Professional Documents
Culture Documents
Influence of A Weak Layer On The Tunnel Face Stability 2020 Computers and
Influence of A Weak Layer On The Tunnel Face Stability 2020 Computers and
Research Paper
A R T I C LE I N FO A B S T R A C T
Keywords: In this study, reliability and sensitivity analyses were employed to investigate the effects of a weak interlayer on
Tunnel face stability tunnel face stability. Due to the irregular geology process and the movement of tunnelling machine, the situa-
Weak interlayer tions of weak soils encountered during tunnel construction are complex and variable. To address this problem,
Reliability analysis this paper adopted reliability and sensitivity analysis to evaluate the face stability of a tunnel considering a weak
Sensitivity analysis
soil layer. An active learning method which combines Kriging and Monte Carlo Simulation (AK-MCS) was used
Kriging
for the reliability analysis. Then, the importance of each stochastic soil property was estimated by coupling the
surrogate model constructed in the AK-MCS with the global sensitivity analysis (GSA). Compared with tradi-
tional approaches, the employed technique (termed as AK-MCS-GSA) is able to provide a variety of interesting
results (failure probability, model response statistics and sensitivity index) but with a greatly reduced compu-
tational time. The obtained results permitted to show that a weak soil layer over the tunnel cross-section can lead
to a significant increase of failure probability. Additionally, the sensitivity indices in GSA revealed that the soil's
frictional angle of the weak layer has the greatest influence on the face stability. Moreover, the impacts of
varying the weak layer’s position and thickness are discussed. The critical position for the weak layer which
leads to the biggest failure probability is remarked.
1. Introduction in recent years for the tunnel face stability. Mollon et al. [3–5] proposed
a 3D rotational failure mechanism model for homogeneous soils and
To prevent failures in geotechnical engineering, there has been an carried out a probabilistic analysis using respectively the collocation-
increasing demand of probabilistically evaluating the risk of en- based stochastic response surface method and the first-order reliability
gineering projects by considering the complexity of the ground and the method (FORM). The 3D rotational mechanism model was then ex-
uncertainties related to soil properties. High failure probabilities can tended to study the effect of anisotropic and nonhomogeneous soils on
result in many casualties and economic damage, while lower values the tunnel face stability by Pan and Dias [6]. Pan et al. [7] conducted a
result in overly expensive systems. Especially, for the accidents occur- probabilistic stability analysis for rock tunnels using the Polynomial
ring during a tunnel construction as shown in Fig. 1, it would require Chaos Expansion. The failure probabilities were discussed in terms of
expensive cost for pavement and building rehabilitation, and pose a different coefficients of variation, distribution types and correlation
threat to life. Therefore, it is important to take into account the un- coefficients. By using the sparse polynomial chaos expansion (SPCE),
certainties of soil properties which are unavoidable for tunnel face Pan and Dias [8] discussed the tunnel face stability in spatially random
stability analysis. soils [9], as well as the influence of the cross-correlation between the
Using a reliability evaluation provides a means of assessing the cohesion and frictional angle on the probability density function of the
degree of the model response uncertainty for the tunnel design. To required face pressure. In addition, Zeng et al. [10] conducted a relia-
rationally deal with the uncertainties in geotechnical engineering bility analysis for rock tunnels through the FORM, the response surface
analysis, many studies of reliability analysis [1,2] have been performed method (RSM) and the Importance Sampling (IS), with the
⁎
Corresponding author.
⁎⁎
Co-corresponding author.
E-mail addresses: xiangfeng.guo@3sr-grenoble.fr (X. Guo), zhangqian@tju.edu.cn (Q. Zhang).
https://doi.org/10.1016/j.compgeo.2020.103507
Received 24 October 2019; Received in revised form 18 February 2020; Accepted 20 February 2020
Available online 28 February 2020
0266-352X/ © 2020 Elsevier Ltd. All rights reserved.
S. Zhou, et al. Computers and Geotechnics 122 (2020) 103507
Fig. 1. Accident due to face instability in mechanized tunnelling. computed by a three dimensional numerical model.
2
S. Zhou, et al. Computers and Geotechnics 122 (2020) 103507
Table 1
Input probabilistic parameters for characterizing the uncertainties in soil properties.
