See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/282196697

The Random Finite Element Method (RFEM) in Probabilistic Slope Stability

Analysis with Consideration of Spatial Variability of Soil Properties

Article in Geotechnical Special Publication · March 2015

DOI: 10.1061/9780784479087.178

CITATIONS READS

18 1,472

3 authors, including:

Pooya Allahverdizadeh Sheykhloo D. V. Griffiths

AECOM Colorado School of Mines
11 PUBLICATIONS 38 CITATIONS 330 PUBLICATIONS 16,054 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Pooya Allahverdizadeh Sheykhloo on 21 October 2016.

The user has requested enhancement of the downloaded file.

IFCEE 2015 © ASCE 2015 1946

The Random Finite Element Method (RFEM) in Probabilistic Slope Stability

Analysis with Consideration of Spatial Variability of Soil Properties

Pooya Allahverdizadeh1, S.M. ASCE, D.V. Griffiths2,3, F. ASCE,

and G.A. Fenton4, M. ASCE.

1
PhD Candidate, Department of Civil and Environmental Engineering, Colorado
School of Mines, Golden, CO, USA, [email protected]
2
Professor, Department of Civil and Environmental Engineering, Colorado School of
Mines, Golden, CO, USA
3
Partner Investigator, Australian Research Council Centre of Excellence for
Geotechnical Science and Engineering, University of Newcastle, Callaghan NSW
2308, Australia
4
Professor, Department of Applied Mathematics, Dalhousie University, Halifax, NS,
Canada

ABSTRACT: This paper presents the results of probabilistic analyses in slope

stability problems using the Random Finite Element Method (RFEM). The influence
of spatially variable soil properties on design outcomes relating to slope stability
analysis has been assessed through parametric studies, with focus on the “worst case”
(critical) spatial correlation length that leads to a minimum reliability of the soil mass.
This critical value is of particular interest, because it could be used for design in the
absence of good site specific data.

INTRODUCTION

Probabilistic methods have been used in slope stability analysis since 1970s, and
have received considerable attention in the literature. Starting in the early 90’s, a new
method called the Random Finite Element Method (RFEM), which combines random
field theory and the finite element method, was developed for use in probabilistic
geotechnical engineering (e.g. Griffiths and Fenton 1993). The method was
subsequently applied to several areas of geotechnical engineering including
probabilistic slope stability analysis by Griffiths and Fenton 2000, 2004. The Local
Average Subdivision method (LAS) proposed by Fenton and Vanmarcke (1990) was
used for generating the random fields. It was shown that traditional probabilistic
analyses, in which spatial variability is ignored by implicitly assuming perfect
correlation, does not necessarily result in a conservative estimates of the probability
of failure. Later on, Griffiths et al. (2009) studied the influence of spatial variability
of soils more precisely.

Page 1
IFCEE 2015
IFCEE 2015 © ASCE 2015 1947

One of the advantages of using elastic-plastic finite elements for stability analysis is
that the failure mechanism is allowed to find the weakest path through the soil. The
ability of the FE approach to model the shape and location of the failure mechanism
offers many benefits over traditional methods in which the shape of the failure
mechanism is fixed a priori. Slope stability analysis is a good example this, in which
commonly used methods such as Bishop’s method, require the failure mechanism to
be circular.
This paper investigates the influence of soil spatial correlation length on
probabilistic slope stability analysis. Numerical results show that for a given value of
the coefficient of variation of soil strength parameters, there is a critical value of the
spatial correlation length which leads to a minimum reliability of the soil mass. In
other words, if spatial variation is ignored or implicitly assumed to be infinite, the
probability of failure can be underestimated resulting in an unconservative design.
The slope studied in this paper is shown in Figure 1 with consideration of both
undrained ϕu = 0, cu and drained c′, tanϕ′ slopes. The slope inclination, height, and
foundation ratio is given by β, H and D respectively. The saturated unit weight of the
soil, γ are held constant, while the shear strength cu of the undrained slope and c′ and
tanϕ′ of the drained slope are assumed to be random variables. For the both drained
and undrained slopes, cu and c′ were expressed in dimensionless forms as Cu and C′
respectively where Cu = cu /(γsat H) and C′ = c′ /(γ H).

