Probabilistic and Sensitivity Analysis

Module IV
Objectives
 Review of basic probability and statistics concepts
 Uncertainty
 Motivation for probabilistic analysis
 Statistical characteristics of data
 Probability of failure
 Simulation – Monte Carlo and Latin Hypercube
 Sensitivity analysis – fundamentals and uses
 Advanced correlation between material parameters
 Spatial variability of materials

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Probabilistic Slope Stability
 Principal areas of uncertainty
in slope stability analysis
 Soil properties and
behaviour
Uncertainty in
 Groundwater conditions Soil Properties

 Mechanisms of failure
 Probabilistic slope analysis Data Scatter
Systematic
Error
DOES NOT account for bad
choice of failure mechanism
Statistical Bias in
 More rational approach to Real Spatial
Variation
Random
Testing Errors
Error in the Measurement
Mean Procedures
design

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Data for various soil types (Harr, 1977)
 Unit weight for clays has a coefficient of variation (COV) ranging from 1.9 to
12.3
Material Parameter Number of Mean Standard Coefficient of
Samples Deviation Variation (%)
Clay (High Water 98 0.206 0.027 13.1
Plasticity) Content
Unit Weight 97 113.30 2.80 2.5

Clay (Med Water 99 0.131 0.0082 6.3

Plasticity) Content
Unit Weight 99 115.80 14.20 12.3

Clay (Low Water 97 0.138 0.0092 6.7

Plasticity) Content
Unit Weight 97 112.50 2.090 1.9

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Data for various soil types (Harr, 1977)
 Angle of friction for sands has a coefficient of variation (COV) ranging from 7.0
to 11.0 percent.
Material Number of Mean Standard Coefficient of
Samples Deviation Variation (%)
Gravel 38 36.22 2.160 6.0

Sand 73 38.80 2.800 7.0

Sand 136 36.40 4.050 11.0

Sand 30 40.52 4.560 11.0

Gravelly Sand 81 37.33 1.970 5.3

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability

“Correlation of Mechanical Properties of Some Rocks in South-

Eastern Nigeria” by Ngerebara O. D, Youdeowei P

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Sources of uncertainty
 Site topography
 Site stratigraphy
 Geologic origins and characteristics
 Groundwater conditions
 Engineering properties
 Material behaviour

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Basic Concepts
 Probability density function (pdf): probability random has value x

Total area equals 1

F(X)
Probability (5<X<8)=
Shaded Area

X X=5 X=8
Probabilistic Analysis in Slide – © 2017 Rocscience Inc.
Probabilistic Slope Stability
 Basic Concepts
1 n
 Mean m   xi
n i 1

n
 Variance Var  1  (x  m) 2
i
n 1 i 1

 Standard deviation
s  Var
 Coefficient of variation s
COV 
m
Probabilistic Analysis in Slide – © 2017 Rocscience Inc.
Probabilistic Slope Stability
 Two similar slopes with different uncertainties with the same mean Fs
Probability

Probability of
failure = area
COV = 0.5
COV = 0.2

Fs = 1 Mean Fs Factor of safety

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Basic Concepts
 General roles of mean and standard deviation

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Coefficient of Variation
 COV of soil inherent variability for various test measurements
(Phoon and Kulhawy 1999)

Test Type Property Soil type Mean Coefficient of

Variation (%)
Lab strength (UC) Su (UC) Clay 10-400 kN/m2 20-55

CPT f Clay and sand 20-40o 5-15

VST Su (VST) Clay 5-400 kN/m2 10-40

SPT N Clay and sand 10-70 blows/ft 25-50

Lab Index Dr Sand 30-70% 10-40

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

The 3 Rule
 Described by Dai and Wang (1992) – uses the fact that 99.73% of all
values of a normally distributed variable fall within plus or minus three
standard deviations of the average
 If HCV is the highest conceivable value of the parameter and LCV is the lowest
conceivable value of the parameter, these are approximately three standard
deviations above and below the average value
 Can be used to estimate a value of standard deviation by first
estimating the highest and lowest conceivable values of a parameter
and then dividing the difference by 6:

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

The 3 Rule


Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Risk and Probability
 Annual probability of failure versus
$LOST and LIVES LOST

 For different types of structures

 Slopes are similar to DAMs and

Probability of failure for design is 10-4
or 0.01% (Baecher 1982)

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Risk and Probability
 Category I—facilities designed, built, and operated
with state-of-the-practice engineering. Generally
these facilities have high failure consequences.
 Category II—facilities designed, built, and operated
using standard engineering practice. Many ordinary
facilities fall into this category.
 Category III—facilities without site-specific design and
substandard construction or operation. Temporary
facilities and those with low failure consequences
often fall into this category.
 Category IV—facilities with little or no engineering.