Frictional angle (°) Cohesion (kPa) Unit weight (kN/m3)
Distribution Beta (Min-Max: 0–45) Log-normal Log-normal
3.1. Deterministic model of tunnel face stability Fig. 4. Computation of the rates of work for tunnel face pressure.
Ẇγ = ∭V →γ ·→v dV = ∑ (→γ ·→vi,j·Vi,j + →γ ·→v i′,j·Vi′,j) In Eq. (3), c is the cohesion of the soil and φ represents the soil’s
i, j frictional angle.
= ω· ∑ (γi, j·Ri, j ·Vi, j·sin βi, j + γi′, j Ri′, j ·Vi′, j·sin βi′, j ) By equating the total rate of external force to the total rate of in-
i, j (2) ternal energy dissipation, the face pressure for collapse is expressed as
where γ is the soil’s unit weight. Vi, j is the volume involved by the
3
S. Zhou, et al. Computers and Geotechnics 122 (2020) 103507
∑i, j (γi, j Ri, j Vi, j sin βi, j + γi′, j Ri′, j Vi′, j sin βi′, j ) second term is a stationary Gaussian process with zero mean. In the
present work, the ordinary Kriging is chosen, thus βT f (x ) is a scalar
− ∑i, j (ci, j Ri, j Si, j cos φi, j + ci′, j Ri′, j Si′, j cos φi′, j ) value to determine. The covariance in the second term between two
σc =
∑j (Σj ·Rj ·cos βj ) (4) points x and w is defined in Eq. (7):
where γi, j, ci, j, φi, j and γi′, j, ci′, j, φi′, j are the average values for the trian- CoV (z (x ), z (w )) = σ 2R (x , w ) (7)
gular facet (Si, j, Si′, j ) illustrated in Fig. 5, respectively. In Eq. (7), σ 2 is the process variance and R (x , w ) represents a cor-
Then, a two-dimensional failure mechanism that is able to account relation function. The correlation function selected in this paper is the
for the heterogeneity of the soil shear strength properties was proposed Gaussian model. A detailed description of these parameters can be
by Mollon et al. [23], in which the slip surfaces of the failure me- found in [29].
chanism are not described by log-spirals but by non-standard curves
that are dependent to the spatial varying soil friction angles using a 3.2.2. Monte Carlo Simulation
spatial discretization technique. In the generation process of this failure Monte Carlo Simulation (MCS) permits to estimate the failure
mechanism, it is constrained to respect the normality condition of the probability by repeatedly running a deterministic model with different
limit analysis theory at each point of the failure surfaces. Through the sets of input parameters and then counting the samples that belong to
comparison with numerical simulation did in Mollon et al. [23], this the failure domain. The CoV for assessing the convergence of MCS is
two-dimensional failure mechanism for nonhomogeneous soils was expressed in Eq. (8) and the Latin Hypercube sampling (LHS) strategy is
shown to give reasonable results in terms of the critical collapse pres- employed to generate the Monte Carlo samples. In this work, the size of
sure and the failure shape. samples NMC is set to be large enough in order to obtain the CoV of
Moreover, by combining the 3D Collapse-Failure mechanism of [5] failure probability smaller than 5%.
and the two-dimensional failure mechanism for heterogeneous soils
[23], a failure mechanism was further established by Pan and Dias [6], 1 − Pf
CoV (Pf ) =
in which different values of soil properties can be assigned to the tri- NMC Pf (8)
angular facets during the spatial discretization. The results obtained by
For reliability analysis, the MCS is considered as a standard re-
this 3D KA model were also compared with the existing solutions [5],
ference to verify the results of other methods due to its versatility and
which showed a good agreement. It is noted that the cases of layered
robustness [30]. In the present study, the results obtained by the direct
soils were considered in the study of Pan and Dias [6]. Additionally, the
MCS are used to evaluate the accuracy and efficiency of the selected
applications of the KA model for heterogeneous cases can be found in
methodology which is introduced as follows:
[8,24,25].
