Computers and Geotechnics 68 (2015) 38–53

Contents lists available at ScienceDirect

Computers and Geotechnics

journal homepage: www.elsevier.com/locate/compgeo

Research Paper

A probabilistic analysis of subsoil parameters uncertainty impacts

on tunnel-induced ground movements with a back-analysis study
S. Miro a,⇑, M. König a, D. Hartmann a, T. Schanz b
a
Ruhr-Universität Bochum, Faculty of Civil and Environmental Engineering, Chair of Computing in Engineering, 150, Universitätsstr., 44780 Bochum, Germany
b
Ruhr-Universität Bochum, Faculty of Civil and Environmental Engineering, Chair for Foundation Engineering, Soil and Rock Mechanics, 150, Universitätsstr., 44780 Bochum, Germany

a r t i c l e i n f o a b s t r a c t

Article history: In this paper a probabilistic-based analysis is presented for evaluating the influences of subsoil parameter
Received 5 January 2015 uncertainties on tunnel-induced ground movements in mechanized tunneling. The procedures of the
Received in revised form 25 February 2015 tunneling process using Slurry Shield Tunnel Boring Machine are numerically modeled and simulated
Accepted 28 March 2015
by utilizing a finite element code. To keep the computational cost of the presented simulation model
Available online 11 April 2015
low, an efficient and reliable surrogate modeling technique is used to substitute the original simulation
model. The input parameter uncertainties are mathematically represented by adequately chosen proba-
Keywords:
bility density functions within their extreme lower and upper bounds. Subsequently, a variance-based
Mechanized tunneling
Finite element simulation
global sensitivity analysis is conducted for quantifying the impact of each uncertain parameter on
Surrogate modeling different system responses that are considered in this study. Afterwards, the propagation of parameter
Global sensitivity analysis uncertainties are evaluated by performing a Monte Carlo-based simulation using the computationally
Reliability analysis inexpensive surrogate model. At this stage, the variations of system responses, which result from input
Bayesian updating parameters propagating uncertainties, are compared with predetermined threshold values and, based
on that, failure criteria of the tunneling system are defined as well as probabilistically quantified. In a last
step, a Bayesian updating procedure is employed for reducing subsoil parameter uncertainties by
utilizing recorded synthetic measurements.
Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction by several researchers for the prediction of these movements

The safety and stability of tunnels as well as surface and subsur-
face adjacent structures and facilities are of utmost importance for
a successful tunnel project. Therefore, the ground movements
initiated by the tunneling process should be minimized and not
be larger than admissible limits. In particular, large movements
of the ground next to the tunnel can result in differential displace-
ments of nearby structures and, consequently, can cause damages
and fractures with the risk of failure.
In shallow tunnels, in particular, the tunneling-induced ground
movements represent dangers and are a challenge that requires
control and mitigation. Hereby, the movements of the ground are
consequences of the redistribution of stresses and changes in pore
water conditions around the tunnel face and tunnel cavity [1,2].
These movements are functions of the hydro-geological conditions
of the ground, tunnel depth, tunnel geometry, and the excavation
method [3]. A variety of analytical methods have been developed
⇑ Corresponding author.
E-mail address: shorash.miro@rub.de (S. Miro).
[4–11,2,12]. Furthermore, with the advancement of computational
methods and computing power, numerical approaches are being
widely used in simulating tunnel excavations and the consequent
ground behavior. Among others, the reader is referred to [13–18].
In comparison with the analytical methods, the numerical meth-
ods, in particular 3D finite element simulations, are able to provide
a comprehensive view of stress and strain distributions in the
ground domain affected by the tunnel. Moreover, the numerical
methods can explore the impact of additional supporting measures
(such as anchoring and soil exchange) that are employed to further
stabilize the tunnel.
In mechanized tunneling using tunnel boring machines (TBM)
different methods and technologies are employed for reducing
the ground displacements by supporting the tunnel face and cav-
ity. Technologies such as mechanical support, compressed air,
earth pressure balance, and slurry support are employed to back
up the tunnel face [1]. The tunnel cavity, on the other hand, is sup-
ported by a TBM-shield followed by precast concrete segments
together with grouting material injected to fill the gap between
the segments and the excavated ground [1]. With these