FIG. 1. Slope profile

The shear strength parameters of the soil Cu and C′ and tanϕ′ are treated as random
variables, characterized statistically by lognormal and normal distributions for
undrained and drained slopes respectively. For the lognormal distribution the
logarithms of the properties are normally distributed. The lognormal distribution is
one of many possible choices (e.g. Fenton and Griffiths, 2008) that has been
advocated and used by several other investigators as a reasonable model for variable
soil properties (e.g. Massih et al., 2008). Lognormal distributions guarantee that the
random variable will never have negative values.
The lognormal distribution is defined by a mean μ and a standard deviation σ. The
probability density function of Cu is given by Equation 1 and an equivalent equation

Page 2
IFCEE 2015
IFCEE 2015 © ASCE 2015 1948

is applied to C′ and tanϕ′.

= − (1)
√

The mean and standard deviation can conveniently be combined in terms of the
dimensionless coefficient of variation defined as:

= (2)

RANDOM FINITE ELEMENT METHOD

The RFEM implementation used in this study combines elastic-plastic finite

element analysis with random field theory in slope stability analysis. The
methodology has been described in details elsewhere (e.g. Fenton and Griffiths,
2008).
The RFEM is used in conjunction with Monte Carlo simulations in which the
stability analysis is repeated until the probabilities relating to output quantities of
interest become statistically reproducible. In the case of slope stability analysis, the
probability of failure is defined by dividing the number of realizations in which the
slope failed by the total number of realizations.

Spatial Correlation

Generally, the mean and standard deviation of a variable are well understood by
engineers. However, the spatial correlation length θ of a random property is less well
understood. This property, called the “scale of fluctuation” or “spatial correlation
length”, has units of length, and represents the distance over which the soil or rock
property in question is reasonably well-correlated to its neighbors. In this research, a
“Markovian” correlation function is used where the spatial correlation is assumed to
decay exponentially with distance (Vanmarcke 1984).

| |/
= (3)

In Eq (3) which is for an isotropic material, τ is the absolute distance between any two
points in the random field, and ρ is the correlation coefficient between properties
assigned to two points in the random field separated by τ.
Since the actual undrained shear strength field is lognormally distributed, its
logarithm yields an underlying normal distributed (or Gaussian) field. The spatial
correlation length is measured with respect to this underlying field, that is, with
respect to lnCu. In particular, the spatial correlation length (θlnCu) describes the
distance over which the spatially random values tend to be significantly correlated in
the underlying Gaussian field. Thus, a large value of θlnCu implies a smoothly varying
field, while a small value will imply a ragged field. In this study, the spatial

Page 3
IFCEE 2015
IFCEE 2015 © ASCE 2015 1949

correlation length has been non-dimensionalized by dividing it by the height of the

slope H and will be expressed in the form:

= (4)

The influence of θ on a wide range of geotechnical systems has been assessed

through parametric studies (e.g. Griffiths and Fenton 2004, Griffiths et al. 2009,
Huang et al. 2010, Kasama and Whittle 2011, Al-Bitar and Soubra 2013) and has
been shown to have a significant influence on probabilistic quantities under
considerations.
Figure 2 (a and b) indicate two random field realizations and the associated failure
mechanisms. Figure 2(a) shows a relatively low spatial correlation length of ΘC′ =0.5
and Figure 2(b) shows a high spatial correlation length of ΘC′ =100. The figures depict
the variation in lnC′ and have been scaled in such a way that the dark and light
regions depict “strong” and “weak” soils, respectively. Black represents the strongest
element and white is the weakest in a particular realization. Although both cases
shown in Figure 2 had the same mean and variance, the different spatial correlation
lengths have led to quite different failure characteristics. In the case with a high
spatial correlation length, a much smoother mechanism was observed, more like the
type of mechanism that would be observed in a homogeneous soil.

