You are on page 1of 18

Arabian Journal of Geosciences (2022) 15: 1490

https://doi.org/10.1007/s12517-022-10710-5

ORIGINAL PAPER

Application of different stochastic numerical procedures in rock tunnel


lining design
Masoud Mazraehli1   · Shokrollah Zare1

Received: 8 December 2021 / Accepted: 12 August 2022 / Published online: 6 September 2022
© Saudi Society for Geosciences 2022

Abstract
This paper aims to utilize different stochastic numerical procedures for designing a rock tunnel lining, namely, strength clas-
sification (SCM) and random field (RFM) methods. The procedures are implemented in a finite difference scheme, while a
stochastic process based on the Monte Carlo simulation technique is followed for rock mass characterization, considering
its property variability. Moreover, a performance function is defined to investigate the lining serviceability based on the
interactions between bending moment, thrust, and shear forces. The effects of coefficient of variation and correlation length
on the results are also investigated. Finally, the probabilistic analysis results are compared with the deterministic ones and
the in situ measurements. The results show that the effect of the coefficient of variation on the spatial variability is more
significant than the effect of the correlation length. It can also be observed that the SCM models calculate lower support
requirements than the RFMs when CoV is smaller than 30%. Moreover, the cross-correlated models lead to lower PoF, which
can be related to the assumed negative cross-correlation between c and 𝜑.

Keywords  Tunnel lining · Probabilistic stability analysis · FDM · RFM · SCM · MCS

Introduction one of the principal uncertainty sources in rock engineering,


the others being measurement and transformation error (Cai
Stability analysis and lining design are significant issues 2011; Hoek 1998; Phoon and Kulhawy 1999; Riedmüller
during the construction of underground excavations in and Schubert 1999). Ignorance of these uncertainties can
rock. There are a large number of researches related to this remarkably influence the analysis results and may lead to
topic which can be divided into deterministic (Chehade applying too conservative or unfavorable safety factors in
and Shahrour 2008; Funatsu et al. 2008; Han et al. 2021; the design (Song et al. 2011).
Pelizza et al. 2000; Zhu et al. 2003) and probabilistic stud- The probabilistic stability analysis methods have been
ies (Behnia and Seifabad 2018; Cai 2011; Heidarzadeh et al. established to capture a more realistic perspective on how
2020; Idris and Nordlund 2019; Idris et al. 2011, 2012, 2016; over- or under-estimated response variables can affect the
Jiale and Xiaohong 2021; Seguini and Nedjar 2016; Tiwari remedial requirements and design aspects (Yu et al. 2019).
et al. 2017, 2018). The deterministic methods have been The Monte Carlo simulation (MCS), first-order (FORM) and
widely used, but the probabilistic studies are restricted in second-order reliability methods (SORM), and point esti-
quantity (Fortsakis et al. 2011; Kroetz et al. 2018; Lang- mate method (PEM) might be utilized as the general scheme
ford and Diederichs 2013; Napa-García et al. 2017; Oreste for probabilistic analysis, though there are some differences
2005). Despite the popularity of the deterministic techniques in their applications. For instance, the methods can be clas-
in practice, these methods are impotent to reckon with the sified into two classes considering rock mass heterogeneity
inherent randomness of geo-mechanical properties, which is (Idris et al. 2016). Based on the assumption, while the MCS
is categorized in a group that takes account of the spatial
variability of rock mass properties, the other methods are
Responsible Editor: Zeynal Abiddin Erguler
placed in another group because they cannot consider the
* Masoud Mazraehli variability in the analysis.
In recent years, the MCS has started a scientific revolu-
1
Faculty of Mining, Petroleum and Geophysics Engineering, tion. It is now practical to get insight into how a problem
Shahrood University of Technology, Shahrood, Iran

