Probabilistic Analysis in Phase2 31-1

Probabilistic Analysis in Phase2

Often the parameters required for analysis are not well known. In these
cases it is desirable to perform a parametric study where model
behaviour is examined for a range of possible inputs. This tutorial
describes how to do this with Phase2.

Topics covered

• Principles

• Statistics

• Error plots

Phase2 v.8.0 Tutorial Manual

 Probabilistic Analysis in Phase2 31-2

Principles

In order to deal with probabilistic inputs and outputs in a

computationally efficient method, the point estimate method (PEM) has
been added to Phase2. The purpose of the method is to be able to combine
probabilistic input variables (also called random variables) and to
evaluate the distribution of the output variables. The principle of PEM is
to compute solutions at various estimation points and to combine them
with proper weighting in order to get an approximation of the
distribution of the output variables.

Probabilistic input variables

x1 x2 xn
− σx 1
+ σx 1
− σx
2
+ σx2
− σx n
+ σx n
...

n
2 combinations

Output variable distribution

y
− σy + σy

The version of PEM implemented in Phase2 is the two-point estimate

method for the first and second moment of uncorrelated variables. The
evaluation points of the random variables in a PEM analysis are located
at ±1σ of the mean.

The method requires 2n evaluations of the solution, where n is the

number of random variables. The distribution of the solution for
y = f ( x1 , x2 ,..., xn ) is given by:

2n
y = ∑ wf i (1)
i =1

Phase2 v.8.0 Tutorial Manual

 Probabilistic Analysis in Phase2 31-3

2
2n 2 
n

σ y = ∑ wf i − ∑ wf i 
2
(2)
i  i 
where the weights w are given by 1/2n. fi are successive evaluations of f at
the 2n possible combinations of the random variables at the point
estimate locations, i.e. at x n − σ xn and xn + σ xn . In the solution
presented here, all input and output variables are assumed to follow a
normal distribution given by their mean x and standard deviation σx .

Applicability
When considering statistical distributions of input parameters in
geomechanics problems, three different concepts must be considered:
uncertainties, variability and heterogeneities. These three concepts must
be treated separately as they have various impacts on the rock mass
behaviour and as different approaches must be used to tackle them.

Uncertainties arise from the difficulty in measuring key geomechanical

properties like rock stresses, rock modulus or rock strength. Any of these
measurements is smudged with some error due to the sampling process,
sample preparation or sensitivity and calibration of the measuring
devices. This uncertainty is usually evaluated and reduced by acquiring
repeated measurements during the development of a project.

Demand

Final design
Probability density

Detailed design
Preliminary design

Capacity

Potential for
failure

Design parameter

Considering a given design criteria, the probability of failure is given by

the gray areas on the Figure above: combining it with the consequences of
failure allows the computation of the risk of a given design (taking the
standard definition of risk being probability of occurrence times
consequence of occurrence) and to evaluate if this risk is acceptable.
There is always some residual risk remaining even at the final design

Phase2 v.8.0 Tutorial Manual

 Probabilistic Analysis in Phase2 31-4

stage (darker grey on the Figure above). This residual risk needs to be
handled by using mitigation measures, e.g. safety or worker exposure
reduction procedures. The PEM method implemented in Phase2 is
adapted to handle this kind of situation, i.e. it will allow one to track how
uncertainties in the input parameters are propagated through the
analyses and results in uncertainty in the design parameters. It allows
the engineer to not limit the design to a single deterministic analysis
with the most probable parameters (the mode of the distribution), but to
evaluate the reliability of the design by considering the dispersion of the
design parameters.

Variability is an inherent property of natural materials and rocks or rock

masses are no exception to the rule. It arises from the various formation
and transformation processes of rock and rock masses, including
diagenesis, fractioned crystallisation, alteration through fluid circulation,
metamorphism; all of which have local influence on the mechanical
parameters of rock and rock masses. Due to this variability, rock mass
properties will vary within a rock unit for example along the trace of a
tunnel. Thus, a failure mechanism will affect more or less severely
various locations along this tunnel.

