Computers and Geotechnics 69 (2015) 540–550

Contents lists available at ScienceDirect

Computers and Geotechnics

journal homepage: www.elsevier.com/locate/compgeo

Research Paper

Assessment of the application of point estimate methods in the

probabilistic stability analysis of slopes
Morteza Ahmadabadi ⇑, Rainer Poisel 1
Technical University of Vienna, Karlsplatz 13, 1040 Wien, Austria

a r t i c l e i n f o a b s t r a c t

Article history: This study examines the applicability of point estimate methods (PEMs) in the probabilistic stability anal-
Received 26 August 2014 ysis of slopes. For illustrative purposes, two slopes indicating two different mechanisms of failure are
Received in revised form 18 June 2015 selected and probabilistic analyses are performed employing three PEMs that have been widely cited
Accepted 19 June 2015
in different literature reports. To evaluate the benefits and shortcomings of each PEM, various cases
Available online 4 July 2015
are studied in which random parameters may be dependent or independent and can have arbitrary dis-
tributions. Furthermore, the efficiency and accuracy of the considered PEMs are assessed by comparing
Keywords:
the results with those obtained by Monte Carlo Simulation (MCS). The results indicate that the considered
Probabilistic analysis
Point estimate method
PEMs are sufficiently precise and, in case an intricate geotechnical structure must be probabilistically
Monte Carlo Simulation analysed, they can readily be incorporated with numerical methods.
Reliability index Ó 2015 Elsevier Ltd. All rights reserved.
Probability of failure

1. Introduction PEM often constitutes a more practical alternative because it

requires less computation time along with requiring only statisti-
Numerical analyses are becoming increasingly popular in cal moments as inputs. In this study, the applicability of PEMs in
geomechanics for the investigation of the stability of geotechnical probabilistically analysing the stability of slopes is investigated
structures. However, soil and rock are known to be among the for three PEMs that have been widely discussed in different litera-
most variable of all engineering materials, leading to uncertainties ture reports. In the following report, first, the basics of the consid-
in the outputs of various analyses; hence, probabilistic analyses are ered point estimate methods are presented (i.e., Section 2). For
performed to take these uncertainties into account. The first step in some PEMs, shortcomings arise when random parameters are cor-
performing such analyses is the selection of an appropriate proba- related or non-normally distributed. The proposed solution to
bilistic tool. Probabilistic tools, such as the first-order reliability these drawbacks is addressed in Section 3. Finally, two illustrative
method (FORM) [13], the second-order reliability method (SORM) examples comprising two slopes with two different modes of fail-
[2], the Monte-Carlo Simulation (MCS) [37] and the point estimate ure are represented in Section 4.
methods (PEMs) [12,14,15,23,28,35,36,42,43] are some alterna-
tives. In combination with numerical methods, the MCS and 2. Point estimate methods (PEMs)
PEMs may be more practical because they do not require the def-
inition of the performance function in a closed form, as is required In general, PEMs consist of replacing a continuous density distri-
when FORM and SORM are used. The MCS is based on several bution function i.e., f x ðxÞ by specifically defining discrete probabili-
deterministic analyses, and better precision in the model output ties that are supposed to model the same low-order moments of that
is achieved if the selected sample is of sufficient size; thus, when distribution function. The determination of these moments is per-
the model computation is time-consuming, the Monte Carlo formed by adding up the weighted discrete realizations. In Fig. 1, this
Simulation would not be practical for uncertainty analysis. relationship is shown in a simplified way by two realizations at xþ
Furthermore, the MCS requires the specification of the marginal and x , represented by the weights wþ and w , respectively.
distributions of the random variables. Under such circumstances, This method is straightforward, easy to use and has become a
staple of geotechnical reliability analyses. This method was first
⇑ Corresponding author. Tel.: +43 6509440594. proposed by Rosenblueth [35,36]. Several researchers in the field
E-mail addresses: [email protected] (M. Ahmadabadi), rainer.poisel@
of reliability, e.g., Lind [23], Zhou and Nowak [43], Harr [12],
tuwien.ac.at (R. Poisel). Hong [15] and Zhao and Ono [42], have commented on the primary
1
Tel.: +43 1 58801 20301. point estimate method (i.e., Rosenblueth’s PEM) and implemented

http://dx.doi.org/10.1016/j.compgeo.2015.06.016
0266-352X/Ó 2015 Elsevier Ltd. All rights reserved.
 M. Ahmadabadi, R. Poisel / Computers and Geotechnics 69 (2015) 540–550 541

well. Rosenblueth’s algorithm considers two realization points for

each random variable. When random variables are symmetric or
x-,w- x+,w+ normally distributed, the realization points are located at l  r,
in which l and r are the mean value and standard deviation of
random variable, respectively. By permuting the two points in each
dimension, 2n (n: number of random parameters) combinations of
PDF

realization points exist, i.e., the model should be evaluated 2n

times. Describing the details of the concept and formulation of this
method is not the objective of this paper; for further studies, refer
to Refs. [35,36]

