Proceedings of the 11th International Conference on Geosynthetics

The basics of probabilistic internal stability analysis and design of

reinforced soil walls explained
Richard J. Bathurst
GeoEngineering Centre at Queen’s-RMC, Canada

ABSTRACT: This paper explains the basics of probabilistic analysis and design for the internal stability
limit states of tensile rupture and pullout in mechanically stabilized earth (MSE) walls. The concepts are
general and can be applied to any soil-structure interaction problem which can be expressed by a simple
linear limit state performance function. The general approach considers uncertainty in the choice of
nominal values and links probability of failure to the underlying accuracy of the load and resistance
models that appear in each limit state equation using bias statistics. A useful closed-form solution is
presented to calculate margins of safety in terms of reliability index. The formulation is easily
implemented in a spreadsheet to facilitate sensitivity analysis during design and thus is an alternative
approach to Monte Carlo simulation.

Keywords: probabilistic analysis, MSE walls, internal limit states, tensile rupture, pullout

1 INTRODUCTION

Geosynthetic reinforced soil walls are an accepted category of mechanically stabilized earth (MSE)
technologies and their presence is ubiquitous on the geotechnical civil engineering landscape. The
analysis and design tools for these systems are most often deterministic in nature and can be traced to
adaptations of classical soil mechanics earth pressure theory and limit equilibrium methods.
The internal stability design for MSE walls in the UK is based on a partial factor approach in which
factors are applied to soil and reinforcement material properties and to load contributions in different
combinations to ensure safe designs. Geotechnical foundation design codes in North America adopt a load
and resistance factor design (LRFD) approach which has been used by structural engineers for decades in
Canada and the USA. In this approach, load terms are multiplied by load factors (magnitude of one or
more) and the resistance side is multiplied by a single resistance factor (with a magnitude of one or less).
The intent of a properly calibrated limit state design equation expressed in a LRFD framework is to ensure
that a target maximum probability of failure will not be exceeded. However, the load and resistance
factors that appear in LRFD codes have been selected largely by fitting to factors of safety used in
allowable (working) stress design (ASD) past practice. Whether a designer uses a partial factor approach
as in the UK or a LRFD approach as in North America, the margin of safety expressed probabilistically is
unknown. This leads to the conundrum of a limit state being satisfactory when viewed from a factor of
safety point of view but unsatisfactory from a probability of failure perspective. This point is
demonstrated for the case of the rupture and pullout limit states for the geosynthetic reinforcement layers
in the MSE wall in Figure 1. The nominal tensile load Qn can be calculated using one of a number load
models found in the literature and design codes. Similarly, the nominal resistance Rn can be calculated for
tensile rupture and pullout limit states using equations found in design guidelines. In conventional
allowable stress design the ratio of nominal resistance to nominal load defines the factor of safety; the
resistance term is adjusted so that the factor of safety satisfies a minimum acceptable value. However,
both nominal resistance and nominal load sides have uncertainty as visualized by the idealized frequency
distributions in Figure 2. Notionally, the area of the overlap of the lower tail of the resistance distribution
on the right with the upper tail of the load distribution on the left indicates a non-zero probability of
16-21 September 2018, Seoul, Korea
 Proceedings of the 11th International Conference on Geosynthetics

Figure 1. Nominal load and resistance equations for reinforcement rupture and pullout limit states for internal
stability of geosynthetic reinforced MSE walls

Figure 2. Factor of safety and probability of failure concepts

failure. Clearly, two different combinations of load and resistance distributions can have different
probabilities of failure (Pf) while the factor of safety remains the same. The distributions for nominal
resistance and load values can be characterized by mean and coefficient of variation (COV) where the
latter is simply the standard deviation divided by the mean. These distributions may have different forms
but lognormal distributions are the most common for the quantities that are the focus of this paper.
An alternative preferred parameter to quantify margins of safety in geotechnical engineering is
reliability index b. The relationship between probability of failure and reliability index is Pf = 1- F(b)
where F is the standard normal cumulative distribution function (CDF) (NORM.DIST in Excel).