Tunnel sections that are beyond the failure threshold

Distance along tunnel

Design parameter
Failure threshold

Here again the PEM approach implemented in Phase2 is well-suited, and

will for example allow the engineer to anticipate what percentage of a
tunnel section will be affected by a failure mechanism for a given severity
level. It will also allow for an evaluation of the range of severity of a given
failure mechanism that should be anticipated and thereby permit the
inclusion of flexibility in the design to handle the less probable but
potentially more severe situation. Having an estimate of the distribution
of the severity of a potential failure mechanism will also permit the
optimisation of the support systems and to obtain a better estimation of
the cost and thus the economical risk of a project. However, there are
some limitations with this approach. Firstly, the severity of the failure
mechanism evaluated must be continuous and monotonic: as will be
illustrated later in this manual, if abrupt changes in behaviour occur the
PEM method can be misleading. Thus, defining domains in which
behaviour is similar and the severity of failure is continuous is critical.
Finally, the wavelength of the variability must be relatively large
compared to the engineering problem evaluated, so that homogeneous
domains can be defined. In the case of short wavelength variations, the
effect will be different and needs to be treated, not as variability as
defined here, but as heterogeneities as discussed in the following.

Phase2 v.8.0 Tutorial Manual

 Probabilistic Analysis in Phase2 31-5

Heterogeneities need to be treated separately as they will not only

influence the severity of a failure mechanism, but will also change the
behaviour of the rocks or rock masses. For example, increasing modulus
heterogeneities in a rock will promote the development of local tensile
stress even in an overall compressive field which will affect the failure
mode, e.g. change from shear mechanism to tensile dominated
mechanism like spalling. The PEM approach implemented in Phase2 does
not simulate the effect of heterogeneities. Heterogeneities must be
handled differently, either by the use of a classification system and
equivalent homogeneous properties (e.g. defining material properties
using GSI) or by explicit modelling of the heterogeneities (e.g. using
“Discrete Function” material type).

Limitations
The main limitation of the current implementation of PEM in Phase2 is
that it deals with normal distribution (mean and standard deviation)
only. If either input or output random variables differ from a normal
distribution they will be approximated by a normal distribution which
will lead to inaccuracies.

The user of the PEM implementation must also be aware that even if all
the inputs follow normal distributions the outputs may not be normally
distributed and the PEM approximation can be inaccurate. This occurs
particularly when different behaviours are occurring (e.g. elasticity,
plasticity). An example of this situation is presented in the following
Figure where a plastic analysis was performed including three random
variables all normally distributed: cohesion c, modulus E and maximum
far field stress S1. In parallel to the PEM analyses, a Monte-Carlo
analysis was performed to evaluate the probabilistic outputs. The
modeled case was a circular opening as displayed in the center part of the
Figure with the plastic zone of two realisations (in red and orange) of the
Monte-Carlo simulation. The probabilistic outputs presented here are the
total displacement at two locations, one in the right wall of the opening
and the other some distance in the opening roof. At the location located in
the wall, the PEM output is very similar to the distribution obtained with
the Monte-Carlo analyses. At this location, the behaviour remains elastic
for all the runs (no behaviour change). This is not the case for the point
located in the tunnel roof, where mixed behaviour (elastic or plastic) can
occur. In this later situation, the distribution obtained from the Monte-
Carlo analyses differs significantly from a normal distribution, therefore
the PEM output doesn’t capture it accurately, though in this case the
overall range of the distributions are probably captured, which may not
be always the case. However it is noteworthy that if not entirely precise
the output of the PEM analyses provides valuable insight in the general
shape of the output distribution and this with a computational effort two
orders of magnitude smaller than with the Monte-Carlo approach (8
Phase2 runs for the PEM analyses vs. 1000 Phase2 runs for the presented
Monte-Carlo results).

Phase2 v.8.0 Tutorial Manual

 Probabilistic Analysis in Phase2 31-6

Clearly, this example illustrates the advantages of the PEM approach but
also the caution that must be taken when applying it, particularly when
mixed behaviours are occurring. In this latter case the results will
probably be inaccurate to some degree.

Monte Carlo

60 PEM

‰ 40
Probabilistic outputs
20

0
0 1 2 3 4 5 6 7 8
-4
total displacement u [m] x 10
S2
S1
Monte Carlo PEM
60

‰ 40

20

0
0 1 2 3 4 -3 5
total displacement u [m] x 10
c =5.1MPa E =23.5 GPa S1 =31.3 MPa

c=15.8 MPa E =19.5 GPa S 1 =29.0 MPa

100 100 100

‰ 50 ‰ 50 ‰ 50

10±3 MPa 20±2 GPa 30±5 MPa

0 0 0
0 10 20 10 20 30 0 50
c [MPa] E [GPa] S1 [MPa]
Probabilistic inputs

Recommendations
In light of the material presented above the following recommendations
can be made:

• PEM analyses are applicable for reliability analyses and

variability analyses, but don’t deal with the effect of
heterogeneities.
• A fundamental assumption is that all random variables are
normally distributed. When a situation differs from this
assumption inaccuracy has to be expected.
• When mixed behaviour occurs (e.g. elasticity/plasticity) one has to
expect that the output variables will not be normally distributed
and that PEM output will be inaccurate.
• It is recommended to use initially a limited number of random
variables and to evaluate their impact on the output

Phase2 v.8.0 Tutorial Manual

 Probabilistic Analysis in Phase2 31-7

distributions. Keeping the modeling exercise initially simple will

help to understand and properly interpret the analyses output.
• It is also recommended to run in addition to a PEM analyses
some deterministic extreme cases (±2σ or ±3σ) models. Indeed, the
evaluation points in a PEM analyses are located at ±1σ of the
mean case and will not capture behaviours that could occur
further off the mean.

Model

Start the Phase2 Model program. For this tutorial, we will start with a
model that has already been created. Go to File → Recent Folders →
Tutorials Folder and open Tutorial 03 Support1.fez. You should see
the following model:

This is a model of a tunnel excavated in phyllite with no support. We will

assume that there is uncertainty regarding the material properties so we
will perform statistical parametric study to observe the possible range of
behaviour for this model, using the point estimate method (PEM).

Materials
The material properties have already been assigned for this model. Go to
the Properties menu and select Define Materials. You can see that the
phyllite has been assigned a Hoek-Brown failure criterion. Change the
failure criterion to Generalized Hoek Brown as shown.

Now we could enter our best guess for the Hoek-Brown parameters and
run the analysis. However, the properties of this material are not well
known, so we will run a statistical analysis by varying the parameters in
a systematic way to see the range of possible behaviours.

Phase2 v.8.0 Tutorial Manual

 Probabilistic Analysis in Phase2 31-8

This can be done by selecting the Statistics button at the bottom of the
dialog. You can see that the button is currently grey so we cannot select
it. We therefore have to first activate the statistical analysis in Phase2.

Click OK to close the dialog.

Statistics
To activate the statistical analysis, open the Project Settings dialog
from the Analysis menu and select the Statistics page. Select the
checkbox for Probabilistic Analysis as shown.

Phase2 v.8.0 Tutorial Manual

 Probabilistic Analysis in Phase2 31-9

Click OK to close the Project Settings dialog.

We could now go back to the material dialog and set the Generalized
Hoek-Brown parameters. Alternatively we can get there through the
Statistics menu.

Go to the Statistics menu and you will see three options:

• Material Property Statistics

• Joint Property Statistics

• Field Stress Statistics

With these options it is possible to perform parametric studies on the

input field stress, material properties or joint properties. For this
tutorial, we will assume that we know the field stress well, so we will
only perform a statistical study on the material properties.

Choose Material Properties from the Statistics menu. You will now
see a dialog which allows you to choose the material property random
variables you wish to define for your probabilistic analysis.

The phyllite material is already selected by default. Click the Add

button. By default Young's Modulus will be added to the random variable
list. You can change this by clicking on the property and selecting the
desired parameter from the drop-list of available parameters. For this
example, we will choose Intact Compressive Strength as shown.

Phase2 v.8.0 Tutorial Manual

 Probabilistic Analysis in Phase2 31-10

Now we can set the mean (average) and standard deviation for this
property. Set the Mean to 50 MPa and the standard deviation to 3 as
shown.

Now, let's assume we would rather use GSI, mi and D to define the Hoek-
Brown failure envelope instead of mb, s and a. Select the Use GSI, mi, D
checkbox. We will also assume that the material is perfectly plastic, so we
only need to enter the peak values (not residuals). Select the Residual =
Peak checkbox.

Phase2 v.8.0 Tutorial Manual

 Probabilistic Analysis in Phase2 31-11

Now let’s add two more random variables. Click the Add button twice.
For the variable 2, choose Hoek-Brown GSI parameter (peak). Set the
mean to 40 and the standard deviation to 4. For variable 3, choose Hoek-
Brown D parameter (peak). Set the mean to 0.5 and the standard
deviation to 0.2.

Click OK to close the dialog.

NOTE: the number of models that need to be run when testing n random
variables is 2n + 1. For this example, n = 3 therefore 8 +1 = 9 separate
models will be run. (The 1 additional analysis uses the mean value of
each variable and is equivalent to a deterministic Phase2 analysis). So
the user should be aware that testing variations in many parameters
simultaneously can generate a very large number of models when using
the point estimate method (e.g. 10 random variables would require over
1000 separate analyses).