2.2. Harr’s method

μ

Fig. 1. Simplified graphical representation of the PEM. Using Harr’s algorithm (1989) [12] to solve problems involving
n normally distributed and correlated random variables; a hyper-
pffiffiffi
sphere with a radius n centered at the origin of standardized
some improvements. Regardless of the concept, efficiency and parameters space is defined as representing the correlation matrix.
accuracy of the above-mentioned PEMs, they mainly differ in the The realization points at which the model should be evaluated are
number and location of realization points and in the method of located at intersections of the hypersphere and eigenvectors of the
treating asymmetric or correlated variables. In this study, three correlation matrix. Because each of the n eigenvectors has two
PEMs (i.e., Rosenblueth’s, Harr’s and Hong’s methods) are chosen intersections with the hypersphere, 2n realization points are pro-
to model the uncertainties that exist in the stability analyses of duced, i.e., the model should be evaluated 2n times. Although
the two slopes. Table 1 summarizes a general comparison of these Harr’s PEM is computationally more efficient compared with
three PEMs. Rosenblueth’s PEM, Harr’s PEM has the disadvantage that it cannot
This comparison provides an initial idea of the applicability, model non-normally distributed variables. In addition, when the
advantages and impediments of each method. For example, irre- number of variables n is large, the realization points are located
spective of the accuracy of each method investigated in this study, farther from the mean, thereby causing the inaccuracy or inappli-
it is clear that when the number of random variables, i.e., n, is large, cability of Harr’s PEM [6]. Further explanations of Harr’s PEM can
Rosenblueth’s method may be computationally inefficient because be found in Refs. [1,4,6,12,14,28].
it must evaluate the model m ¼ 2n times. In addition, deficiencies
in Harr’s and Hong’s approaches arise when they address asym- 2.3. Hong’s method
metric or correlated variables, respectively. In this study, the
Nataf transformation [21,24,27] (see Section 3) is proposed to When using Hong’s method [15], more than two realization
address the drawback of these methods with respect to the depen- points can be selected. In addition, Hong’s method takes into
dent non-normal variables. As is clear from the table above, con- account higher-order moments of random variables; as a result,
trary to Reosenblueth’s and Harr’s PEMs, the Hong’s PEM has the it is possible to calculate higher-order moments of the model out-
capability to consider more than two realization points for each put. In the case that, for each variable, k estimation points are con-
random variable. In this study, Hong’s approach is categorized into sidered, the model should be evaluated a total of n  k times. In the
four schemes depending on the number (i.e. k ¼ 2; 3; and 5) that is present study, this method is categorized into four schemes based
selected for each random parameter. on the number of realization points considered. For example, the
4n þ 1 scheme corresponds to the case that for each variable, five
realization points are selected, such that one point is located at
2.1. Rosenblueth’s method
mean value and two points are located at each side of the mean.
Note that Hong’s PEM also has some shortcomings. An important
The primary point estimate method proposed by Rosenblueth
drawback of Hong’s PEM is that it is unable to model correlated
[35,36] is widely used in geotechnical practice for reliability calcu-
random parameters. Furthermore, sometimes some schemes of
lations [6]. An advantage of this method is its simplicity in compar-
this method are unusable because realization points are located
ison with the other PEMs, although it has some disadvantages as
beyond the marginal distribution or they are imaginary numbers.
This alternative is not discussed in detail, and further details can
Table 1 be found in Refs. [3,6,15].
Comparison of the three employed PEMs (n: number of variables).

Author’s No. of realization No. of Capability for variables 3. Nataf transformation

method points for each model with
variable (k) evaluations
As mentioned in Section 2, Harr’s PEM is incapable to consider
(m)
Asymmetry Correlation asymmetric variables; conversely, Hong’s schemes are only appli-
cable for uncorrelated variables. To circumvent such limitations,
Rosenblueth 2 2n U U
the Nataf transformation [27] is implemented when dealing with
Harr 2 2n  U
Hong U  correlated non-normal variables. In the following, the Nataf trans-
2n 2 2n formation is briefly described.
scheme Let the correlated random variables vector X be
2n + 1 3 2n þ 1
scheme X ¼ ðx1 ; x2 ; . . . ; xn Þ ð1Þ
3n 3 3n
scheme Assuming the marginal CDF F xi ðxi Þ; i ¼ 1; . . . ; n; of each random
4n + 1 5 4n þ 1 variable xi ; i ¼ 1; . . . n; of each random variable X i and the correla-
scheme
tion matrix q ¼ ðqij Þnn are known, and applying the
542 M. Ahmadabadi, R. Poisel / Computers and Geotechnics 69 (2015) 540–550

Fig. 2. Geometry of slope I.

Table 2
Material properties for slope I. f X1 ðx1 Þf X2 ðx2 Þ; . . . ; f Xn ðxn Þ
f x ðxÞ ¼ /n ðy; q0 Þ ð4Þ
Material Parameters /ðy1 Þ/ðy2 Þ; . . . ; /ðyn Þ
c ðkN=m3 Þ rt (kPa) c (kPa) / (°) E (MPa) m (–) Generally, this distribution model is referred to as the Nataf
MH Weathered A 21 8 8 34 50 0.4 distribution.
M Weathered A 21 7 7 24 50 0.4 To solve q0 ¼ ðq0ij Þnn in Eq. (4), any two random variables
Sed. Marine. 19 4 4 17 50 0.4
Fresh Andesite 21 100 100 30 50 0.4
ðxi ; xj Þ are considered. The linear correlation between these random
Mat 7 19 0 0 37 50 0.4 variables is calculated by
Mat 8 21 100 100 30 50 0.4
Z Z ! !
þ1 þ1
F 1
i ½Uðyi Þ  lxi
F 1
j ½Uðyj Þ  lxj
qij ¼ /2 ðyi ; yj ; q0ij Þdyi dyj
1 1 rxi rxj
isoprobabilistic transformation, the random vector X can be trans-
ð5Þ
formed to the standardized normal random vector
Y ¼ ðy1 ; y2 ; . . . ; yn Þin y space as follows: where lxi and lxj are the means of correlated random variables xi
( and xj , respectively; rxi and rxj are the standard deviation of corre-
Uðyi Þ ¼ F xi ðxi Þ
ð2Þ lated random variables xi and xj , respectively. Given the marginal
yi ¼ U1 ½F xi ðxi Þ distributions and the correlation for a pair of random variables,
the equivalent correlation coefficient q0ij can be determined by
in which UðÞ and U1 ðÞ are the standard normal CDF and the inverse
standard normal, respectively, with joint probability density func- solving the nonlinear equation, as shown in Eq. (5). Because a prac-
tion Un ðy; q0 Þ having zero means, unit standard deviation and a cor- tical solution of this equation is not easily obtained, Liu and Der
relation matrix q0 ¼ ðq0ij Þnn that is given by Kiureghian [24] developed a set of semi-empirical formulae.
  However, such empirical formulae may not always be valid for ran-
1 1 T 1 dom variables following arbitrary distributions. In this study, the
Un ðy; q0 Þ ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp y q0 y ð3Þ
ð2pÞn jq0 j 2 solution of this equation, i.e., Eq. (5), is performed using a trial
and error procedure by means of a MATLAB code that is presented
Then, according to Nataf transformation theory and the differ- in Appendix A.
entiation rules of implicit functions, the approximate joint proba- By obtaining the equivalent correlation matrix q0 , which can be
bility density function f x ðÞ in x space is expressed as decomposed as

