Simulation Modelling Practice and Theory 100 (2020) 102061

Contents lists available at ScienceDirect

Simulation Modelling Practice and Theory

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

Comparative Approaches to Probabilistic Finite Element Methods

T
for Slope Stability Analysis
Ashley P. Dysona, Ali Tolooiyanb,
⁎

a
Geotechnical and Hydrogeological Engineering Research Group (GHERG), Federation University, Australia
b
Geotechnical Engineering, School of Engineering, University of Tasmania, Hobart, TAS 7001, Australia

ARTICLE INFO ABSTRACT

Keywords: Probabilistic slope stability analyses are often preferable to deterministic methods when soils are
Random Finite Element Method inherently heterogeneous, or the reliability of geotechnical parameters is largely unknown. These
Monte Carlo simulation methods are suitable for evaluating the risk of slope failure by producing a range of potential
point estimate method scenarios for the slope stability factor of safety. Several probabilistic methods including the Point
spatial variation
Estimate Method, Monte Carlo Method and Random Finite Element Method, can be combined
slope stability analysis
with the Finite Element technique. In this study, various shear strength distributions are con-
sidered for three different probabilistic Finite Element Methods to determine Factor of Safety and
Probability of Failure distributions, based on the associated method of slope stability analysis.
Results are presented for a case study of an Australian open-cut coal mine, with a range of shear
strength parameter distributions for coal and interseam cohesive materials considered. Coal and
interseam shear strength parameters are varied independently, to determine the effects of each
material on the slope Factor of Safety.

1. Introduction

Probabilistic methods are an important aspect of risk assessment for slope stability of open-cut mines. In many cases, probabilistic
methods are able to capture soil variability and uncertainty when a deterministic analogue is deemed inappropriate. Soils often
exhibit highly heterogeneous properties which can have a profound impact on the geotechnical behaviour of large geostructures such
as slopes and embankments [1–3]. Uncertainty and heterogeneity are particularly important in regions displaying appreciable ani-
sotropy or weakened materials [4–6]. Even when a range of material parameters are available, inaccuracies due to measurement and
laboratory conditions compared with in-situ field-scale behaviour can produce significant errors [7]. Hence, probabilistic slope
stability methods have become an increasingly prevalent tool for geotechnical analysis [1, 8–12]. A range of applicable probabilistic
methods exist, including First Order Second Moment (FOSM) [8, 10, 13–15], the Point Estimate Method (PEM) [16-21], the Modified
Point Estimate Method (MPEM) [22], Monte Carlo Simulation (MCM) [23–25], and the Random Finite Element Method (RFEM) [26].
RFEM differs from FOSM, PEM and MCM by including the effects of spatial variability, incorporating a distance-based correlation
length. In certain cases, probabilistic methods may be favourable to deterministic analyses. However, a common drawback is the
often excessive computational resources required [27]. Griffiths and Fenton noted that one thousand simulations were necessary to
achieve a Factor of Safety (FoS) convergence for a 2:1 undrained clay slope [28]. For this reason, prediction and optimisation
techniques are sometimes required to investigate large numerical models with high levels of complexity [29, 30]. Two of the most
common slope stability analysis methods for incorporating material heterogeneity are Finite Element Method (FEM) [28, 31] and the

⁎
Corresponding author.
E-mail addresses: a.dyson@federation.edu.au (A.P. Dyson), ali.tolooiyan@utas.edu.au (A. Tolooiyan).

https://doi.org/10.1016/j.simpat.2019.102061
Received 28 October 2019; Received in revised form 18 December 2019; Accepted 19 December 2019
Available online 23 December 2019
1569-190X/ © 2019 Elsevier B.V. All rights reserved.
A.P. Dyson and A. Tolooiyan Simulation Modelling Practice and Theory 100 (2020) 102061

Nomenclature and probabilistic

ρ(τ) autocorrelation function
Symbols σ1 major principal stress
σ3 minor principal stress
c cohesion τ separation distance
d normal standard deviate for a desired confidence τx separation distance parallel to the x axis
level with Monte Carlo simulation τy separation distance parallel to the y axis
E elastic modulus φ friction angle
F force ψ dilation angle
G(x) standard normal random field ACF Autocorrelation function
k permeability CoV Coefficient of Variation
K random field cluster FEM Finite Element Method
m number of random variables in Monte Carlo si- FoS Factor of Safety
mulation LAS Local Average Subdivision
s standard deviation LEM Limit Equilibrium Method
t(x) deterministic term of soil variability model MC Mohr-Coulomb
w(x) fluctuation term for soil variability model MCM Monte Carlo Method
Xi sequence of independent lognormal random vari- PDF Probability Density Function
ables PEM Point Estimate Method
γ unit weight PMF Probability Mass Function
θ scale of fluctuation PoF Probability of Failure
θ1 major principal scale of fluctuation RFEM Random Finite Element Method
θ2 minor principal scale of fluctuation RoS Regions of Significance
θx scale of fluctuation parallel to the x axis RSS Representative Slip Surface
θy scale of fluctuation parallel to the y axis SoF Scale of Fluctuation
μ mean value of a geotechnical parameter SRF Strength Reduction Factor
ν Poisson's ratio SRM Strength Reduction Method
ξ(x) soil variability model, consisting of a deterministic

Limit Equilibrium Method (LEM) [32, 33]. The Finite Element Method can be used in combination with the Point Estimate Method,
(known as the Finite Element Method – Point Estimate Method), Monte Carlo simulation, and the Random Finite Element Method
(RFEM).
In this study, the MCM, PEM and RFEM methods are compared for a range of distributions to determine slope stability safety
factors, thereby highlighting the robustness of each method for varying distributions. A case study of the Yallourn open-cut brown
coal mine in Victoria, Australia is presented, with a focus on determining factors of safety for varying probabilistic methods, when
presented with varying distributions of input shear strength parameters. Furthermore, analysis of the variation of each material
distribution considered independently, allows for a greater understanding of the impacts of each material layer.