FIG. 2. Typical random field realizations and deformed mesh at slope failure for
two different spatial correlation length, a) ΘC′ =0.5, b) ΘC′ =100

Page 4
IFCEE 2015
IFCEE 2015 © ASCE 2015 1950

RFEM RESULTS

The results of the RFEM analysis for the undrained and drained slopes are presented
in this section. Using this method, random fields are generated and assigned to each
element. Gravity loads are then applied, and if the algorithm could not converge
within a specific number of iterations, failure is said to have occurred. Lack of
convergence means that no stress redistribution could be found that is simultaneously
able to satisfy both the Mohr Coulomb failure criterion and global equilibrium. The
analysis is repeated numerous times, using Monte Carlo simulations, using the same
mean, standard deviation, and spatial correlation length of soil properties. The spatial
distribution of properties, however, varies from one realization to the next. Following
a “sufficient” number of realizations, the probability of failure pf is estimated by
dividing the number of failures by the total number of simulations.
A typical finite element mesh used for this problem is shown in Figure 3. The
majority of the elements are 8-node square except the elements adjacent to the slope
which are degenerated into triangles. The slope model has 910 total elements which
results in 910 random variables for the undrained slope with Cu and 1820 random
variables for the drained slope with C′ and tanϕ′ in each simulation. 4000 simulations
were used to ensure the reproducibility of the model.

2H H/tanβ 2H

FIG. 3. Typical mesh used for the RFEM slope stability analysis

Table 1 and 2 show the strength parameters and dimensions of the undrained and
drained slopes respectively. Young’s modulus (E) and Poisson ratio (υ) are set to 105
(kPa) and 0.3 respectively for all analyses. The unit weight of the soil is also
considered as a deterministic parameter equal to 20 (kN/m3).
The value of μ Cu and V was fixed at 5 (kPa) and 0.3 respectively for the undrained
slope. The 0.3 value for the coefficient of variation was selected, as the typical V
values for undrained shear strength lie in the range 0.13-0.5 (e.g. Duncan 2000). Four
slopes with different slope angles were modeled with different spatial correlation
lengths to study the effect of soil spatial variability and slope angle on the probability
of failure of slopes. The height of the slope, H is equal to 1. Figure 4 illustrates the
variation of the probability of failure with non-dimensionalized spatial correlation
length and slope angle for the undrained slope.

Page 5
IFCEE 2015
IFCEE 2015 © ASCE 2015 1951

Table 1. Parameter used in this study for the undrained slope. The coefficient of
variation was kept constant, V= σCu/μCu =0.3.

μCu (kPa) β (slope angle) ΘCu = θ/H

5 26.6 0.1, 0.25, 0.5, 1, 2, 4, 8
5 45 0.1, 0.25, 0.5, 1, 2, 4, 8
5 55 0.1, 0.25, 0.5, 1, 2, 4, 8
5 63.5 0.1, 0.25, 0.5, 1, 2, 4, 8

Table 2. Parameter used in this study for the drained slope. Slope angle is kept
constant to, β = 26.6° (2:1 slope).

μC′ (kPa) tanϕ′ V = σC′/μC′ = σtan ϕ′/μtan ϕ′ Θ = θ/H

5 0.364 0.25 0.01, 0.05, 0.1, 0.2, 0.5, 1, 2, 10, 100
5 0.364 0.3 0.01, 0.05, 0.1, 0.2, 0.5, 1, 2, 10, 100
5 0.364 0.35 0.01, 0.05, 0.1, 0.2, 0.5, 1, 2, 10, 100
5 0.364 0.45 0.01, 0.05, 0.1, 0.2, 0.5, 1, 2, 10, 100

0.7

0.6

0.5

0.4
pf

0.3

0.2
β=26.6
β=45
0.1 β=55
β=63.5
0
0.1 1 10
ϴCu

FIG. 4. Probability of failure vs. spatial correlation length for different slope
angles for the undrained slope, ϕu =0.