2 FORMULATION OF LIMIT STATE EQUATIONS

If all uncertainty in the true magnitude of load and resistance terms for the limit states introduced above
was solely due to the estimation of the soil friction angle (f) and soil unit weight (g), then the probability
of failure of each limit state can be computed using the following equation and Monte Carlo simulation:

16-21 September 2018, Seoul, Korea

 Proceedings of the 11th International Conference on Geosynthetics

Rn
g= -1 (1)
Qn
Specifically, random values of f and g are simulated from probability distributions for these parameters
for a total of N times, and each pair of values are used to compute values of Rn and Qn. Unlike naturally
deposited soils, the backfill soil in a MSE wall is an engineered material and therefore the variability in
soil unit weight and friction angle are low. For Monte Carlo simulation, COV values of 0.10 and 0.05 for
f and g are reasonable. For large N values, the number of times g < 0 divided by N trials (realizations) is
the probability of failure.
However, the true probability of failure is strongly influenced by the accuracy of the load and resistance
models that appear in the formulation of a limit state equation. The expectation that any design equation
for Qn and Rn in geotechnical foundation engineering design will give the actual load or resistance is not
reasonable. Furthermore, we expect that different models will have different accuracy when predicted
values are compared to measured values. The ratio of measured load or resistance to the corresponding
predicted load or resistance is called method bias (or bias for brevity) and is denoted by l in this paper.
The nominal load and resistance terms in Equation 1 can be transformed to give a performance function
that is expressed in terms of true (i.e. measured) values of load (Qm) and resistance (Rm) terms as follows:

λR R n R
g= -1 = m -1 (2)
λ QQn Qm

To compute the true probability of failure, the same Monte Carlo simulation technique described above is
used with the additional steps of also sampling resistance and load bias values from random distributions
of lR and lQ during each realization.

3 LIMIT STATE BIAS STATISTICS AND INTERPRETATION

For the rupture (tensile strength) limit state, the value of lR is simply the ratio of estimated tensile
strength at end of design life divided by the nominal estimated value. In North American practice the
nominal value is the long-term allowable strength Tal computed as:

Tult
Tal = (3)
RF

Here, Tult is the reference ultimate strength of the reinforcement and RF is a reduction factor that accounts
for the loss of tensile strength over the design life of the reinforcement due to creep, installation damage
and possible chemical/biological processes. Bias values for tensile strength have been computed by
comparing available strength at end of design life (e.g. 75 years) to nominal values using results of
product-specific long-term creep rupture testing and installation damage testing. Bathurst and Miyata
(2015) developed bias statistics for the combined effect of installation damage and creep on geogrid
reinforcement tensile strength at the end of design life for different geogrid classification types. The COV
of tensile rupture bias values ranged from 0.04 to 0.14. A value of COVlR = 0.10 was judged to be a
reasonable typical value. The mean of tensile rupture bias for the same materials was taken as lR = 1.10
(Bathurst et al. 2011). This value follows from North American practice to conservatively report the
ultimate tensile strength as the minimum average roll value (MARV) which is computed as the average
ultimate tensile strength of a roll that is two standard deviations below the average ultimate tensile
strength from multiple rolls. The value of lR = 1.10 means that the true long-term tensile strength of a
reinforcement geosynthetic is 10% higher, on average, than the certified product value reported for Tult
and used in Equation 3.
Bias statistics for the pullout limit state can be gathered from conventional laboratory pullout box
testing. In this case, each measured maximum pullout load is compared to the predicted value using a
particular model and the same vertical stress used in the test; the bias value is computed as the measured
value divided by the predicted value. Bias values computed in this manner can be expected to vary

16-21 September 2018, Seoul, Korea

 Proceedings of the 11th International Conference on Geosynthetics

between different models due to: 1) differences in the ability of the underlying deterministic pullout
model (of which there are many) to capture the mechanics of soil-reinforcement interaction during
pullout; 2) uncertainty due to variation in pullout box testing protocols; 3) quantity and quality of data,
and; 4) consistency in interpretation of data gathered from multiple sources (the typical case) (Allen et al.
2005). As just one example, bias statistics for the same soil and reinforcement type acquired from a
laboratory pullout box text may be different from the bias statistics for the same materials and vertical
confining stress for tests performed in-situ (Miyata and Bathurst 2012).
Bias values for two different pullout models are plotted as cumulative distribution function (CDF) plots
in Figure 3. A total of n = 318 test results were taken from Huang and Bathurst (2009). These plots are
created in the following steps: 1) bias values are sorted in rank order from 1 to n by magnitude (lowest
bias value first); 2) each value is then assigned a cumulative probability from p = 1/(n+1) to 1; 3) the
standard normal variable z is computed as z = F-1(p) where F-1 is the inverse standard normal
distribution function (NORM.S.INV in Excel), and; 4) plot z versus bias. If bias values are normally
distributed the data will appear as a straight line on this plot with linear axes. If the bias values are
lognormally distributed the data will present as a straight line on a log-linear plot which is the case here
for Figure 3. In this figure, the data have also been plotted with cumulative probability as the vertical axis
which may be more familiar to readers. For plots of this type (and using generic bias notation l), the mean
of Ln values of l, denoted as mLNl, will intersect the horizontal axis at z = 0 and the slope of the line will
be the inverse of the standard deviation (1/ sLNl) of the ln(l) distribution. The following transformations
can be used for sLNl and mLNl:

σ LNλ = LN(1 + COVλ2 ) (4a)

1 2
μ LNλ = LN(μ λ ) - s LNλ (4b)
2

Figure 3 shows pullout bias values using two different pullout models. The pullout model proposed by
Huang and Bathurst (2009) is judged to be better than the current AASHTO (2017) model with respect to
accuracy because the mean of bias values (mlR) is closer to one and the spread in bias values expressed by
COVlR is much less.
Figure 4 shows the results of bias analysis for the load side for the two limit states introduced in this
paper. Load model 1 is the current AASHTO (2017) Simplified Method and load model 2 is the
Simplified Stiffness Method (Allen and Bathurst 2015, 2018). The former is a combined soil strength-
and reinforcement type-based method to compute maximum tensile reinforcement loads (Tmax in Figure 1)
under operational conditions, while the latter is a reinforcement stiffness-based approach. The load bias
values plot as roughly straight lines over most of the data range in Figure 4 confirming, at least visually,
that the bias values may be reasonably assumed to be lognormally distributed. There is a discrepancy
between the measured data and the fitted approximations at the lower tails, but this is not a practical
concern because it is the upper tail that has the most influence on probability of failure as illustrated
notionally in Figure 2.
The bias statistics for the two load models demonstrate that the AASHTO (2017) Simplified Method is
very conservative because the measured loads are only 40% of predicted loads on average. Furthermore,
there is large uncertainty in the accuracy of the model which is quantified by the COV of load bias values
(i.e. COVlQ= 95%). In contrast, the Simplified Stiffness Method is much more accurate on average with
only 5% overestimation of maximum tensile loads and less spread in accuracy of predictions quantified by
COVlQ = 36%. The reason for the better agreement for the Simplified Stiffness Method is that it accounts
for the influence of reinforcement stiffness on magnitude of maximum load in a reinforcement layer
(Tmax) and was empirically calibrated against a large database of measured Tmax values to give a mean bias
value close to one and a minimum COV of load bias values (Allen and Bathurst 2015, 2018).
So far in this paper, possible correlations between nominal values and between bias values and
predicted values have not been addressed. These correlations (dependencies) are possible depending on
the limit state and the models that appear in the limit state design equations. For example, consider Figure
5 which shows pullout bias values plotted against predicted (nominal) (Pmax) pullout values. The bias
values are the same data discussed with Figure 3. Log axes have been used in Figure 5 to spread the data
points out for better visual clarity, but this choice of axes does not prejudice the quantitative interpretation
described next. The data show that for the poorer pullout model (PM1) the accuracy of the model changes

16-21 September 2018, Seoul, Korea

 Proceedings of the 11th International Conference on Geosynthetics

Figure 3. Cumulative distribution function (CDF) plots for pullout bias using two different pullout models

Figure 4. Cumulative distribution function (CDF) plots for load bias using two different load models
with the magnitude of the predicted pullout capacity. This is an undesirable outcome in principle for both
allowable stress design and probabilistic design approaches. A preferred behaviour is that model accuracy
does not change with predicted pullout capacity which is the case for pullout model (PM2) in this paper.
The strength of a linear correlation between X and Y in a plot of data such as Figure 5 can be quantified
by Pearson’s correlation coefficient (r) which is simply the R in R2 used in linear regression analyses.
Positive values are computed when Y increases with X, and negative values when Y decreases with X
which is the case for pullout model (PM1) in Figure 5. Parameter r is an input parameter in Monte Carlo
simulations using Equation 2.
Similar correlations can occur between nominal load (Qn = Tmax) and nominal resistance (Rn = Pmax);
such a case is illustrated in Figure 6 for the combination of poor load model (LM1) and poor pullout
model (PM1) discussed previously. In this example, the correlation is because friction angle (f) and unit
weight (g) appear in equations for Tmax and Pmax , and both soil parameters are sampled from the same
populations during Monte Carlo simulation (i.e. the same friction angle and unit weight of soil contribute
to tensile load and resistance in the pullout limit state equation).