2. Probabilistic finite element methods

2.1. Point Estimate Method (PEM)

The Point Estimate Method, as developed by Rosenblueth [19], is a multivariate technique for estimating the expected value of a
performance function (e.g. the Factor of Safety). In the PEM, the continuous Probability Density Functions (PDFs) of input random
variables are replaced by sets of discrete points by evaluating PDFs at locations determined by the standard deviation and Pearson's
coefficient of skewness of the input parameters. Thus a two-point Probability Mass Function (PMF) is produced. These point values
are then weighted to provide an estimate of the first two moments of the performance function. Given a performance function y of n
variables (Eq. 1), each variable is evaluated at two points x i + and x i (Eq. 2) [34]:
y = f (x ) = f (x1, x2, …, xn) (1)

x i + = x¯i + xi +. xi, x i = x¯i + xi xi x i = x¯i + xi xi (2)

where, x̄ i is the mean of the distribution of xi; σxi the standard deviation; and xi +and xi are so-called standard deviation units,
derived from the Pearson skewness coefficient νxi, as shown by Eq. 3. In the case where symmetric distributions are considered, both
xi + and xi are equal to unity. When considering a single random variable, the weights (named probability concentrations) are
determined for each point estimate (Eq. 4). For uncorrelated variables, the weights i and j, are given by Eq. 5. When the random
variable distribution is symmetric, both weights are equal to 0.5. Finally, the Factor of Safety performance function mean and

2
A.P. Dyson and A. Tolooiyan Simulation Modelling Practice and Theory 100 (2020) 102061

standard deviation can be evaluated by Equations 6 and 7, respectively [34].

2 2
xi xi
xi + = + 1+ , xi = xi + xi
2 2 (3)

xi
Px+ = , Px = 1 Px +
xi + + xi (4)

Pi, j = Pxi. Pxj (5)

where, Pxi and Pxj are evaluated as the weights of a single random variable. For each variable, there are two points evaluated,
meaning y is evaluated for a total number of 2n combinations. Although the PEM is a simple procedure, the number of required
computations grows exponentially with the number of input variables, often beyond what is reasonable for practical purposes [27].
2n
FoSµ = Pi FoSi
i=1 (6)

2n
FoS = Pi (FoSi FoSµ )2
i=1 (7)

2.2. Monte Carlo Method (MCM)

The Monte Carlo Method, when coupled with Finite Element Method simulation, can be a powerful tool for probabilistic slope
stability analysis. By sampling a set of material properties from joint PDFs, a Factor of Safety distribution can be determined. The
general iterative procedure of a Monte Carlo method for Finite Element Method slope stability is covered in the four steps described
below:

1) Estimate the probability distribution for each input parameter for a given soil layer/region;
2) Generate random values for each parameter;
3) Finite Element Method slope stability simulation with the Strength Reduction Method, using parameters from (2);
4) Repeat the process N times (N > 100) to determine the converged statistical properties of interest.

The Probability of Failure (PoF) for a given slope is given by Eq. 8 [27], where M is the number of slope safety factors less than
1.0.
N M
PoF =
N (8)

2.3. Random Finite Element Method (RFEM)

One of the principal sources of material heterogeneity is the inherent spatial variability of soils due to their stress history and
depositional nature [35]. As such, common statistical parameters such as the mean and standard deviations do not capture the
relationships of material properties based on their spatial structure [36]. A common method for describing the correlation structure at
given locations is provided by random field theory initially developed by Vanmarcke [37]. An important parameter for describing
spatial structure is the autocorrelation function, which provides a correlation measure as a function of distance. Although many
correlation functions exist, the most common two-dimensional correlation function is the Markovian exponentially decaying cor-
relation function [38]:

2 x 2 y
( x, y) = exp
x y (9)

Fig. 1. Slope geometry dimensions and mesh distribution.

3
 A.P. Dyson and A. Tolooiyan

Table 1
Comparative Significance of Probability of Failure [53].
Probability of Failure (%) Design Criteria on Basis of which Probability of Failure is Established Aspects of Natural Situation in Terms of which Probability of Failure can be
Assessed
Serviceable Life Public Liability Minimum Surveillance Required Frequency of Evident Slope Rate of Evidently Unstable Movements
Failures

50 to 100 Effectively zero Public access is forbidden Serves no purpose (excessive probability Slope failures generally evident Abundant evidence of creeping valley
tantamount to failure) sides

4
20 to 50 Very short term (temporary open Public access forcibly Continuous monitoring with intensive A significant number of unstable Clear evidence of creeping valley sides
pit mines) prevented sophisticated instruments slope works
10 to 20 Very short term Public access actively Continuous monitoring with sophisticated Somewhat unstable slopes Some evidence of slowly creeping valley
prevented instruments evident sides
5 to 10 Short term Public access prevented Conscious superficial monitoring No read evidence of unstable Extremely slowly creeping valley sides not
slopes readily evident
0.5 to 5 Medium-term (semi-permanent Public access allowed Incidental superficial monitoring No unstable slopes evident No unstable movements evident
slopes)
Less than 0.5 Very long term (permanent) Public access free No monitoring required Stable slopes No movements
Simulation Modelling Practice and Theory 100 (2020) 102061
A.P. Dyson and A. Tolooiyan Simulation Modelling Practice and Theory 100 (2020) 102061

Fig. 2. A cluster dendrogram of groups of similar random field realisations, associating random field instances based on levels of similarity. The
strong level of similarity of instances 12 and 23 is highlighted in this case.

Fig. 3. RFEM analysis procedure, as described by Dyson and Tolooiyan [30].

5
A.P. Dyson and A. Tolooiyan Simulation Modelling Practice and Theory 100 (2020) 102061

Fig. 4. Location of the Yallourn open-cut mine.