http://dx.doi.org/10.1016/j.compgeo.2015.03.012
0266-352X/Ó 2015 Elsevier Ltd. All rights reserved.
 S. Miro et al. / Computers and Geotechnics 68 (2015) 38–53
technologies the natural stress state of soil can be retained to a
large extent and, as a result, the settlements can be reduced.
In spite of the aforementioned supporting approaches, ground
movements are especially significant in shallow tunnels in soft
soils. Here, the mechanical properties of the soil play a substantial
role on these movements and their predictions. In particular, the
soil characteristics are associated with inevitable uncertainties
that result from: (i) the natural variability of soils, (ii) the measur-
ing errors in quantifying the soil properties by site and laboratory
tests, and (iii) the diminutive fraction of the investigated soil
volume in comparison with the whole affected soil domain.
According to [19], the fraction of the investigated soil using bore-
holes is only about 106 to 109 of the total site volume. In addition
to that, accessibility to sufficient site exploration might be
restricted because of surface overly or lack of permission [20]. To
this end, studying the propagation of uncertainties in subsoil
mechanical properties to tunnel-induced ground movements is
essential for a reliable assessment of these movements.
Furthermore, a systematic reduction of the aforementioned
uncertainties by employing an appropriate inverse analysis meth-
odology is vital for the tunnel project success. In this respect, back
analysis methods in geotechnical engineering were first
formulated and utilized in [21,22]. They were used, moreover, for
estimating and identifying system parameters in many geotechni-
cal applications, such as [23–28].
Mollon et al. [29] present a comprehensive study and metho-
dology for evaluating the influences of soil parameter uncertainties
on tunnel-induced ground movements that were predicted using a
3D finite element simulation. In this study we use an alternative
approach for predicting the tunnel-induced ground movements
with an advanced elasto-plastic soil constitutive model with iso-
tropic hardening, namely the Hardening Soil Model (HS-Model)
[30,31]. Moreover, the representation of the uncertain soil proper-
ties as random variables with probability density functions is dis-
cussed on the basis of introducing lower and upper extreme
values. Additionally, a Bayesian back analysis approach [32] is
adopted to reduce the uncertainties of soil properties during tunnel
excavation and, consequently, to reduce the potential risk of uncer-
tainty propagation to the ground movements. This work extends a
recent paper [18] focused on a global sensitivity analysis of the
same mechanized tunnel computational model.
The structure of this paper is outlined as follows: Section 2
includes a detailed description of the 3D finite element simulation
of the shallow tunnel along with the predicted ground settlements
and displacements. The probabilistic representation of input
parameter uncertainties is discussed in Section 3. In Section 4 a
surrogate model based on quadratic polynomial regression is intro-
duced along with the variance-based global sensitivity analysis. In
Section 5, a reliability analysis is performed for evaluating the
impact of subsoil parameter uncertainty representations on the
probability of failure through limit state functions. As a last step,
in Section 6, a sequential Bayesian back analysis approach is
conducted to update the subsoil parameter uncertainties based
on synthetic data. The paper ends with a summary and conclusions
in Section 7.
2. Mechanized tunnel 3D finite element simulation
2.1. Description of the simulation model
A three dimensional finite element software PLAXIS 3D, version
2012, is employed for simulating the mechanized tunneling
process. In this simulation the excavation of homogeneous soil
by means of a Slurry Shield Tunnel Boring Machine is modeled.
Herein, an advanced elasto-plastic soil constitutive model with
39
isotropic hardening, the Hardening Soil Model (HS-Model), is used.
The parameters of the HS-Model are described in Table 1.
Fig. 1 illustrates the 3D FE-model of a shallow tunnel 90 m long
in the X-axis direction, 45 m wide in the Y-axis direction, and 45 m
deep in the Z-axis direction. These dimensions represent only one-
half of the subsoil model, due to the symmetry of geometry, mate-
rial properties, and initial and boundary conditions with respect to
the vertical plane X–Z normal to the Y-axis. A discretization with a
total number of 22,542 10-node tetrahedral elements and 34,841
nodes has been adopted after a preliminary study of the impact
of different mesh sizes. The excavation length of this tunnel equals
60 m and its diameter D is 8.5 m. As the overburden is equal to
1  D, the considered tunnel corresponds to a very shallow tunnel
that is expected to result in large ground movements in real world
projects [29]. The groundwater table is not considered in this
study, however, it can easily be included in the numerical sim-
ulation. The shield of the TBM is taken to be 9 m long and simu-
lated, besides the tubing that consists of annular precast concrete
segments (1.5 m), as circular plate elements with a linear elastic
model. The material properties of both the lining elements and
the TBM-shield are given in Table 2.
Modeling the effect of the grouting material, which is injected
in the gap between the excavated soil and the erected lining seg-
ments, is performed by a non-uniformly distributed load increas-
ing with the depth. This load is directly applied to the soil
elements that are located between the TBM shield and the last
assembled tubing section. In a similar manner, the face support
pressure, which is applied to counteract any active failure ahead
of the TBM, is modeled as a non-uniformly distributed load that
grows linearly from the tunnel crown to its invert. The contact area
between the shield skin and the tubing with the adjacent soil is
modeled by a 40% reduced shear strength of the enclosing soil.
Hereby interface elements [33] are utilized. The tunnel excavation
is performed by a step-wise procedure (45 steps). Further details
about the developed simulation model can be found in [18,34].
2.2. Tunneling-induced ground movements
Most of the existing literature dealing with tunneling-induced
ground movements discusses the distribution of surface settle-
ments as a result of the tunneling process. These settlements, in
the case of tunnel excavation in homogeneous soil, shape a trough
that can be described by the longitudinal settlement along the tun-
nel axis, as well as the transversal settlement along a cross-section
perpendicular to the tunnel axis. This is illustrated in Fig. 2. The
maximum value of these settlements occurs at the tunnel axis
behind the TBM-face in the stabilized area of the soil. In addition
to surface settlements that constitute the major part of the ground
movement induced by the tunneling process, other movements
Table 1
The description of hardening soil model parameters.
HS-model parameters Description
/ (°) Friction angle
w (°) Dilatancy angle
c (kN/m2) Cohesion
Tangent stiffness for primary oedometer loading
Eref
oed
(kN/m2)
Eref 2 Secant stiffness in standard drained triaxial test
50 (kN/m )
Unloading reloading stiffness
Eref
ur
2
(kN/m )
p (kN/m2)
ref Reference stress level
m (–) Exponent of the Ohde/Janbu law
Rf (–) Failure ratio
m (–) Poisson’s ratio
cunsat (kN/m3) Unit weight (unsaturated)
Rinter (–) Strength reduction factor for interfaces
40 S. Miro et al. / Computers and Geotechnics 68 (2015) 38–53

Y X
8.5m

D = 8.5m
45m

Fig. 1. Three dimensional finite element model for mechanized tunnel simulation.

Table 2 and deformations were investigated in [29]. Specifically, the slopes

Linear elastic material properties that are used for the TBM-shield and tunnel lining. and distortions of the settlement trough as well as the horizontal
Parameter Description Tunnel lining TBM-shield displacements and deformations (compression and extension) in
linear elastic linear elastic direction of and perpendicular to the tunnel axis were discussed.
E (kN/m2) Young’s modulus 3  107 21  107 In this paper, the settlements, slopes, displacements, and deforma-
m (–) Poisson’s ratio 0.2 0.3 tions in direction of and transversal to the tunnel axis are
c (kN/m3) Unit weight 24 38 considered.
d (m) Thickness 0.2 0.35 Using PLAXIS 3D software and the HS-Model, the adopted
ground movements are computed and depicted in Figs. 3 and 4.

Settlement Longitudinal
Transversal trough cross section
cross section

Z
Y
X

Fig. 2. Illustration of the ground surface settlement trough with the longitudinal and transversal cross sections.
 S. Miro et al. / Computers and Geotechnics 68 (2015) 38–53 41

Longitudinal direction
0.002

0
Displacement (m) −0.002 S2
−0.004

−0.006

−0.008

−0.01
S1 uX
uZ
−0.012
15 25 35 45 55 65 75 85
X (m)
0.0005

0.0004
Deformation or Slope (m/m)

S4
0.0003

0.0002
S7
0.0001

−0.0001
S6 XX
ZZ
−0.0002
15 25 35 45 55 65 75 85
X (m)
Fig. 3. Vertical and horizontal displacements as well as deformations with their extreme values along the longitudinal direction.

These illustrated settlements and horizontal displacements repre-
sent the total values at the 35th excavation step, i.e. when the tun-
nel face is 52.5 m away from the tunnel start. Boundary influences
are excluded by considering a transversal cross-section in 15 m (10
excavation steps) distance from the tunnel start. The stabilization
of the soil is reached in around 22.5 m (15 excavation steps)
beyond the TBM face, and the decompression of the soil is seen
approximately 15 m ahead of the tunnel front. The slopes and
deformations are estimated by numerical differentiation of the dis-
placements since the strains are not large and only compression
and extension are considered. Because the described movements
are repeated in further excavation increments, only 35 steps need
to be utilized for predicting and assessing these movements.
The maximum values of the adopted ground displacements and
strains are symbolized by nine output variables S1 ; . . . ; S9 , as
indicated in Figs. 3 and 4. Herein, S1 is the maximal surface
settlement, and S2 as well as S3 are the maximal horizontal
displacements in the planes: normal to X-axis at Y ¼ 0 and normal
to Y-axis at X ¼ 15 respectively. S4 and S5 are the extreme values of
the slopes along the longitudinal and transversal directions, and
S6 ; . . . ; S9 are the ultimate compression and extension values of
the ground surface in the aforementioned planes.
3. Representation of subsoil parameter uncertainties
As already mentioned in the introduction, the tunneling-
induced ground movements substantially depend on, among other
factors, the mechanical properties of the excavated soil. These
properties, however, are subject to different sources of uncertain-
ties due to natural variabilities, measuring errors, transformation
42 S. Miro et al. / Computers and Geotechnics 68 (2015) 38–53

Transversal direction
0.004
0.002 S3
0
Displacement (m)

−0.002
−0.004
−0.006
−0.008
−0.01 S1 uY
uZ
−0.012
0 10 20 30 40
Y (m)
0.001
S5
0.0008
Deformation or Slope (m/m)

0.0006 S8

0.0004
0.0002
0
−0.0002
−0.0004
−0.0006 S9 Y Y
ZZ
−0.0008
0 10 20 30 40
Y (m)
Fig. 4. Vertical and horizontal displacements as well as deformations with their extreme values along the transversal direction.