13
Vol.:(0123456789)
1490 Page 2 of 18 Arab J Geosci (2022) 15: 1490

solution is affected by the input parameter variation using periods, stating that the topic is becoming more demand-
the method (Binder and Heermann 2010; Sari 2009; Sari ing over time. One of the advantages of these methods is
and Karpuz 2006). In geotechnical engineering, the MCS that they can determine the statistical distribution of the
was first applied in many aspects, such as single random response. In this regard, it would be possible to calculate
geo-mechanical variables (SRV) (e.g., Behnia and Seifabad the probability of failure (PoF) for stability analysis. Fur-
2018; Tiwari et al. 2017, 2018). Although these researches thermore, there is a lack of study for the lining design of the
furthered the MCS applications in geotechnical engineer- tunnel excavated in spatially variable rocks based on interac-
ing, they neglected spatial randomness. Subsequently, some tions between the acting moments and forces.
researchers attempted to fill the gap by utilizing the MCS This paper implements the SCM and RFM for analyzing
coupled with numerical software packages. Hsu and Nelson the stability of a supported tunnel, considering variability
(2006) and Srivastava (2012) incorporated numerical meth- in geo-mechanical properties. A procedure is followed to
ods and the MCS to analyze slope stability in spatially vari- take account of spatial rock mass variability in the stability
able weak rocks. Idris and Nordlund (2019) and Idris et al. analysis, combining the MCS and FDM methods. The vari-
(2011, 2012, 2016) used the finite difference method (FDM) ance reduction technique is utilized in the RFM, while the
for analyzing the stability of underground mine stopes con- stochastic mapping approach is adopted in the SCM models.
sidering spatial variability. Yu et al. (2019) took advantage The stochastic results are compared to each other. Another
of stochastic numerical modeling to investigate tunnel liner comparison is also made between the stochastic, determin-
performance and concluded that the procedure could lead to istic, and monitoring data.
a more equitable and economical design.
The above-mentioned studies established the MCS
applications in modeling the inherent randomness of rock Probabilistic numerical modeling
mass properties. There are, however, some aspects that still
require more research. Uncertainty in the distribution of the A square FDM model with a 70 m side length was built
input parameters is one of these aspects that needs more using FLAC and, as said before, was discretized with 78,400
surveillance. Tiwari et al. (2017) used the PEM for stability finite difference quadrilateral zones, with 280 zones in both
analysis of underground structures, implementing normal x and y directions. It means that each zone has a side length
distribution functions for input variables. Analyzing consid- of 0.25 m. As a result, it provides an acceptable resolution
erable number of geo-mechanical laboratory and field data, for random property mapping and reasonable numerical
Mazraehli and Zare (2020) demonstrated that distributions accuracy.
of rock mass properties do not necessarily follow normal and The roller and fixed boundary conditions were applied to
lognormal distribution function rules. different boundaries. The upper side was fixed in the y direc-
The stochastic numerical method was first adopted for tion, while the displacements of the other sides were fixed in
the stability analysis of soil slopes. The MCS was combined both x and y directions. Before modeling the tunnel excava-
with the numerical analysis to introduce the spatially vari- tion, the numerical calculation was performed to approach
able soil properties as random field models (RFMs) (Fenton the balanced condition for initial in situ stresses. Figure 1
1999; Kim and Major 1978). Compared to the soil, rock is illustrates the model geometry and boundary conditions. The
a more complicated environment due to the effect of dif- field stresses for the studied section were set, defining a con-
ferent parameters such as strength properties (e.g., uniaxial stant hydrostatic in situ stress field ( k = 1 ) (Geodata 2016).
compressive strength, elastic modulus, cohesion, friction), The sequential excavation method is being utilized in the
joint properties (e.g., roughness, spacing, and orientation), construction phase through top heading and bench technique
weathering, and so on (Cai et al. 2004; Cai 2011; Chen et al. (i.e., two stages). The top heading of the tunnel is in the form
2019; Ching et al. 2011). Similar researches were carried out of an arc with a diameter of 13 m, and the rectangular bench
for rocks afterward (Haldar and Babu 2008; Lü et al. 2018; has a side length of 3.3 m with a total tunnel height of 9.8 m.
Pandit and Babu 2021; Pandit et al. 2018; Song et al. 2011). After excavation of each sequence, the support system was
Song et al. (2011) investigated the effect of spatial variability installed using liner elements. Three numerical monitoring
of rock mass properties on the ground deformation due to points were selected based on the predefined measurement
tunnelling. Idris et al. (2016) proposed the strength classifi- points on the crown and sidewalls. The excavation sequences
cation method (SCM) for analyzing tunnel stability in spa- and measurement points are also presented in Fig. 1.
tially variable rocks, which requires a lower computational The MCS method was implemented in FLAC using a FISH
time than the RFM. Yu et al. (2019) evaluated the tunnel function. FISH is a scripting language embedded within the
liner performance using conditional and unconditional RFM software to define new variables and functions (Itasca 2015).
models. Zhang et al. (2022) compared the number of stud- The scripted function involves an iterative process of assign-
ies conducted in the field of spatial variability over different ing random variables to the numerical zones. As said before,

13
Arab J Geosci (2022) 15: 1490 Page 3 of 18 1490

Fig. 1  a Boundary conditions and model geometry; b excavation sequences (S1 and S2) and measurement points (C, L, and R)

two different procedures were followed for property mapping, l = int ((j).urand) + 1 lcount ≤ number of calculated class zones
which are described below. (4)
where the pair (k.l) shows a zone located in the kth row and
Stochastic mapping approach
lth column. The pair is selected randomly, and its property
variables are assigned based on a certain property class
This method divides probability density functions into diverse
weaker or stronger than the mean values. The function int
classes based on the mean value and standard deviation. Sub-
rounds the product of i and urand to its nearest integer, while
sequently, it selects the class representatives and assigns them
urand is a stochastic uniform value between 0 and 1. The
stochastically to the model zones. The schematic illustration
process continues until the pair number approaches the cal-
and the flowchart of the mapping process are presented in
culated frequency for each class zones (a–f in Fig. 2):
Figs. 2 and 3, respectively. Each model realization might be
The process of random properties realization might be
specified by a matrix in the form of [zij ]280×280 , where zij rep-
summarized as follows:
resents a zone located in the ith row and jth column whose
centroid coordinates are distinguished by
(1) Construction of model grid with n + 2 nodes in both x
xi = l∕(n + 1) + i.sl i = 0.1.2 … .279 (1) and y directions
(2) Determination of zone centroid coordinates (xi , yi ) using
Eqs. (1) and (2)
yj = l∕(n + 1) + j.sl j = 0.1.2 … .279 (2)
(3) Calculating proportional frequency for every property
where sl is the side length of each zone and n + 1 represents class of the variables
the number of zones in horizontal or vertical directions. (4) Assigning mean values of the properties based on their
After determining the zone centroid, random properties are distribution functions to all numerical zones
assigned to the corresponding zones denoted by zij . To this (5) Allocating random numerical zones using Eqs. (3) and
end, the zones are selected randomly using the following (4)
equations: (6) Assigning the values that belong to the other classes
(weaker or stronger than the mean) to the specified
k = int ((i).urand) + 1 kcount ≤ number of calculated class zones zones in the last step
(3)