Table 2
Yallourn brown coal properties [56].
Property Value

Rank ASTM lignite B

moisture content (% wb) 58.0
ash yield (% db) 1.7
volatile matter content (% bd) 50.3
Fixed carbon (% db) 48.0
sulphur content (% db) 0.28
Gross dry specific energy (MJ/kg) 26.1
net wet specific energy (MJ/kg) 6.9

Table 3
Geotechnical parameters for coal and non-coal materials.
Material γunsat (kN/m3) γsat (kN/m3) E (MPa) ν k (m/day)

Coal 11.4 11.5 40 0.2 7.6e-3

Non-coal 20.15 20.2 52 0.3 4.7e-2

Table 4
Shear strength parameter statistics.
Coal Non-coal
Parameter c (kPa) φ (°) c (kPa) φ (°)

Mean (μ) 150.72 27.28 31.81 23.67

Standard Deviation (s) 69.78 4.91 5.00 4.51
COV (%) 46 18 16 19

where, ρ is the autocorrelation coefficient, defined at lag distances τx and τy, in the x and y directions, respectively. Similarly, θx and
θy are the scales of fluctuation (SoF) for both x and y. The scale of fluctuation is defined as the measure of distance within which
points are significantly correlated [39]. Li et al. noted that the sensitivity to slope failure was insensitive to the type of autocorrelation

6
A.P. Dyson and A. Tolooiyan Simulation Modelling Practice and Theory 100 (2020) 102061

Table 5
Common Coefficient of Variation shear strengths [58].
Parameter Coefficient of Variation (%)

Cohesion 20 – 80
Friction angle 4 – 20

Table 6
Shear strength length scales.
Coal Non-coal
Scale of Fluctuation (m) c φ c φ

θx 320.3 772.4 812 627.8

θy 23.1 8.7 35 13.1

Table 7
Shear strength correlations.
Coal Non-coal
c φ C φ

Correlation with depth 0.129 -0.260 -0.027 -0.060

rc,φ 0.046 -0.003

function [40]. Often, variables must be considered simultaneously when variables are correlated. The covariance function Cov[X,Y]
describes the strength of the relationship between the random variables X and Y as:
Cov [X , Y ] = E [XY ] E [X ] E [Y ] = E [XY ] µX µ Y (10)

A range of slope stability parameters may be considered for random field generation. However, the most frequently implemented
soil parameters are material cohesion c, internal friction angle φ, dilation angle ψ, elastic modulus E, unit weight γ, and permeability
k. Alonso noted that the variation of material strength parameters c and φ are a significant contributing factor in the Factor of Safety
distributions compared with the variation of soil unit weight [8]. Furthermore, Griffiths and Fenton observed that the variation of
cohesion, friction angle and unit weight provided the greatest impact on safety factors when using random finite element methods
[28].
In many cases, geotechnical parameters follow lognormal distributions, especially in cases where parameters require strictly non-
negative values, e.g. cohesion, unit weight, etc. The lognormal random field X is specified by a transformation from a standard
Gaussian random field G(x) by
Xi = exp[µlnX + slnX G (x i )] (11)
th th
where, Xi is the lognormal parameter for the i random field element; xi is the centre point of the i element; and μlnX and slnX are the
mean and standard deviation of the normal distribution ln X, respectively. Both mean μlnX and standard deviation slnX parameters of
the distribution are determined by the following transforms:
slnX 2
µlnX = ln(µ )
2 (12)

s2
slnX 2 = ln 1 +
µ2 (13)

while the lognormal PDF is defined as:

2
1 1 ln X µln X
f (X ) = exp
X s ln X 2 2 s ln X (14)

In many cases, material parameter distributions may be linked or jointly influenced by another parameter. In particular, weak
negative cross-correlations between cohesion and the internal friction angle of materials have been observed. Cherubini noted
correlations in the range of -0.7 to -0.24 [41] while Yucumen et al. noted values from -0.49 to -0.24 [42].
A common method of random field generation for a Finite Element geometry is the Local Average Subdivision (LAS method) [43].
Global averages are subdivided into local regions with the constraint that the local averages preserve the global statistical properties.
LAS leads to a variance reduction, as fluctuations in the parameters tend to cancel each other out when averaged across the spatial
region. As such, the variance of a LAS region is commonly less than the variance of a point field and requires a dimensionless variance
adjustment to rectify the cell-to-cell variance [44].

7
A.P. Dyson and A. Tolooiyan Simulation Modelling Practice and Theory 100 (2020) 102061

Fig. 5. (a) Yallourn North-East batter with 2D cross-section location (b) unexcavated cross-section geometry (c) excavated cross-section geometry.

Table 8
Number of realisations required per mean/CoV parameter.
Method Number of simulations Mean computation time per simulation instance (mins) Total computation time (hours)

PEM 16 12 3.2
MCM 100 12 20
RFEM 100 36.5 60.8

2.4. Finite Element Method (FEM) for slope stability

The Finite Element Method has become one of the prevailing tools for numerical analysis of slope stability, requiring few a priori
assumptions, especially regarding the slope failure mechanism [45, 46]. FEM slope failure develops “naturally” in zones where the
soil shear strength is unable to provide sufficient support to resist the shear stresses [45]. Unlike Limit Equilibrium Methods, FEM
does not incorporate the concept of slices and requires no assumptions of inter-slice forces and the shape of the potential slip surface.
FEM is capable of progressive failure simulation up to and including overall shear failure, detailing information regarding de-
formations at working stress levels.
In this study, FEM slope stability analysis is performed with a two-dimensional plane strain, linear elastic model combined with

8
A.P. Dyson and A. Tolooiyan Simulation Modelling Practice and Theory 100 (2020) 102061

Table 9
List of distribution mean parameters tested.
Mean value multiplier Coal Non-coal interseam
Cohesion (kPa) Friction angle (°) Cohesion (kPa) Friction angle (°)

1.2 180.8 32.7 38.1 28.4

1.1 165.7 30.0 34.9 26.0
1 150.72 27.2 31.8 23.6
0.9 135.6 24.5 28.6 21.3
0.8 120.5 21.8 25.4 18.9