In this work, the Adaptive Kriging Monte Carlo Simulation (AK-
In conclusion, the KA model proposed by Mollon et al. [5] was
MCS) proposed by Echard et al. [28] is adopted for the reliability
firstly applied to the cases of homogeneous soils. Then, a spatial dis-
analysis. This method can guarantee that the constructed Kriging model
cretization technique [23] was introduced, which permits to take into
is highly accurate in the vicinity of the limit state surface. This is rea-
account the heterogeneity of soil properties. In order to describe the
lized by gradually adding selected samples to the experimental design
failure shape in the model with a more rational way, the log-spiral
of the Kriging metamodel. In addition, this method requires a lower
assumption was replaced by non-standard curves determined by the
number of evaluations of the limit state functions compared with the
soil’s frictional angle at each discretized point. Moreover, based on the
direct MCS, thus reducing the total computational time. The flow chart
comparisons performed in [6–8,23–26], between the KA model and
of the adaptive experimental design algorithm [31] is shown in Fig. 6
numerical model for both homogeneous and heterogeneous cases, the
and its main characteristics are summarized as follows:
adopted deterministic model is expected to be reliable in the studied
To select the best next candidate sample for the experimental de-
case.
sign, the U-function [28] is chosen as the learning function:
To estimate the failure probabilities against the collapse of a tunnel
face, the performance function in a reliability analysis can be defined |μg (x )|
as: U (x ) =
σg (x ) (9)
g = σT − σc (5)
where μg (x ) and σg (x ) are the mean and standard deviation of g in Eq.
where σT is the applied face pressure and σc is the critical face pressure (5). The next candidate sample, from a MCS candidate pool:
calculated with the current soil condition. The safe domain is where S = {s(1), …, s(NMC) } , is chosen as s∗ = argmins ∈ S U (s ) .
g > 0 , while g < 0 refers to the failure case. In the subsequent analysis, The stopping criterion proposed in [32] is adopted, which directly
the applied face pressure is set as a constant. focuses on the uncertainty of the failure probability estimation:
P+ −
f − Pf
3.2. Probabilistic analysis methods ⩽ ε Pf
P f0 (10)
3.2.1. Kriging where = (μg (x) ⩽ 0),
P f0 = (μg (x) ∓ 1.96 × σg (x) ⩽ 0) , and ε Pf is
P±f
Kriging, based on a Gaussian Process assumption, is a powerful the threshold for the convergence criterion.
method for constructing surrogate models [27,28] and it has interesting
characteristics compared with other meta-modeling techniques. First, 3.2.3. Global sensitivity analysis
the prediction in a point belonging to the design of experiments is the A sensitivity analysis aims at quantifying the importance of each
exact value of the performance function in this point. Furthermore, it input parameter with respect to the output. Global sensitivity analysis
provides not only predicted values, but also estimations of the local [33] is a useful approach to implement sensitivity analysis that con-
variance of the predictions. This variance defines the local uncertainty siders the whole input domain. Specifically, with regard to the case of
on the prediction. The basic function of the Kriging theory is shown in dependent input variables, the Kucherenko [34] sensitivity indices are
Eq. (6). used in this study.
If the input variables X (Sample with a certain value in X is denoted
M (x) = β T f (x) + z(x) (6)
as x ) are divided into two complementary subsets Xv and Xw = X v ,̃
where the first term on the right part of Eq. (6) represents the mean then the following moments can be estimated through N (large enough)
value of the Gaussian process and βT is the regression coefficients. The samples of X :
4
S. Zhou, et al. Computers and Geotechnics 122 (2020) 103507
Fig. 7. The ground formation of the tunnel face in the reference case.
5
S. Zhou, et al. Computers and Geotechnics 122 (2020) 103507
(a) (b)
Fig. 8. (a) PDF of the required face pressure. (b) Convergence of the failure probability in the AK-MCS analysis.
Table 2 the direct GSA which is combined with the deterministic model is il-
Estimation of statistical moments and failure probability (Reference case). lustrated in Fig. 9. With regard to the sensitivity indices in this figure,
Failure Probability Statistical moments of the face pressure (kPa)
the estimates given by the Kriging-based GSA are in good agreement
with the results of the direct GSA, revealing a good accuracy of the
Pf CoV (Pf ) Mean Standard deviation employed technique in global sensitivity analysis.