uncertainties, or statistical uncertainties related to lack of available
information [35,36]. Therefore, an appropriate description of these
uncertainties and evaluation of their propagation to ground move-
ments are substantial for a reliable shallow-tunnel project. A com-
prehensive and thorough review of the variability of soil properties
has been presented and discussed in [35–38].
The probabilistic modeling of soil properties is customarily per-
formed to take into account the uncertainties of relevant proper-
ties. This modeling is carried out by using random variables or
random fields [38]. In the first case, the random variables are
defined by their mean, coefficient of variation COV, and their
probability distribution function. These quantities are considered
to be constant within the same layer and, therefore, the layer is
regarded as homogeneous. On the other hand, random fields are
utilized to capture the spatial variabilities (spatial correlations) of
soil properties, i.e., to model inhomogeneous soils. As a first step,
the homogeneous case has been considered in this work.
Nevertheless, the authors are aware of spatial correlations impor-
tance, in particular, the affected soil volume is relatively large in
tunnel projects.
The simulation models that are utilized for predicting the
ground movements are functions of different factors that can be
treated as input variables in the models. These variables can be,
for instance, related to the geometry of the ground, the numerical
procedure for solving the problem, modeling the excavation pro-
cess, the material model used for describing the soil behavior, or
material model parameters that often represent soil properties.
Indeed, all the input factors are associated with uncertainties that
affect the predictions obtained from the models. Therefore, to
evaluate the influence of soil properties variations on the ground
 S. Miro et al. / Computers and Geotechnics 68 (2015) 38–53 43

movements, these properties are modeled as random variables,
while the other inputs are considered to be deterministic.
Nevertheless, the uncertainty of the inputs regarded as deter-
ministic should be taken into consideration. For instance, the influ-
ences of the discretized mesh along with the dimensions of the
ground model on the system responses can be evaluated in a pre-
liminary study similar to [18]. Hereby, the responses obtained from
a model with relatively large dimensions and a fine mesh are taken
as reference values. Subsequently, the degree of mesh refinement
and the ground dimensions of the adopted model are taken such
that its system responses are sufficiently close to the reference val-
ues. Of course, this strategy can be followed up to investigate and
minimize the influence of the size and the number of the excava-
tion steps.
The constitutive parameters of HS-Model that are considered to
be uncertain are /; w; c; Eref ref ref
50 ; Eoed ; Eur ; m; m and c. Here, m is the expo-
nent of the Janbu law and does not represent a physical quantity.
However, because its value is uncertain, it is also included in the
study. Two equality constraints are considered: Eref ref
50 ¼ Eoed and
w ¼ /  30 if / > 30; w ¼ 0 otherwise. Consequently, the number
of the independent parameters reduces from 9 to 7. The uncertain-
ties of these parameters are modeled by probability density func-
tions (PDFs), namely normal (N) and lognormal (LN)
distributions, along with their mean values and coefficients of
variations. Hereby, the PDFs themselves are regarded within
extreme lower and upper bounds. These bounds are chosen such
that they represent, based on engineering judgement, realistic pos-
sible extreme values for the properties of the soil under explo-
ration. Moreover, the adopted bounds must encompass all the
samples that are obtained from field and laboratory tests. As this
study is based on a synthetic case, the bounds of the parameters
are selected to represent realistic extreme values of a medium
dense sand with some amount of capillary cohesion. Two proba-
bilistic scenarios, an optimistic and a pessimistic one, are proposed
in this study to convey the degree of knowledge that is available for
the explored soil (in [29] a neutral case was also proposed to repre-
sent an average knowledge that lies between the two extreme sce-
narios). Hereby, the normal and lognormal distributions of the
parameters are considered in both scenarios. In the pessimistic
scenario an additional case of uniform distributions is also consid-
ered to be an excessively pessimistic case (with no prior knowl-
edge) that might result in areas with no access to soil
exploration. In Table 3 the mean values, coefficients of variation,
and the ranges of the uncertain parameters are given for the
adopted scenarios.
presented in the previous section, performing this probabilistic
analysis becomes prohibitive. This is because of the large amount
of evaluations of the forward model that are required to estimate
failure probabilities or to perform a variance-based sensitivity
analysis for the uncertain soil properties. Therefore, utilizing a sur-
rogate modeling technique to substitute the computationally
expensive tunnel simulation is inevitable. In this paper, a response
surface methodology based on quadratic polynomials is utilized.
The procedure for constructing the surrogate model works as fol-
lows: the system responses that are used to estimate the displace-
ment curves in Figs. 3 and 4 are the displacement values at the
discretization nodes located at the ground surface in the planes
at Y ¼ 0 and X ¼ 15 for the 35th excavation step. From these sys-
tem responses the output variables S1 ; . . . ; S9 are assembled in an
output vector that needs to be approximated using surrogate mod-
eling. For that, a response surface (a surrogate) is built for each
component of the output vector using the quadratic polynomial
regression technique [39]. In total, 100 support (training) points
uniformly distributed in the parameter ranges along with their
corresponding outputs are utilized to construct the individual sur-
rogate models. After that, further 100 forward runs for different
input values are applied to validate their accuracy. This validation
is based on calculating the coefficient of determination R2 [39],
which equals to 1 for a perfect representation of the original
model. In Table 4, R2 values for all output variables are shown.
4.2. Global sensitivity analysis
A variance-based global sensitivity analysis is conducted to
evaluate the importance of the input parameters with respect to
the nine output variables S1 ; . . . ; S9 . For this purpose, the total effect
sensitivity index STi [40] is employed and computed according to
[41]:
ðyB  yCi ÞT ðyB  yCi Þ
STi ¼ ; ð1Þ
B Þ2
2yTB yB  2nðy
where yB and yCi are vectors containing model evaluations for
matrices B and Ci respectively. Matrix B contains n randomly gener-
ated samples of the input parameters, and the matrix Ci is con-
structed such that its columns are copied from the matrix B
except its ith column copied from the corresponding column in
matrix A, which contains another n randomly generated samples
of the input parameters. y B is the mean value of the components
of yB .

4. Surrogate modeling and sensitivity analysis

Table 4
4.1. Surrogate modeling with quadratic polynomial regression The accuracy measure of the constructed surrogate models.

S1 S2 S3 S4 S5 S6 S7 S8 S9
Due to the computational effort needed in the 3D mechanized
R2 0.997 0.991 0.999 0.997 0.999 0.998 0.993 0.999 0.994
tunnel finite element simulation (forward model), which has been

Table 3
Uncertainty representation of HS-model parameters that convey a medium dense sand with some amount of capillary pressure.