Table 10
List of Coefficient of Variation shear strength parameters for each material ranging from ± 50% of the initial parameters tested.
CoV multiplier Coal Non-coal interseam
Cohesion Friction angle Cohesion Friction angle

1.5 0.69 0.27 0.24 0.285

1.25 0.58 0.125 0.20 0.237
1 0.46 0.18 0.16 0.190
0.75 0.35 0.135 0.12 0.142
0.5 0.23 0.09 0.09 0.095

the Mohr-Coulomb (MC) perfectly plastic failure criterion. Considering the necessary mesh size and number of probabilistic instances
required, also the fact that the possible three-dimensional geometric effects are not among the objectives of this study, two-di-
mensional plane strain analyses were implemented. Furthermore, an elastic model with the Mohr-Coulomb failure criterion was
considered as an appropriate model based on material behaviour determined from in-house laboratory tests on coal and interseam
materials. Due to gravity loading of the slope materials, forces are generated, creating normal and shear stresses. In terms of the
principal stresses with the convention of a compression-negative sign, the stresses are compared with the MC criterion can be defined
as follows:

1 + 3 1 3
F= sin c cos
2 2 (15)
where, 1 and 3 are the major and minor principal effective stresses, respectively; c′and ϕ′ are the effective cohesion and internal
friction angle parameters, respectively; and F is the failure function which is interpreted as:

F < 0 elastic (stresses inside the failure envelope)

F = 0 yielding (stresses are on the failure envelope)
F > 0 yielding but must be redistributed (stresses outside the failure envelope)

In the case of RFEM, the FEM geometry is partitioned into distinct element sets, where each element is allocated a set to material
shear strength parameters (c and ϕ) from an associated random field distribution. The Abaqus element type CPEP (4-node plane strain
quadrilateral, with bilinear displacement, bilinear pore pressure) was implemented for all finite elements within the slope model
geometry. In the case of RFEM modelling, complexities associated with the spatially variable shear strength parameters require a
particularly fine mesh distribution (in the case of this research, 40,000 elements) as depicted in Fig. 1. Due to the magnitude of the
fluctuations in shear strength parameters, elements must be sufficiently fine to accommodate the spatial variation detailed by the
random fields. As such, the dimensions of each element cannot exceed the half the material correlation length in both “across the
bore” and “down the bore” directions. For this reason, coupled with the mesh required to model the variation of the stratigraphy and
surface geometry, a fine mesh is required to consider the spatial variation of shear strength parameters using RFEM.

2.5. Strength Reduction Method (SRM)

The stability Factor of Safety (FoS) of a slope is defined as the factor which the original slope shear strength parameters must be
reduced by in order to bring the slope to the point of failure. The Strength Reduction Method (SRM), sometimes referred to as the
Shear Strength Reduction Method, iteratively reduces shear strength parameters by a Strength Reduction Factor (SRF) until the
failure state is achieved. The method defined by Bishop [47] defines the coupled reduction of cohesion and friction angles as follows:
c0
c=
SRF (16)

1
tan 0
= tan
SRF (17)
in which c0 and ϕ0 are the initial material cohesion and friction angle parameters, respectively; c and ϕ are the reduced parameters;

9
A.P. Dyson and A. Tolooiyan Simulation Modelling Practice and Theory 100 (2020) 102061

Fig. 6. Typical random field realisations and slope slip surface (a) cohesion random field realisation illustrating long horizontal correlation lengths
(b) typical friction angle realisation illustrating minimal difference between correlation lengths between layers (c) typical slope slip surface, in-
dicating the region of deformation.

10
A.P. Dyson and A. Tolooiyan Simulation Modelling Practice and Theory 100 (2020) 102061

Fig. 7. Mean safety factors for various mean value input parameters, in each case identifying monotonically increasing FoS values with respect to
shear strength parameters. RFEM slopes are noted to fail at lower safety factors than PEM and MCM for all examples. Lowest safety factors were
observed for interseam shear strength parameters.

and SRF is the iterative Strength Reduction Factor. Once the shear strength parameters are reduced to precipitate the onset of slope
failure, the FoS is defined as by the ratio of the initial shear strength parameters with respect to the final shear strength parameters:

c0 tan 0
FoS = =
cf tan f (18)

where, cf and ϕf are the shear strength parameters required to bring the slope to the verge of failure. There are several definitions of
slope failure as discussed by Zienkiewicz and Taylor [48] including bulging of the slope line [49] and limiting of the shear stresses on
the failure surface [50]. The most common attributes for determining the critical SRF includes:

1 Development of the shear surface through plasticised elements [46].

2 Large scale slope strain.
3 Non-convergence within the FEM solver, indicating that no stress distribution can be determined that can satisfy both the MC
failure criterion and global equilibrium [51]. Non-convergence is associated with a rapid increase in nodal displacements.

Although the Strength Reduction Method is the predominant FEM technique for determining slope safety factors, all existing
versions of the Finite Element package Abaqus do not implement an inbuilt Strength Reduction Method. In this research, an optimised
SRM was considered by incorporating Fortran and Python codes in conjunction with Abaqus scripts by the procedure detailed by
Dyson and Tolooiyan [52].

2.6. FEM for MCM and PEM vs. RFEM

The Finite Element Method can be applied to many probabilistic slope stability analysis methods, including Point Estimate
Methods, Monte Carlo Simulation and Random Finite Element Methods. In this research, two categories exist – probabilistic methods

11
A.P. Dyson and A. Tolooiyan Simulation Modelling Practice and Theory 100 (2020) 102061

Fig. 8. Factor of Safety standard deviation with respect to mean value input parameters. Greatest variation of FoS values occurred for PEM
simulation, decreasing with respect to an increase in shear strength parameters.

