle
effective friction angle
fill fill friction angle
variable representing soil stratigraphy
bulk unit weight
fill fill unit weight
soil parameter
Bjerrum vane correction factor
L , M Bjerrum vane correction factor for SuL and SuM, respectively
correlation coefficient between two variables
[] standard deviation
2 [] variance
2 z variance of X(z)
z interval of length z
z interval of length z
(z) variance function

Appendix A
This appendix illustrates the general structure of the
spreadsheet model used in the probabilistic analysis of the
James Bay dyke.

2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:19 AM

Color profile: Generic CMYK printer profile


Composite Default screen

682

Can. Geotech. J. Vol. 39, 2002

Fig. A1. Spreadsheet model, section 1: dyke geometry, stratigraphy, and input variables.

2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:19 AM

Color profile: Generic CMYK printer profile


Composite Default screen

El-Ramly et al.

683

Fig. A2. Spreadsheet model, section 2: tables of factor of safety computations. G.S., ground surface; S.S., slip surface.

2002 NRC Canada

I:\cgj\Cgj39\Cgj-03\T02-034.vp
Tuesday, May 14, 2002 10:37:20 AM