As it can be seen in Figure 4, by increasing the slope inclination, the probability of

Page 6
IFCEE 2015
IFCEE 2015 © ASCE 2015 1952

failure for the slope increases. This fact was expected based on traditional theories.
An important observation from Figure 4 is the influence of spatial correlation length
on probability of failure of the slopes. By increasing the spatial correlation length for
the 2:1 (β = 26.6°) slope, the pf increases. This observation was addressed in previous
studies on slopes (e.g. Griffiths and Fenton 2004). For the steeper slopes, however, a
maximum pf happens when the ΘCu is between 1 and 0.5. This spatial correlation
length called the “worst case” correlation length has been reported by other
researchers for bearing capacity and retaining wall problems (e.g. Griffiths and
Fenton 2001, Fenton and Griffiths 2003, Fenton et al. 2005, Massih et al. 2008).
However, for slopes with small coefficient of variation (V ≤ 0.5) the maximum
probability of failure was observed to happen when the slope has an infinite spatial
correlation length.
Another 2:1 slope was modeled with drained soil properties. The spatial correlation
and coefficient of variation of C′ and tan ϕ′ were assumed to be the same:

= = (5)

V = σC′/μ C′ = σtanϕ′/μ tanϕ′ (6)

The height of the slope H, is equal to 10m. Three correlations ρ, between C′ and tan
ϕ′ were considered to evaluate the effect of ρ on pf for the drained slope using normal
distribution. Figure 5 shows the variation of the pf with spatial correlation length and
ρ.
According to Figure 5, a slope with a positive correlation between C′ and tan ϕ′
leads to the highest probabilities of failure. It has been suggested by some
investigators (e.g. Cherubini 2000) that C′ and tan ϕ′ may have a negative correlation
which results in lower values for the pf. Thus, modeling the slope with no correlation
between the C′ and tan ϕ′ would be on the conservative side.

0.5

0.4

0.3
pf

0.2

ρ =- 0.5
0.1 ρ =0
ρ =0.5
0
0.01 0.1 1 10 100
ϴ

FIG. 5. Probability of failure vs. spatial correlation length for different cross-
correlations for the drained slope.

Page 7
IFCEE 2015
IFCEE 2015 © ASCE 2015 1953

For the drained slope, the slope angle β, was kept constant while the coefficient of
variation V, and spatial correlation length varied to investigate the worst case Θ based
on different V values. Figure 6 illustrates the variation of the pf with Θ and V. The
mean of C′ and tan ϕ′ is kept constant while the standard deviation of these variables
changed with coefficient of variation.

0.7

0.6

0.5

0.4
pf

0.3

0.2 V=0.4
V=0.35
0.1 V=0.3
V=0.25
0
0.01 0.1 1 10 100
ϴ

FIG. 6. Probability of failure vs. spatial correlation length for different

coefficient of variations for the drained slope.

By increasing the coefficient of variation, the pf of the slope increases. A maximum

probability of failure also occurs by increasing V similar to that observed in the
undrained slope. The worst case spatial correlation length also has a value between
0.5H and H. This means that when the slope has a spatial correlation length close to
the height of the slope, it has the highest value for the probability of failure or lowest
reliability.
When θ/H is large, the field is more strongly correlated, so that it appears smoother
with less variability in each realization. The slope consequently tends to behave as a
homogeneous slope more like that predicted by traditional methods. Conversely,
when θ/H is small, the random field is typically rough in appearance; however, as the
variability is high, the soil behaves like a uniform mass (with properties approaching
the median). This results in having a slope with homogeneous soil in each realization
as well. Thus, for very large and very small spatial correlation lengths, fewer failures
are expected.
In finite element analysis of slope stability, as mentioned before, failure is free to
seek out the weakest path through the soil. For intermediate correlation lengths within
the scale of the slope height, the soil properties measured at one location may be quite

Page 8
IFCEE 2015
IFCEE 2015 © ASCE 2015 1954

different from those actually present at other locations. It gives the failure the
opportunity to find the weakest path through the soil which could have a non-circular
or non-linear shape. Figure 2(a) shows a failure mechanism for a drained slope with
intermediate spatial correlation length, θ/H= 0.5 which is the worst case Θ for this
slope. As it can be seen, failure doesn’t follow a specific path. The failure is be able to
find its path where the soil has the weaker parameters (lighter color). Therefore, for
intermediate correlation lengths, more failures are observed. Following this reasoning,
the maximum probability of failure occurs when the slope has an intermediate spatial
correlation length as shown in Figures 4 and 6.