Parameter Range Mean COV (%) Type of PDF

Optimistic Pessimistic
/ (°) ½30; 45 36 5 15 N/LN
c (kN/m2) ½0; 10 5 10 30 N/LN
Eref (kN/m2) ½20; 000; 40; 000 30,000 10 20 N/LN
oed
Eref 2 ½100; 000; 200; 000 150,000 10 20 N/LN
ur (kN/m )
m (–) ½0:4; 0:7 0.55 5 10 N/LN
m (–) ½0:1; 0:3 0.2 10 20 N/LN
cunsat (kN/m3) ½14; 20 17 5 15 N/LN
44 S. Miro et al. / Computers and Geotechnics 68 (2015) 38–53

By the adopted sensitivity index, the contribution of each input
parameter to the overall variance of the regarded system response
is estimated taking into consideration its interaction effects to the
remaining input parameters. In other words, the model response
uncertainty is apportioned to the input parameter uncertainties.
As the sum of total effect indices of all input parameters is larger
or equal to 1 (equal to 1 only in case of additive models with no
interactions between the input parameters), these indices are nor-
malized in order to obtain the contribution of each input parame-
ter to the total output variance as a percentage. More details about
employing variance-based sensitivity analysis in mechanized tun-
nel simulation can be found in [18].
5. Probabilistic reliability analysis
The steps of the probabilistic analysis conducted in [29] are
adopted here for evaluating the influences of ground model
parameter uncertainties on the tunneling induced ground
movements. At first, the coefficients of variation of the output vari-
ables, which result from the input variations, are estimated.
Subsequently, the estimated variations are assigned to the input
parameter variations through total effect sensitivity indices.
Finally, the probabilities of failure are computed through limit
state functions.
5.1. Coefficients of variation
The propagation of uncertainties from the input parameter
space to the spaces of the nine output variables is determined by
performing a Monte Carlo simulation. Starting from the scenario
that portray our knowledge about the input parameters, a sample
of 105 points from the input space is drawn and their correspond-
ing output variables are computed through the constructed surro-
gate models. The coefficients of variation (COVs) of the output
variables are calculated for the five considered cases, i.e., optimistic
and pessimistic scenarios with normal and lognormal distributions
as well as uniform distributions for the extreme pessimistic case.
This is illustrated in Fig. 5. Apparently, the variations of the output
variables are larger in the pessimistic scenario than in the
optimistic one. However, the differences between output varia-
tions in the cases of normally and lognormally distributed input
variables are significantly small in both optimistic and pessimistic
scenarios. In the optimistic case the adopted lower and upper
extreme values encompass at least three standard deviations from
both sides of the mean values, i.e., more than 99% of the generated
samples lie within their extreme bounds. In spite of that, changing
the shape of distribution of the input parameters from normal to
lognormal causes merely minor reductions in the output varia-
tions. However, in the pessimistic case the shape of distribution
affects the output variations differently, which might be due to
cut tails that lie outside the defined borders. Altogether, the influ-
ence of the distribution type (normal or lognormal) of the input
parameters on the output variations is low in the present model.
Nevertheless, this influence can be more significant based on
the COV values of the input variables. In this case, it is essential
that the larger the COV is, the more significant the difference in
shape between lognormal and normal distributions becomes.
Consequently, the change of distribution type from normal to log-
normal can result in larger or smaller output variations depending
on the underlying computational model. In the case of no available
knowledge about the input variables, i.e., uniformly distributed
variables, maximal variations of the output variables are unsur-
prisingly obtained.
The variations of the output variables S1 ; . . . ; S9 (see Fig. 5)
respond differently to the input uncertainties. In particular, the
maximum settlement S1 shows a larger variation in comparison
to the horizontal displacements (S2 and S3 ) as well as to the slopes
in the longitudinal and transversal directions (S4 and S5 ). On the
other hand, the deformations along the longitudinal and transver-
sal directions demonstrate large variations compared with their
corresponding horizontal displacements. Most likely, this is due
to the change of position of the maximum displacements S2 and
S3 whereas their amplitudes display no significant deviations.
5.2. Sensitivities
Having estimated the variations of the output variables, the
total effect sensitivity indices for the input parameters are

45
Optimistic Normal
Optimistic Lognormal
40 Pessimistic Normal
Pessimistic Lognormal
35 Uniform
COV (%)

30

25

20

15

10
S1 S2 S3 S4 S5 S6 S7 S8 S9
System responses
Fig. 5. The coefficients of variation of the output variables based on knowledge scenario and input variable distributions.
 S. Miro et al. / Computers and Geotechnics 68 (2015) 38–53
computed based on knowledge scenario and the applied probabil-
ity distribution functions. Here, the variance of the output vari-
ables are apportioned to the contributions of the input parameter
uncertainties taking into consideration their coupling terms.
Since the differences between the results obtained by normal and
lognormal distributions for each scenario are small, only the
sensitivity results of (a) the optimistic scenario with normal dis-
tributions and (b) the pessimistic scenario using both normal as
well as (c) uniform distributions are depicted in Fig. 6.
The obtained sensitivity results indicate that input parameter
sensitivities can only be slightly affected by the probabilistic sce-
nario, which is the case in the output variable S1 . On the other
hand, a non-sensitive parameter in the optimistic case might
become relatively important in the case of uniform distributions,
such as the sensitivity of cohesion c in variable S8 . This clearly
demonstrates that the sensitivity of a parameter depends on its
uncertainty together with the uncertainties of the other parame-
ters. The maximum ground settlement S1 is primarily affected by
the variations of the friction angle / along with the primary load-
ing stiffness Eoed . This can be anticipated by the loading condition
of the soil. The maximum horizontal displacement in the longitu-
dinal direction S2 demonstrates a significant sensitivity to the unit
weight c followed by the primary loading stiffness Eoed and the
unloading reloading stiffness Eur . Even the poison ratio m shows a
noticeable sensitivity compared to its uncertainty. This means that
the elastic parameters have large sensitivities to S2 . It implies that
the horizontal displacement along the longitudinal axis is primar-
ily a result of the elastic response to the face support pressure.
Herein, the sensitivity of Eur is explained by the unloading process
in the longitudinal direction as the face support pressure is pass-
ing. The maximum horizontal displacement in the transversal
direction S3 , on the other hand, shows a large sensitivity to Eur fol-
lowed by /. This can be explained by the settlement curve along
the transversal direction for ten random samples of the input
parameters in Fig. 7. One can notice that, for most of these curves,
an upheave occurs at a distance of around 15 m from the origin.
This behavior, which is related to the influence of the applied grout
100
80
60
ST i (%)
40
20
0
abc abc abc abc abc abc
S1 S2 S3 S4 S5
System responses
Fig. 6. Normalized total effect sensitivity indices of the input parameters for: (a) Optimistic scenario with normal distributions. (b) Pessimistic scenario with normal
distributions. (c) Uniform distributions.
45
pressure and the relative shallow depth of the tunnel, is apparently
the reason for the horizontal displacement curve along this direc-
tion (see Fig. 4). By this, it can be concluded that S3 is substantially
led by this upheave and the unloading condition of the soil.
Therefore, the unloading reloading stiffness Eur together with the
dilatancy angle w (which is embedded here in the friction angle
because of the equality condition w ¼ /  30 for / P 30) have
the largest sensitivity measures. It can be also observed that the
maximum slopes S4 and S5 have the same sensitive parameters
as the settlements. Moreover, in longitudinal deformations the
large influence of the friction angle / is expected to be due to
the impact of plastic deformations around the tunnel on the shape
of the displacement curve in this direction. In the transversal case,
on the other hand, the influence of / is probably related to the
embedded dilatancy angle. In both cases, the sensitivity of Eoed is
larger in the compression state (S7 and S9 ) than in the tension state
(S6 and S8 ), because of the loading condition of the soil.
To this end, the sensitivity results indicate that solely four
parameters, i.e. Eur ; Eoed ; /, and c (implicitly E50 and w), have the
largest sensitivity measures. As a consequence, the variation of
the output variables can be lessened by reducing the uncertainties
of the corresponding sensitive input parameters. This can be
achieved by further exploration and testing of the soil under exca-
vation or by an appropriate back analysis procedure as explained
later in Section 6. If only the settlement of the ground surface is
of interest, then its variation can be drastically reduced by decreas-
ing the uncertainty of the friction angle and the primary loading
stiffness. It should clearly be noted that the sensitivity results are
essentially dependent on the relative dispersions of input parame-
ters, i.e. the sensitivity of a parameter will decrease by reducing its
uncertainty [18].
5.3. Failure probabilities
This section investigates the impact of subsoil parameter uncer-
tainties on the failure probability of the tunnel. In reliability analy-
sis, the failure of a given system is defined through a failure event F
Eoed
m
γ
c
φ
ν
Eur
abc abc abc
S6 S7 S8 S9
46 S. Miro et al. / Computers and Geotechnics 68 (2015) 38–53
Fig. 7. Settlement curves along the transversal direction for ten random samples of the input parameters.
described by a functional relation, the limit state function gðxÞ, as
follows [42]:
F ¼ gðxÞ < 0;
where x is a set of values of the uncertain system inputs that are
modeled as random variables. In other words, the failure event F
is the set of negative realizations of the limit state function.
Subsequently, the probability of failure PF is obtained by the integral
Z
PF ¼ f X ðxÞdx;
gðxÞ<0
where f X ðxÞ is the joint probability density function of the random
variables X (subsoil properties in this study). This integral is esti-
mated through a Monte Carlo simulation.
Herein, five failure modes are defined based on the output vari-
ables and are assigned different limit state functions. The first
mode A describes the event in which the maximal ground surface
settlement exceeds an admissible threshold f:
g A ðxÞ ¼ f  S1 ðxÞ:
The second failure mode B is related to the maximal horizontal dis-
placements, along both longitudinal and transversal directions, that
overreach an allowable threshold d:
g B ðxÞ ¼ d  maxðS2 ðxÞ; S3 ðxÞÞ:
By the third failure mode C the event in which the maximal slopes
of the settlement trough S4 and S5 surpass an admissible value m is
described:
g C ðxÞ ¼ m  maxðS4 ðxÞ; S5 ðxÞÞ:
The last two failure modes D and E are linked to the maximal hori-
zontal deformations in the longitudinal and transversal directions.
Here, two different thresholds t and c are assigned to the exten-
sion and compression states respectively, taking into account that
structures respond differently to extension and compression condi-
tions. The corresponding limit state functions are
g D ðxÞ ¼ t  maxðS6 ðxÞ; S8 ðxÞÞ; ð7Þ
g E ðxÞ ¼ c  maxðjS7 ðxÞj; jS9 ðxÞjÞ: ð8Þ
ð2Þ
Each of the above defined failure modes represents only one aspect
of the total system under investigation. Nevertheless, the system as
a whole fails if only a single mode fails. In other words, the system
failure is equal to the union of the five single failure modes. That is,
mathematically speaking, if at least one of the five single limit state
functions is negative then the system fails. This behavior is inter-
preted by an overall failure event described by the following limit
ð3Þ state function g System ðxÞ:
g System ¼ minðg A ; g B ; g C ; g D ; g E Þ: ð9Þ
A Monte Carlo simulation is performed to estimate the failure
probabilities for each defined single mode and for each scenario
and type of distribution of the random variables. For that, a range
of 100 admissible thresholds are defined for each mode. Then the
corresponding five individual failure probabilities are computed
by using 105 samples. In Fig. 8, the threshold ranges are plotted
ð4Þ against failure probabilities on a logarithmic scale. Obviously, as
the plots demonstrate, the failure probabilities increase drastically
from the optimistic scenario to the pessimistic one. However, the
influence of distribution type between normal and lognormal, as
it is the case in the coefficients of variation (Fig. 5), is not notice-
ð5Þ able. Only in the failure mode B, associated with the optimistic sce-
nario, do the lognormal distributions yield a slightly lower failure
probability.
For arbitrarily chosen thresholds in the five single mode limit
state functions, the corresponding failure probabilities as well as
ð6Þ the failure probability of the system are computed by using 105
samples for the pessimistic case with normal PDFs, and given in
Table 5. As can be seen, the failure probability of the system is lar-
ger than the maximum of the five single failure probabilities. This
is because the sample points that cause the system failure equals
the union of the samples causing the failure of each single mode.
That is to say, by evaluating Eq. (9) for N samples in the Monte
Carlo simulation, the sample set N system # N that causes the system
 S. Miro et al. / Computers and Geotechnics 68 (2015) 38–53 47