Fig. 9. Factor of Safety convergence for PEM, MCM and RFEM techniques. While PEM requires 16 instances (24 realisations for N = 4 variables),
convergence for MCM and RFEM methods is reached prior to 100 simulation instances.

whereby shear strength parameters are sampled, then applied as properties for the whole layer, without intra-layer variation (PEM
and MCM); and probabilistic spatial variation (RFEM), where numerous values are sampled from shear strength distributions, then
applied within each layer with a given parameter correlation length. As noted by Dyson and Tolooiyan [29, 30] RFEM analysis
requires significant computational power compared to traditional Finite Element Method slope stability simulations where material
parameters are fixed throughout each soil layer. Due to the computational demands associated with RFEM, in many cases, optimi-
sation techniques are required to limit simulation time to acceptable levels. Although PEM and MCM slope stability analyses also

12
A.P. Dyson and A. Tolooiyan Simulation Modelling Practice and Theory 100 (2020) 102061

Fig. 10. Mean safety factors for ranging CoVs, exhibiting tight mean FoS bounds, declining in value for all test methods with respect to change in
shear strength parameter CoV. RFEM FoS parameters are observed as lower than PEM and MCM safety factors for all material shear strenghth
parameters. Interseam material CoVs produced lower FoS values for RFEM simulations with highly spread input shear strength distributions.

necessitate numerous simulation instances, computation times are significantly decreased, as the fine mesh requisite for RFEM is in
many cases unnecessary for PEM and MCM. In this research, the comparison of PEM, MCM and RFEM slope stability analyses
considers the impact of computational requirements, establishing the performance of each method based on the respective Factor of
Safety results.
The parameter distributions of cohesion and friction angle are identical for both the aforementioned cases and provide two
different paradigms for probabilistic analysis – the combination of layers with parameters sampled from a distribution can be thought
of as the uncertainty in selecting the “true” parameter value; while spatially varying elements sampled from shear strength dis-
tributions can be considered as potential realisations of spatially variable patterns within soils.
As such, the analysis of these two methods for a range of input distributions allows for the comparison of methods in determining
Factors of Safety (FoS) and Probabilities of Failure (PoF), while also considering how robust the results are, given different input
distributions. Given random field analysis for slope stability, regions within random fields that provide the greatest impact to slope
stability (named Regions of Significance or RoS), can be determined [29, 30]. In the study of RFEM simulation within this research,
the impact of varying shear strength distributions on the locations of Regions of Significance is considered. Table 1 highlights the
importance of the Probability of Failure when considering slope design and assessment. PEM, MCM and RFEM are all amenable to the
calculation of such a parameter, given the requirement of numerous probabilistic realisations necessary for each method.
The similarity of random fields may be considered by the element-wise comparison of random field matrices, in conjunction with
the Frobenius norm [29, 30]:

1/2
m n
D F = |aij bij |2
i=1 j=1 (19)

where, aij and bij are the elementwise components of two random field matrix instances A and B, respectively, each of size m x n.
When the random field similarity ||D||F is small, the two random field instances are considered as similar, and produce comparable
Factors of Safety. As ||D||F increases, the random fields become dissimilar, with FoS comparison no longer possible. Random field
realisations can be clustered based on their similarity values, depicting the level of association between random fields considered

13
A.P. Dyson and A. Tolooiyan Simulation Modelling Practice and Theory 100 (2020) 102061

Fig. 11. FoS standard deviation for ranging CoVs with respect to shear strength parameter CoV. PEM FoS variation is noted as drastically larger than
MCM and RFEM and functionally dependent on the level of shear strength input parameter variation.

(Fig. 2). After the classification of random field instances deemed as similar, comparisons of the element-wise random field char-
acteristics can be used to determine the regions of greatest impact on slope safety factors and slip surface geometries. Comparisons of
random field characteristics within cluster groups are determined by the total absolute difference between each realisation within
each cluster:

Ck = Mi Mj
i j K (20)
th th
where, Ck is the overall random field similarity matrix C of the cluster K; and Mi and Mj are the i and j random fields of the cluster
K. Elements within the matrix Ck with values approaching zero are locations with parameters common to the random fields within
the cluster. Once the locations of significance within the cluster are identified, the individual random fields are analysed to determine
the values within the region of interest. The clusters deemed of greatest importance are those with safety factors at the extremities of
the FoS distribution.

3. Field conditions - Victorian brown coal and Yallourn open-cut mine site conditions

Fig. 3
The Latrobe Valley Depression is an extension of the Gippsland Sedimentary Basin in Victoria, Australia, and covers an area of
800km2 [54], containing some of the world's thickest continuous brown coal seams [55]. The Yallourn open-cut (Fig. 4), is one of
three large brown coal mines in the region. Table 2 presents Yallourn brown coal properties as detailed by Perera et al. [56]. Beneath
the coal seam at Yallourn lies an interseam layer consisting of a clayey material much weaker in strength [57], with geotechnical
parameters of both coal and non-coal interseam materials shown in Table 3. Geotechnical parameters detailed in Tables 3–7 are
collated from in-house laboratory tests conducted at Federation University Australia's Geotechnical and Hydrogeological Engineering
Research Group (GHERG) coupled with data provided from the defunct State Electricity Commission of Victoria (SECV). Further
details of material parameter distributions are provided by Dyson and Tolooiyan [30]. A two-dimensional cross-section (Fig. 5) was
chosen for Finite Element Method slope stability analysis of the Yallourn mine North-Eastern batter, with associated shear strength
parameter distributions necessary for probabilistic analyses given in Table 4. A common parameter for describing the variation of a
distribution is the Coefficient of Variation (CoV), defined as the ratio of the distribution standard deviation (s) with respect to the

14
A.P. Dyson and A. Tolooiyan Simulation Modelling Practice and Theory 100 (2020) 102061

Fig. 12. Probability of failure for ranging CoVs. With each method, probabilities are equal to zero for low valued CoVs and all shear strength
parameters. In particular, Probabilities of Failure are exaggerated for the Point Estimate Method due to the minimal number of simulation instances,
coupled with the process of sampling at locations away from mean parameters. Results are highlighted by the high spread of FoS results detailed in
Fig. 11.

distribution mean (μ). Commonly observed CoVs are detailed in Table 5. As a comparison, typical values for the Coefficient of
Variation observed by Queiroz et al. [58] are presented in Table 6.
Of particular importance for RFEM simulation are parameter Scales of Fluctuation (θ), defining the spatial variability of the
random field. Coal and interseam shear strength SoF values in both the x and y directions are shown in Table 5. Of particular note is
the anisotropy of both coal and non-coal interseam materials (a mixture of clay and silt), with a rapid variation of shear strengths in
the “down the bore” direction, compared with “across-the bore”. Victorian Brown Coal exhibits no significant shear strength para-
meter correlations, no depth-dependent trends (Table 6). As such, parameter correlation has been neglected in this study.