Phase2 v.8.0 Tutorial Manual

 Probabilistic Analysis in Phase2 31-12

The model definition is now finished. Save the model using the Save As
option in the File menu. The program actually now saves multiple
versions of the file with different material parameters - so it may take a
few seconds for the save to complete.

Compute
Run the model using the Compute option in the Analysis menu.

As described above, the statistical analysis is performed by running

multiple models with different material properties. If you have a multi-
core computer, Phase2 uses a multi-core parallel processing compute
engine by default, so the analysis should not take too long.

Once the model has finished computing select Interpret to view the
results.

Interpret
The Interpret program starts and reads the results of the probabilistic
analysis. You are now looking at the Maximum Compressive Stress
(Sigma1).

You can see how there are low stresses around the tunnel where failure
has occurred, and high stresses further out.

By default, this plot is showing the model run with the mean value for
each parameter. This is called the Base File. Select Statistics from the
menu and you can choose different plots:

• Base File Values: shows the model results when each parameter
is set to its mean value.

Phase2 v.8.0 Tutorial Manual

 Probabilistic Analysis in Phase2 31-13

• Mean Values: shows the average of all the models

• Standard Deviation Values: shows the standard deviation for

all the models

• Coefficient of Variation Values: shows the ratio of standard

deviation to the mean (dimensionless)

Try choosing different options from the Statistics menu and observe the
range in calculated Sigma1.

NOTE: the Interpret program calculates the mean and standard

deviation 'on the fly' for the different quantities when you plot them. If
you have an analysis with many elements and many changing variables,
you may notice a significant pause when you change plots. You can solve
this by choosing Compute All Statistics from the Statistics menu. This
will calculate everything so that plots will appear instantly.

Go back to showing the base file and show the failed elements using the
‘Display Yielded Elements’ button. Zoom in on the excavation (using the
middle mouse wheel) and the model should look like this:

This shows the failure that occurs in the base file, but perhaps you wish
to see the extent of failure as parameters vary. Press the 'Display Yielded
Elements' button again to turn off the yielded elements. From the
Statistics menu, choose Show Model Yield Zones. This option will
show yielded zones from all component files of the probabilistic analysis.
The plot should now look like this:

Phase2 v.8.0 Tutorial Manual

 Probabilistic Analysis in Phase2 31-14

The darkest colour indicates elements that fail in every model. Lighter
colours indicate elements that are less likely to fail. This plot is useful
when trying to determine the extent of possible failure when designing
rock bolt support, for example.

The shading of the yield zones can be customized in the Show Yield
Options dialog available in the Statistics menu.

Error plots
You can also get statistical information for queries. To test this, first turn
off the yielded zone display by unselecting Show Model Yield Zones
from the Statistics menu. Now we want to look at the floor heave for this
tunnel. Change the plot to show Total Displacement contours.

Phase2 v.8.0 Tutorial Manual

 Probabilistic Analysis in Phase2 31-15

Go to Query → Add Material Query. Select a point at the bottom left of

the tunnel (the mouse should snap to a point where the bottom flat part
starts to curve). Now choose the corresponding point on the bottom right.
Hit Enter to finish entering points. You will now see the Query Point
dialog. Choose to generate values at 20 locations as shown.

Click OK to close the dialog. You can now generate an error plot for this
query. Right click on the query line and select Error Plot (or else go to
Statistics → Error Plot and click on the query line). You will now see a
plot showing the mean displacement along the bottom of the tunnel with
error bars indicating one standard deviation of displacement.

The error plot indicates the range of possible floor displacements that can
be expected for the given uncertainty in material properties.

Note: to customize the chart display, right-click on the plot and select
Chart Properties. Experiment with the various chart properties. You can
also zoom and pan the chart display with the mouse wheel.

Close the error plot view.

Phase2 v.8.0 Tutorial Manual

 Probabilistic Analysis in Phase2 31-16

Info Viewer
Finally let’s have a look at the Info Viewer.

Select the Info Viewer option from the toolbar or the Analysis menu.

Scroll down the Info Viewer until you see the probability and statistics
sections.

For a probabilistic analysis, the Info Viewer presents a summary of:

• The number of random variables

• The mean and standard deviation of each random variable

• The actual values of the random variables used for each

component file of the probabilistic analysis

Since this example uses 3 random variables, there are 8 possible

combinations (2x2x2) of variables generated by the point estimate
method.

This concludes the tutorial. Please exit the Phase2 Interpret and Model
programs.

Phase2 v.8.0 Tutorial Manual