Uncertainty modelling and sensitivity analysis of tunnel face stability

Wolfgang Fellina , Julian Kingb , Ansgar Kirscha , Michael Oberguggenbergerc,∗

a Unit of Geotechnical and Tunnel Engineering, University of Innsbruck, Innsbruck, A-6020, Austria
b Breath Research Unit, Austrian Academy of Sciences, A-6850 Dornbirn, Austria
c Unit of Engineering Mathematics, University of Innsbruck, Innsbruck, A-6020, Austria

Abstract
This paper proposes an approach to the choice and evaluation of engineering models with the aid of a typical application in
geotechnics. An important issue in the construction of shallow tunnels, especially in weak ground conditions, is the tunnel face
stability. Various theoretical and numerical models for predicting the necessary support pressure have been put forth in the literature.
In this paper, we combine laboratory experiments performed at the University of Innsbruck with current methods of uncertainty and
sensitivity analysis for assessing adequacy, predictive power and robustness of the models. The major issues are the handling of the
twofold uncertainty of test results and of model predictions as well as the decision about what are the influential input parameters.
Keywords: Tunnel face stability, Model error, Sensitivity analysis, Monte Carlo simulation

1. Introduction
This article addresses the question of model choice and
model adequacy in engineering design, especially in geotech-
nics. Experimental and mathematical methods will be com-
bined to achieve this task. In fact, various types of simplifi-
cations and assumptions have to be introduced in geotechnical
calculations. This can lead to different models for the same
geotechnical problem. These models do not predict the same
system behavior in general.
The question arises how to assess adequacy, predictive power
and robustness of the models. We set out to investigate this
issue using laboratory data on the one hand and methods from
uncertainty analysis on the other hand. Predictive power can be
assessed by comparison of experimental results and theoretical
prediction. Here the uncertainty lies in the experimental results,
the input data of the models and the propagation of uncertainty
to the theoretical output.
Robustness and adequacy of the models can be best under-
stood by means of sensitivity analysis [1, 2, 3, 4]. When
combining experimental data and theoretical models, sampling
based sensitivity analysis – with its recently developed power-
ful statistical indicators – suggests itself as a suitable approach
[5, 6, 7, 8, 9, 10]. As a further important tool in the assessment
of the joint uncertainty of the model parameters, we employ
bootstrap resampling techniques [11, 12, 13, 14].
The construction of shallow tunnels is an engineering chal-
lenge up to the present day. Tunnels with low cover are often
headed using the shield technique. In this context the face sta-
bility is an important issue. In order to minimize settlements
at the ground surface and to prevent failure of the soil ahead of
∗ Correspondingauthor. Tel.: +43 512 507 6824; fax: +43 512 507 2941.
Email address: michael.oberguggenberger@uibk.ac.at (Michael
Oberguggenberger)
Preprint submitted to Structural Safety
the face, the tunnel face must be supported. It has been a long-
standing topic of research how to predict the necessary support
pressure for shield tunnelling. A variety of theoretical and nu-
merical models for estimation of the minimum required support
pressure have been proposed. The theoretical approaches can
be subdivided into kinematic approaches with failure mecha-
nisms (e.g. [15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27])
and static approaches with admissible stress fields (e.g. [23,
28]). Some additional approaches are neither purely kinematic
nor purely static ([29, 30]). We will use some of these models
to exemplify the proposed strategy for assessing the predictive
power of a geotechnical model.
Experimental investigations of face stability range from ex-
periments at single gravity, so-called 1g model tests (e.g. [31,
32, 33, 34]) to centrifuge tests at multiples of g (e.g. [35, 36,
37, 38, 39, 40]). Large scale tests are rare, e.g. [41]). We use
a series of 1g model tests [33, 42] for comparison with the pre-
diction of the chosen theoretical models.
In the theoretical models under scrutiny, the output param-
eter was the necessary support pressure p s . The input (soil)
parameters possessing the largest degree of random variabil-
ity were identified as the actual void ratio e and the loose and
dense state void ratios el and ed , respectively. All these parame-
ters were estimated in small-scale laboratory experiments. An-
other important model parameter is the friction angle ϕ of the
soil. This parameter is estimated by means of a linear model
ϕ ≈ β0 + β1 Id , with the relative density Id (which in turn is
a function of e, el and ed ). In order to assess the influence of
the regression coefficients β0 , β1 on the output p s , we needed
to determine their statistical distribution. We achieved this by
means of the so-called resampling technique, producing a large
bootstrap sample of the experimental data and thereby simu-
lating the joint distribution of β0 and β1 . We believe that this
is a novel method for obtaining joint distributions – including
May 26, 2010
correlations – of geotechnical data.
As a first application of the statistical data model, we could
assess the ranges of the output parameter p s by means of the
First-Order-Second-Moment-Method and compare them with
the experimental results. The model with the best fit was then
scrutinized further: we calculated the sensitivities of the output
p s with respect to the five input parameters described above.
Here we used Monte Carlo simulation based on the input dis-
tributions obtained before. Going beyond the rather crude pic-
ture obtained by scatterplots, we computed stronger statistical
measures of sensitivity, such as partial correlation coefficients.
These indicators are designed so as to remove hidden influences
of co-variates. In addition, this method lends itself to a further
application of resampling, allowing to determine the statistical
significance of the resulting sensitivities. Further, these meth-
ods are applicable in numerical models as well – accordingly,
we included a Finite Element calculation in our list of models.
In short, the goal of the paper is to propose an approach to
model choice and model assessment with the aid of a typical ap-
plication in geotechnics. Experiments play a twofold role here.
On the one hand, 1g-model tests are performed to investigate
the behavior of a tunnel face close to failure. On the other hand,
the outcome of these tests are contrasted with the predictions of
theoretical models. These theoretical models contain material
parameters that in turn are determined from (different) exper-