16-21 September 2018, Seoul, Korea

 Proceedings of the 11th International Conference on Geosynthetics

Figure 5. Example of pullout models with and without correlation between bias and predicted pullout capacity

4 CALCULATION OF PROBABILITY OF FAILURE (OR RELIABILITY INDEX)

The previous sections have described the formulation of a general limit state equation suitable for Monte
Carlo simulation (Equation 2) and the random variables that appear in these analyses. However, engineers
may not be comfortable with Monte Carlo techniques and the tools necessary to carry out these
simulations. Fortunately, for the conditions described here and the observation that random variables for
bias and nominal values are lognormally distributed, the following closed-form solution is available for
the calculation of reliability index b introduced earlier (Bathurst and Javankhoshdel 2017):

é μ 2
(1+COVQn 2
)(1+COVλQ )ù
ln ê( λR Fn ) 2 2
ú
êë μ λQ (1+COVRn )(1+COVλR ) ú
û
β=
2 2 2 2
(5)
é (1+COVQn )(1+COVλQ )(1+COVRn )(1+COVλR )(1+ρ R COVRn COVλR ) 2 (1+ρ Q COVQn COVλQ ) 2 ù
ln ê ú
ë (1+ρ n COVRn COVQn ) 2 û

Here, Fn = mRn/mQn is the nominal factor of safety which is the ratio of the mean of nominal resistance (Rn)
and mean of nominal load (Qn). This ratio is assumed equal to the ratio of nominal resistance (Rn) and
nominal load (Qn) used at time of design. This is the quantitative connection to factor of safety used in
ASD. Quantities mlR and mlQ are mean values of nominal resistance and load method bias values (lR and
lQ), respectively, and have uncertainty quantified by COVlR and COVlQ as discussed earlier. The nominal
resistance (Rn) value and nominal load value (Qn) used at design time in the limit state design equations
also have uncertainty quantified by coefficients of variation denoted as COVRn and COVQn.
For reliability-based design the magnitudes of these COV values can be linked to the notion of level of
understanding found in the Canadian Highway Bridge Design Code for LRFD of foundations (Fenton et
al. 2016). Three levels of understanding are identified as high, typical and low; they are used to select
matching resistance factors that increase in the order of low to high level of understanding. The choice of
level of understanding captures the confidence of the designer with the choice of the model used to
compute the nominal values for the project conditions, the amount and quality of the project data
including material properties and knowledge of ground conditions, familiarity with the wall technology
proposed and the applicability of the selected wall technology for the project works. Bathurst and
Javankhoshdel (2017) mapped values of COV = 0.10, 0.20 and 0.30 to high, typical and low levels of
understanding, respectively. Bathurst et al. (2017) gave an example how the engineer can self-evaluate to
select the level of understanding that is appropriate for the case of the pullout limit state in a MSE wall
project.

16-21 September 2018, Seoul, Korea

 Proceedings of the 11th International Conference on Geosynthetics

Figure 6. Example of negative correlation between nominal resistance (pullout) and nominal load (maximum
tensile load) using combination of pullout model PM1 and load model LM1

Figure 7. Reliability index b versus nominal factor of safety for tensile rupture and pullout limit states using
different load and resistance models and typical level of understanding
Parameters rR and rQ are Pearson’s correlation coefficients between variables Rn and lR, and between
Qn and lQ, respectively, and represent bias dependencies with nominal values described earlier.
Parameter rn is the correlation coefficient between Rn and Qn and is called nominal correlation following
the terminology introduced by Lin and Bathurst (2018). For the pullout limit state the soil material
properties and their statistical characteristics are the same for the load equation associated with the active
wedge in Figure 1 and the pullout equation associated with the resistant zone. Hence, rn ≠ 0 and will vary
with changes in the distributions for friction angle and unit weight assumed at the location of each
reinforcement layer as demonstrated by Lin and Bathurst (2018).
The mathematical relationship between probability of failure Pf and b has been given earlier in the
paper. However, the advantage of Equation 5 is that it is easily implemented in a spreadsheet and
generates smooth distributions of b when plotted against independent parameters of interest for sensitivity
analyses during design.
Example calculation results are shown in Figure 7. For these calculations the level of understanding
was taken as typical and the nominal correlation coefficient was taken as rn = -0.79 from Figure 6 for the
pullout limit state, and rn = 0 for the tensile strength limit state (i.e. the load equation for Tmax and