As shown in Table 4, the direct MCS and GSA need nearly 40 h in
1.91 × 10 -2 7.16 × 10 -3 42.31 3.50
total for providing the results in Table 3 and Fig. 9, using a Xeon E5-
2695v2@2.4 GHz server. It is noted that the computational procedure
was implemented in Matlab R2016.b in this study. However, the AK-
MCS-GSA significantly reduces this time to 0.5 h with the same com-
puter and the same simulation platform. The Kriging model training
needs around 30.6 mins while the time required for Kriging-based re-
liability and sensitivity analyses is almost negligible. These comparisons
highlight the accuracy and efficiency of the employed AK-MCS-GSA in
estimating failure probability, face pressure statistics and sensitivity
indices.
For the kinematical approach (KA) used in this study, Mollon et al.
Fig. 9. Sensitivity analysis results – Comparison between Kriging-based GSA [36–38] studied the collapse failure mechanism in homogeneous soil
and direct GSA. mass and found that the results were close to the solutions given by
complex 3D numerical simulations. Then, Mollon et al. [23] and Pan
the risk estimation and make more reasonable decisions. and Dias [6] extended the original model to nonhomogeneous soils by
using a spatial discretization technique to describe the failure surfaces
4.2. Comparison between the AK-MCS-GSA and traditional approaches and consider the heterogeneity of soil properties.
Therefore, the adopted KA model, based on the mechanism of Pan
In this section, the accuracy and efficiency of the AK-MCS-GSA and Dias [6,8] which can account for soil heterogeneity, is expected to
technique is highlighted by comparing the results with the direct MCS be reliable in the case of layered soils. To further validate the employed
and GSA which are based on the original deterministic model. deterministic model considering a weak soil layer, this section is de-
The comparison with the direct MCS is given in Table 3. The results voted to perform 3D numerical simulations, thus making a comparison
provided by the direct MCS require a higher number of evaluations of between the KA model and the numerical analysis which is based on a
the limit state functions (around 21,000 calls), compared with the AK- finite difference software FLAC3D [39].
MCS with only 149 calls. For the failure probability calculated by the According to the reference case defined in Section 4.1, a numerical
AK-MCS, the outcome is very close to the direct MCS. Moreover, the model of a circular tunnel with diameter D = 10 m under buried depth
error of the mean value and the standard deviation of the applied C = 10 m was created, with three layers as shown in Fig. 7. Due to the
pressure are respectively 0.12% and 1.13%. The CoV of failure prob- symmetry, only half of the system was modelled. The adopted mesh was
ability obtained by AK-MCS is much smaller than the direct MCS. composed by approximately 85,000 elements and 90,000 nodes. As
Furthermore, the comparison between the Kriging-based GSA and
Table 4
Comparison of computational efficiency between AK-MCS-GSA and direct si-
Table 3
mulations.
Reliability analysis results – Comparison between AK-MCS and direct MCS.
MCS GSA Metamodel construction Total time
Method Ncall Pf COV (Pf ) Mean (kPa) S.D. (kPa)
Direct calculation 19.7 h 18.7 h / 38.4 h
Direct MCS 21,000 1.76 × 10 -2 5.16 × 10 -2 42.31 3.54 Kriging-based <1 s <1 s 30.6 min 0.5 h
AK-MCS 18 + 131 1.91 × 10 -2 7.16 × 10 -3 42.26 3.50 Relative error < 9% < 7% / /
6
S. Zhou, et al. Computers and Geotechnics 122 (2020) 103507
Fig. 11. The velocity field of the case 5 in Table 5 (σc = 43.57 kPa ).
Fig. 10. The mesh used in the 3D numerical model for the reference case. [41] or multi-layered soils [25] and reported an error around 10%.