iments. Thus both the outcome of the 1g-model tests and the
predictions are uncertain. In the presence of this twofold uncer-
tainty, the assessment of model quality requires sophisticated
methods from data analysis and uncertainty analysis. Bootstrap
resampling techniques are used to assess the statistical distribu-
tions of the input parameters, resulting in variability intervals
for the model predictions that can be compared with confidence
intervals of the test results. This enables a comparison of the
range of the predicted output with the range of the measured
output and thus allows to assess the model quality.
We take one further step, once the fittest model has been cho-
sen. This step is sensitivity analysis, determining a ranking of
the input parameters according to their influence on the model
output. The significance of the ranks is again assessed by boot-
strap methods. Highly influential input parameters should be
known more precisely than less influential parameters. This
aids in deciding where to focus the effort in further experimen-
tal, in situ or laboratory investigations. In addition, if the influ-
ence of an input parameter is classified as non-zero, this sup-
ports the structure of the model that considers it as a factor to
be accounted for.
The paper is organized as follows. In Section 2, we briefly
present the theoretical models under investigation. In Section
3, we describe the experimental set-up. Section 4 is devoted to
uncertainty and sensitivity analysis. We formulate the statistical
data models and explain how joint distributions were obtained
by resampling. Then we do the FOSM calculation that leads
to an overall assessment of the predictive power of the mod-
els. This section concludes with the sensitivity analysis based
on Monte Carlo simulations and the statistical indicators men-
tioned above. In Section 5, we discuss the Finite Element model
and the corresponding uncertainty/sensitivity analysis. The fi-
2
nal section summarizes our conclusions.
The methods of uncertainty/sensitivity analysis are based on
our earlier paper [4]; for a general survey of sampling based
sensitivity analysis we recommend [8].
2. Theoretical models
A lot of researchers have put forward theoretical models and
empirical relations to predict the necessary support pressure p s
for tunnels in soft ground. We use five models for our com-
parison, namely those of Horn, Kolymbas, Krause, Léca and
Dormieux, Vermeer and Ruse.
Horn was the first to present a kinematic mechanism with
a sliding wedge for the given problem (Fig. 1). The silo
theory [43] is used to calculate the vertical force acting on
the top of the wedge. Force equilibrium yields the support
force as function of the wedge geometry. The necessary sup-
port force is the maximum value of this function. We use
the original version of the Horn model [15] in our study.
Note that there are various variations of the original model
[16, 17, 18, 19, 44, 20, 22, 45, 22]. They differ by assump-
tions about the lateral earth pressure coefficients used in the
silo theory, the vertical distribution of vertical earth pressure,
cohesional and frictional forces on the top of the wedge, the ap-
proximation of the non-rectangular tunnel cross section and the
shape of the basal boundary of the sliding wedge. An overview
of these variations can be found in [33].
wid
th w gth
l
len
cover C
cuboid
D
prismatic
wedge
diameter D
Figure 1: Geometry of Horn’s rigid block mechanism.
Kolymbas derived a single equation for the support pressure,
which follows from the equilibrium condition by assuming an
admissible stress field above the tunnel and full mobilization of
soil strength at the tunnel crown [29].
Krause considered a single hemispherical body of soil as
three-dimensional failure mechanism. Cohesion and frictional
forces in the contact area build up a resistance against a rota-
tional type of failure. By equilibrating the moments acting on
the body, he derived an expression for the necessary support
pressure p s [22].
Léca and Dormieux [23] presented a sliding wedge mech-
anism of one or two solid conical wedges with circular cross
sections. Making use of the upper bound theorem, they calcu- steel frame model tunnel
lated the support pressure as a function of the inclination angle
α of the cone axis with respect to the horizontal. By maximiz-
ing the resulting pressure over a range of α, expressions for the
necessary support pressure were obtained.
Vermeer and Ruse presented a simple equation [46, 47, 48],
which was developed using Finite Element calculations. The
soil was modelled with a linear elastic, perfectly plastic consti-
tutive model with a Mohr-Coulomb failure condition. Failure
was triggered by a reduction of the pressure on the tunnel face. model tunnel
piston
The principal input parameters of the above models are the
unit weight γ, the cohesion c and friction angle ϕ of the soil, as
Figure 2: Box for 1g-model tests.
well as the overburden C and the diameter D of the tunnel. The
cohesion is set to zero in this investigation, as dry sand was used
in the laboratory experiments [33]. The geometry parameters Table 1: Properties of the applied Ottendorf-Okrilla sand [54].
C and D can be determined with high accuracy, do not change Property Sand S1
throughout the test and are independent of the other parameters. Mean grain size d50 0.58 mm
The dry unit weight γd = γ s /(1 + e) is a function of the unit Coeff. of uniformity U 3.6
weight of the solids γ s and the actual void ratio e. The friction Grain shape (sub)angular
angle ϕ depends generally on the stress level and the density of Particle density ρs 2.635 g/cm3
the soil. While the stress level in the small scale experiments
does not change significantly with the depth we assume the fric-
The tunnel was modelled by a 4 mm thick hollow aluminium
tion angle to be a function of the relative density only ϕ = ϕ(Id ).
cylinder with an inner diameter of Di = 10 cm, which reached
The relative density is
approximately 7 cm into the soil domain. The face of the model
el − e tunnel was supported by an aluminium disk with a slightly
Id = (1)
el − ed smaller diameter than the inner diameter of the tunnel to ex-
wherein the void ratios el (loose state) and ed (dense state) fol- clude friction between disk and tunnel. The piston rod was sup-
low from standardized laboratory experiments. We applied a ported by a linear roller bearing, embedded in the side wall. It
linear relationship ϕ ≈ β0 + β1 Id . allowed for a horizontal displacement of the piston into (and
In summary, we use the d = 5 parameters e, el , ed , β0 and β1 out of) the cylinder, which is denoted by s. The support force
as input for the theoretical models. was measured by a load cell at the outside end of the piston rod
(Fig. 3) while slowly retracting the piston into the cylinder.