Failure mode A Failure mode B

1 1

0.1 0.1

0.01 0.01
Pf

Pf
0.001 0.001

0.0001 0.0001
4 6 8 10 12 1.75 2.25 2.75 3.25 3.75
ζ (mm) δ (mm)

Failure mode C Failure mode D

1 1

0.1 0.1

0.01 0.01
Pf

Pf

0.001 0.001

0.0001 0.0001
0.4 0.55 0.7 0.85 1 0.3 0.4 0.5 0.6 0.7
p (mm/m) t (mm/m)

Failure mode E
1

0.1 Legend

0.01 Normal Opt.

Pf

Lognormal Opt.
Normal Pes.
0.001 Lognormal Pes.
Uniform
0.0001
0.1 0.2 0.3 0.4 0.5
c (mm/m)
Fig. 8. Single-mode probability failures plotted in a logarithmic scale against a range of admissible thresholds and based on the type of distribution as well as the probabilistic
scenario.

failure (i.e., returning negative values in Eq. (9)) is given by reliable estimation of the probability failure of a system, an
N System ¼ N A [ N B [ N C [ N D [ N E . Consequently, N system is larger or adequate description and selection of single mode failures is
equal to the largest single mode sample set. Based on that, for a decisive.
48 S. Miro et al. / Computers and Geotechnics 68 (2015) 38–53

Table 5 samples from the posterior PDF. The implementation of MCMC is

A comparison between the probability failure of the tunnel as a whole system and its conducted by the Metropolis–Hastings algorithm [43,44], which
single-mode probability failures for arbitrarily chosen admissible thresholds and a
pessimistic scenario with normal PDFs.
works as follows: Given the target probability distribution pðxjdÞ,
a Markov chain with a given sample xk is considered, then the next
Failure mode Threshold value Probability of failure
sample xkþ1 is obtained from xk according the subsequent steps:
A 8.3 mm 0.33
B 2.8 mm 0.3369
C 0.7 mm/m 0.4527
1. Generate a candidate sample x from a jumping probability den-
D 0.53 mm/m 0.43982 sity function Pðx jxk Þ.
E 0.32 mm/m 0.3514 2. Calculate
System 0.5625
 