16-21 September 2018, Seoul, Korea

 Proceedings of the 11th International Conference on Geosynthetics

Equation 3 for Tal do not have common random variables). For the tensile rupture limit state the resistance
side is the same for both cases identified in the figure (using Equation 3). However, the two load models
for Tmax (LM1 and LM2) described earlier are used in the tensile rupture limit state and in combination
with the two pullout models (PM1 and PM2). Regardless of the choice of combination of load and
resistance models for each limit state, as nominal factor of safety (Fn) increases so does the reliability
index (or equivalently the probability of failure decreases), as expected. However, different load and
resistance model combinations will give different margins of safety expressed as reliability index (or
probability of failure) for the same nominal factor of safety and all other conditions remain the same.
Shown in Figure 7 are values of b = 2.33 and 3.09 that correspond to probabilities of failure of 1/100
and 1/1000, respectively. The smaller b value is recommended as the target minimum reliability index for
internal limits states design and LRFD calibration for MSE walls. This value may appear small but MSE
walls are highly strength-redundant systems. In other words, if one reinforcement layer fails, other layers
can compensate and thus system failure is unlikely (Allen et al. 2005; Bathurst et al. 2008). Figure 7
shows that if a reinforcement layer is designed to just satisfy a target reliability of b = 2.33, the
corresponding factor of safety using the poorer load model (LM1) is 1.70 while the same design with the
more accurate load model (LM2) is higher at 2.50.

5 CONCLUSIONS

This paper is an attempt to describe a rational reliability theory-based approach for analysis and design of
internal limit states for geosynthetic reinforced MSE walls. Nevertheless, the same concepts apply to
other MSE wall types including soil nails and steel reinforced MSE wall systems. The application of the
concepts described in this paper are particularly well-suited to MSE wall internal limit states analysis and
design because the wall backfill soil and soil reinforcement materials are engineered materials, and the
soil-structure limit states can be expressed by simple linear equations with known accuracy. This is not
the case for other soil-structure problems in geotechnical foundation engineering (e.g. piles) where the
soils are natural deposits and thus their properties are much more variable.

REFERENCES

Allen, T. M. and Bathurst, R. J. 2015. An improved simplified method for prediction of loads in reinforced soil
walls. ASCE Journal of Geotechnical and Geoenvironmental Engineering 141(11): 04015049.
Allen, T. M. and Bathurst, R. J. 2018. Application of the Simplified Stiffness Method to design of reinforced soil
walls. ASCE Journal of Geotechnical and Geoenvironmental Engineering 144(5): 04018024.
Allen, T. M., Nowak, A. S. and Bathurst, R. J. 2005. Calibration to Determine Load and Resistance Factors for
Geotechnical and Structural Design. Transportation Research Board Circular E-C079, Washington, DC, 93 p.
AASHTO. 2017. LRFD Bridge Design Specifications, 8th Ed., American Association of State Highway and
Transportation Officials (AASHTO), Washington DC. 73.
Bathurst, R. J., Allen, T. M. and Nowak, A. S. 2008. Calibration concepts for load and resistance factor design
(LRFD) of reinforced soil walls. Canadian Geotechnical Journal 45(10): 1377-1392.
Bathurst, R. J., Huang, B. and Allen, T. M. 2011. Analysis of installation damage tests for LRFD calibration of
reinforced soil structures. Geotextiles and Geomembranes 29(3): 323-334.
Bathurst, R. J. and Javankhoshdel, S. 2017. Influence of model type, bias and input parameter variability on
reliability analysis for simple limit states in soil-structure interaction problems. Georisk 11(1): 42-54.
Bathurst, R.J., Javankhoshdel, S. and Allen, T.M. 2017. LRFD calibration of simple soil-structure limit states
considering method bias and design parameter variability. ASCE Journal of Geotechnical and
Geoenvironmental Engineering 143(9): 04017053-1-14.
Bathurst, R. J. and Miyata, Y. 2015. Reliability-based analysis of combined installation damage and creep for the
tensile rupture limit state of geogrid reinforcement in Japan. Soils and Foundations 55(2): 437-446.
Fenton, G. A., Naghibi, F., Dundas, D., Bathurst, R. J. and Griffiths, D. V. 2016. Reliability-based geotechnical
design in the 2014 Canadian Highway Bridge Design Code. Canadian Geotechnical Journal 53(5): 236-251.
Huang, B. and Bathurst, R. J. 2009. Evaluation of soil-geogrid pullout models using a statistical approach. ASTM
Geotechnical Testing Journal 32(6): 489-504.
Lin, P. and Bathurst, R.J. 2018. Influence of cross-correlation between nominal load and resistance on reliability-
based design for simple linear soil-structure limit states. Canadian Geotechnical Journal 55(2): 279-295.
Miyata, Y. and Bathurst, R. J. 2012. Analysis and calibration of default steel strip pullout models used in Japan.
Soils and Foundations 52(3): 481-497.

16-21 September 2018, Seoul, Korea