3. Experimental investigation
carriage goniometer
In a research project at the Unit of Geotechnical and Tunnel
Engineering 1g-model tests were developed to investigate the
behavior of a tunnel face close to failure [33, 42].
We note that scaling is always a problem in geotechnical ex-
periments. For the face stability experiments the main issues
are: similarity and deterministic size effect (grain size in re-
lation to dimension of the experiment [49, 50] and soil non- load cell

linearity [51]) and stochastic size effects (decreasing shear re- knob
sistance with increasing sample size [52, 53]). A detailed dis-
cussion of these issues is given in [33, 42]. However, mechan-
Figure 3: Carriage construction with load cell, goniometer and turning knob.
ically sound face stability models should perform well at all
scales as long as appropriate input parameters are used. We
put effort into the choice of appropriate parameter values for a
3.2. Sand properties
stress level comparable to the one prevailing in the small scale
experiment. Obviously, for large scale experiments other input Commercially available Ottendorf-Okrilla sand with grain
parameters must be used. The experimental set-up we used is diameters between 0.1 and 2.0 mm was used. Some properties
described in the following subsection. are listed in Table 1.

3.1. Model box 3.3. Preparation of the sandbox experiments

The model box (Fig. 2) consisted of a steel frame with inner The tests were performed with dry sand for various cover-to-
dimensions 37.2×37.2×41.0 (width × depth × height in cm) and diameter ratios (C/D-ratios) and different initial relative densi-
wooden walls. ties Id . Dense samples were prepared by pouring sand into the
3
box with a funnel. With a drop height of approximately 10 cm,
Table 2: Results for the necessary dimensionless support pressure.
layers of 5 cm thickness were installed with an approximately
homogeneous density distribution. These layers were then Test No. C/D material ND
compacted by means of a tamper until a high density was 1-4 0.25 S1 0.11
reached. Loose samples were prepared by carefully putting the 5-13 0.50 S1 0.11
14-17 0.75 S1 0.12
sand into the box with a small shovel, trying to prevent any 18-23 1.00 S1 0.12
compacting action. The void ratio was determined from the to- 24-27 1.50 S1 0.12
tal mass of the sand inside the box and the occupied volume. 28-31 2.00 S1 0.13
The resulting void ratios e and relative densities Id are therefore
average values for the material in the box. Spatial variability
of void ratio and therefore friction angle can have an impact on surrounding ground in such a way that the soil in front of the
the failure mechanism and, hence, face pressure. However, this piston face builds up less pressure on the face.
is currently beyond the ability of the experimental setup. A first
attempt of a kinematic model which considers spatial variability 3.5. Determination of the friction angle ϕ
of the friction angle can be found in [27]. In soil mechanics, it is an accepted fact that ϕ depends on
stress level and initial density of the soil. Standard soil mechan-
3.4. Results of the sandbox experiments ics tests are not suited for pressure levels below, say, 50 kPa.
For each combination of C/D and relative density Id two sep- In the performed experiments stresses around 2 kPa prevailed.
arate tests were performed to check the reproducibility of the Therefore, a series of 75 sand chute experiments were per-
results, which were in total 31 tests. Load-displacement curves formed with the applied sand (Fig. 5). The chute was slowly
for loose specimen and different C/D-ratios can be seen in Fig- tilted until the material slid down. The sand chute has a width
ure 4. The resulting force on the piston was normalized with of 11.7 cm. This is large compared to the grain size of the sand
the area of the piston A, the dry unit weight γd and the piston (0.1 to 2 mm). Therefore, we do assume that the side walls do
diameter D to F/(A γd D) = p/(γd D). not affect the results. All experiments showed an initial slip sur-
face inside the sand layer. So, the influence of the rough chute
bottom should also be negligible. Therefore, we assumed that
0.40
C/D=0.5, I =0.34 an infinite slope model is applicable and interpreted the tilt an-
d
0.35 C/D=0.5, I =0.25
d
gle α at the initiation of the sand avalanche as the peak angle
0.30
C/D=1.0, I =0.33
d ϕpeak of the sand at very low stress level.
C/D=1.0, I =0.33
d We used a simplified method for the estimation of the critical
0.25
state friction angle. The angle of repose of a loose tip of dry soil
p/(γ D)

subjected to toe excavation can serve as an approximation of the

d

0.20

0.15
critical-state angle ϕc [55, 56]. Herle [57] suggested to slowly
ND pour a pile of sand from a funnel and measure the slope of the
0.10
pile. This method is standard for the determination of ϕc for
0.05 the hypoplastic constitutive model of sand [58] and is used in
0.00
the following. From the average angle of repose we estimated
0.0 0.5 1.0 1.5 2.0
s/D (%)
2.5 3.0 3.5 4.0 ϕc = 32.5◦ (from 44 separate measurements, standard deviation
1.1◦ ) for the applied sand, which fits into the typical range for
Figure 4: Evolution of the normalized support pressure ND = p/(γd D) with quartz sand ϕc = 33◦ ± 1◦ [55].
the normalized piston advance s/D.

Figure 4 shows that all curves (for different overburden)

achieve a constant (residual) value after relative displacements
of 2 to 3 %. This value is interpreted as (dimensionless) neces-
sary support pressure ND , because application of a lower pres-
sure (in a pressure-controlled test) would lead to infinite dis-
placements. As ND could not be determined precisely, the re-
sults must be understood as mean values with a precision of
± 0.01. Table 2 compiles results of experiments with different
C/D ratios.
The results convey that the cover-to-diameter ratio does not α
significantly influence the results. This is supported by exper-
imental evidence [23] and numerical results [46]. Arching has
been observed in a Particle Image Velocimetry (PIV) analy-
sis of the failure mechanism [33, 42]. The arch redirects the
weight above the sliding wedge onto the tunnel lining and the Figure 5: Sand chute experiments: tilt angle α at the onset of the sand sliding.