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Basic Concepts
 Common probability density functions for geotechnical engineering
 Normal
 Lognormal
 Uniform
 Triangular
 Beta
 Gamma
 Exponential

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Basic Concepts: Normal distribution
 Most widely used; symmetrical
 Example: cohesive strength of material

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Lognormal distribution
 Random variable cannot have negative values due to physical aspects of
problem
 Positively skewed distribution with large spike near zero
 Example: elastic modulus of soil

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Exponential distribution
 Used to model the risk of the
occurrence of some event (e.g.
earthquake) during life time of
engineering structure
 Example: spacing among rock joint
traces in outcrop

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Uniform distribution
 Used as ‘first’ model for quantity that lies
between a and b but about which little else
is known
 Good for problems with known bounds but
no information on relative likelihood of the
random variable
 Does not assume preferential likelihood of
random variable – each value has equal
chance

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Probability of failure
Relative Relative
Bin Bin
Frequency Frequency Hilighted Data = Factor of Safety - bishop sim plified < 1 (110 points)
0.786400 0.044476 0.786400 0.000000
0.808885 0.088951 0.808885 0.000000
0.831369 0.177903 0.831369 0.000000 4 4
0.853853 0.133427 0.853853 0.000000
0.876337 0.444757 0.876337 0.000000
0.898821 0.266854 0.898821 0.000000
0.921305 0.355806 0.921305 0.000000
0.943790 0.889515 0.943790 0.000000
3 3
0.966274 1.111894 0.966274 0.000000
0.988758 1.378748 0.988758 0.000000
1.011242 0.000000 1.011242 1.867981

Relative Frequency
1.033726 0.000000 1.033726 2.001408
1.056210 0.000000 1.056210 2.401690
1.078695 0.000000 1.078695 2.935399 2 2
1.101179 0.000000 1.101179 2.668545
1.123663 0.000000 1.123663 3.602535
1.146147 0.000000 1.146147 3.424632
1.168631 0.000000 1.168631 3.469108
1.191115 0.000000 1.191115 3.469108
1 1
1.213600 0.000000 1.213600 2.668545
1.236084 0.000000 1.236084 2.713020
1.258568 0.000000 1.258568 1.601127
1.281052 0.000000 1.281052 1.779030
1.303536 0.000000 1.303536 1.734554
1.326020 0.000000 1.326020 0.978466 0 0
0.8 0.9 1.0 1.1 1.2 1.3 1.4
1.348505 0.000000 1.348505 0.978466
1.370989 0.000000 1.370989 0.355806 Factor of Safety - bishop simplified

1.393473 0.000000 1.393473 0.489233

1.415957 0.000000 1.415957 0.266854
1.438441 0.000000 1.438441 0.088951 SAMPLED: m ean=1.145 s.d.=0.1171 m in=0.7865 m ax=1.461 (PF=11.000% RI=1.24048, best fit=Norm al distribution)
FIT: Norm al m ean=1.145 s.d.=0.1171 m in=0.7865 m ax=1.461
1.460925 0.000000 1.460925 0.088951
1.483410 0.000000 1.483410 0.000000

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Basic Concepts
 Monte Carlo simulation
 In slope stability used to evaluate probability of failure
 Can be used with only little background in statistics and probability
 Relatively simple
 Robust
 Very general (compared to analytical methods)

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Basic Concepts
 Probabilistic information from Monte Carlo simulations
 Probability of failure
 Mean factor of safety
 Standard deviation
 Best-fit distribution

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Basic Concepts
 Input required for simulation of correlated variables
 Probability distribution for each random variable (mean, standard
deviation, shape, min, max)
 Coefficient of correlation
 Slide allows for advanced correlation between material parameters between
different materials