The idea of using the KA model instead of a numerical one is mainly
for alleviating the computational burden for a probabilistic analysis
shown in Fig. 10, the size of the numerical model was taken as 48 m in
which requires normally many deterministic simulations. For one de-
the X direction, 50 m in the Y direction, and 43 m in the Z direction. For
terministic calculation, the numerical model of Fig. 11 needs around 5 h
the displacement boundary conditions, the bottom boundary was as-
while the adopted KA model can reduce it to 6 s. Although an advanced
sumed to be fully fixed and the lateral boundaries were constrained in
reliability method (Adaptive Kriging) is used in this study, it still may
the normal direction. Besides, the Mohr-Coulomb failure criterion was
need about one month to provide an accurate Pf estimate if the required
assumed for the soils, and the assumption of a non-associated flow rule
deterministic simulation number is similar to the one of Table 3 (i.e.
was adopted [23]. The considered type of soils for the model is sandy
149). In conclusion, the adopted approach is able to give a relatively
clay.
accurate estimate of σc for the soils with a weak layer but with a greatly
For the comparison, ten different input sets, as shown in Table 5,
reduced calculation time, therefore it could be used in this study for the
were considered. These ten samples were randomly generated from the
reliability and sensitivity analysis. However, it should be noted that an
joint distribution defined in Table 1 together with a correlation coef-
analytical model (e.g. the adopted KA) is usually less accurate than a
ficient of −0.5 for c - φ by using the LHS method. Then, the two models
numerical one, since the boundary conditions cannot be fully con-
were both employed to predict the critical collapse pressure σc for each
sidered and assumptions for the failure shape should be made. In fact,
sample. Concerning the soil elastic properties and the tunnel support
to manage risk and achieve a reliable geotechnical design, a multi-stage
lining parameters which are necessary for a numerical simulation, the
safety assessment is often carried out for a geotechnical project. The
values were taken from a previous study [40]. The stress-control
presented procedure (i.e. KA-Kriging-MCS-GSA), being efficient and
method [40] was coded herein using the FISH language in FLAC3D to
relatively accurate (Tables 4 and 5), is recommended for a first stage.
estimate the critical collapse pressure σc at tunnel face, in which
The multiple results (failure probability, sensitivity indices and model
bracketing and bisection are repeated until the difference between the
response statistics) provided by this procedure can be used for guiding a
upper and lower brackets becomes smaller than 0.1 kPa in this work.
second stage analysis which will potentially be based on a more so-
As an illustration, Fig. 11 presents the total velocity of case 5
phisticated numerical model.
(Table 5) when the support pressure is equal to 43.57 kPa.
Table 5 presents the comparison results. It can be seen that the σc
obtained from the adopted KA model agree reasonably well with those
5. Discussions
provided by the 3D numerical model. The relative errors between the
two models are all lower than 13% and the average is equal to 10% for
The aim of this section is to investigate the effects of several factors
the ten random cases. This indicates that the KA model is effective for
on the critical face pressure, the failure probability and the sensitivity
the soils with a weak layer since the predicted σc is reasonable and
indices. The discussion considers three factors: (1) The applied face
acceptable with an absolute difference being about 4 kPa compared to
pressure; (2) The thickness of the weak layer; (3) The position of the
the 3D numerical results. This finding is consistent with the previous
weak layer.
studies which also performed such a comparison for homogeneous soils
Table 5
The ten random input sets and the comparison between the two models.
Input parameters Output σc
cbot cmid ctop φbot φmid φtop γbot γmid γtop FLAC KA model Relative Error (%)
3D
kPa kPa kPa ° ° ° kN/m3 kN/m3 kN/m3 kPa kPa
5.23 4.18 6.15 22.30 14.89 21.60 19.33 18.47 18.70 38.30 42.32 10.50
14.52 3.60 6.32 16.22 16.21 19.36 20.13 19.29 19.85 41.86 45.6 8.93
10.39 3.97 8.22 19.47 15.44 19.17 18.67 20.06 18.52 39.14 43.14 10.22
6.21 5.08 10.41 23.17 13.01 20.18 18.97 19.20 19.16 37.89 42.12 11.16
6.65 4.71 8.36 16.77 12.99 23.06 20.24 20.63 18.88 43.57 48.53 11.38
8.33 5.49 12.20 17.95 13.10 16.58 17.82 19.44 18.39 39.88 42.96 7.72
5.85 5.97 7.52 20.59 16.23 17.26 19.24 18.10 19.46 38.48 42.05 9.28
8.07 4.40 6.96 18.32 14.56 20.29 19.66 18.35 19.29 40.94 44.85 9.55
6.77 6.51 7.16 20.65 13.35 18.97 18.09 17.31 20.95 42.34 45.92 8.46
9.38 5.32 9.01 18.59 14.66 20.81 18.74 19.91 20.05 38.91 43.85 12.70
7
S. Zhou, et al. Computers and Geotechnics 122 (2020) 103507
Fig. 13. Global sensitivity for the soil properties considering different applied face pressures.