13
1490 Page 4 of 18 Arab J Geosci (2022) 15: 1490

Fig. 2  Schematic illustration of the property assignment process in SCM

(7) Repeating steps 5 and 6 to the point that all the zones that the variability would be different from one realization
are assigned to another. It is, hence, required to generate an adequate
(8) Repeating steps 1 to 7 to the point that the quantity of number of realizations for the analysis. In this regard, there
simulations does not significantly affect the response are some differences between the SCM and RFM techniques
required as follows:

In this regard, each realization corresponds to a possible • All the values in the variation range are probable to occur
arrangement of geo-mechanical properties of the ground. in a random field, while the SCM uses the representative
Moreover, using the procedure, the average values for the values for each probability class.
model parameters remain very close to the mean values. • The RFM considers the correlation between different
Furthermore, analyzing several realizations would result in points in space, while it is ignored in the SCM.
a preferable perspective on the tunnel response. • In contrast to the SCM, the cross-correlation between
cohesion ( c ) and friction angle ( 𝜑 ) is considered in the
Random field mapping approach random fields.

The random field theory provides a practical tool for inves- Figure 4 presents the flowchart of the random field map-
tigating the spatial variability of geo-mechanical properties. ping process. The detailed formulation of the random field
The essential principles of the method have been introduced generation process is described in the following subsections.
by Vanmarcke (1977, 1983), based on which three statisti-
cal features are required to define a random field: (1) mean Correlation structure
value; (2) variance (or standard deviation or coefficient of
variation); and (3) scale of fluctuation or correlation length. Geo-materials, especially sedimentary rocks, demonstrate
Due to the inherent randomness of the analysis, it is clear spatial randomness in their properties due to the formation

13
Arab J Geosci (2022) 15: 1490 Page 5 of 18 1490

START

ith realization construction


j = j+1 Yes j < nicount

Model descretization

No
Assigning mean values to all
zone

Calculating the class frequency


ni = ni+1 Yes ni < nn
(nicount)

Allocating zone j centroid No

Assigning representative value to


i = i+1 Yes i<n
zone centroid

END

Fig. 3  Flowchart of the property assignment process in SCM

process. In this study, the Markov autocorrelation function the model to consider the sample size effect (Fenton and
has been used in defining the correlation between different Griffiths 2008). 𝜎T2 can be determined by Eq. (6) as follows:
points in the space. The two-dimensional (2D) form of this
function might be presented as 𝜎T2 = 𝜎 2 Γ(T) (6)

⎛ � where  Γ(T) denotes the variance reduction factor and
�� 2𝜏 �2 � 2𝜏y �2 ⎞
𝜌(𝜏x , 𝜏y ) = exp ⎜−� x
+ ⎟ (5) reduces the point variance (𝜎 2 ) under local averaging. Fur-
⎜ 𝜃x 𝜃y ⎟ thermore, T is the averaging domain of the random field.
⎝ ⎠
The 2D form of  Γ(T) in a squared finite difference mesh
where 𝜌 is the correlation coefficient between two points would be:
separated by known distances ( 𝜏x and 𝜏y ) in the x and y direc- T T
tions, respectively. The parameters 𝜃x and 𝜃y are the correla- 4
Γ(T) = (T − 𝜏x )(T − 𝜏y )𝜌d𝜏x d𝜏y (7)
tion lengths in these directions as well. In this paper, it is T4 ∫ 0∫ 0
assumed that the correlation lengths are identical in the both Due to the complexity of Eq. (7), it is often practical to
directions (i.e., 𝜃x = 𝜃y = 𝜃). define a dimensionless spatial correlation length ( Δ ) as

Local averaging and variance reduction 𝜃


Δ= (8)
T
The given variance 𝜎 2 of inherent spatial variability is at the
The numerical integration of Eq. (7) leads to the graphi-
point level, while it is needed to calculate the variance of an
cal illustration of the variance reduction function, presented
element size ( 𝜎T2  ) in the random field generation. The local
in Fig. 5. It can be observed that Γ(T) increases with the
averaging process is adopted to obtain 𝜎T2  , which enables
increase of the dimensionless correlation length.