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Basic Concepts
 Features Latin Hypercube sampling
 Considerably reduces chance of under-representing low probability regions
of distributions
 Requires fewer samples to give accuracy similar to Monte Carlo method

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Basic Concepts
 Latin Hypercube sampling. Example of N=10
 Divide cdf into N probability intervals
 Sample randomly from each interval
Cumulative Distribution

1
0.9
0.8
0.7
Probability

0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.5 1 1.5 2 2.5 3 3.5
X

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Monte Carlo vs Latin Hypercube sampling

Monte Carlo (left) and Latin Hypercube (right) cohesion 1000 samples

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability

Define the slope geometry

Specify slip surfaces in the slope

Probabilistic Slope Stability

Define the slope geometry

Specify slip surfaces in the slope

Assign the probability distribution to soil properties

Probabilistic Slope Stability
Define the slope geometry

Specify slip surfaces in the slope

Assign the probability distribution to soil properties

Generate n sets of soil properties by sampling from their distribution curves using MC simulation

Calculate factor of safety using deterministic limit equilibrium method for all slip surfaces using each
set of variables generated by MC (or LHS) simulation and find the lowest factor of safety for each
set

Calculate probability of failure

Pf = (No. of slip surfaces with Fs ≤ 1)/(Total No. of MC (or LHS) simulations)
Probabilistic Slope Stability
 Global Minimum vs. Overall Slope

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 The Global Minimum analysis type first performs the regular stability
analysis to determine the slip surface with the overall minimum factor
of safety
 The probabilistic analysis is then carried out on the global minimum
slip surface, using the samples generated for each random variable.
The probability of failure is then simply the number of analyses which
result in a safety factor less than 1, divided by the total number of
samples

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 This is done for each analysis method (e.g. Bishop or Janbu)
independently
 The Global Minimum analysis type assumes that the probability of
failure calculated for the overall minimum slip surface is representative
of the probability of failure for the entire slope

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 The Overall Slope analysis type does not make this assumption
 Instead, the entire search for a global minimum slip surface, is
repeated N times (where N = number of samples). For each search
iteration, a new set of random variable samples is first loaded, and the
search is carried out
 The critical probabilistic slip surface is the surface resulting in the
lowest FS among all the surfaces and samples.

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 A global minimum slip surface is then determined for each search
iteration.
 This might result in the location of 10 to 50 different Global Minimum
slip surfaces
 The definition of the probability of failure is the same as for the Global
Minimum method

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Overall Slope Example
 We will use Overall Slope analysis to determine the probability of
failure of the global minimum slip surfaces.

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Method
 First, switch the probabilistic analysis
type to Overall Slope.

 Then set the material properties to the

appropriate distributions.

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Method
 Choose a number of surfaces to
compute; Overall Slope analysis can
take a very long time. We will only
choose 500 surfaces, so the Compute
should only take a few minutes.

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Results
 Interpret gives us a large amount of information on the slope – the global
minimum slip surfaces, the mean safety factor, the probability of failure, etc.

Probability of Failure Factor of Safety

(%) (mean)

Overall Slope 3.7 1.177

Critical Probabilistic 2.8 1.201

Surface (Normal)

Critical Probabilistic 2.8 1.201

Surface (Lognormal)

Critical Deterministic 2.9 1.19

Surface

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Overall Slope example compare to literature
  2  
  1 + COVsu  
 ln  / Fs  
2
  1 + COVγ  
P = p[Fs < 1] = Φ    

 
f   2  2  
 ln  1 + COVsu  1 + COVγ   
    
 
 

• Probabilistic design chart

• Cohesive soil slopes
26
• PoF vs Fs

1.5

Javankhoshdel and Bathurst (2014)

(Javankhoshdel and Bathurst 2014)

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Overall Slope example compare to literature

COVsu = 0.5
COVg = 0.1
MC = 5000 (Overall Slope option)

Fs = 1.5
Pf = 26.2 %

(Javankhoshdel and Bathurst 2014)

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Sensitivity Analysis
 Basics
 Change an input parameter slightly while holding all others constant
 Observe impact on computed outcomes (factor of safety)
 If minor input parameter change causes fairly large factor of safety change
then design (F) is said to be highly sensitive to parameter
 Powerful procedure for establishing how sensitive factor of safety is to
changes in input parameters
 Powerful for studying range of possibilities