pðx jdÞPðxk jx Þ
a ¼ min 1; : ð14Þ
pðxk jdÞPðx jxk Þ
6. Probabilistic back analysis
3. Generate a uniformly distributed sample U 2 ð0:0; 1:0Þ.
4. If U 6 a accept xkþ1 ¼ x , otherwise xkþ1 ¼ xk .
Apparently, as it is shown in the probabilistic analysis of failure,
5. Check convergence. If not fulfilled, return to step 1.
the larger the uncertainties of the soil properties, the larger the
failure probabilities. Therefore, lowering relevant uncertainties is
This algorithm generates a sequence of samples fx1 ; x2 ; . . . ; xn g,
essential for improving the reliability of a tunnel project, and for
which represent the posterior PDF. The jumping probability dis-
reducing the potential risk of uncertainty propagation to the
tribution is taken to be Gaussian in this work, therefore,
ground movements. Based on that, the utilization of inverse analy-
sis techniques, by employing recorded measurements during tun- Pðxk jx Þ=Pðx jxk Þ is equal to one due to the symmetry of the dis-
neling, serves as a vital methodology for improving engineering tribution. In this case the algorithm is known as the Metropolis
knowledge about the values of soil properties. For this purpose, sampler.
the Bayesian back analysis technique (Bayesian updating) is In order to ensure that the obtained Markov chain is a represen-
conducted. tation of the probabilistic solution (the posterior PDF), it is essen-
In Bayesian back analysis the state of knowledge that we have tial to monitor the convergence of the sequence. For this
about the uncertain input parameters is represented through a purpose, a convergence diagnostics based on [45] is implemented.
prior multivariate probability distribution function pprior ðxÞ. This This convergence monitoring works as follows:
prior knowledge is combined with the likelihood function LðdjxÞ
 Start m parallel sequences, each of length n (after discarding
(herein d is a vector of recorded system responses) that gives a
half of the samples which correspond to the so-named burn-in
measure of how good a model is in explaining the measurements
period).
d. As a result, the probabilistic solution of the inverse problem,
the posterior probability density function pposterior ðxÞ, is obtained  For the drawn samples xij (i ¼ 1; . . . ; n; j ¼ 1; . . . ; m) and for each
by [32]: estimated parameter compute the between- and within-se-
quence variances B and W:
pprior ðxÞ  LðdjxÞ
pposterior ðxÞ ¼ pðxjdÞ ¼ ; ð10Þ
n X m
1X n
1X m
pðdÞ B¼ ðxj  xÞ2 ; where xj ¼ xi ; x ¼ xj ;
m  1 j¼1 n i¼1 j m j¼1
where the denominator is the marginal density function over the
data space, which works as a normalization constant for the poster- ð15aÞ
ior probability. 1X m
1 X n
2
Here, the back analysis is considered particularly for the pes- W¼ s2 ; where s2j ¼ ðxi  xj Þ : ð15bÞ
m j¼1 j n  1 i¼1 j
simistic scenario with independent and normally distributed input
parameters. For this, pprior ðxÞ is a multivariate normal distribution  Estimate the marginal posterior variance varðxjdÞ for each
function: parameter by:
1 n1 1
pprior ðx1 ; . . . ; xp Þ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi d
v ar þ ðxjdÞ ¼ W þ B: ð16Þ
ð2pÞp jRx j n n
 
1  As the sample size n increases, the within-sequence variance
 exp  ðx  lÞT R1
x ðx  lÞ ; ð11Þ
2 W will approach varðxjdÞ for each parameter. This is moni-
tored by the potential scale reduction
herein, Rx is the covariance matrix of the input parameters. For a
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
given state of the input variables x the difference (error) between
d
b¼ v ar þ ðxjdÞ
a measured and a simulated system response is: R ; ð17Þ
W
ui ðxÞ  di ¼ i ; i ¼ 1; . . . ; q; ð12Þ
which declines to 1 as n ! 1.
where q is the number of measurements. The errors are considered
to be independent and normally distributed Nð0; r2i Þ and, conse- b becomes close to 1 for all parameters, the mn sim-
As soon as R
quently, the likelihood function becomes:
ulations from the second halves of all sequences are taken into
 
1 1 T 1 consideration as a representative sample of the posterior PDF.
LðdjxÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi exp  ðd  uðxÞÞ R d ðd  uðxÞÞ ; ð13Þ From sensitivity analysis results in Section 4.1, it is proved that
ð2pÞq jRd j 2
the four parameters Eur ; Eoed ; /, and c (implicitly E50 and w) have
where Rd is the observation covariance matrix. the largest impact on the variation of the output variables.
Due to the difficulty of determining the normalization constant Therefore, and in order to reduce the dimension of the identifica-
pðdÞ, a computational sampling technique, namely Markov chain tion problem, only these parameters are taken into consideration
Monte Carlo MCMC, is used to draw a sequence of dependent in the back analysis process. The system responses that are utilized
 S. Miro et al. / Computers and Geotechnics 68 (2015) 38–53 49

as measurement data d (observations) for the identification pur-
pose are the vertical displacements at four specified observation
points. These points are located at the cross-sections A and B that
are illustrated in Fig. 9. The displacements are considered starting
from the 30th excavation step, in order to exclude small values and
to avoid boundary effects, and ending at the 42nd where they
reach a still stand (see Fig. 10). A convenient distance from the tun-
nel start (36 m) is taken for cross-section A in order to avoid
boundary effects. The distance between the two sections is held
large enough (9 m) to cover a sufficiently ample domain. The
observation points are chosen based mainly on the study in [18].
That is, the uncertain parameters show sensitivity to the system
response of interest at the observation points and, consequently,
can be identified. In Fig. 11 the total effect sensitivity indices of
the seven parameters are plotted in the range of the excavation
steps 30 to 42. At the upper point A1 , which is subject to settle-
ment, the friction angle / has a large sensitivity, followed by the
primary loading stiffness Eoed , while the remaining parameters
show almost no sensitivity. By contrast, the unloading reloading
stiffness Eur shows a significant sensitivity at the lower point A2 ,
due the unloading condition at the tunnel invert, followed by /
and c. The same behavior of the settlements and sensitivities

B1
A1

B2
A2

Fig. 9. Observation points for recording vertical displacements during TBM advancement.

0.008
A1
0.006 A2

0.004
0.002
0
UZ (m)

−0.002
−0.004
−0.006
−0.008
−0.01
5 10 15 20 25 30 35 40 45
Excavation Step
Fig. 10. Total vertical displacements at observation points A1 and A2 .
50 S. Miro et al. / Computers and Geotechnics 68 (2015) 38–53

A1
Eur
0.8 φ
γ
Eoed
ν
0.6 c
m
ST i

0.4

0.2

0
30 32 34 36 38 40 42
Excavation Step

A2
Eur
0.8 φ
γ
Eoed
ν
0.6 c
m
ST i

0.4

0.2

0
30 32 34 36 38 40 42
Excavation Step
Fig. 11. Total effect sensitivity index of the parameters calculated for the total vertical displacements at points A1 and A2 and excavation steps between 30 and 42.

occurs at the points B1 and B2 with a 6 steps (9 m) delay. Its worth are saved. Subsequently, the saved system responses are exposed
mentioning that, for each excavation step, the system responses at to two cases of random noise in order to associate measurement
the observation points are replaced by surrogate models in a simi- uncertainties. In the first case, a coefficient of variation
lar fashion to Section 4.1. COV d ¼ 5% is considered to add low noise to the measurements.
The synthetic measurements for the identification problem are In the second case, a significant noise with COV d ¼ 15% is added
generated as follows: the tunneling simulation model is run for a to mimic distinct measurement uncertainties. In a subsequent
chosen set of input parameters provided in Table 6. After that, step, the non-sensitive input parameters (m; c and m) are arbitrarily
the corresponding vertical settlements at the observation points fixed within their defined ranges, while the remaining four sensi-
tive parameters (Eur ; Eoed ; /, and c) are regarded as unknown vari-
Table 6 ables. These unknown variables are estimated from the two
Input parameter values that are used for generating the synthetic measurements. generated synthetic measurements by conducting the back analy-
Parameter c m m cunsat sis procedure. Here, a successful back calculation will give estima-
/ Eref Eref
ur
(°) (kN/m2)
oed
(kN/m2) (–) (–) (kN/m3) tions of the four unknown parameters close to their reference
(kN/m2)
values (the values for which the synthetic measurements are gen-
Value 38 5 25,000 135,000 0.6 0.22 18
erated) in Table 6.
 S. Miro et al. / Computers and Geotechnics 68 (2015) 38–53 51

Likelihood
Monitored data
function

In-situ &
Prior PDF Bayes Theorem Posterior PDF
laboratory tests

Fig. 12. A sequential Bayesian back analysis for reducing the uncertainties of soil properties using recorded system responses.