13
1490 Page 6 of 18 Arab J Geosci (2022) 15: 1490

START Is the variable auto-


No
correlated?

ith realization construction


Yes

Model descretization to nicount


zones Implementing auto-correlation
function
Generating standard normal
number

Are the variables


No
Generating variable based on the cross-correlated?
statistics

Yes
Allocating zone j centroid

Implementing cross-correlation
function
Assigning variable to zone centroid

i = i+1 Yes i<n


j = j+1 Yes j < nicount

No
No

END

Fig. 4  Flowchart of the property assignment process in RFM

Effect of the variance reduction on the random field


fluctuation

The Gaussian random field might be generated by Eq. (9)


using the mean value ( 𝜇 ) and the product of normalized
correlated random field ( U2 ) and standard deviation in ele-
ment size ( 𝜎T ):
U = 𝜇 + 𝜎T .U2 (9)
The normalized correlated random field U2 can be
expressed as follows:
U2 = L.U1 (10)
where L is the lower triangular of the covariance matrix
decomposition and U1 is the standard Gaussian random
field. The Cholesky decomposition was used to obtain L.
Substituting Eqs. (10) and (6) in Eq. (9) would result in Fig. 5  Variance reduction factor versus Δ for a Markov correlation
Eq. (11): function

13
Arab J Geosci (2022) 15: 1490 Page 7 of 18 1490


U = 𝜇 + 𝜎 Γ.L.U1 (11) 2016). Figure 6 presents the longitudinal profile of the tun-
nel. According to the profile, the studied section is located in
The Cholesky decomposition might also be used to derive the Shemshak Formation, close to the north portal (chainage
the cross-correlation between c and 𝜑 . The cross-correlation 0 + 293.20 m). The formation mainly consists of argillite
matrix is defined as follows: and sandstone sequences with coal lenses and dacite dykes.
[ ] The highest uniaxial strength values are associated with the
1 𝜌c𝜑 � �
dacite samples, while the median and the lowest are related
𝜌cc = = L TL (12)
𝜌c𝜑 1 to the sandstone and the argillite samples, respectively. Fur-
thermore, the groundwater condition is in the form of drip-
where 𝜌c𝜑 = −0.6 is the cross-correlation factor and L and

ping (e.g., 0.1 to 0.3 lit/s) in this section.


L T are the lower triangular of the cross-correlation matrix

The coefficients of variation of rock mass properties


decomposition and its transpose, respectively.
were considered based on the literature (Mazraehli and Zare
The cross-correlated standard normal matrix might be defined as
2020). Table 1 presents the coefficients of variation ( CoVs )

U3 = L .U2 (13) of different properties. Different correlation lengths were
assumed for the random fields which include 0.1, 0.25, 0.5,
Finally, the cross-correlated random field can be expressed 1, 2, 3, 4, and 5 m. Based on this table and the mean values
using Eq. (14) as follows: of the properties, the statistical parameters were calculated
√ (Table 2).
U = 𝜇 + 𝜎 Γ.U3 (14)

It should be noted that this study is focused on the uncer-


tainty of the geo-mechanical properties, while discontinuities Spatial variability and statistical features
are the other aspects of uncertainty in rock engineering. In the
current paper, the effect of discontinuities has been implicitly As stated above, it is required to introduce three different
considered in the GSI, which is used for rock mass classifica- statistics for random field generation: mean value, disper-
tion purposes. On the other hand, it is also possible to explic- sion index, and scale of fluctuation. The mean value is the
itly consider the effect of discontinuities by utilizing stochastic constant part of the variable, and the others determine its
modeling techniques such as discrete fracture network (DFN). variation around it. As a non-dimensional parameter, it is
Based on the available literature, it has been reported that dis- sometimes beneficial to use CoV instead of standard devia-
continuities cause some irregularities in stress and displace- tion ( 𝜎 ) to describe relative data dispersion. Figure 7 com-
ment distribution so that the larger displacements occur adja- pares different realizations developed using the RFM and
cent to these structures (e.g., Jia and Tang 2008; Maazallahi SCM methods. It also shows the effect of 𝜃 and CoV on
and Majdi 2020; Taghizadeh et al. 2020; Xing et al. 2018). the random field realizations of the geo-mechanical proper-
ties. According to the figure, the variability increases with 𝜃
Case study and modeling specifications and CoV  , but the effect of the latter one is more significant.
Contrarily, the spatial variation level remains almost con-
The Alborz twin tunnels include an essential part of the Teh- stant for an SCM realization since there are only a certain
ran–North expressway with 6300 m in length each (Geodata number of classes for each property. In this case, the spatial

Fig. 6  Longitudinal profile of the Alborz twin tunnels (Geodata 2016)