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Sensitivity Analysis
 Useful for comparing effectiveness of alternative remedial measures,
optimizing tradeoff between max safety & min cost
 Engineering decisions based on sensitivity results
 Determine input parameter more accurately
 Redesign slope to be less sensitive to input parameter

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Sensitivity Analysis
 Example – reduce failure mechanism
sensitivity to f 1.231

2.8 Modified design

2.6

2.4
Factor of Safety - bishop simplified

2.2

2.0
soil 1 : Cohesion (kN/m2)
soil 1 : Phi (deg)
1.8

1.6

1.4

1.2
0 10 20 30 40 50 60 70 80 90 100
Percent of Range (mean = 50%)

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Sensitivity Analysis
 Sensitivity of model parameters
 Material properties
 Support
 Identifying most influential parameters
 Determination of remedial measures
 Allocation of resources and effort in site investigations

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Tension Crack Sensitivity Analysis
 We want to determine the tension
crack depth that minimizes the factor
of safety
 Sensitivity analysis
 Only need to specify minimum and
maximum tension crack and Slide will
test 50 possible tension crack
boundary locations

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Tension Crack Sensitivity Analysis

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Advanced Statistical Correlation
Probabilistic Slope Stability
 Basic Concepts
 Correlation between two random variables X and Y
 Correlation coefficient, r positive

̅ is the sample mean

is the standard deviation

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Basic Concepts

Plot of X against Y

25

20
r = 0.97
15
Y

10

0
0 5 10 X 15 20 25

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Basic Concepts

Plot of X against Y

15
10
5
0
0 5 10 15 20 25
Y

-5
-10
-15
-20
X

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Probabilistic Slope Stability
 Basic Concepts
 Rule of thumb for practical analysis

 Assume weak correlation

0 < r < 0.3

 Assume perfect correlation

r > 0.9

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Advanced Material Statistics Options
 Equating: Sampling will be identical for the specified variables. Useful
when you have different materials which have some common
properties (e.g. cohesion and friction angle) and some properties
which are different (e.g. anisotropic bedding angle)
 Correlating: Define correlation coefficients between any two random
variables in any materials. This is useful when you have different
materials which have some properties (e.g. cohesion and friction
angle) which are correlated
 Provides more realistic results
 Particularly important when modeling anisotropic material properties
in a probabilistic analysis

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Example 1: Anisotropic Material
 Anisotropic material with linear
bedding
 The cohesion and friction angle will be
the same for the three areas, but the
angle of bedding plane orientation
(theta) from the horizontal will be A
different in each case
B
 We will define the anisotropic material
using three sub-materials: A, B, C C

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Example 1 : No Correlation
 Probabilistic analysis with Latin-
Hypercube sampling and 10 000
samples
 The random variables are defined
 No correlation is defined

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Example 1 : No Correlation
 Probability of failure = 7.65% (the slope failed in 7.65% of the 10 000
trials)
 During the analysis, the random variables in each trial were sampled
from the distributions defined. This means that in Trial 1 (for example)
the cohesion in A may have had a value of 1 kPa, while the cohesion in B
may have had a value of 19 kPa.
 Because A, B, and C are the same material, this would likely not occur in
the field.

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Example 1 : No Correlation

Cohesion in B vs cohesion in A

 To correct this, we need to equate or correlate the random variables between the
three sub-materials
Probabilistic Analysis in Slide – © 2017 Rocscience Inc.
Example 1 : Equating Materials
 We equate cohesion, cohesion 2, phi, phi 2 for all three sub-materials.
 This means that if in Trial 1, say, the cohesion sampled for Sub A is 10
kPa, the cohesion in Sub B (and Sub C) will also be 10 kPa, for that trial.
 Pof = 14.59%

Cohesion in B vs cohesion in A

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Example 1 : Correlating Materials
 We correlate the variables with a correlation coefficient = 0.8. PoF =
13.63%
 Note that we correlate A to B, B to C, and A to C.

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Example 1 : Correlating Materials
 Notice that failing to correlate the anisotropic sub-materials resulted in
a low PoF (unconservative)!