q0 ¼ C0 CT0 ð6Þ
(where C0 is the lower triangular matrix obtained from Cholesky
Table 3 decomposition), the dependent standard normal vector Y can be
Statistical information of random parameters of slope I.
obtained from the independent standard normal random vector U
Material Random parameters as follows:
c (kPa) / (°) Y ¼ C0 U ð7Þ
l r l r
The original non-normal random variables are given by
MH Weathered A 8 2 34 3
M Weathered A 7 2 24 3
xi ¼ F 1
i ½Uðyi Þ; i ¼ 1; 2; . . . ; n ð8Þ
Sed. Marine. 4 1 17 2
 M. Ahmadabadi, R. Poisel / Computers and Geotechnics 69 (2015) 540–550 543

Fig. 3. A typical result of the model evaluation of slope I (i.e., safety factor) using the shear strength reduction method.

Table 4
Results of the probabilistic stability analysis of slope I using PEMs.

Employed numerical code PEM No. model evaluations m Factor of safety Reliability index Probability of failure
l r ðbÞ P f ð%Þ

FLAC Rosenblueth 64 1.036 0.0893 0.4009 34.42

Harr & Hong’s 2n scheme 12 1.038 0.0889 0.4252 33.53
Hong
2n + 1 & 3n schemes 13 & 18 1.029 0.0873 0.3276 37.16
4n + 1 scheme 25 1.021 0.0842 0.2458 40.29

Table 5 different mechanisms of failure incorporating the

Results of probabilistic stability analysis of slope I performed by Rocscience [31]. above-mentioned PEMs with different numerical methods.
Employed Probabilistic No. model Factor of Probability
numerical code tool evaluations safety of failure
m
4.1. Example 1  slope I with circular failure mode
l r P f ð%Þ
This example is taken from Rocscience [31] and aims to reveal
Finite element Rosenblueth’s 64 1.05 0.092 30.40
(phase 2) PEM how uncertainties that exist in the shear strength of slope materi-
Analytical (Slide) MCS 10,000 1.07 0.091 22.0 als can be taken into consideration by means of the aforemen-
tioned PEMs. In addition, the examined slope indicates the case
that a cut is made in weak materials, such as highly weathered
or closely fractured rock and rock fills. In such materials, a strongly
Table 6
Number of simulations required for calculating pf ¼ 35% with 10% error using MCS in
defined structural pattern (wedge or toppling) no longer exists, and
example 1. the slide surface is free to find the line of least resistance through
the slope. Observations of slope failures in these materials suggest
Standard deviation (d) Confidence (%) No. simulations
that this slide surface generally takes the form of a circle [41]. In
1.28 80 305 this case, the failure of rock slope is not controlled by geological
1.64 85 500
features, such as bedding planes and joints that divide the rock into
1.96 95 714
discontinues mass; thus, the slope is supposed to be as continuum
and it is analysed by FLAC [16] using the strength reduction tech-
4. Illustrative examples nique [39].

In this section, the applicability and usefulness of the aforemen- 4.1.1. Slope geometry
tioned PEMs in probabilistic stability analysis of slopes is consid- The examined slope depicted in Fig. 2 has a height of H ¼ 60 m
ered. The examples are used to investigate probabilistically two and a general face inclination of wf  58 . It is assumed that water
544 M. Ahmadabadi, R. Poisel / Computers and Geotechnics 69 (2015) 540–550