4. Probabilistic simulation results and comparison

To compare safety factor results PEM, MCM and RFEM, a range of mean parameters and CoVs for coal and non-coal shear strength
parameters (cohesion and internal friction angle) were considered. The number of simulation realisations for each test method is
shown in Table 8. While PEM only requires 2N simulations (in this case, a total of 16 simulations), MCM and RFEM require numerous
simulation instances to achieve convergence. An initial estimate of 100 simulations was chosen for both methods to reach the Factor
of Safety convergence. Table 8 presents mean and total computation times for the methods considered. Of particular note is the
difference in PEM and MCM computational requirements compared with RFEM simulation. As RFEM simulation adopts a fine mesh
due to the length scale of material variation, significant computational resources are necessary. Due to the limited number of PEM
simulation instances, completion of the full simulation set in achieved in approximately 5% of the time compared to RFEM simu-
lation. In the case of RFEM simulation, FoS prediction techniques based on the first 50 random field instances are used to predict all
further RFEM realisations, without the need for cumbersome full FEM simulation [29]. Mean and CoV multipliers were implemented
to consider a number of distributions, altering parameters by a given scalar with respect to the initial mean and standard deviation
parameters. The mean multipliers considered are shown in Table 9, while the CoV multipliers are presented in Table 10. Initially,
mean and CoV parameters are altered individually while all other parameters are kept constant to assess the impact of each

15
A.P. Dyson and A. Tolooiyan Simulation Modelling Practice and Theory 100 (2020) 102061

distribution change on the FoS. Typical cohesion and friction angle random fields are shown in Fig. 6a and b, respectively, while
Fig. 6c shows an example of commonly occurring slope failure slip surface.
Fig. 7 shows a range of FoS values, for mean shear strength input parameters. In each case, RFEM produced the lowest safety
factors due to the numerous weak elements in each spatially varying random field. Of particular note is the conversative RFEM safety
factors compared to both PEM and MCM. Due to the spatially variable nature of RFEM simulation, locations containing weak shear
strengths allow shear bands to plasticise at lower FoS values, compared to PEM and MCM. As such, the RFEM Factors of Safety in
Fig. 7 are observed to be significantly lower for all test cases. PEM and MCM provided similar safety factors across both shear strength
parameters and both materials. Due to the computational expense of simulating numerous MCM realisations, PEM is preferable of the
two methods for determining the FoS in this particular instance. The interseam cohesion and friction angle produced the greatest
levels of slope instability, with the FoS dropping below 1.3 with the variation of both parameters. The variability of FoS values for a
range of mean input parameters is shown in Fig. 8. In all cases, PEM overpredicts the FoS variability, with respect to MCM and RFEM.
The convergence of mean safety factors is sparticularly important to ensure the minimum number of simulation realisations are used
(Fig. 9). PEM two-point samples for four variables produces only 16 realisations, while both MCM and RFEM were implemented with
100 probabilistic instances. In the case of RFEM, only the first 50 realisations were simulated with full FEM, while 100 MCM full FEM
simulations were required.
Similar to the mean value multipliers of Table 9, CoV multipliers are presented in Table 10. In the case of the Coefficient of
Variation, the distribution mean remains constant, while the distribution standard deviation is the unconstrained variable. For all
three methods, Factors of Safety decrease with respect to the Coefficient of Variation (Fig. 10). As with Fig. 7, the lowest levels of the
slope FoS occurred when the interseam distributions were varied. As the CoV increases, RFEM FoS values decrease at a greater rate
than PEM and MCM. Moreover, the variability of PEM safety factors is observed to be significantly greater than both RFEM and MCM,
with respect to input shear strength variation (Fig. 11). Similarly, the Probability of Failure for PEM is decidedly higher than RFEM
and MCM (Fig. 12). As only 16 PEM simulation realisations are computed, the failure of a slope when considering a combination of
slow valued strengths will produce an overpredicted PoF, suggesting the method has yet to converge. In the case where CoV mul-
tipliers of 1.5 were used, indicating a 50% increase in CoVs compared with the initial dataset of observed material parameters, PoF
values did not exceed 0.05 for MCM and 0.08 for REM. As such, even the most conservative of shear strength parameter distributions
produce Factors of Safety within the categories of Table 1 deemed as short and medium-term semi-permanent slopes. Given these
Probabilities of Failure, Kirsten [53] states that these slopes are likely to exhibit no evidently unstable movements.

5. Summary and conclusions

Probabilistic slope stability analysis methods are often necessary for evaluating soil variability and uncertainty. Although a
number of numerical methods may be incorporated within the Finite Element technique, output Factor of Safety and Probability of
Failure distributions can be dissimilar and method dependent. In the case of the MCM and RFEM methods, numerous simulation
instances are necessary to attain parameter convergence, requiring a significant computational capacity. A case study of the Yallourn
open-cut mine in Victoria, Australia, is presented, with the variability of input shear strength parameter distributions presented for
both coal and interseam materials. The variation of interseam shear strength parameters produced lower safety factors compared
with the variation of coal parameters, suggesting the interseam materials are more likely to impact the slope failure mechanism than
the coal. Although the sampling procedure for each probabilistic method was different, the presented mean Factors of Safety with
respect to the input mean shear strength parameters and coefficient of variation distributions were relatively similar for both PEM
and MCM. In the case of RFEM, zones containing weakened shear strength parameters allow slopes to fail at lower Factors of Safety,
as noted by Griffiths and Fenton [28]. When considering slope Probabilities of Failure, PEM significantly overpredicts failure, due to
the relatively few instances simulated for this method as well as the high level of FoS distribution variability. Although RFEM is the
most complex method considered, requiring additional knowledge of the spatial variability of cohesion and friction angle parameters,
the method produces the most conservative FoS parameters for all materials and input distributions. RFEM requires significant
computational time and resources with respect to both PEM and MCM. However, in cases where the effects of slope failure pose a
significant risk to life and economic productivity, the addition of the spatial variability captured in RFEM (resulting in lower safety
factors) is a necessary consideration in many cases of slope stability.