PoF: No Correlation PoF: Equating PoF: Correlation (0.8)

7.65% 14.59% 13.63%

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Note on Equating Correlation
 Equating:
 The distributions between the random variables being equated must be the
same.
 The ratios of the relative min divided by the standard deviation and the
relative max divided by the standard deviation have to be the same.
 Simply, the distributions have to look the same. Truncating the distributions at
different points causes problems.

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Note on Equating Correlation
 Equating:
 If you have different relative truncations, you must use the Advanced
Correlation dialog to correlate the variables. Setting the correlation to 1 will
give you a correlation coefficient very close to 1, but not the exact 1.0 you
would get by equating the variables.

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Example 2: Equating Materials
 5 models in a multi-scenario Slide file:
 1 material
 2 materials
 4 materials
 8 materials
 27 materials
 27 materials defined, with identical properties.

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Example 2: Equating Materials

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Example 2: Project Settings

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Example 2: No Correlation

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Example 2: No Correlation

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Example 2: No Correlation

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Example 2: Equate Dialog
 There is no correlation in the scenarios in
the “No Correlation” group.
 In each scenario in the “Correlation”
group, the material properties (cohesion,
cohesion 2, phi, phi 2) have been equated.
 The result is the same for all scenarios:

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Example 2: Summary
 No correlation: As more materials are
added, each on is sampled individually, Scenario Probability of Failure
from different parts of the distribution. (%)
 Results in a heterogeneous material No Correlation Equating
where each material averages to the
mean. 1 material 15.780 16.820

 At 27 materials, the PoF = 0%. 2 materials 7.840 16.820

 This example demonstrates the
4 materials 1.960 16.820
importance of correlation
8 materials 0.420 16.820

27 materials 0.000 16.820

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Example 14 – PROBABILISTIC EQUATE INITIAL
 Open Example 14 Probabilistic Equate Initial

Material Models in Slide – © 2017 Rocscience Inc.

 Example 3 : compare to literature
Recommended soil properties correlations:

Slide 6.0 Software Slide 7.0 Software

c c or f or su

f g
-1 < r < 0 0<r<1

(Javankhoshdel and Bathurst 2016)

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Example 3 : compare to literature

f = 30 ̊
Fs = 1.3
Pf = 21% (no correlation)
Pf = 3% (with correlation)

COVc = 0.5
COVf = 0.2
0.44
COVg = 0.1

Maximum practical values of cross-

correlation
rcf = -0.7 and rcg = rfg = +0.7

Probabilistic Analysis in Slide – © 2017 Rocscience Inc. Slope angle Javankhoshdel and Bathurst (2016)
Example 3 : compare to literature
COVc = 0.5
COVf = 0.2
COVg = 0.1
MC = 5000

Fs = 1.3
Pf (without correlation) = 21.2%
Pf (with correlation) = 2.95%

(Javankhoshdel and Bathurst 2016)

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Spatial Variability in Slide
Introduction
 A deterministic analysis assumes the input parameters are exact values.
 A probabilistic analysis takes into account variability or uncertainty in the
input parameters.
 Example: Let’s keep all parameters constant and only look at cohesion.

Deterministic Analysis
Cohesion = 2 kPa
Factor of Safety = 1.02

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Probabilistic Analysis in Slide
 To run a Probabilistic Analysis, a
random variable (a variable with
uncertainty) must be defined.
 A random variable is defined with the
following input:
 Distribution type (e.g. normal,
uniform, etc.)
 Mean, μ
 Standard Deviation, σ 0.5 1.5 2 3.5

 We will define cohesion as a random

variable with:
 Distribution = Normal
 Mean = 2 kPa
 Standard Deviation = 0.5 kPa

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Probabilistic Analysis in Slide
 To run the analysis, a Sampling
Method and Number of Samples are
defined.
 The Global Minimum option will use
the deterministic slip surface
throughout the analysis.
 The Overall Slope option will repeat
the search for the global minimum
1000 times (number of samples) –
once in each simulation.