The back analysis process is performed in a sequential manner.
This is elucidated in Fig. 12, and can be summarized in a few steps:
firstly, the prior PDF is formulated based on the state of knowledge
about the uncertain soil properties. After that, the recorded mea-
surements of two subsequent excavation steps (starting from step
30), which constitute a vector d of 8 components, are used to con-
struct the likelihood function. Subsequently, the posterior PDF is
obtained by Eq. (10) and sampled by using MCMC method. The
generated sample is used to update the prior PDF, while the
recorded measurements for two subsequent tunnel advancements
are utilized to build a new likelihood function. Then, consequently,
the posterior PDF is updated. This procedure is repeated until the
last regarded excavation stage.
By carrying out the adopted back analysis approach the input
parameter uncertainties are reduced sequentially according to
the tunnel advancement. This is illustrated in Fig. 13, assuming
that COV d equals 5%. The results achieved after the last back analy-
sis sequence are given in Table 7. As expected, it can be seen from
Table 7 that for more accurate measurements the uncertainties of
the identified parameters are smaller. Furthermore, the estimated
means are generally closer to the reference values. In addition to
that, the uncertainty of parameters with a large sensitivity (Eur
and /) is reduced more than in the case of less sensitive ones (c
and Eoed ). Of course, the convergence of the implemented MCMC
method is achieved in all stages of the back analysis process. As
an example, the convergence of the first sequence of the identifica-
tion, where COV d equals 5%, is plotted in Fig. 14. It should be men-
tioned that the lower and upper bounds of the input parameters
are included in the sampling procedure by rejecting the samples
that lie outside the ranges.
The back analysis strategy introduced above has proved to be a
powerful tool in employing field observations for the purpose of
reducing subsoil parameter uncertainties. In particular, this
approach is able to take into consideration the noise associated
with the recorded measurements in the parameter identification
process. Nevertheless, the advancements in sensing and monitor-
ing technologies as well as the data-rich environment of modern
mechanized tunnels can be utilized in providing more accurate
and precise measurements of the system response. In addition to
that, the progressively developing computational capabilities of

Uncertainty reduction by Bayesian updating

Eur
20 φ
γ
Eoed
15
COV (%)

10

0
30 32 34 36 38 40 42
Excavation Step
Fig. 13. A sequential reduction of parameter uncertainties using Bayesian back analysis for the case of low measurement uncertainties COV d ¼ 5%.
52 S. Miro et al. / Computers and Geotechnics 68 (2015) 38–53

Table 7
The identification of the uncertain parameters achieved after the last back analysis sequence for two cases of measurement uncertainties COV d ¼ 5% and COV d ¼ 15%.

Parameter Prior Reference COV d ¼ 5% COV d ¼ 15%

Dist. Mean COV% values Mean COV% Mean COV%
Eur N 150,000 20 135,000 135,966 2.2 131,765 5
/ N 35 15 38 38.5 1.2 38.5 2.6
c N 16 15 18 16.6 2.5 16.6 5.2
Eoed N 30,000 20 25,000 25,521 3.2 26,210 6.1

MCMC convergence
1
Eur
φ
γ
Eoed
0.1
−1
R

0.01

0.001
2500 5000 7500 10000 12500 15000
Sample size
b 6 1:01 for all parameters.
Fig. 14. Monitoring the convergence of the implemented MCMC method. Here the algorithm is considered to be converged when R