Acknowledgments

Financial support for this research has been provided by Earth Resources Regulation of the Victorian State Government
Department of Economic Development, Jobs, Transport and Resources. The first author is funded by the Australian Government
Research Training Program (RTP) and the GHERG scholarship programme.

References

[1] M. Abdulai, M. Sharifzadeh, Uncertainty and Reliability Analysis of Open Pit Rock Slopes: A Critical Review of Methods of Analysis, Geotech. Geol. Eng. (2018)
1–25.
[2] F Allan, T Yacoub, J Curran, On using spatial methods for heterogeneous slope stability analysis, 46th US Rock Mechanics/Geomechanics Symposium, American
Rock Mechanics Association, 2012.
[3] DV Griffiths, GA. Fenton, Risk assessment in geotechnical engineering, John Wiley & Sons, Inc, Hoboken, New Jersey, 2008.
[4] W.-F. Chen, N. Snitbhan, H.-Y. Fang, Stability of slopes in anisotropic, nonhomogeneous soils, Can. Geotech. J. 12 (1) (1975) 146–152.

16
A.P. Dyson and A. Tolooiyan Simulation Modelling Practice and Theory 100 (2020) 102061

[5] A. Mouyeaux, C. Carvajal, P. Bressolette, L. Peyras, P. Breul, C. Bacconnet, Probabilistic stability analysis of an earth dam by Stochastic Finite Element Method
based on field data, Comput. Geotech. 101 (2018) 34–47).
[6] K.-K. Phoon, F.H. Kulhawy, Characterization of geotechnical variability, Can. Geotech. J. 36 (4) (1999) 612–624.
[7] A. Sekhavatian, A.J. Choobbasti, Comparison of Point Estimate and Monte Carlo probabilistic methods in stability analysis of a deep excavation, Int. J. Geo-Eng.
9 (1) (2018) 20.
[8] EE. Alonso, Risk analysis of slopes and its application to slopes in Canadian sensitive clays, Geotechnique 26 (3) (1976) 453–472.
[9] S.E. Cho, Probabilistic assessment of slope stability that considers the spatial variability of soil properties, J. Geotech. Geoenviron. Eng. 136 (7) (2009) 975–984.
[10] R. Liang, N. b, M.A. OK, A reliability based approach for evaluating the slope stability of embankment dams, Eng. Geol. 54 (3–4) (1999) 271–285.
[11] E.H. Vanmarcke, Reliability of earth slopes, J. Geotech. Eng. Div. 103 (11) (1977) 1247–1265.
[12] TF. Wolff, Probabilistic slope stability in theory and practice, Uncertainty in the geologic environment: From theory to practice, ASCE, 1996, pp. 419–433.
[13] J.T. Christian, C.C. Ladd, G.B. Baecher, Reliability applied to slope stability analysis, J. Geotech. Eng. 120 (12) (1994) 2180–2207.
[14] K. Li, P. Lumb, Probabilistic design of slopes, Can. Geotech. J. 24 (4) (1987) 520–535.
[15] G Mostyn, K. Li, Probabilistic slope analysis—state of play, Proceedings, Conference on Probabilistic Methods in Geotechnical Engineering, Canberra, Australia,
AA Balkema, 1993, pp. 89–110 Rotterdam.
[16] SM Miller, JK Whyatt, EL McHugh, Applications of the point estimation method for stochastic rock slope engineering. Gulf Rocks, 2004, the 6th North America
Rock Mechanics Symposium (NARMS), American Rock Mechanics Association, 2004.
[17] V. Nguyen, R. Chowdhury, Probabilistic study of spoil pile stability in strip coal mines—two techniques compared, Int. J. Rock Mech. Min. Sci. Geomech.
Abstracts: Elsevier (1984) 303–312.
[18] H. Park, T. West, Development of a probabilistic approach for rock wedge failure, Eng. Geol. 59 (3–4) (2001) 233–251.
[19] E. Rosenblueth, Point estimates for probability moments, Proc. Natl. Acad. Sci. 72 (10) (1975) 3812–3814.
[20] E. Rosenblueth, Two-point estimates in probabilities, Appl. Math. Modell. 5 (5) (1981) 329–335.
[21] M. Valerio, C. Clayton, S. D’Ambra, C. Yan, An application of a reliability based method to evaluate open pit slope stability, Slope Stabil. (2013).
[22] M.E. Harr, Probabilistic estimates for multivariate analyses, Appl. Math. Modell. 13 (5) (1989) 313–318.
[23] DS Chandler, Monte Carlo simulation to evaluate slope stability, Uncertainty in the Geologic Environment: From Theory to Practice, ASCE, 1996, pp. 474–493.
[24] H. El-Ramly, N. Morgenstern, D. Cruden, Probabilistic slope stability analysis for practice, Can. Geotech. J. 39 (3) (2002) 665–683.
[25] V. McGuffey, D. Grivas, J. Iori, Z. Kyfor, Conventional and probabilistic embankment design, J. Geotech. Eng. Div. 108 (10) (1982) 1246–1254.
[26] D. Griffiths, G.A. Fenton, Influence of soil strength spatial variability on the stability of an undrained clay slope by finite elements, Slope Stabil. 2000 (2000)
184–193.
[27] W. Gibson, Probabilistic methods for slope analysis and design, Aust. Geomech. 46 (3) (2011) 29.
[28] D.V. Griffiths, G.A. Fenton, Probabilistic slope stability analysis by finite elements, J. Geotech. Geoenviron. Eng. 130 (5) (2004) 507–518.
[29] A.P. Dyson, A. Tolooiyan, Prediction and classification for finite element slope stability analysis by random field comparison, Comput. Geotech. 109 (2019)