table is very low, meaning that its effects could be disregarded in distribution of the model output (i.e., safety factor). In this exam-
the stability calculations. Note that for the sake of consistency ple, it is supposed that safety factor has a normal distribution;
and to compare the results of analyses with those presented by hence, Pf and b are computed using Eq. (9).
[31], the slope geometry and the properties of slope materials
are represented without any manipulation. l1
b¼ ; Pf ¼ 1  UðbÞ; ð9Þ
r
4.1.2. Material properties
In Fig. 2, slope consists of six materials. The properties of the where UðÞ denotes the cumulative density function (CDF) of the
slope materials are represented in Table 2, in which c; rt ; c; /; E standard normal distribution. Note that when random variables
and m denote unit weight, tensile strength, cohesion, internal fric- are uncorrelated and normally distributed, Hong’s 2n scheme and
tion angle, elasticity module and Poisson’s ratio, respectively. Harr’s algorithm are identical. Similarly, Hong’s 2n þ 1 scheme
As mentioned above, the shear strength of the slope materials is and Hong’s 3n scheme are also the same.
assumed to be uncertain. In addition, the behavior of the materials A comparison of the results of different point estimate methods
is assumed to be modeled by the Mohr–Coulomb failure criterion, indicates that there is good agreement among these methods;
meaning that cohesion ðcÞ and friction angle ð/Þ are treated as ran- however the results of probabilistic stability analyses performed
dom variables, and 6  2 ¼ 12 random variables should be consid- by Rocscience [31] using finite element [32] and analytical meth-
ered in probabilistic stability calculations; to make the ods [34] are provided in Table 5 for further comparison.
probabilistic analyses more efficient, the friction angle and cohe- To highlight the accuracy of the PEMs, the stability of slope is
sion of three materials termed Mat 8, Mat 7, and Fresh Andesite also probabilistically analysed using Monte Carlo Simulation
are assumed to be deterministic because they have higher values (MCS). As mentioned before, the accuracy of MCS is directly corre-
in comparison to other materials, and they do not determine the lated with the number of simulations; nevertheless, the number of
location of critical failure surface (i.e., with minimum safety simulations was chosen to be as small as possible because model
factor). evaluation with a PC that is equipped with a quad core processor
Table 3 contains the statistical information of the random takes approximately 25 min, i.e., if a large sample is selected, then
parameters that are assumed to be uncorrelated and normally probabilistic analyses by means of MCS are infeasible.
distributed. To estimate the number of simulations required, Eq. (10) is used
[7]. This formula indicates that, for calculating the probability of
failure p with a% error, n simulations are required.
4.1.3. Probabilistic stability analyses
As mentioned earlier, probabilistic stability analyses of slope  2
d 1p
are performed by incorporating the above-mentioned point esti- n¼ ð10Þ
mate methods with FLAC. Fig. 3 displays a sample of the model a p
evaluations in terms of the safety factor; for clarity, the velocity Table 6 represents the number of simulations required to calcu-
vectors and maximum shear strains are also displayed. Note that late the probability of failure Pf ¼ 35% with 10% error with differ-
in the numerical model, the boundary conditions correspond to ent confidences. According to the table below, to calculate the
fixed movement in the horizontal direction at the right and the left probability of failure with 10% error and 85% confidence (i.e. stan-
end points of the model (i.e., x ¼ 0 m & x ¼ 200 m), and the fixed dard deviation d = 1.64) with MCS, sample size should be Ns = 500
movements in the horizontal and vertical directions at the lowest (see row 2 of the Table 6). It is apparent that MCS is much more
part of the model (i.e. y ¼ 40 m). inefficient in comparison with PEMs when it is coupled with
numerical modeling codes (e.g., FLAC).
4.1.4. Results To better understand the effect of sample size in the results of
Table 4 summarizes the results of probabilistic stability analy- MCS, two sets of probabilistic stability analyses are conducted for
ses of slope using the above-mentioned PEMs. These results samples with sizes N s1 ¼ 500 and N s2 ¼ 1500. The results of these
include the mean ðlÞ and standard deviation ðrÞ of the safety fac- two probabilistic analyses are compared in Fig. 4 in terms of the
tor, reliability index ðbÞ and probability of failure ðPf Þ. Note that for histogram of safety factor. In Fig. 5, the histograms of input random
calculating Pf and b, it is necessary to know (or estimate) the parameters for a samples with size of Ns2 = 1500 are presented.

80 100
Number of simulations 90 Number of simulations
70 Ns=500 Ns=1500
80
60
70
50
60
Frequency

Frequency

40 50

40
30
30
20
20
10
10

0 0
0.85 0.9 0.95 1 1.05 1.1 1.15 1.2 1.25 0.9 0.95 1 1.05 1.1 1.15
Safety factor Safety factor

Fig. 4. Histogram of the safety factors for slope I with MCS, left graph N s1 ¼ 500, right graph N s2 ¼ 1500.
 M. Ahmadabadi, R. Poisel / Computers and Geotechnics 69 (2015) 540–550 545

100 120
Number of samples=1500 Number of samples=1500
90
100
80

70
80
60
Frequency

Frequency
50 60

40
40
30

20
20
10

0 0
4000 5000 6000 7000 8000 9000 10000110001200013000 30 32 34 36 38 40
Cohesion1 (Pa) Friction angle1

100
90
Number of samples=1500
Number of samples=1500 90
80
80
70
70
60
60
Frequency

Frequency

50
50
40
40
30
30
20 20

10 10

0 0
13 14 15 16 17 18 19 20 21 1000 2000 3000 4000 5000 6000 7000
Friction angle 2 Cohesion2 (Pa)

120 90
Number of samples=1500
Number of samples=1500
80
100
70

80 60
Frequency

Frequency

50
60
40

40 30

20
20
10

0 0
2000 3000 4000 5000 6000 7000 8000 9000 1000011000 19 20 21 22 23 24 25 26 27 28 29
Cohesion 3 (Pa) Friction angle 3

Fig. 5. Histograms of the random parameters of slope I, with normal distribution fits.

In addition, a quantitative comparison is also demonstrated in d ¼ 1:99, which means that with a sample size of n ¼ 1500, the
Table 7. An increase in the number of simulations is found to lead probability of failure, pf , would be 21%, with more than 95% confi-
to a considerable change in P f , although the mean value of the dence (see Table 6). Note that these sample sizes are large for this
safety factor changes slightly. From the data in Table 7, and by sub- type of numerical calculation, and this makes the probabilistic sta-
stituting n ¼ 1500; a ¼ 10% and p ¼ 21% in Eq. (10), we obtain bility analyses expensive.
546 M. Ahmadabadi, R. Poisel / Computers and Geotechnics 69 (2015) 540–550

Table 7 and wp is the dip of the sliding plane. b denotes the location of ten-
Results of the probabilistic stability analysis of slope I with MCS.
sion crack behind the slope crest, and B is the location of tension
Employed numerical No. Factor of safety Probability of crack behind the slope toe. The following assumptions are made
code simulations failure in the plane failure analysis:
l r P f ð%Þ

FLAC N s1 ¼ 500 1.035 0.0815 32  Both sliding surface and tension cracks strike parallel to the
N s2 ¼ 1500 1.040 0.0524 21.3
slope.
 The tension crack is vertical and is filled with water to a depth
zw .
 Water enters the sliding surface along the base of the tension
crack and seeps along the sliding surface, escaping at atmo-
spheric pressure where the sliding surface daylights in the slope
face.
 The shear strength s of the sliding surface is defined by cohe-
sion c and friction angle /, which are related by the Mohr–
Coulomb failure criterion.