4
4. Uncertainty/Sensitivity analysis
4.1. Input distributions
A linear relationship between ϕ and Id was suggested by [55,
59, 60] and [61]. We thus fitted the linear model
ϕ = β0 + β1 Id + ε, E{ε} = 0, cov{ε} = σ2 I
where E{ε} is the expectation value of the error and cov{ε} its
covariance matrix. The regression model was fitted to the sam-
ple (ϕi , Id,i ), i = 1, . . . , n of size n = 75 obtained by the sand
chute and angle of repose experiments outlined above. Here, it
was assumed that the critical void ratio ecrit at a very low stress
level (as in our small scale tests) can be approximated by the
loose state void ratio el [57, 58, 62, 63, 64, 65]. For e < ecrit
(Id > 0) the friction angle ϕ is a peak friction angle ϕpeak ; for
e = ecrit (Id = 0) peak and critical friction angle are assumed
to be equal, thus ϕpeak = ϕcrit = ϕ. In Figure 6, the measured
friction angles are plotted against the relative density Id .
42
matrix). This is repeated B = 1000 times. We perform a linear
regression on each of these matrices and obtain a population of
n = 1000 pairs of regression coefficients (b0 , b1 ). This popu-
lation is an approximation to the empirical joint distribution of
(β0 , β1 ) and depicted in the upper panel of Figure 7.
From the given scatterplot as well as from the form of
(2) the corresponding marginal distributions we expect a bivari-
ate Gaussian distribution (β0 , β1 ) ∼ N(β, Σβ ), where β and
Σβ are the sample mean, respectively the sample covariance of
(b0 , b1 )i , i = 1, . . . , B. This assumption might be tested by ex-
amining the set
di = ((b0 , b1 )i − β)T Σ−1
β ((b0 , b1 )i − β) (3)
which is well known to follow a χ2 (2)-distribution given
normality [66]. In our case this can be consolidated either
quantitatively (passing a KS-test) or graphically via quantile-
quantile-plots: if di is a sample of a χ2 (2)-distribution, then
 the
sorted values d[i] are estimators of the quantiles χ2 (2)−1 i−0.5
B ,
i.e. the corresponding points should lie on a straight line (see
Figure 7, lower panel).

40
9
38 8
β1

7
ϕ [° ]

36
6
34 32 32.5 33 33.5
β0
32 20
quantiles χ2

p−value KS: 0.568

30 10
0 0.2 0.4 0.6 0.8 1
Id
0
0 5 10 15 20
Figure 6: Sand chute results and linear regression. d[i]

The aforementioned scheme leads to the estimators
b0 = 32.72, b1 = 7.26. Testing for the significance of
regression by means of the standard F-statistic leads to a
clear rejection of the null hypothesis H0 : β1 = 0. Taking
into account the high coefficient of determination (R2 = 0.8)
we hence have strong indications that the linear model is
appropriate.
Figure 7: Empirical joint distribution of (β0 , β1 ) and multivariate normality test.
From the viewpoint of stochastic simulation we hence are
able to realize (β0 , β1 ) ∼ N(β, Σβ ) by employing the usual di-
rect scheme [67], i.e. generation of an uncorrelated bivariate
Gaussian, multiplication by the Cholesky factor of Σβ and addi-
tion of β.

We turn to assessing the distribution of the regression pa- 4.2. FOSM

rameters β0 , β1 . Regressing on the data obtained from the sand
chute experiments supplies us with a single realization of β0 , β1 .
To obtain a sample of the values (β0 , β1 ) and thus have a possi-
bility to assess the bivariate distribution of (β0 , β1 ) we employed
a resampling technique called bootstrapping cases [13]. For
background on resampling in general we refer to [14]. More
specifically, B = 1000 bootstrap samples of size n = 75 each
are generated by the following procedure: From the (n × 2) data
matrix (ϕi , Id,i ) we randomly draw n-times, so that each row has
How do different models propagate the uncertainties of the
input parameters? To answer this question, we employed the
classical First-Order-Second-Moment-Method, thereby obtain-
ing an approximation for mean and variance of the predicted
dimensionless support pressure. More specifically, let f be a
sufficiently smooth model function depending on random vari-
ables Xi (i = 1, . . . , d) with corresponding nominal (mean) val-
ues µi and standard deviations σi . By linearizing f around
µ = (µ1 , µ2 , . . . , µd ), i.e.
equal probability of being drawn. The results are combined to 
Xd
∂ f 

produce a bootstrap (n × 2) matrix (note that some rows of the f ≈ f (µ) +  Xi − µi (4)
bootstrap matrix may be repetitions of rows of the original data i=1
∂Xi µ
5
and exploiting the properties of mean and variance, we arrive at of ND for the Krause model is much narrower than that for the
µ f ≈ f (µ) and Horn model, although they predict roughly the same nominal
 value µND .

X X ∂ f ∂ f 
d  d Xi−1
∂ f  2 2 It becomes also clear from Figure 8 that some of the models
σ2f ≈  σi + 2  σi j , (5)
i=1
∂Xi µ i=1 j=1
∂Xi ∂X j µ fail to predict the obtained experimental results on a 95% con-
fidence level. The best agreement is achieved by the Vermeer-
where σi j denotes possible covariances between the input vari- Ruse model, which was investigated further by a Monte Carlo
ables. Here, the required partial derivatives were determined by study.
numerically differentiating the model function (finite difference Due to the non-linearity of the face pressure models, the best
approximation). model for our small scale tests may not be the best for large
scale tests. However, the decision process would be as outlined
For the given task, the three void ratios were assigned inde- here if large scale tests were used as bases for any model com-
pendent, univariate normal distributions. As mentioned above, parison.
the void ratio e was determined with the total mass of sand and
its occupied volume in the sandbox to be e = 0.65. The mean 4.3. Monte Carlo study
values and standard deviations for the regression parameters β0 To determine the influence of each parameter on the neces-
and β1 were extracted from the estimated bivariate distribution. sary dimensionless support pressure ND predicted by the re-
The mean values and standard deviations of the limit void ra- spective model, we conducted a Monte Carlo study of size
tios el and ed are based on 20 separate laboratory experiments N = 1000, with the common input samples generated accord-
for each limit. We assume that the variation of e is similar to ing to the postulated probability distributions. Care was taken
that of el . All values are summarized in Table 3. As β0 and β1 to ensure that the physical constraints ed,i ≤ el,i and ed,i ≤ ei
are correlated, also the covariance σβ0 β1 and correlation coeffi- hold for each sample. Correlation control was applied in order
cient ρβ0 β1 were important; from the empirical joint distribution to keep the sample correlation matrix associated with (ed , el , e)
they were quantified as σβ0 β1 = −0.0388 and ρβ0 β1 = −0.6, re- close to diagonal [68].
spectively. Generally, the output statistics obtained for each model
closely resemble those calculated via FOSM, thus reconfirm-
Table 3: Statistical parameters for the five input parameters.
ing our previous ranking.
Consequently, results will be presented for the model of
Par. mean µi std. dev. σi distribution type Vermeer-Ruse only. We shall stress the fact, however, that all
ed 0.48 0.03 normal schemes presented below are equally applicable for all models
el 0.71 0.03 normal
discussed and give very similar qualitative results.
e 0.65 0.05 normal
β0 32.7 0.16 normal A first scan of the influence of the individual input parame-
β1 7.28 0.4 normal ters on the output can be done by means of scatterplots (Fig. 9).
A decidedly positive correlation between e and ND is visible, a
Figure 8 illustrates the result of the FOSM calculations for negative correlation between el and ND appears to be present,
the different models. The error bars indicate approximate 95% but is less obvious. However, such plots might be misleading as
confidence intervals based on the calculated model variance. they cannot take into account hidden interactions between the
input parameters under scrutiny.
0.20