8
S. Zhou, et al. Computers and Geotechnics 122 (2020) 103507
Fig. 14. The thickness of weak layer increases from centre to upper and lower ends.
the unit weight of the top layer have bigger sensitivity indices com-
pared with other layers’ soil properties, this might be attributed to the
important volume of the top soil which drives the system failure.
Fig. 16. Global sensitivity indices for the thickness of weak layer.
9
S. Zhou, et al. Computers and Geotechnics 122 (2020) 103507
Fig. 17. Influence of the position of the weak layer on failure probability.
dramatically as the weak layer moves down from T/D = 1/8 (near the combines Kriging with Monte Carlo Simulation. Through the estab-
centre of the tunnel cross-section) to T/D = 3/8, with the failure lished metamodel and global sensitivity analysis, engineers are able to
probability changes from 0.0069 to 0.0327. identify the significant soil properties, and thus have better under-
The GSA results are plotted in Fig. 18. It is found that the Kucher- standings of the tunnel face instability with rich information provided
enko indices of the bottom layer decrease as the position of the weak in this work. Furthermore, the deterministic model used in this work is
layer moves from the ground surface to the tunnel invert especially for based on an extension of the classical rotational failure mechanism,
the frictional angle while the ones of the three soil properties in the top which is able to effectively account for multi-layered soils. A compar-
layer are similar both in values and the growth trend since the growth ison with 3D numerical results indicates that the adopted deterministic
trends of these three curves are basically the same. The above- model can give a good estimation of the critical pressure in layered soils
mentioned trend for the top and the bottom layer can be attributed to for tunnel face stability. The following conclusions can be drawn from
the proportion of these two layers in terms of the tunnel face. this study.
Additionally, the situation of the weak interlayer is more inter-
esting, as the sensitivity indices peak around 1/8, which is exactly the 1. According to the comparative study of Section 4.2, the employed
suddenly increasing position observed in Fig. 17, revealing that the AK-MCS-GSA can provide a variety of useful results (e.g. Pf, model
variation of failure probability is closely related to the sensitivity in- response distribution and sensitivity indices) with a good accuracy.
dices of the weak layer’s soil properties. The total computational time can be significantly reduced by using
this technique,
2. For the reference case, the sensitivity indices reveal that the soil’s
6. Conclusions
frictional angle of the weak layer has the greatest influence on the
face stability, followed by the cohesion of the top layer, and unit
Along the tunnel construction path, different scenarios of the soil
weight of the weak layer,
layers over the tunnel cross-section are usually encountered in real
3. The failure probability increases rapidly when the thickness of the
cases. Moreover, the intermediate weak layer at the tunnel face has a
weak layer changes in a range from 1/4D to 1/2D, as well as when
great influence on the system performance due to its lower shear
the weak layer moves down from T/D = 1/8 to T/D = 3/8,
strength parameters. Therefore, to discuss the effect of a weak layer on
4. The most dangerous position of the weak layer is highlighted in this
the tunnel face stability, this work introduces reliability analysis and
work. It is located slightly higher than the tunnel bottom. This can
global sensitivity analysis based on an active learning method which
10
S. Zhou, et al. Computers and Geotechnics 122 (2020) 103507
11
S. Zhou, et al. Computers and Geotechnics 122 (2020) 103507
[25] Ibrahim E, Soubra A-H, Mollon G, Raphael W, Dias D, Reda A. Three-dimensional models with dependent variables. Comput Phys Commun 2012;183(4):937–46.
face stability analysis of pressurized tunnels driven in a multilayered purely fric- https://doi.org/10.1016/j.cpc.2011.12.020.
tional medium. Tunn Undergr Sp Tech 2015;49:18–34. [35] Marelli S, Schöbi R, Sudret B. UQLab user manual–Reliability analysis. Journal;
[26] Sun Zhibin, Li Jianfei, Pan Qiujing, Dias Daniel, Li Shuqin, Hou Chaoqun. Discrete 2015.
kinematic mechanism for nonhomogeneous slopes and its application. Int J [36] Mollon Guilhem, Dias Daniel, Soubra Abdul-Hamid. Probabilistic analysis and de-
Geomech 2018;18(12):04018171. https://doi.org/10.1061/(ASCE)GM.1943-5622. sign of circular tunnels against face stability. Int J Geomech 2009;9(6):237–49.