4.2.2. Material properties

Again, for the sake of consistency and to compare the results of
the probabilistic analyses with those presented by Wyllie & Mah
[41], the input parameters and variables are considered without
any manipulation. The deterministic and statistic parameters that
are utilized in the probabilistic stability analysis of slope II are
Fig. 6. Geometry of slope II with plane failure. listed in Tables 8 and 9.
The statistical parameters of slope 2 are listed in Table 9. In this
table, wp , /, c and zw symbolise the dip of the sliding plane, the fric-
Table 8 tion angle, the cohesion of the joint and the depth of water in the
Deterministic parameters of slope II. tension crack, respectively.
Parameter Unit Value
H m 30.50 4.2.3. Probabilistic stability analyses
cw kN
m3
10
Before evaluating the model, the realization points should be
crock kN
m3
27
computed based on the desired PEMs. In this example, the random
wf deg. 58
variables are assumed to be uncorrelated, but two random vari-
b m 12.50
ables, i.e., c and zzw have non-normal distributions; consequently,
B m 31.56
z m 19.01 m as mentioned in Sections 2 and 3, Harr’s PEM should be manipu-
lated by means of the Nataf transformation, i.e., the non-normal
variables are first transformed to the normal space and then
4.2. Example 2  rock slope II with plane failure Harr’s PEM is used to obtain the realization points. In this research,
this procedure that is a combination of Harr’s PEM and a normal
This example is taken from chapter 6 of Wyllie & Mah [41] and transformation is called the modified Harr’s PEM. Table 10 presents
aims to probabilistically investigate the stability of a rock slope in a comparison among the realization points that are calculated
the context of plane failure. Although plane failure is a compara- using the modified Harr’s and the primary Harr’s PEMs. Patently,
tively rare sight in rock slopes, it would not be appropriate to in the case that the non-normal random parameter has a substan-
ignore the two-dimensional case because there are many valuable tial effect on the model output, the modification of Harr’s method
lessons to be learned from a consideration of the mechanics of this is crucial.
simple failure mode. Plane failure is particularly useful for demon-
strating the sensitivity of the slope to changes in shear strength 4.2.4. Results
and ground water conditions – changes that are less obvious when Although, in a 2D model of rock slope with plane failure, there is
dealing with the more complex mechanics of three-dimensional an analytical solution for the safety factor (see chapter 6 of [41]),
slope failure [41]. probabilistic stability analyses of rock slopes are accomplished by
coupling the aforesaid point estimate methods with UDEC [17].
4.2.1. Slope geometry The distinct element model for the rock slope is shown in Fig. 7.
The slope geometry and ground water condition considered in In the figure, the results of the model evaluation in terms of safety
this analysis are defined in Fig. 6. In this figure, z is the depth of factor along with the vectors of displacement of the sliding block
the tension crack, H is the slope height, wf is the slope face angle are also observed.

Table 9
Statistical parameters of slope II.

Variable Distribution Mean ðlÞ Standard deviation ðrÞ Lowe bound Upper bound
wp (deg.) Normal 20 2.4 – –
/ (deg.) Normal 20 2.7 – –
c (kPa) Triangular 80 – 40 130
zw
z (%)
Triangular 50 – 0 100
 M. Ahmadabadi, R. Poisel / Computers and Geotechnics 69 (2015) 540–550 547

Table 10
Comparison of the realization points that are computed by Harr’s and modified Harr’s PEMs in the probabilistic stability analyses of slope II.

Realization points Harr’s PEM Modified Harr’s PEM

zw zw
wp (deg.) / (deg.) c (kPa) z (%) wp (deg.) / (deg.) c (kPa) z (%)

1 24.8 20 83.33 50 24.8 20 82.56 50

2 15.2 20 83.33 50 15.2 20 82.56 50
3 20 25.4 83.33 50 20 25.4 82.56 50
4 20 14.6 83.33 50 20 14.6 82.56 50
5 20 20 120.15 50 20 20 119.88 50
6 20 20 46.51 50 20 20 49.05 50
7 20 20 83.33 90.8 20 20 82.56 89.33
8 20 20 83.33 9.20 20 20 82.56 10.66

Fig. 7. Distinct element model of slope II with UDEC.

Table 11
Results of the probabilistic analyses of slope II using PEMs.

Probabilistic tool No. model evaluations Safety factor Model output distribution
Normal Log-normal
Mean (l) Std. ðrÞ Safety index ðbÞ P f (%) P f (%)

Rosenblueth 16 1.3581 0.2561 1.3983 8.10 6.13

Harr 8 1.3604 0.2463 1.4633 7.17 5.22
Hong 2n scheme 8 1.3583 0.2513 1.4258 7.70 5.73
Hong 2n + 1 scheme 9 1.3590 0.2524 1.4223 7.75 5.78
Hong 3n scheme 12 1.3535 0.2462 1.4358 7.56 5.62
Hong 4n + 1 scheme 17 1.3364 0.2531 1.3291 9.20 7.34
548 M. Ahmadabadi, R. Poisel / Computers and Geotechnics 69 (2015) 540–550