4.4. Partial correlation coefficients

0.15 We therefore turn to a method which intends to remove the
influence of co-variates on the correlation between a given input
variable Xi and the output variable Y. This method is based on
ND

0.10
the partial correlation coefficient (PCC). In our case, the model
output is Y = ND , the five input variables X1 , . . . , X5 are ed , el ,
0.05 e, β0 and β1 .
In order not to overload the notation, we use the generic
0 X and Y for random variables (rather than the specific de-
Horn Leca Ruse notations ed etc.). The partial correlation between two ran-
Krause Kolymbas
dom variables Xi and Y given a set of co-variates Xri =
{X1 , . . . , Xi−1 , Xi+1 , . . . , Xd } is defined as the correlation between
Figure 8: Variation of predicted ND for distributed input data; experimental the two residuals sXi ·Xri and sY·Xri obtained by regressing Xi on
results shaded in grey.
Xri and Y on Xri , respectively. More precisely, one first con-
structs the two regression models
Apart from the different predictions of the mean µND (central X X
points in Figure 8), it becomes obvious that the different models bi = ξ0 +
X ξ j X j, b Y = η0 + η j X j, (6)
process the input uncertainties differently: e.g. the distribution j,i j,i

6
 already available are used in this calculation, empirical confi-
dence intervals for the PCCs can be obtained with a minimum
0.16 0.16
of additional computational cost.
0.14 0.14

0.12 0.12 1

0.1 0.1
0.5
0.4 0.45 0.5 0.55 0.65 0.7 0.75 0.8
ed el
0
0.16 0.16

0.14 0.14 −0.5

0.12 0.12

0.1 0.1 −1
ed el e β0 β1

0.5 0.6 0.7 0.8 32.2 32.7 33.2

e β0 Figure 10: PCCs and corresponding 95% bootstrap confidence intervals of five
input parameters ed , el , e, β0 , β1 with output ND (Vermeer-Ruse).
0.16

0.14 Accordingly, all PCCs test to be nonzero and hence can be

regarded as assertive. As above, el and e are recognized as the
0.12
most influential factors. On could suppose that the smallness
0.1 of the PCCs of β0 and β1 (though they are still significantly
different from zero) could originate from the fact that they enter
6.5 7 7.5 8
β1 as correlated variables in the model. However, in this case the
scatterplots (Figure 9) confirm the weakness of the correlation
with the output ND .

Figure 9: Scatterplots of five input parameters ed , el , e, β0 , β1 versus output ND