117–129.
[30] A.P. Dyson, A. Tolooiyan, Probabilistic investigation of RFEM topologies for slope stability analysis, Comput. Geotech. 114 (2019) 103129.
[31] M.A. Hicks, J.D. Nuttall, J. Chen, Influence of heterogeneity on 3D slope reliability and failure consequence, Comput. Geotech. 61 (2014) 198–208.
[32] F Chen, L Cheng, T Zhou, X Chen, W Zhang, Probabilistic Assessment on Stability of Slopes Reinforced with Piles Considering Spatial Variability of Soil
Properties, IOP Conference Series: Earth and Environmental Science, IOP Publishing, 2019042023.
[33] S. Javankhoshdel, N. Luo, R.J. Bathurst, Probabilistic analysis of simple slopes with cohesive soil strength using RLEM and RFEM, Georisk: Assess. Manage. Risk
Eng. Syst. Geohazards 11 (3) (2017) 231–246.
[34] J.T. Christian, G.B. Baecher, Point-estimate method as numerical quadrature, J. Geotech. Geoenviron. Eng. 125 (9) (1999) 779–786.
[35] T. Elkateb, R. Chalaturnyk, P.K. Robertson, An overview of soil heterogeneity: quantification and implications on geotechnical field problems, Can. Geotech. J.
40 (1) (2003) 1–15.
[36] X. Emery, Properties and limitations of sequential indicator simulation, Stochastic Environ. Res. Risk Assess. 18 (6) (2004) 414–424.
[37] EH Vanmarcke, Random fields: analysis and synthesis, MIT Press, 1983.
[38] E.H. Vanmarcke, Probabilistic modeling of soil profiles, J. Geotech. Eng. Div. 103 (11) (1977) 1227–1246.
[39] M. Lloret-Cabot, G.A. Fenton, M.A. Hicks, On the estimation of scale of fluctuation in geostatistics, Georisk: Assess. Manage. Risk Eng. Syst. Geohazards 8 (2)
(2014) 129–140.
[40] D.-Q. Li, S.-H. Jiang, Z.-J. Cao, W. Zhou, C.-B. Zhou, L.-M. Zhang, A multiple response-surface method for slope reliability analysis considering spatial variability
of soil properties, Eng. Geol. 187 (2015) 60–72.
[41] C. Cherubini, Data and considerations on the variability of geotechnical properties of soils, Proceedings of the international conference on safety and reliability,
ESREL, 1997, pp. 1583–1591.
[42] MS Yucemen, WH Tang, A-S Ang, A probabilistic study of safety and design of earth slopes, University of Illinois Engineering Experiment Station. College of
Engineering, University of Illinois at Urbana-Champaign., 1973.
[43] G.A. Fenton, E.H. Vanmarcke, Simulation of random fields via local average subdivision, J. Eng. Mech. 116 (8) (1990) 1733–1749.
[44] GA Fenton, DV. Griffiths, Random Field Generation and the Local Average Subdivision Method, in: DV Griffiths, GA Fenton (Eds.), Probabilistic Methods in
Geotechnical Engineering, editors, Springer Vienna, Vienna, 2007, pp. 201–223.
[45] D Griffiths, P. Lane, Slope stability analysis by finite elements, Geotechnique 49 (3) (1999) 387–403.
[46] T. Matsui, K.-C. San, Finite element slope stability analysis by shear strength reduction technique, Soils Found. 32 (1) (1992) 59–70.
[47] AW. Bishop, The use of the slip circle in the stability analysis of slopes, Geotechnique 5 (1) (1955) 7–17.
[48] O Zienkiewicz, R. Taylor, 4th edn., The finite element method 1 Basic Formulation and Linear, 1989.
[49] N Snitbhan, W. Chen, Finite element analysis of large deformation in slopes, Proc 2nd Conf on Num Methods in Geomechanics, ASCE, 1976, p. 7171.
[50] JM Duncan, P. Dunlop, Slopes in stiff-fissured clays and shales, University of California, Berkeley, 1968.
[51] OC Zienkiewicz, C. Humpheson, R.W. Lewis, Associated and non-associated visco-plasticity and plasticity in soil mechanics, Geotechnique (1975).
[52] A.P. Dyson, A. Tolooiyan, Optimisation of strength reduction finite element method codes for slope stability analysis, Innov. Infrastruct. Sol. 3 (1) (2018) 38.
[53] H. Kirsten, Significance of the probability of failure in slope engineering, Civil Eng.= Siviele Ingenieurswese 1983 (v25i1) (1983) 17–29.
[54] C. Gloe, The Latrobe Valley coal measures. Contribution to Tertiary Section of the Geology of Victoria, Geol. Soc. Aust. (1974).
[55] C. Barton, C. Gloe, G. Holdgate, Latrobe Valley, Victoria, Australia: a world class brown coal deposit, Int. J. Coal Geol. 23 (1–4) (1993) 193–213.
[56] M Perera, P Ranjith, M Peter, Effects of saturation medium and pressure on strength parameters of Latrobe Valley brown coal: carbon dioxide, water and nitrogen
saturations, Energy 36 (12) (2011) 6941–6947.
[57] J Xue, A. Tolooiyan, Reliability analysis of block sliding in large brown coal open cuts, The 2012 World Congress on Advances in Civil, Environmental, and
Materials Research (ACEM’12), 2012, pp. 1578–1587.
[58] I.M. Queiroz, Comparison between Deterministic and Probabilistic Stability Analysis, Featuring Consequent Risk Assessment. World Academy of Science,
Engineering and Technology, International Journal of Environmental, Chemical, Ecological, Geol. Geophys. Eng. 10 (6) (2016) 636–643.

17