Table 12 probabilistic stability calculations. Again, note that to interpret

Results of the probabilistic stability analyses of slope II with MCS. the results of probabilistic analyses by means of PEMs in terms
Numerical code No. simulations Safety factor P f (%) of probability of failure, an appropriate distribution of the model
Mean (l) Std. ðrÞ output (safety factor) should be considered. In this example,
because two random variables have non-normal distributions, it
RocPlane [18,38] 10,000 1.3559 0.2558 6.26
might be thought that they affect the distribution of the safety fac-
UDEC 1000 1.3797 0.2563 4.80
tor; as a result, a lognormal distribution of the safety factor is also
6000 1.3435 0.2260 5.60
included in the analyses.
As in the first example, the mean and standard deviation of the Note that probability of failure for lognormal distribution is cal-
safety factor and the safety index ðbÞ and the probability of failure culated using Eq. (11).
ðP f Þ are calculated using three PEMs, the results of which are sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2 ffi
included in Table 11. It is remarkable that although the two vari-
l2Fos lFos þ r2Fos
mu ¼ ln qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; sigma ¼ ln
ables are non-normal, all of the methods produce different realiza- l2Fos þ r2Fos
2
lFos
tion points, and they should be considered separately in the

700 140
No. of simulations No. of simulations
Ns=6000 Ns=1000
600 120

500 100
Frequency

400 Frequency 80

300 60

200 40

100 20

0 0
0.6 0.8 1 1.2 1.4 1.6 1.8 2 0.8 1 1.2 1.4 1.6 1.8 2 2.2

Safety factor Safety factor

Fig. 8. Histograms of the safety factor for slope II with MCS: left graph N s1 ¼ 6000, right graph N s2 ¼ 1000.

Table 13
qc;/
Variation of F ¼ q 0 for different quantities of qc;/ in the probabilistic stability analysis of slope II.
c;/

qc;/ qc;/0 qc;/ qc;/ qc;/0 qc;/

F¼ q 0 F¼ q 0
c;/ c;/

0.10 0.1005 1.0050 0.60 0.6026 1.0043

0.20 0.2009 1.0045 0.70 0.7030 1.0043
0.30 0.3013 1.0043 0.80 0.8035 1.0044
0.40 0.4018 1.0045 0.90 0.9039 1.0043
0.50 0.5022 1.0044 1.0 1.0043 1.0043

Table 14
Comparison of the realization points that are produced by the original and the modified Hong’s PEM (3n scheme) in the probabilistic stability analysis of slope II.

Realization points Original Hong’s PEM ðqc:/ ¼ 0Þ Modified Hong’s PEM (qc:/ ¼ 0:50)

wp (deg.) / (deg.) c (kPa) zw (%) wp (deg.) / (deg.) c (kPa) zw (%)

1 20 20 83.33 50 20 20 83.23 50
2 24.16 20 83.33 50 24.16 20 83.23 50
3 15.84 20 83.33 50 15.84 20 83.23 50
4 20 20 83.33 50 20 20 83.23 50
5 20 24.67 83.33 50 20 24.67 67.03 50
6 20 15.32 83.33 50 20 15.32 101.24 50
7 20 20 113.46 50 20 20 59.90 50
8 20 20 56.81 50 20 20 79.04 50
9 20 20 79.04 50 20 20 83.23 50
10 20 20 83.33 50 20 20 83.23 50
11 20 20 83.33 81.60 20 20 83.23 81.60
12 20 20 83.33 18.39 20 20 83.23 18.39
 M. Ahmadabadi, R. Poisel / Computers and Geotechnics 69 (2015) 540–550 549

/ are negatively correlated. A series of correlation coefficients,

2 Rosenblueth's PEM
Harr's PEM qc;/ ¼ 0 1:0; is used to model the common shear test results,
Hong 2n scheme
1.9 in which the cohesion generally decreases as the friction angle
Hong 2n+1 scheme
Hong 3n scheme increases, and vice versa [25,26]. The Nataf transformation is used
1.8
Safety Index (β)

Hong 4n+1 scheme to consider the correlation where the reliability analyses are per-
Monte Carlo Simulation.
1.7 formed by means of Hong’s method. Note that, although cohesion,
c, has a triangular distribution, because its distribution is approxi-
1.6 mately symmetric about the mean value, the coefficient factor
q0ij
1.5 F¼ qij , as it is described by [21,24], would be equal to 1.0. This
matter is readily proved by solving Eq. (5), the results of which
1.4
are given in Table 13. According to this table, when the random
1.3 variables are transformed to the normal space, the variation in cor-
0 -0.1 -0.2 -0.3 -0.4 -0.5 -0.6 -0.7 -0.8 -0.9 -1 relation coefficient ðq0ij  qij Þ is less than 0.50%.
Correlation coefficient (ρc,φ) The table below, Table 14, indicates a comparison between the
original Hong’s PEM and the modified Hong’s PEM in terms of the
Fig. 9. Variation of the reliability index versus the correlation coefficient in the
probabilistic stability analyses of slope II.
realization points that are used for the model evaluations. The
importance of this manipulation by means of the NATAF transfor-
mation is better appreciated by examining Figs. 9 and 10, in which
the variation of the safety index, b, and the probability of failure, Pf ,
8.2 versus the correlation coefficient are shown, respectively.
Rosenblueth's PEM
It is evident from the figures above that the results obtained
Harr's PEM
Probability of failure (Pf ) %