(Vermeer-Ruse). 5. Finite Element model

For reasons of comparison, we subjected a Finite Element

obtaining the residuals
bi ,
sXi ·Xri = Xi − X sY·Xri = Y − b
Y.
Since sXi ·Xri and sY·Xri are those parts of Xi and Y that remain
after subtraction of the best linear estimates in terms of Xri , the
partial correlation coefficient
ρXi ,Y·Xri = ρ(sXi ·Xri , sY·Xri )
quantifies the linear relationship between Xi and Y after removal
of any part of the variation due to the linear influence of Xri .
As compared to the usual correlation coefficient, the partial cor-
relation coefficient is a much more discriminating, if not to say
radical, measure of influence. In fact, if the input-output map is
a truly linear function, the PCC of an input variable that enters
with a non-zero coefficient is equal to one. On the other hand, if
two input variables are multiples of each other, their PCCs with
the output are zero. In reality, input-output maps are not ideally
linear functions and so the effect is somewhat moderated. Still
the PCCs are an accentuating measure of influence. For further
background on PCCs and computational issues see [8, 69].
Calculation of the PCCs is based on the original Monte Carlo
sample and the associated support pressures. Again, we em-
ployed bootstrapping to gain further statistical information on
this sensitivity index (Fig. 10). As only support pressure values
model of the sandbox experiments to the same uncertainty and
sensitivity analysis. As the FOSM cannot yield any information
(7) about the distribution of the predicted ND values, we turned to
Monte Carlo simulation performing N = 1000 Finite Element
calculations with the same input sample as in the preceding sec-
tion.
For the FE calculations, described in the following, the linear
(8) elastic, perfectly plastic Mohr-Coulomb model was used. This
model is frequently applied to geotechnical boundary value
problems. The elastoplastic Mohr-Coulomb model is imple-
mented in the Finite Element code A, which was used for
the simulations.
It requires five input parameters:
• Young’s modulus E and Poisson’s ratio ν describe the ma-
terial behavior in the elastic domain.
• The friction angle ϕ, cohesion c and dilation angle ψ gov-
ern the plastic behavior of the material.
The self-weight of the soil γd and the friction angle ϕ were
calculated from the input sample exactly in the same way as
described in the previous section. The Young modulus was es-
timated at E = 825 kPa, which accounts for the low stress levels
prevailing in the sand specimen. The cohesion c was neglected
for dry sand, the dilation angle assumed to be ψ = 4◦ . Poisson’s
7
ratio was derived from ϕ with the help of Jaky ’s empirical for-
mula, K0 ≈ 1 − sin ϕ, and an expression for K0 from elasticity
theory
ν 1 − sin ϕ
K0 = Ã ν≈ .
1−ν 2 − sin ϕ
In the model (Fig. 11) soil and piston were modelled as in-
dependent parts, which interacted via frictionless contact el-
ements. This way the laboratory conditions were best repro-
duced.
Figure 11: Finite element mesh with 1438 elements (piston not shown).
The tunnel lining was not modelled explicitly; instead the
nodes on the perimeter of the tunnel were fixed in all directions,
thus representing a rigid lining. During the numerical simula-
tion the piston was retracted into the tunnel and the resultant
force on the piston was logged.
The ND values of the Finite Element calculations were de-
rived from the support force after 0.5 mm piston displacement,
when all the curves had reached a well-defined plateau.
As before, we calculated the partial correlation coefficients
between input and the necessary dimensionless support pres-
sure estimated by the FE model (Fig. 12).
1
0.5
0 probability methods for sensitivity analysis in engineering: A case study,
−0.5
−1
ed el e β0 β1
Figure 12: PCCs and corresponding 95% bootstrap confidence intervals of five
input parameters ed , el , e, β0 , β1 with output ND (FE model).
As can readily be seen, the FE model and the Vermeer-Ruse
model exhibit very similar sensitivities with respect to the un-
certain soil parameters.
8
6. Conclusions
In this paper, we demonstrated how statistical estimators in
conjunction with laboratory experiments can be used to assess
the predictive power and robustness of models in geotechnical
engineering. We performed a case study with a number of the-
oretical and numerical models for the required support pressure
for shield tunnelling. Ultimately, this may serve as a basis for
determining the safety of the tunnel face during construction
and for gaining deeper understanding of the relevant mecha-
nisms leading to failure.
In particular, we showed that novel statistical methods con-
stitute a powerful tool for identifying the most influential model
parameters, and we showed that a comparison of models is pos-
sible in terms of analyzing the uncertainty of the model answer
and contrasting it with test results, which also contain uncer-
tainties. It turned out that bootstrapping is a suitable method
for producing samples of multi-dimensional random variables
and for assessing the significance of the computed sensitivities.
This is particularly convenient in a Finite Element analysis, be-
cause it does not require additional computational cost beyond
the original Monte Carlo simulation.
We remark that the sensitivities obtained in the Vermeer-
Ruse model and the FE-model turned out to be quite similar.
This may be explained by the fact that the Vermeer-Ruse model
was originally calibrated on a similar FE-model.
Acknowledgements
The first two authors (W. F., A. K.) gratefully acknowledge
support by the Tyrolean Science Foundation, the third au-
thor (J. K.) is recipient of a DOC-fellowship of the Austrian
Academy of Sciences at the Breath Research Unit.
References
[1] C. Bucher, Robustness analysis in structural optimization, Structure and
Infrastructure Engineering 5 (4) (2009) 287–293.