13
1490 Page 8 of 18 Arab J Geosci (2022) 15: 1490

Table 1  Coefficients of Characteristic CoV (%) fewer than the RFM, the same number of simulations was
variation of the geo-mechanical considered sufficient to produce the corresponding dataset.
properties 70
Gm After estimating the required sample size, it was tried to
Km 70 select the optimum number of classes for the SCM mod-
𝜎tm 84 els. Accordingly, seven sets of SCM models were examined
c 25 with different class numbers. Figure 9 presents the repre-
𝜙 11 sentative values for these models. Since there was no need
to repeat the same simulation 100 times, the mean values
variation of the properties specifies the difference among were assigned to a single model for the first case. After cal-
the realizations. culation, the representative values were assigned to the cor-
responding model using the SCM scheme (as illustrated in
Fig. 3). According to Fig. 9b, the second realization set has
Sample size two property classes for each parameter, which comprise
50% of the model zones in each realization. The third one,
The accuracy of the MCS technique largely depends on the however, has three property classes for which the major class
number of simulations (Ang and Tang 1984). It is, therefore, includes 68.2% of the zones, while the minor two possess
clear that an increase in the number of simulations would 15.7% of them each and so on.
result in a higher level of accuracy. In order to achieve an To select the optimum class number, the modeling results
acceptable level of accuracy, it was tried to evaluate the effect were compared for unlined tunnels. Figure 10 shows the
of the sample size on the tunnel displacement before support mean displacements versus the number of property classes.
installation. Figure 8 shows the mean displacements versus It can be observed that the displacement values increase by
the sample size for the RFM procedure (i.e., with different increasing the class numbers. It is mainly because of the
correlation lengths). The displacement increases with the local presence of weaker zones in the tunnel boundary for
sample size up to a definite value, and after that, some fluctua- the models with a higher number of classes. On the other
tions were recorded. Afterward, the mean displacement con- hand, the displacements for the last set with seven property
verges to a representative value that remains almost constant classes are almost identical to its lower set. It was, therefore,
up to 110 simulations. Hence, it is decided to use the samples decided to select the set with six property classes as the
with 100 simulations as the representative sample size for representative set for the SCM procedure.
the RFM. Since the influencing parameters for the SCM are

Stress relaxation factor


Table 2  PDFs and their statistical parameters of rock mass
characteristics
The concept of stress relaxation was used to consider the effect
Characteristic Statistic Value of three-dimensional condition in the models. The value of
stress relaxation can be calculated using Eq. (15), presented
Gm(GPa) Mean 1.40
by Panet (1995):
Standard deviation 1.00
PDF Gamma � �2
ur ⎡ ⎤
0.75
Km(GPa) Mean 1.70 𝜆= = 0.25 + 0.75 ⎢1 − ⎥ (15)
Standard deviation 1.20 urmax ⎢ 0.75 + x

⎣ r ⎦
PDF Lognormal
𝜎tm(MPa) Mean 0.75 where x is the advance rate, r denotes the tunnel radius, ur
Standard deviation 0.63 is the displacement after the advance rate is completed, and
PDF Lognormal urmax denotes the final displacement of the tunnel. Here, the
c(MPa) Mean 0.44 advance rate of the tunnel was 3 m, and its equivalent radius
Standard deviation 0.11 ( r ) was calculated as (13 + 9.8)∕4 = 5.7 m.
PDF Normal
𝜙(◦) Mean 40.19
Standard deviation 4.36 Performance function
PDF Normal
GSI Mean 45.00 The liner elements were utilized to simulate the support sys-
Standard deviation 8.30 tem. The elements are capable of tolerating bending moment
PDF Normal ( M  ), thrust ( N  ), and shear ( Q ) forces. Since the paper was

13
Arab J Geosci (2022) 15: 1490 Page 9 of 18 1490

Fig. 7  Stochastic realization
for a RFM with 𝜃 = 0.5m
and CoV = 10% ; b RFM with
𝜃 = 0.5m and CoV = 50%
c RFM with 𝜃 = 5m and
CoV = 10% ; d RFM with
𝜃 = 5m and CoV = 50% ; e SCM

not intended to study the support–ground interface, the inter- axial stress resulting from the thrust would be constant
action was ignored. The liner elements were installed after throughout the liner cross-section and can be determined as
excavating each sequence (top or bench) and stress relaxa- follows (e.g., Timoshenko 1976):
tion. Furthermore, the stresses generated in the lining were
N
used to define the tunnel liner performance function. 𝜎N = (16)
A
In this equation, 𝜎N denotes the axial stress induced by the
Stresses induced in the support system thrust N  , and A is the area of the liner cross-section. The bend-
ing stress induced by the bending moment M can be expressed
As shown in Fig. 11, the stresses counteract the internal by Eq. (17) (e.g., Timoshenko 1976):
forces to maintain the equilibrium state. In this regard, the

13
1490 Page 10 of 18 Arab J Geosci (2022) 15: 1490

Mean right displacement (m) 0.035 0.040

Mean roof displacement (m)


0.030 0.035

0.030
0.025
0.025
0.020
0.020
0.015
0.015
0.010
0.010
Theta 0.1 Theta 0.1
0.005 Theta 2.0 Theta 2.0
0.005
Theta 5.0 Theta 5.0
0.000 0.000
0 50 100 0 50 100
Number of simulations Number of simulations
(a) (c)
0.035 0.080

Mean floor displacement (m)


Mean left displacement (m)