7.2 Hong 2n scheme using the different methods are not much different, demonstrating
Hong 2n+1 scheme the accuracy of proposed solution by means of the Nataf transfor-
Hong 3n scheme
6.2
Hong 4n+1 scheme mation. In these figures, the Monte Carlo Simulations are per-
Monte Carlo Simulation formed using [33] a sample size of n ¼ 100; 000. Evidently, by
5.2 considering the correlation coefficient between cohesion c and
the internal friction angle of rock joints /, the safety index, b,
4.2
and the probability of failure, P f , change considerably. In addition,
in this example, disregarding the correlation coefficient between
3.2
cohesion c and internal friction angle of rock joints / clearly leads
to a conservative geotechnical design because the probability of
2.2
-0.05 -0.15 -0.25 -0.35 -0.45 -0.55 -0.65 -0.75 -0.85 -0.95 failure decreases as the correlation coefficient increases.
Correlation coefficient (ρc,φ)
5. Summary and conclusions
Fig. 10. Variation of the probability of failure versus the correlation coefficient in
the probabilistic stability analyses of slope II.
Application of point estimate methods in the reliability analysis
of slopes was examined. Three PEMs were investigated using two
Pf Lognormal ¼ logncdf ð1; mu; sigmaÞ ð11Þ examples, which include two slopes with two modes of failure.
The following conclusions can be made from the results of this
where logncdf ðÞ denotes the cumulative distribution function corre- study.
sponding to a lognormal distribution, which can be calculated using
MATLAB code. 1. Although Rosenblueth’s PEM is applicable for both skewed and
To estimate the accuracy of the results of the PEMs, the results asymmetric variables, it may be infeasible due to the computa-
of three probabilistic analyses based on MCS are given in Table 12. tional effort required when the number of random parameters
These analyses are conducted with different sample sizes and with is large.
different numerical methods. The first results are those of Wyllie & 2. In problems dealing with normally distributed random vari-
Mah [41], in which the stability of the discussed slope is analysed ables and when the number of random variables is not large,
using an analytical method [33]. The other two Monte Carlo Harr’s algorithm is not only very efficient but also provides suf-
Simulations are performed with UDEC for N s1 ¼ 1000 and ficiently accurate results. For problems involving random
N s2 ¼ 6000 simulations. Fig. 8 exhibits a comparison between the parameters with a non-normal distribution, the modified
histograms of the safety factor based on the results of Monte Harr’s PEM is proposed as an alternative. In the second example,
Carlo Simulations with UDEC. the asymmetry of the variables changes the probability of fail-
A comparison of the results of Monte Carlo Simulations and the ure from 7.0% to 7.17%. However, undoubtedly, when
results calculated by the PEMs proves the accuracy of the PEMs; non-normal variables play an important role in the perfor-
furthermore, it illustrates that assuming a log-normal distribution mance function, considering the asymmetry of non-normal
for the safety factor is more accurate. variables may drastically change the results.
3. In contrast to Rosenblueth’s and Harr’s methods, Hong’s PEM is
4.2.5. Correlation of the random variables capable of theoretically considering more than two realization
As mentioned in Section 2, Hong’s approach was originally points for each random parameter. In the case that all variables
established for uncorrelated random variables. This section repre- are uncorrelated and normally distributed, Harr’s PEM and
sents how to treat correlated random parameters when using Hong’s 2n scheme, as well as Hong’s 3n and 2n þ 1 schemes,
Hong’s method; moreover, the effect of correlation coefficient vari- are the same.
ation on reliability index, b, and probability of failure, pf , is studied. 4. When Hong’s PEM is utilized for problems involving correlated
Among the four uncertain parameters (see Table 9), it is variables, the Nataf transformation is proposed to resolve the
assumed that cohesion c and internal friction angle of rock joints drawbacks of Hong’s PEM with respect to the correlated
550 M. Ahmadabadi, R. Poisel / Computers and Geotechnics 69 (2015) 540–550