[2] H. Konietzky, R. Schlegel, Using sensitivity analysis and optimization
for civil engineering and geotechnical applications, in: Proceedings Op-
timization and Stochastic Days 4.0, Weimar, http://www.dynardo.com,
2007.
[3] M. Oberguggenberger, W. Fellin, Reliability bounds through random sets:
Non-parametric methods and geotechnical applications, Computers &
Structures 86 (2008) 1093–1101.
[4] M. Oberguggenberger, J. King, B. Schmelzer, Classical and imprecise
International Journal of Approximate Reasoning 50 (2009) 680–693.
[5] G. Archer, A. Saltelli, I. Sobol, Sensitivity measures, ANOVA-like tech-
niques and the use of the bootstrap, Journal of Statistical Computation
and Simulation 58 (1997) 99–120.
[6] S. Ferson, W. Tucker, Sensitivity analysis using probability bounding, Re-
liability Engineering & System Safety 91 (2006) 1435–1442.
[7] J. Hall, Uncertainty-based sensitivity indices for imprecise probability
distributions, Reliability Engineering & System Safety 91 (2006) 1443–
1451.
[8] J. C. Helton, J. D. Johnson, C. J. Sallaberry, C. B. Storlie, Survey of
sampling-based methods for uncertainty and sensitivity analysis, Relia-
bility Engineering & System Safety 91 (2006) 1175–120.
[9] V. Kreinovich, J. Beck, C. Ferregut, A. Sanchez, G. Keller, M. Aver-
ill, S. S.A., Monte-Carlo-type techniques for processing interval uncer-
tainty, and their potential engineering applications, Reliable Computing
13 (2007) 25–69.
[10] I. Sobol, Global sensitivity indices for nonlinear mathematical models and
their Monte Carlo estimates, Mathematics and Computers in Simulation
55 (2001) 271–280.
[11] P. Good, Resampling methods: a practical guide to data analysis, 2nd
Edition, Birkhäuser, Boston, MA, 2001.
[12] C. Lunneborg, Data analysis by resampling: concepts and applications,
Duxbury, Pacific Grove, CA, 2000.
[13] D. Montgomery, E. Peck, G. Vining, Introduction to linear regression
analysis, Wiley, New York, 2001.
[14] J. Shao, D.-S. Tu, The Jackknife and the Bootstrap, Springer-Verlag, New
York, 1995.
[15] M. Horn, Horizontaler Erddruck auf senkrechte Abschlussflächen von
Tunneln, Landeskonferenz der ungarischen Tiefbauindustrie (Deutsche
Überarbeitung STUVA), Düsseldorf (1961).
[16] G. Anagnostou, K. Kovári, Ein Beitrag zur Statik der Ortsbrust beim Hy-
droschildvortrieb, in: Symposium ’92, Probleme bei maschinellen Tun-
nelvortrieben?, Gerätehersteller und Anwender berichten., 1992.
[17] G. Anagnostou, K. Kovári, Face stability conditions with earth-pressure-
balanced shields, Tunnelling and Underground Space Technology 11 (2)
(1996) 165–173.
[18] G. Girmscheid, Tunnelbohrmaschinen - Vortriebsmethoden und Logistik,
in: Betonkalender, Chap. 1.3, Ernst & Sohn, Berlin, 2005, pp. 119–256.
[19] P.-M. Mayer, U. Hartwig, C. Schwab, Standsicherheitsuntersuchungen
der Ortsbrust mittels Bruchkörpermodell und FEM, Bautechnik 80 (2003)
452–467.
[20] J. Holzhäuser, Problematik der Standsicherheit der Ortsbrust beim TBM-
Vortrieb im Betriebszustand Druckluftstützung, in: Beiträge anlässlich
des 50. Geburtstages von Herrn Professor Dr.-Ing. Rolf Katzenbach,
no. 52 in Mitteilungen des Institutes und der Versuchsanstalt für Geotech-
nik, Darmstadt University of Technology, 2000, pp. 49–62.
[21] P. Vermeer, N. Ruse, Z. Dong, D. Härle, Ortsbruststabilität von Tunnel-
bauwerken am Beispiel des Rennsteig-Tunnels, in: Tagungsband 2. TAE
Kolloquium ”Bauen in Boden und Fels”, 2000, pp. 195–202.
[22] T. Krause, Schildvortrieb mit flüssigkeits- und erdgestützter Ortsbrust,
Mitteilung des Instituts für Grundbau und Bodenmechanik, Technische
Universität Braunschweig, no. 24, 1987.
[23] E. Leca, L. Dormieux, Upper and lower bound solutions for the face
stability of shallow circular tunnels in frictional material, Géotechnique
40 (4) (1990) 581–606.
[24] A.-H. Soubra, Three-dimensional face stability analysis of shallow circu-
lar tunnels, in: Int. Conf. on Geotechnical and Geological Engineering,
Melbourne, Australia, November 19-24, 2000, pp. 1–6.
[25] A.-H. Soubra, Kinematical approach to the face stability analysis of shal-
low circular tunnels, in: 8th Int. Symp. on Plasticity, British Columbia,
Canada, 2000, pp. 443–445.
[26] A.-H. Soubra, D. Dias, F. Emeriault, R. Kastner, Three-dimensional face
stability analysis of circular tunnels by a kinematical approach, in: Geo-
Congress 2008: Characterization, Monitoring, and Modeling of GeoSys-
tems (GSP 179), New Orleans, 2008, pp. 894–901.
[27] G. Mollon, K. Phoon, D. Dias, A.-H. Soubra, A new 2d failure mecha-
nism for face stability analysis of a pressurized tunnel in spatially variable
sands, in: GeoFlorida 2010: Advances in Analysis, Modeling & Design
(GSP 199), ASCE, Reston, 2010, pp. 2052–2061.
[28] J. H. Atkinson, D. M. Potts, Stability of a shallow circular tunnel in cohe-
sionless soil, Géotechnique 27 (2) (1977) 203–215.
[29] D. Kolymbas, Tunnelling and Tunnel Mechanics, Springer-Verlag, Berlin,
2005.
[30] H. Balthaus, Standsicherheit der flüssigkeitsgestützten Ortsbrust bei
schildvorgetriebenen Tunneln, in: Festschrift H. Duddeck, Institut für
Statik der Technischen Universität Braunschweig, Springer, Berlin, 1988,
pp. 477–492.
[31] D. Takano, J. Otani, S. Fukushige, H. Natagani, Investigation of interac-
tion behavior between soil and face bolts using x-ray ct, in: J. Desrues,
G. Viggiani, P. Bèsuelle (Eds.), Advances in X-ray Tomography for Geo-
materials, ISTE Ltd., London, 2006, pp. 389–395.
[32] D. Sterpi, A. Cividini, A physical and numerical investigation on the sta-
bility of shallow tunnels in strain softening media, Rock Mech. and Rock
Engng. 37 (4) (2004) 277–298.
[33] A. Kirsch, On the face stability of shallow tunnels in sand, Logos Ver-
lag, Berlin: Advances in Geotechnical Engineering and Tunneling (ed. D.
Kolymbas), 2009.
9
[34] D. Subrin, D. Branque, N. Berthoz, H. Wong, Kinematic 3D approaches
to evaluate TBM face stability: Comparison with experimental labora-
tory observations, in: EURO:TUN 2009 – Computational methods in tun-