0.030 0.070

0.060
0.025
0.050
0.020
0.040
0.015
0.030
0.010
0.020
Theta 0.1 Theta 0.1
0.005 Theta 2.0 Theta 2.0
0.010
Theta 5.0 Theta 5.0
0.000 0.000
0 50 100 0 50 100
Number of simulations Number of simulations
(b) (d)

Fig. 8  Mean values of displacement versus number of simulations in the a right wall; b left wall; c crown; d invert

Mz In this equation, 𝜎 max and 𝜎min are the maximum and


𝜎M = − (17)
I minimum axial stresses at the intrados and extrados of the
liner, respectively. The parameter t also denotes the lining
where 𝜎M is the bending stress resulted from the bending moment
thickness. It should be noted that since shotcrete is an engi-
M , I denotes the moment of inertia, and z is the perpendicular
neered material, its properties were considered constant
distance to the neutral axis of the liner which is positive above
over the occupied space. The resulting shear stress 𝜏 can
the axis. It is, therefore, concluded that M causes compressive
also be calculated using Eq. (19) (e.g., Timoshenko 1976):
stress in the area below the neutral axis and tensile stress in the
[ ( )2 ]
opposite side. Moreover, the maximum bending stresses occur at 3Q 2z
the intrados and extrados of the lining cross-section. 𝜏= 1− (19)
2A t
The maximum and minimum axial stresses induced in
the tunnel lining might be determined by superposition The minimum shear stress is equal to zero at the intra-
of the N − M interaction, as presented by Eq. (18) (e.g., dos and extrados of the liner, while the maximum values
Timoshenko 1976): occur at the neutral axis and equals 3Q∕2A . It is, therefore,
obvious that the effect of Q on the stress state reaches its
N Mt highest level at the neutral axis. The principal stresses can
(18)
max
𝜎min = ±
A I 2

13
Arab J Geosci (2022) 15: 1490 Page 11 of 18 1490

(a) (b) (c)

(d) (e) (f)

(g)

Fig. 9  Representative values of each property class for different SCM models

be calculated in terms of the maximum shear stress ( 𝜏max ) Lining failure definition
and 𝜎N as follows (e.g., Timoshenko 1976):
√ Since the purpose of this paper was to design shotcrete lin-
( 𝜎 )2
1
𝜎3 =
𝜎N
± N 2
+ 𝜏max (20) ing, it is necessary to investigate the safety and reliability of
2 2 the support system. The compressive and tensile strengths
of the lining were supposed as 20 and − 4 MPa, respectively.
where 𝜎 1 and 𝜎3 are the major and minor principal stresses at
Its original design thickness ( w ) was also assumed to be
a point located at the neutral axis, respectively.
0.1 m. The resistances provided by the lining support can
be described as follows:

13
1490 Page 12 of 18 Arab J Geosci (2022) 15: 1490

0.025 shotcrete liner is often accompanied by local supports such


as lattice girders and rock bolts.
Mean displacement (m)

0.020 Based on the interactions due to the induced stresses, it


can be inferred that there are four expected failure modes
0.015 for the support system, including (1) failure in compression
due to N − M interaction; (2) failure in tension due to N − M
0.010 interaction; (3) failure in compression due to N − Q inter-
Roof action; and (4) failure in tension due to N − Q interaction.
0.005
Floor In the case of the N − M interaction, the tunnel liner would
Right fail in compression when 𝜎 max exceeds Rc . On the other hand,
Left
tensile failure can occur when 𝜎min is less than Rt . Similarly,
for the N − Q interaction, the compressional liner failure takes
0.000
0 2 4 6 8
Number of SCM classes place when 𝜎 1 at the neutral axis exceeds Rc . Contrarily, the
tensile failure, in this case, would be the result of minimum
Fig. 10  Mean displacements versus number of property classes principal stress ( 𝜎3 ) less than Rt (considering its negative sign).

Performance function formulation


Rc = FoS.𝜎c (21)
It is required to define a performance function to inves-
Rt = FoS.𝜎t (22) tigate the serviceability of the lining. This function was
used here to determine the PoF, defined using Eq. (23):
where Rc and Rt are the compressive and tensile resistances,
respectively, 𝜎c and 𝜎t also denote the compressive and ten- g(x) = U − R(x) (23)
sile strengths of the lining, and FoS is the factor of safety
provided by the lining. The FoS was considered since a where U is the liner resistance ( Rc or Rt ), and R(x) denotes
the variable induced stress.