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Probabilistic Analysis Computation
Simulation #1
• Cohesion = 0.7 kPa
• FS (deterministic) = 0.95

Simulation #2
• Cohesion = 2.1 kPa
• FS (deterministic) = 1.02

… …
Cohesion Distribution Simulation #1000
• Cohesion = 3.3 kPa
• FS (deterministic) = 1.08

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Probabilistic Analysis Results
#
#
x 100%  PF = 20.10%. Approximately 200 of the
simulations resulted in a FS<1
 This is called the Single Random
Variable (SRV) Approach.
 Although it is an improvement over
the Deterministic approach, it still
assumes that the slope is
homogenous, within each simulation.
 Also notice that the FS and PF don’t
fully agree. A slope with an FS=1.02
could be acceptable. One with a 20%
chance of failure is completely
unacceptable.

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Spatial Variability
 A more realistic model should take into account the variability of the material,
within each simulation. For example, Simulation 1 out of 1000 could look as shown:

it is unlikely that a cohesion of 0.7 would change suddenly to 3.3 in

nature. This is where Correlation Length comes in.

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Spatial Variability
 Correlation Length, ϴ is the distance over which the values of a
random variable will be significantly correlated, or similar.
 In the SRV approach, ϴ = ∞. The values of the random variable are
completely correlated for a homogeneous slope – they are the same
throughout.
 In the purely random process in the previous slide, where randomly
sampled values are placed in the grid, ϴ -> 0.
 Correlation length visualization:

ϴ = 0.2 ϴ = 2.0
Probabilistic Analysis in Slide – © 2017 Rocscience Inc.
Fenton, G. A., & Griffiths, D. V. (2008). Risk assessment in geotechnical engineering. Wiley.
Correlation Length

qH = 1000m, qV = 1000m

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Correlation Length

qH = 5m, qV = 1000m

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Correlation Length

qH = 1000m, qV = 5m

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Correlation Length

qH = 5m, qV = 5m

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Spatial Variability Implementation
 The first step to implementing spatial variability into the Slide probabilistic
analysis is creating a 2D random field that satisfies the following
requirements:
1. Accounts for anisotropy.
 Horizontal correlation length ranges between 10-60 metres.
 Vertical correlation length ranges between 1-3 metres due to bedding.
2. Accounts for local averaging.
 The original input mean and variance for the entire slope, change values when
assigned to a smaller area of the slope (in a grid box, for example).
 We will use the Local Average Subdivision Method (LAS) derived by Fenton
& Vanmarcke (1990).
 By applying Limit Equilibrium methods on top of the Random field, the
method that is going to be used in Slide will be called RLEM (Random Limit
Equilibrium Method)
Probabilistic Analysis in Slide – © 2017 Rocscience Inc.
Why RLEM? (PF vs q/H)
/H=0.5

=8≈∞

Javankhoshdel et al. (2016)

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Implementation in Slide
 We are currently working to implement Random Field theory (Fenton and
Vanmarcke 1990) into Slide.
 The original program (tailored for an SSR analysis) will be modified for a
limit equilibrium analysis.
 With the introduction of spatial variability into the probabilistic analysis, the
only additional input form the user will be the correlation length.

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Local Average Subdivision Method (LAS)
 The Local Average Subdivision
Method (LAS) is the best 2D random
field generator method for our case
in terms of efficiency and accuracy.
 It works by recursively dividing the
slope into 4 sections.
 In the figure, the means from the
parent cells are used to compute the
mean of child cells Z1i+1, Z2i+1, Z3i+1,
and Z4i+1. This continues until the
desired resolution is reached.
 Zi values are standard normal to
transfer to normal, lognormal or
bounded distribution

Probabilistic Analysis
Fenton, G.inA.,
Slide – © 2017 D.
& Griffiths, Rocscience Inc.
V. (2008). Risk assessment in geotechnical engineering. Wiley.
 Visualization of Random Field: Cohesion
 Cohesion:
 Mean: 10 kPa
 Standard Deviation: 5.0
 Distribution: Lognormal
 Correlation Length: 20 m
(isotropic)
 Each color represents a
region with different value
for cohesion (spatially
variable)

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 The advantages of RLEM in Slide
1) Easy to use and better visualization.
2) More than 10 times faster analysis compare to available commercial and non-
commercial software packages.
3) The ability of handling any types of geometry.
4) Ability of modeling layered slopes.
5) Ability of carrying out slope stability analysis with circular and non-circular slip
surfaces.
6) Ability of considering cross-correlated random fields.
7) The ability of handling CPT data and measuring spatial correlation length with
traditional and more advanced approaches.