nelling, 2009, pp. 801–807.
[35] P. Chambon, J.-F. Corté, J. Garnier, D. König, Face stability of shallow
tunnels in granular soils, in: H.-Y. Ko., F. McLean (Eds.), Centrifuge 91,
Balkema, Rotterdam, 1991, pp. 99–105.
[36] P. Chambon, J.-F. Corté, Shallow tunnels in cohesionless soil: stability of
tunnel face, ASCE Journal of Geotechnical Engineering 120 (7) (1994)
1148–1165.
[37] R. Al Hallak, J. Garnier, E. Léca, Experimental study of the stabil-
ity of a tunnel face reinforced by bolts, in: O. Kusakabe, K. Fujita,
Y. Miyazaki (Eds.), Geotechnical Aspects of Underground Construction
in Soft Ground, Balkema, Rotterdam, 2000, pp. 65–68.
[38] J. W. Plekkenpol, J. S. van der Schrier, H. J. Hergarden, Shield tun-
nelling in saturated sand – face support pressure and soil deformations,
in: A. Bezuijen, H. van Lottum (Eds.), Tunnelling: A Decade of Progress,
GeoDelft 1995-2005, Taylor & Francis, London, 2006.
[39] H. Kamata, H. Mashimo, Centrifuge model test of tunnel face rein-
forcement by bolting, Tunnelling and Underground Space Technology 18
(2003) 205–212.
[40] T. Kimura, R. J. Mair, Centrifugal testing of model tunnels in soft soil, in:
Proc. 10th Int. Conf. Soil Mech. Found. Engng., Stockholm, Vol. 1, 1981,
pp. 319–322.
[41] C. Renpeng, L. Jun, B. Xuecheng, C. Yunmin, Large scale experimental
investigation on face stability of shallow tunnels in dry cohesionless soil,
in: EURO:TUN 2009 – Computational methods in tunnelling, 2009, pp.
809–816.
[42] A. Kirsch, Experimental investigation of the face stability of shallow tun-
nels in sand, Acta Geotechnica 5 (1) (2010) 43–62.
[43] H. A. Janssen, Versuche über Getreidedruck in Silozellen, Zeitschrift des
Vereins Deutscher Ingenieure 39 (35) (1895) 1045–1049.
[44] A. Kirsch, D. Kolymbas, Theoretische Untersuchung zur Ortsbruststa-
bilität, Bautechnik 82 (7) (2005) 449–456.
[45] M. Mohkam, Y. W. Wong, Three dimensional stability analysis of the
tunnel face under fluid pressure, in: G. Swoboda (Ed.), Proc. 6th Int.
Conf. on Numerical Methods in Geomechanics, Innsbruck, Vol. 4, 1988,
pp. 2271–2278.
[46] N. M. Ruse, Räumliche Betrachtung der Standsicherheit der Ortsbrust
beim Tunnelvortrieb, Mitteilungen des Instituts für Geotechnik, Univer-
sität Stuttgart, no. 51, 2004.
[47] P. Vermeer, N. Ruse, Die Stabilität der Tunnelortsbrust in homogenem
Baugrund, Geotechnik 24 (3) (2001) 186–193.
[48] P. Vermeer, N. Ruse, T. Marcher, Tunnel heading stability in drained
ground, Felsbau 20 (6) (2002) 8–18.
[49] Technical Commitee 2 of ISSMGE – Physical Modelling in Geotechnics,
Catalogue of scaling laws and similitude questions in centrifuge mod-
elling, 2007.
[50] P. Chambon, A. Couillaud, P. Munch, A. Schürmann, D. König, Stabilité
du front de taille d’un tunnel: Étude de l’effet d’échelle, in: Geo 95, 1995,
p. 3, cited in [49].
[51] D. Muir Wood, Geotechnical Modelling, Spon Press, London, 2004.
[52] F. Tatsuoka, S. Goto, T. Tanaka, K. Tani, Y. Kimura, Particle size ef-
fects on bearing capacity of footing on granular material, in: A. Asaoka,
T. Adachi, O. F. (Eds.), Deformation and Progressive Failure in Geome-
chanics, Pergamon, New York, 1997, pp. 133–138.
[53] J. Tejchman, Fe-simulations of a direct wall shear box test, Soils and
Foundations 44 (4) (2004) 67–81.
[54] A. Laudahn, An Approach to 1g Modelling in Geotechnical Engineering
with Soiltron, Advances in Geotechnical Engineering and Tunnelling, no.
11, Logos, Berlin, 2004.
[55] M. Bolton, The strength and dilatancy of sands, Géotechnique 36 (1)
(1986) 65–78.
[56] D. H. Cornforth, Prediction of drained strength of sands from relative
density measurements. In: Evaluation of relative density and its role
in geotechnical projects involving cohesionless soils, in: ASTM Spec.
Techn. Publ. 523, American Society for Testing and Materials, Philadel-
phia, 1973, pp. 281–303.
[57] I. Herle, Hypoplastizität und Granulometrie einfacher Korngerüste,
Veröffentlichungen des Institutes für Bodenmechanik und Felsmechanik
der Universität Fridericana Karlsruhe, no. 142, 1997.
[58] I. Herle, G. Gudehus, Determination of parameters of a hypoplastic
constitutive model from properties of grain assemblies, Mechanics of
Cohesive–Frictional Materials 4 (1999) 461–486.
[59] S. Fukushima, F. Tatsuoka, Strength and deformation characteristics of
saturated sand at extremely low pressures, Soils and Foundations 24 (4)
(1984) 30–48.
[60] E. De Beer, Influence of the mean normal stress on the shearing strength
of sand, in: Proc. 6th Int. Conf. Soil Mech. Fndn. Eng., Montreal, Vol. 1,
1965, pp. 165–169.
[61] W. Wu, Hypoplastizität als mathematisches Modell zum mechanischen
Verhalten granularer Stoffe, Veröffentlichungen des Institutes für Boden-
mechanik und Felsmechanik der Universität Fridericana Karlsruhe, no.
129, 1992.
[62] A. Alarcon-Guzman, G. Leonards, J. Chameau, Undrained monotonic
and cyclic strength of sands, Journal of Geotechnical Engineering
144 (10) (1988) 1089–1109.
[63] K. Ishihara, Liquefaction and flow failure during earthquak,
Géotechnique 43 (3) (1993) 351–415.
[64] R. Mohamad, R. Dobry, Undrained monotonic and cyclic triaxial strength
of sand, Journal of Geotechnical Engineerin 112 (10) (1986) 941–958.
[65] R. Verdugo, K. Ishihara, The steady state of sandy soils, Soils and Foun-
dations 36 (2) (1996) 81–91.
[66] T. W. Anderson, An introduction to multivariate statistical analysis, Wi-
ley, New York, 1958.
[67] L. Devroye, Non-uniform random variate generation, Springer-Verlag,
Berlin, 1986.
[68] R. Iman, W. Conover, A distribution-free approach to inducing rank cor-
relation among input variables, Communications in Statistics - Simulation
and Computation 11 (1982) 311–334.
[69] R. Iman, M. Shortencarier, J. Johnson, A FORTRAN 77 program and
user’s guide for the calculation of partial correlation and standardized re-
gression coefficients, SANDIA Report SAND85-0044, Tech. rep., Sandia
National Laboratories, Albuquerque (1985).

10