variables. In the second example, the probability of failure is [7] Gibson W. Probabilistic methods for slope analysis and design. Austral
Geomech 2011;46(3):29–40.
observed to vary from Pf ¼ 8:0% to P f ¼ 2:8% when the correla-
[12] Harr ME. Probabilistic estimates for multivariate analyses. Appl Math Model
tion coefficient qc;£ varies from 0 to 1. 1989;13(5):313–8.
5. The more accurate is assumed the distribution of performance [13] Hasofer AM, Lind NC. An exact and invariant first-order reliability format. J Eng
Mech, ASCE 1974;100(1):111–21.
function, the more precise are the results obtained by the [14] He J, Sällfors G. An optimal point estimate method for uncertainty studies.
PEMs; as a result, if an analytical or approximate solution of Appl Math Model 1994;18(9):494–9.
the examined problem is possible, running the MCS is recom- [15] Hong HP. An efficient point estimate method for probabilistic analysis. Reliab
Eng Syst Saf 1998;59(3):261–7.
mended to achieve a better estimation of the distribution of [16] ITASCA consulting group Inc. Fast lagrangian analysis of continua (FLAC)
the performance function. ver.7.0: two-dimensional explicit finite difference program for engineering
mechanics computation; 2015.
[17] ITASCA consulting group Inc. Universal distinct element code (UDEC) ver.5.0:
advanced, two-dimensional distinct element modeling for geotechnical
Acknowledgements analysis of rock, soil and structural support; 2015.
[18] Jiang S, Li D. Reliability analysis of unsaturated slope with spatially correlated
The authors thank the anonymous reviewers for their construc- soil properties. In: Vulnerability, uncertainty, and risk. American Society of
Civil Engineers; 2014. p. 2429–38.
tive comments, which improved this paper. [21] Lemaire M. Structural reliability. John Wiley & Sons; 2009.
The authors acknowledge the funding of this work by the [23] Lind NC. Modelling of uncertainty in discrete dynamical systems. Appl Math
Austrian Society for Geomechanics. Model 1983;7(3):146–52.
[24] Liu P-L, Der Kiureghian A. Multivariate distribution models with prescribed
The first author is grateful to Professors A. Rohatsch and A. Preh marginals and covariances. Probab Eng Mech 1986;1(2):105–12.
for their ongoing supports, advice and suggestions. [25] Low BK. Reliability analysis of rock slopes involving correlated nonnormals. Int
J Rock Mech Min Sci 2007;44(6):922–35.
[26] Low BK. Efficient probabilistic algorithm illustrated for a rock slope. Rock
Appendix A. ‘‘MATLAB code for solving the Eq. (5)’’ Mech Rock Eng 2008;41(5):715–34.
[27] Nataf A. Determination des distributions de probabilites dont les marges sont
donnes. Comptes Renud de l’Academie des Sci 1962;225(5):42–3.
[28] Panchalingam G, Harr ME. Modelling of many correlated and skewed random
variables. Appl Math Model 1994;18(11):635–40.
clc; [31] Rocscience Inc. Probabilistic slope stability analysis. Available: <https://www.
rocscience.com/help/phase2/webhelp/pdf_files/tutorials/Tutorial_32_
% Input information of random variables Probabilistic_Slope_Stability_Analysis.pdf>.
mu1 = 20; std1 = 2.70; [32] Rocscience Inc. Phase 2 ver.8.0: finite element analysis for excavations and
a2 = 40; mu2 = 80; b2 = 130; slopes; 2015.
[33] Rocscience Inc. RocPlane ver.3.0: planar sliding stability analysis for rock
delta = 0.0001;
slopes; 2015.
R = 0.90; [34] Rocscience Inc. Slide ver.6.0: 2D limit equilibrium slope stability analysis;
pd1 = makedist(‘normal’,’mu’,mu1,’sigma’,std1); 2015.
std1 = std(pd1); mu1 = mean(pd1); [35] Rosenblueth E. Point estimates for probability moments. Proc Natl Acad Sci
USA 1975;72(10):3812–4.
pd2 = makedist(‘triangular’, ’a’, a2, ’b’, mu2, ’c’, [36] Rosenblueth E. Two-point estimates in probabilities. Appl Math Model
b2); 1981;5(5):329–35.
std2 = std(pd2); mu2 = mean(pd2); [37] Rubinstein RY, Kroese DP. Simulation and the Monte Carlo method. 2nd
ed. Wiley; 2008.
% This array includes the outputs [38] Shamekhi E, Tannant DD. Probabilistic assessment of rock slope stability using
count = 0; response surfaces determined from finite element models of geometric
a1 = 7; a2 = 7;% Integral Limits realizations. Comput Geotech 2015;69:70–81.
[39] Matsui T, San K-C. Finite element slope stability analysis by shear strength
q = 1; reduction technique. Soils Found 1992;32(1):59–70.
while q > 1e5 [41] Wyllie DC, Mah C. Rock slope engineering. 4th ed. Taylor & Francis; 2004.
count = count + 1; [42] Zhao Y-G, Ono T. New point estimates for probability moments. J Eng Mech
2000;126(4):433–6.
R0 = R + delta ⁄ count [43] Zhou J, Nowak AS. Integration formulas to evaluate functions of random
fun = @(z1, z2) (((icdf(pd1,normcdf(z1,0,1)))m variables. Struct Saf 1988;5(4):267–84.
u1)./std1).⁄ ...
(((icdf(pd2, normcdf(z2, 0,1)))mu2)./std2).⁄ ... Further reading
1./(2 ⁄ pi. ⁄ sqrt(1R0^2)).⁄ ...
exp(1 ⁄ (z1.^2 + z2.^2-2. ⁄ R0. ⁄ z1. ⁄ z2)./... [5] Christian JT, Baeche GB. Point estimate method as numerical quadrature.
Geotech Geoenviron Eng 1999;125(9):779–86.
(2. ⁄ (1R0.^2))); [8] Griffiths DV, Huang J, Fenton GA. Probabilistic infinite slope analysis. Comput
q = integral2(fun, a1, a1, a2, a2)R Geotech 2011;38(4):577–84.
end [9] Griffiths VD, Fenton GA. Probabilistic methods in geotechnical
engineering. Springer; 2007.
F = R0/R [10] Haldar A, Mahadevan S. Probability, reliability, and statistical methods in
engineering design. John Wiley; 2000.
[11] Haldar A, Mahadevan S. Reliability assessment using stochastic finite element
analysis. Wiley; 2000.
[19] Kim JM, Sitar N. Reliability approach to slope stability analysis with spatially
References correlated soilproperties. Soils Found 2013;53(1):1–10.
[20] Le TMH. Reliability of heterogeneous slopes with cross-correlated shear
[1] Baecher GB, Christian JT. Reliability and statistics in geotechnical strength parameters. In: Geotechnical safety and risk IV. CRC Press; 2013. p.
engineering. Wiley; 2003. 161–7.
[2] Breitung K, Hohenbichler M. Asymptotic approximations for multivariate [22] Li D, Chen Y, Lu W, Zhou C. Stochastic response surface method for reliability
integrals with an application to multinormal probabilities. J Multivariate Anal analysis of rock slopes involving correlated non-normal variables. Comput
1989;30(1):80–97. Geotech 2011;38(1):58–68.
[3] Caramia P, Carpinelli G, Varilone P. Point estimate schemes for probabilistic [29] Phoon KK. Reliability-based design in geotechnical engineering: computations
three-phase load flow. Electr Power Syst Res 2010;80(2):168–75. and applications. Taylor & Francis; 2008.
[4] Che-Hao C, Yeou-Koung T, Jinn-Chuang Y. Evaluation of probability point [30] Reale C, Xue J, Pan Z, Gavin K. Deterministic and probabilistic multi-modal
estimate methods. Appl Math Model 1995;19(2):95–105. analysis of slope stability. Comput Geotech 2015;66:172–9.
[6] Christian JT, Baecher GB. The point-estimate method with large numbers of [40] Wang Y, Tung Y-K. Improved probabilistic point estimation schemes for
variables. Int J Numer Anal Meth Geomech 2002;26(15):1515–29. uncertainty analysis. Appl Math Model 2009;33(2):1042–57.