CONCLUSIONS

This paper has investigated the probability of failure for both drained and undrained
slopes using the RFEM. The RFEM combines the FEM with Monte Carlo simulation
in which spatial variability is properly taken into account. The RFEM enables the
failure mechanism to seek out a weakest path through heterogeneous soil which can
lead to higher probabilities of failure than might be predicted spatial variability is
ignored. The influence of the coefficient of variation V, slope angle β, and spatial
correlation length Θ, on the probability of failure pf was studied. It was shown clearly
that a worst case spatial correlation length exists for the both drained and undrained
slopes. This worst case spatial correlation length, leading to a maximum probability
of failure was shown to be of the order of 0.5H to H, where H is the slope height. The
implication of this result is that the spatial correlation length need not be estimated if
there is insufficient data, since the worst case Θ can be used to yield a conservative
design aimed at a target reliability. This result is a practical and important finding, as
the soil spatial variability is generally difficult and expensive to estimate accurately
and requires a large number of samples.

REFERENCES

Al-Bitar, T., and Soubra, A. H. (2013). “Bearing capacity of strip footings on spatially
random soils using sparse polynomial chaos expansion” Int. J. Numer. Anal. Meth.
Geomech., 37:2039–2060.
Cherubini, C. (2000). “Reliability evaluation of shallow foundation bearing capacity
on c′, ϕ′ soils” Can. Geotech. J., 37: 264-269.
Duncan, J. M. (2000). “Factors of safety and reliability in geotechnical engineering”
ASCE J. Geotechnical & Geoenv. Eng., 126(4): 307-316.
Fenton, G. A., and Vanmarcke, E. H. (1990). “Simulation of random fields via local
average subdivision.” J. Eng. Mech., 116(8): 1733–1749.
Fenton, G.A. and Griffiths, D.V. (2008). “Risk assessment in geotechnical
engineering”, John Wiley & Sons, Inc, New York.
Griffiths, D. V. and Fenton, G. A. (1993). “Seepage beneath water retaining structures
founded on spatially random soil.” Geotechnique, 43(4), 577–587.
Griffiths, D.V. and Fenton, G.A. (2001). “Influence of soil strength spatial variability
on the stability of an undrained clay slope by finite elements” ASCE, In Slope
Stability 2000, Proceeding of GeoDenver 2000, 184–193.
Griffiths, D.V. and Fenton, G.A. (2001). “Bearing capacity of spatially random soil:

Page 9
IFCEE 2015
 IFCEE 2015 © ASCE 2015 1955

the undrained clay Prandtl problem revisited” Geotechnique, 51(4): 351–359.

Griffiths, D.V. and Fenton, G.A. (2004). ‘‘Probabilistic slope stability analysis by
finite elements’’, ASCE J. Geotechnical & Geoenv. Eng., 130(5): 507–518.
Griffiths, D.V., Huang, J.S., and Fenton, G.A. (2009). “Influence of spatial variability
on slope reliability using 2D random field” ASCE J. Geotechnical & Geoenv. Eng.,
135(10): 1367-1378.
Huang, J.S., Griffiths, D.V., and Fenton, G.A. (2010). “System reliability of slopes by
RFEM” Soil and Foundations, 50(3): 343-353.
Kasama, K. and Whittle, A.J. (2011). “Bearing capacity of spatially random cohesive
soil using numerical limit analyses”, ASCE J. Geotechnical & Geoenv. Eng., 137:
989-996.
Massih, D. S. Y. A., Soubra, A. H., and Low, B. K. (2008). “Reliability based
analysis and design of strip footings against bearing capacity failure.” ASCE J.
Geotechnical & Geoenv. Eng., 134(7): 917–928.
Vanmarcke, E. H., (1984). “Random fields: Analysis and synthesis.” The MIT Press,
Cambridge, Mass.

Page 10
View publication stats IFCEE 2015