Fig. 11  Stresses in tunnel
support due to the interaction of
a thrust force–bending moment;
b thrust–shear force

(a)

(b)

13
Arab J Geosci (2022) 15: 1490 Page 13 of 18 1490

[ ] [ ]
When the performance function becomes negative pnq = P 𝜎1 > Rc = P Ccnq > FoS (30)
c
( g(x) < 0 ), it implies that the corresponding stress exceeds
the resistance limit and a failure event is probable. Con- nq [ ] [ nq ]
trarily, when g(x) > 0 , the liner would be stable, and its pt = P 𝜎3 > Rt = P Ct > FoS (31)
performance is desirable. The limit state surface is also
where pnmc and c are the compressional PoF in the
nq
p N−M
defined by g(x) = 0 , which is the boundary between the
and N − Q interaction, respectively. Furthermore, pnm and
unstable and the stable conditions. t
pt are also the tensile PoF in the N − M and N − Q interac-
nq
To establish a non-dimensional criterion of failure, the
tion, respectively.
induced stresses were divided by the compressive and ten-
sile strengths of the liner. It would introduce a parameter
that is called strength factor and may be different regarding
Probability of failure (PoF)
the interaction and failure mode type as follows:
Ccnm = 𝜎max ∕𝜎c (24) The FoS is the first parameter needed for determining the PoF.
The primary factor of safety provided by the liner has been
considered as 1.2. The computed probabilities of failure for
Ctnm = 𝜎min ∕𝜎t (25)
the SCM analysis are presented in Fig. 12. It can be observed
that the N − M tensile failure ( pnm t = 72%
 ) would be more
Ccnq = 𝜎1 ∕𝜎c (26) probable than the other modes, and the N − M compressional
failure ( pnm
c = 4% ) stands in the second place among others.
nq
Ct = 𝜎3 ∕𝜎t (27) Moreover, the PoFs due to the N − Q interaction (both in
compression and tension) are negligible (almost zero).
In the above equations, Ccnm and Cc are the compressional
nq
The obtained PoF values for different procedures are pre-
strength factors in the N − M and N − Q interaction, respec- sented and compared in Table 4. It can be observed that the
tively. Furthermore, Ctnm and Ct are also the tensile strength
nq
implementation of cross-correlated parameters has led to a lower
factors in the N − M and N − Q interaction, respectively. PoF. It can be related to the negative cross-correlation factor
The mean values of strength factors obtained from differ- between c and 𝜑 , which equates to − 0.6. It results in a larger
ent procedures are presented in Table 3. As stated in the pre- cohesion value where the friction angle is small and vice versa.
vious sections, these procedures include the SCM and RFM. Besides, the effect of correlation length is obvious in this table.
Four different sets were considered in the RFM models: (1) It can be seen that the PoFs in the N − M interaction are more
𝜃 = 0.5m , cross-correlated c and 𝜑 ; (2) 𝜃 = 5m , cross-cor- significant than in the N − Q interaction. Furthermore, the cor-
related c and 𝜑 ; (3) 𝜃 = 0.5m , independent c and 𝜑 ; and (4) responding SCM values are higher than the RFM probabilities.
𝜃 = 5m , independent c and 𝜑 . According to the table, the
N − Q strength factors are smaller than the others, while the
correlation length effect is significant in the RFM models. Effect of statistical parameters on PoF
As a result, the PoF of the tunnel lining can be defined
for different modes as In this section, the effect of CoV and 𝜃 was investigated on the
[ ] [ ] PoF of the tunnel liner. According to Fig. 13, the tensile PoF in
pnm
c
= P 𝜎max > Rc = P Ccnm > FoS (28) the N − M interaction is more than in the other failure modes.
Even with an increase in CoV and 𝜃 , the N − Q compressional
[ ] [ ]
pnm = P 𝜎min < Rt = P Ctnm > FoS (29) and tensile PoFs are almost zero. It is also observed that the
t
N − M curves steepen beyond the CoV of 30%. It means that
the wider the variation range of geo-mechanical properties,
the higher the liner PoF.

Table 3  Mean values of strength factors from different procedures


PoF SCM Cross-correlated RFM Independent RFM Discussion
𝜃 = 0.5m 𝜃 = 5m 𝜃 = 0.5m 𝜃 = 5m
In the above sections, different procedures were adopted for
Ccnm 0.72 0.78 0.99 0.75 0.88 stochastic–numerical analysis of tunnel lining performance,
Ctnm 1.49 0.66 1.23 0.47 1.20 including the SCM and RFM methods. Although the running
nq
Cc 0.23 0.36 0.40 0.33 0.36 time for the SCM is considerably lower than the RFM, the
nq
Ct 0.03 0.01 0.02 0.01 0.02 former lacks incorporation of the effect of the correlation

13
1490 Page 14 of 18 Arab J Geosci (2022) 15: 1490

Fig. 12  Liner probability of failure in a compressional N − M  ; b tensile N − M  ; c compressional N − Q ; d tensile N − Q