computers can lead to a real time back analysis in parallel to the
tunnel excavation. Consequently, the observational data along
with the back analysis strategy complement the geotechnical
explorations and lead to updated subsoil parameters used in the
design.
7. Summary and conclusions
A probabilistic-based study is presented in this paper to evaluate
the impacts of subsoil parameter uncertainties on tunnel-induced
ground movements in a shallow mechanized tunnel. The applied
case study is a synthetic one and not based on a real world tunnel
project. Moreover, it does not include any possible interactions
between the ground and nearby structures. Nevertheless, the
complete framework can be applied to realistic problems.
Due to the fact that the uncertainties of soil mechanical proper-
ties are inevitable, a systematic representation, quantification,
impact evaluation, and reduction of relevant uncertainties are sub-
stantial for the reliability of geotechnical applications. Customarily,
the uncertain soil properties are defined as random variables that
are described by a probability distribution function, a mean value,
and a coefficient of determination (or a standard deviation).
Herein, the most commonly used types of distribution in geotech-
nics, namely normal, lognormal, and uniform, are discussed.
From this study it is seen that the effect of distribution type
between normal and lognormal on the output variation is small.
This is obvious in the case of small uncertainties, i.e. coefficients
of variation for the input parameters COV 6 10%. In the case of
parameters with large uncertainties, where the shape difference
between normal and lognormal distributions becomes apparent,
the distribution type impact on output variation can be noticeable.
Nonetheless, this depends much on the sensitivity of the system
response to the given input parameters as well as on the com-
putational model itself. The introduction of lower and upper
bounds (extreme values), on the other hand, obviously reduces
the differences between normally and lognormally distributed
parameters even in the case of large uncertainties. These ultimate
bounds, however, necessitate engineering judgement to exclude
the outliers that do not represent realistic properties of the soil
under consideration. All in all, the description of soil variability
should be based on site-specific data and investigations.
Consequently, the type of distribution along with the extreme
bounds should rely on and interpret these data.
Performing a global sensitivity analysis, in particular estimating
the variance-based total effect sensitivity index that considers
parameter interactions, is substantial for decomposing the output
variation in accordance to input parameters contribution. In this
way, parameters with a large impact on output variation can be
investigated more accurately in order to reduce their uncertainties.
This detailed investigation can be performed either by laboratory
and in-situ testing or through a back analysis study.
The sequential Bayesian back analysis methodology presented
in this paper serves as a supportive tool for reducing the uncertain-
ties of the input parameters with a large influence on system
output quantities of interest. Such an approach can employ obser-
vational data to complement the geotechnical investigations and
lead to updated parameters to be used in the design. In particular,
it can be enormously advantageous in pessimistic scenarios of soil
 S. Miro et al. / Computers and Geotechnics 68 (2015) 38–53
exploration, especially for sites with limited accessibility. The glo-
bal sensitivity analysis is required for deciding whether the obser-
vation points are suitably chosen for measurement records or not.
Herein, the identifiability of the uncertain parameters is affected
by their sensitivities to the considered system response at the
observation points.
To this end, similar geotechnical applications, which need to
incorporate uncertainties of soil properties, can employ the
approach suggested in [29] and is extended in this paper. This
includes: (1) defining the uncertain factors of the system, (2)
appropriately describing the uncertainties through adequately
chosen distribution functions, (3) estimating uncertainty prop-
agation from the input space to the output space via the output’s
coefficient of variation, (4) decomposing the output variation to
the contribution of individual input factors by computing total
effect sensitivity indices of the inputs, (5) studying the reliability
of the whole system with its partial components, and in case of
large failure probabilities, (6) performing a systematic Bayesian
back analysis approach by utilizing recorded measurements for
reducing input parameter uncertainties and, consequently,
increasing the system reliability.
Acknowledgement
This research has been supported by the German Research
Foundation DFG through the Collaborative Research Center SFB
837. The authors gratefully acknowledge this support.
References
[1] Maidl B, Herrenknecht M, Maidl U, Wehrmeyer G. Mechanised shield
tunnelling. John Wiley & Sons; 2013.
[2] Pinto F, Whittle AJ. Ground movements due to shallow tunnels in soft ground.
i: Analytical solutions. J Geotech Geoenviron Eng 2013;140(4).
[3] Atkinson JH, Potts DM. Subsidence above shallow circular tunnels in soft
ground. J Geotech Eng Div 1977;103:307–25.
[4] Peck RB. Deep excavations and tunnelling in soft ground. In: Proc. 7th int. conf.
on SMFE; 1969. p. 225–90.
[5] Attewell P. Ground movements caused by tunnelling in soil. Conference on
large ground movements and structures, vol. 948. London: Pentech Press;
1978.
[6] OReilly MP, New BM. Settlements above tunnels in the united kingdom - their
magnitude and prediction. Tech. rep., Transport and Road Research Laboratory,
Crowthorne (England); 1982.
[7] Sagaseta C. Analysis of undrained soil deformation due to ground loss.
Geotechnique 1988;38(4).
[8] Verruijt A, Booker J. Surface settlements due to deformation of a tunnel in an
elastic half plane. Geotechnique 1996;46(4):753–6.
[9] Mair R, Taylor R, Burland J. Prediction of ground movements and assessment of
risk of building damage due to bored tunnelling. In: Fourth international
symposium of international conference of geotechnical aspects of on
underground construction in soft ground, AA Balkema; 1996. p. 713–8.
[10] Verruijt A. A complex variable solution for a deforming circular tunnel in an
elastic half-plane. Int J Numer Anal Methods Geomech 1997;21(2):77–89.
[11] Gonzalez C, Sagaseta C. Patterns of soil deformations around tunnels.
application to the extension of madrid metro. Comput Geotech
2001;28(6):445–68.
[12] Pinto F, Zymnis DM, Whittle AJ. Ground movements due to shallow tunnels in
soft ground. ii: Analytical interpretation and prediction. J Geotech Geoenviron
Eng 2013;140(4).
[13] Bernat S, Cambou B. Soil-structure interaction in shield tunnelling in soft soil.
Comput Geotech 1998;22(3):221–42.
[14] Kasper T, Meschke G. A 3d finite element simulation model for tbm tunnelling
in soft ground. Int J Numer Anal Methods Geomech 2004;28(14):1441–60.
53
[15] Kasper T, Meschke G. On the influence of face pressure, grouting pressure and
tbm design in soft ground tunnelling. Tunnell Underground Space Technol
2006;21(2):160–71.
[16] Mroueh H, Shahrour I. A simplified 3d model for tunnel construction using
tunnel boring machines. Tunnell Underground Space Technol
2008;23(1):38–45.
[17] Migliazza M, Chiorboli M, Giani G. Comparison of analytical method, 3d finite
element model with experimental subsidence measurements resulting from
the extension of the milan underground. Comput Geotech 2009;36(1):113–24.
[18] Miro S, Hartmann D, Schanz T. Global sensitivity analysis for subsoil parameter
estimation in mechanized tunneling. Comput Geotech 2014;56:80–8.
[19] Journel A, Alabert F. Non-gaussian data expansion in the earth sciences. Terra
Nova 1989;1(2):123–34.
[20] Guglielmetti V, Grasso P, Mahtab A, Xu S. Mechanized tunnelling in urban
areas: design methodology and construction control. CRC Press; 2008.
[21] Cividini A, Jurina L, Gioda G. Some aspects of ‘characterization’ problems in
geomechanics. Int J Rock Mech Mining Sci Geomech Abstr
1981;18(6):487–503.
[22] Cividini A, Maier G, Nappi A. Parameter estimation of a static geotechnical
model using a bayes’ approach. Int J Rock Mech Mining Sci Geomech Abstr
1983;20(5):215–26.
[23] Ledesma A, Gens A, Alonso E. Estimation of parameters in geotechnical
backanalysis – i. Maximum likelihood approach. Comput Geotech
1996;18(1):1–27.
[24] Gens A, Ledesma A, Alonso E. Estimation of parameters in geotechnical
backanalysis – ii. Application to a tunnel excavation problem. Comput Geotech
1996;18(1):29–46.
[25] Schanz T, Zimmerer MM, Datcheva M, Meier J. Identification of constitutive
parameters for numerical models via inverse approach. Felsbau
2006;24(2):11–21.
[26] Knabe T, Schweiger HF, Schanz T. Calibration of constitutive parameters by
inverse analysis for a geotechnical boundary problem. Canadian Geotech J
2012;49(2):170–83.
[27] Knabe T, Datcheva M, Lahmer T, Cotecchia F, Schanz T. Identification of
constitutive parameters of soil using an optimization strategy and statistical
analysis. Comput Geotech 2013;49:143–57.
[28] Khaledi K, Miro S, König M, Schanz T. Robust and reliable metamodels for
mechanized tunnel simulations. Comput Geotech 2014;61:1–12.
[29] Mollon G, Dias D, Soubra A-H. Probabilistic analyses of tunneling-induced
ground movements. Acta Geotech 2013;8(2):181–99.
[30] Schanz T, Vermeer P, Bonnier P. The hardening soil model: formulation and
verification. Beyond 2000 in computational geotechnics; 1999. p. 281–96.
[31] Schanz T. Zur modellierung des mechanischen verhaltens von
reibungsmaterialien. Habil thesis, Institut für Geotechnik, Universität
Stuttgart, Vermeer, P.A. (Hrsg.), Mitteilung 45; 1998.
[32] Beck JV, Arnold KJ. Parameter estimation in engineering and science. James
Beck; 1977.
[33] Brinkgreve RBJ, Engin E, Swolfs WM. PLAXIS 3D Version 2011. Reference
manual; 2011.
[34] Schanz T, Zarev V, Barciaga T. Adäkuate modellbildung für die numerische
analyse maschineller tunnelvortriebe. Ruhr-Geo 2012.
[35] Phoon K-K, Kulhawy FH. Characterization of geotechnical variability. Canadian
Geotech J 1999;36(4):612–24.
[36] Lacasse S, Nadim F. Uncertainties in characterising soil properties.
Publikasjon-Norges Geotekniske Institutt 1997;201:49–75.
[37] Phoon K-K, Kulhawy FH. Evaluation of geotechnical property variability.
Canadian Geotech J 1999;36(4):625–39.
[38] Baecher GB, Christian JT. Reliability and statistics in geotechnical
engineering. John Wiley & Sons; 2005.
[39] Myers RH, Montgomery DC. Response surface methodology. 2 ed. John Wiley &
Sons; 2002.
[40] Homma T, Saltelli A. Importance measures in global sensitivity analysis of
nonlinear models. Reliab Eng Syst Safety 1996;52(1):1–17.
[41] Jansen MJ. Analysis of variance designs for model output. Comput Phys
Commun 1999;117(1):35–43.
[42] Faber MH. Statistics and probability theory: in pursuit of engineering decision
support, vol. 18. Springer; 2012.
[43] Metropolis N, Rosenbluth AW, Rosenbluth MN, Teller AH, Teller E. Equation of
state calculations by fast computing machines. J Chem Phys
1953;21(6):1087–92.
[44] Hastings WK. Monte carlo sampling methods using markov chains and their
applications. Biometrika 1970;57(1):97–109.
[45] Gelman A, Carlin JB, Stern HS, Dunson DB, Vehtari A, Rubin DB. Bayesian data
analysis. CRC Press; 2013.