0001303. https://doi.org/10.1061/(ASCE)1532-3641(2009)9:6(237).
[27] Mo S, Shi X, Lu D, Ye M, Wu J. An adaptive Kriging surrogate method for efficient [37] Mollon MG, Phoon KK, Dias D, Soubra AH. Influence of the scale of fluctuation of
uncertainty quantification with an application to geological carbon sequestration the friction angle on the face stability of a pressurized tunnel in sands. Journal
modeling. Comput Geosci 2019;125:69–77. https://doi.org/10.1016/j.cageo.2019. 2011(Issue):225–32. https://doi.org/10.1061/41183(418)14.
01.012. [38] Mollon Guilhem, Dias Daniel, Soubra Abdul-Hamid. Range of the safe retaining
[28] Echard B, Gayton N, Lemaire M. AK-MCS: An active learning reliability method pressures of a pressurized tunnel face by a probabilistic approach. J Geotech
combining Kriging and Monte Carlo Simulation. Struct Saf 2011;33(2):145–54. Geoenviron Eng 2013;139(11):1954–67. https://doi.org/10.1061/(ASCE)GT.1943-
https://doi.org/10.1016/j.strusafe.2011.01.002. 5606.0000911.
[29] Santner TJ, Williams BJ, Notz W, Williams BJ. The design and analysis of computer [39] Group I C. FLAC3D 5.0 manual. Journal 2012.
experiments. Journal. 2003; 1: https://link.springer.com/book/10.1007%2F978-1- [40] Mollon Guilhem, Dias Daniel, Soubra Abdul-Hamid. Probabilistic analysis of cir-
4939-8847-1. cular tunnels in homogeneous soil using response surface methodology. J Geotech
[30] Pan Q, Qu X, Liu L, Dias D. A sequential sparse polynomial chaos expansion using Geoenviron Eng 2009;135(9):1314–25. https://doi.org/10.1061/(ASCE)GT.1943-
Bayesian regression for geotechnical reliability estimations. Int J Numer Anal Meth 5606.0000060.
Geomech 2020. [41] Pan Q, Dias D. Upper-bound analysis on the face stability of a non-circular tunnel.
[31] Pan Q, Dias D. An efficient reliability method combining adaptive Support Vector Tunn Undergr Sp Tech 2017;62:96–102. https://doi.org/10.1016/j.tust.2016.11.
Machine and Monte Carlo Simulation. Struct Saf 2017;67:85–95. https://doi.org/ 010.
10.1016/j.strusafe.2017.04.006. [42] Guo X, Du D, Dias D. Reliability analysis of tunnel lining considering soil spatial
[32] Schöbi R, Sudret B, Marelli S. Rare event estimation using polynomial-chaos kri- variability. Eng Struct 2019;196:109332.
ging. ASCE-ASME J Risk Uncertainty Eng Syst, Part A: Civ Eng 2017;3(2). https:// [43] Li X-Y, Zhang L-M, Gao L, Zhu H. Simplified slope reliability analysis considering
doi.org/10.1061/AJRUA6.0000870. spatial soil variability. Eng Geol 2017;216:90–7.
[33] Fang Y, Su Y. On the use of the global sensitivity analysis in the reliability-based [44] Al-Bittar Tamara, Soubra Abdul-Hamid, Thajeel Jawad. Kriging-based reliability
design: Insights from a tunnel support case. Comput Geotech analysis of strip footings resting on spatially varying soils. J Geotech Geoenviron
2020;117:103280https://doi.org/10.1016/j.compgeo.2019.103280. Eng 2018;144(10):04018071. https://doi.org/10.1061/(ASCE)GT.1943-5606.
[34] Kucherenko S, Tarantola S, Annoni P. Estimation of global sensitivity indices for 0001958.
12