Probabilistic Analysis
Fenton, G.inA.,
Slide – © 2017 D.
& Griffiths, Rocscience Inc.
V. (2008). Risk assessment in geotechnical engineering. Wiley.
 Correlation Coefficient, rc-f  -0.5
Cohesion Friction Angle
Mean 10 kPa Mean 30 degrees
Standard Standard
5.0 kPa 6.0 degrees
Deviation Deviation
Distribution Lognormal Distribution Lognormal
Correlation Correlation
Length 20 m Length 20 m
(isotropic) (isotropic)

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Example 1: Compare to literature (Huang et al. 2010)
Parameter Example 1A Example 1B
Slope Angle, b 27◦ 27◦
Covariance function Markovian Markovian
Number Monte Carlo simulations 2000 2000
Cohesion 7 kPa, COV: 0.3 7 kPa, COV: 0.3
Friction Angle 20◦, COV: 0.3 20◦, COV: 0.3
Correlation Length (isotropic) ∞ 5m

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Huang et al. 2010 Example 1A: Correla on Length ∞

Factor of Safety 1.2

Huang et al. Pf 28.5%

Slide SRV Pf 26.65%

Slide Spatial Pf 25.75%

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Huang et al. 2010 Example 1B: Correlation Length 5m

Factor of Safety 1.2

Huang et al.
10.5%
(RFEM) Pf

Slide Spatial Pf 6.15%

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Example 2: Cross-Correlation
Parameter Example 2A Example 2B Example 2C

Slope Angle, b 27◦ 27◦ 27◦

Covariance
Markovian Markovian Markovian
function
Number Monte
2000 2000 2000
Carlo simulations
Cohesion COV 0.5 0.5 0.5
20 m
Friction Angle COV 0.2 0.2 0.2

Unit Weight COV 0.1 0.1 0.1

Correlation Length 10, 20,
∞ ∞ 60 m
(isotropic) 40, 80 m
rc-f = -0.5, rc-f = -0.5,
Cross-Correlation None rc-g = 0.45, rc-g = 0.45,
rf-g = 0.45 rf-g = 0.45

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Example 2A: θ=∞, no cross-correlation

Factor of Safety 1.3

Slide SRV Pf 16.65%

Slide Spatial Pf 14.5%

No cross-correlation

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

Example 2B: θ=∞, with cross-correlation

Factor of Safety 1.3

Slide SRV Pf 4.5%

Slide Spatial Pf 4.25%

rc-f = -0.5, rc-g = 0.45, rf-g = 0.45

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Example 2C: Varying θ, with cross-correlation
 qH = qV = 10, 20, 40, 80 m
 rc-f = -0.5, rc-g = 0.45, rf-g = 0.45

θ/H Pf (%) q/H = infinity

Pf = 4.25 %
1 0.05
2 0.15
4 0.55
Fs =1.3
8 1.3
∞ 4.25 q/H = 1
Pf = 0.05 %

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Cuckoo Search, Surface Altering

Properties
Method Cuckoo Search (GLE)
COVf 0.2
COVc 0.5
No. simulations (LHS) 4000
No. Slices 200
q/H 2
Mean Fs 1.043
Pf 19.6%

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Auto Refine Search, Surface Altering

Properties
Method Auto Refine Search(GLE)
COVf 0.2
COVc 0.5
No. simulations (LHS) 4000
No. Slices 200
q/H 2
Mean Fs 1.047
Pf 18.425%

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Cuckoo Search, Surface Altering

Properties
Method Cuckoo Search (GLE)
COVf 0.2
COVc 0.5
No. simulations (LHS) 4000
No. Slices 200
q/H 2
Mean Fs 1.058
Pf 6.3%

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

 Auto Refine Search, Surface Altering

Properties
Method Auto Refine Search(GLE)
COVf 0.2
COVc 0.5
No. simulations (LHS) 4000
No. Slices 200
q/H 2
Mean Fs 1.061
Pf 5.875%

Probabilistic Analysis in Slide – © 2017 Rocscience Inc.

End