length on the response of interest. It is also observed that underground excavations during construction and service
the effect of CoV is more significant on the SCM result so periods to control and investigate the ground behavior. The
that it increases beyond 30% of variation around the mean Alborz tunnel was instrumented to monitor inward deforma-
property values. tions using convergence pins and extensometers. Conver-
This section compares the modeling results obtained from gence pins are the most common tools used for tunnel move-
different methods with the in situ measurements to outline ment measurement, which evaluate the relative displacement
the most probable variability characteristics of the studied of two points on the excavation boundary with respect to
section. It is essential to utilize monitoring techniques in each other (Kavvadas 2003). The monitoring results are pre-
sented in Fig. 14. According to the figure, relative displace-
Table 4  Probability of failure from different procedures ( FoS = 1.2) ments were equal to 4 mm for the hypothetical line connect-
ing the left and the right walls (L-R) and 3 mm for the crown
PoF SCM Cross-correlated RFM Independent RFM to both the right wall (C-R) and the left wall (C-L).
𝜃 = 0.5m 𝜃 = 5m 𝜃 = 0.5m 𝜃 = 5m The convergence values from numerical analysis and
the measured values by monitoring surveys are compared
pnm 0.04 0.03 0.22 0.05 0.12
c
in Table 5. In this regard, it would be possible to obtain a
pnm 0.72 0.15 0.51 0.50 0.63
t
perspective on how the geo-mechanical properties vary in
nq
pc 0.00 0.00 0.00 0.00 0.00
the studied section. It must be noted that the accuracy of
nq
pt 0.00 0.00 0.00 0.00 0.00
the monitoring tool was in millimeters, but the numerical

13
Arab J Geosci (2022) 15: 1490 Page 15 of 18 1490

Fig. 13  CoV and 𝜃 effects on probability of failure in a compressional N − M  ; b tensile N − M  ; c compressional N − Q ; d tensile N − Q

displacements were calculated in 0.01 mm. The mean numer- finite difference method. In this regard, the lining perfor-
ical convergences are in good accordance with the moni- mance function was defined based on the compressional and
tored displacements. However, it seems that the determin- tensile failure modes. Besides, the deterministic numerical
istic model underestimates the convergences. Furthermore, analysis was also performed for comparison purposes, and the
it can be observed that a higher correlation length results numerical results were compared to the in situ measurements
in the overestimation of convergence for this case. On the to check the validity. According to the results, the following
other hand, when the cross-correlation between c and 𝜑 is conclusions can be reported from the conducted study:
neglected, the calculated convergence values would be larger
than the cross-correlated case. Accordingly, the cross-cor- • Compared to the correlation length ( 𝜃 ), the coefficient
related models with a lower correlation length reflect more of variation ( CoV  ) has a more significant effect on the
realistic in situ condition. Last but not least, the stochastic spatial variability of a random field. On the other hand,
modeling techniques require more computational efforts than the spatial variability decreases with increasing 𝜃.
the deterministic analyses, but they can be considered more • The SCM models calculate lower support requirements
realistic. than RFMs when CoV is smaller than 30%. After this
point, the design values increase drastically for the
SCM procedure.
Conclusions • The cross-correlation between cohesion ( c ) and friction
angle ( 𝜑 ) affects the probability of failure (PoF), so the
Tunnel lining performance was stochastically analyzed using cross-correlated models have led to lower PoF. It can
different procedures, including strength classification (SCM) be related to the negative cross-correlation between c
and random field (RFM) methods. The spatial variability of and 𝜑 in this study.
rock mass properties was considered using the stochastic

13
1490 Page 16 of 18 Arab J Geosci (2022) 15: 1490

Fig. 14  Monitoring results

Table 5  Calculated and Convergence In situ Deterministic SCM Cross-correlated RFM Independent RFM
measured convergences (mm)
𝜃 = 0.5m 𝜃 = 5m 𝜃 = 0.5m 𝜃 = 5m

L-R 4.00 1.55 3.84 4.11 4.26 5.64 6.11


C-R 3.00 1.55 3.11 3.22 3.71 4.09 4.44
C-L 3.00 1.57 2.78 3.19 3.88 4.12 4.65

• The effect of the axial thrust–bending moment interac- Acknowledgements  The authors would like to express their grateful
tion on the lining failure is more significant than the appreciation to Geodata Engineering SpA and Tehran-Shomal Freeway
Company for providing the required data during the study.
axial thrust–shear force interaction. In this case, it was
also observed that the probability of tensile failure Data availability  The data that support the findings of this study are
mode was much higher than the compressional failure. available on request from the corresponding author.
• The PoFs in both compressional and tensile mode due
to thrust–shear force interaction were almost zero, even Declarations 
with increasing CoV and 𝜃 values.
• In general, it can be concluded that the spatial variabil- Conflict of interest  The authors declare no competing interests.
ity has a remarkable effect on the tunnel lining design,
which must be considered. It would result in a reliable
design process compared to the conventional procedures. References

Finally, it can be stated that defining the performance Ang AHS, Tang WH (1984) Probability concepts in engineering planning
and design. vol 2: Decision, risk, and reliability. Wiley, New York
function based on the strength factors would enable the Behnia M, Seifabad MC (2018) Stability analysis and optimization of
future studies to consider the effect of different factors of the support system of an underground powerhouse cavern consid-
safety ( FoS ) provided by the support system. ering rock mass variability. Environ Earth Sci 77(18):645

13
Arab J Geosci (2022) 15: 1490 Page 17 of 18 1490

Binder K, Heermann DW (2010) Monte Carlo simulation in statistical Itasca (2015) FLAC - Fast Lagrangian Analysis of Continua. Version
physics. Springer, Heidelberg 7.0. Minneapolis, Minnesota
Cai M (2011) Rock mass characterization and rock property variability Jiale H, Xiaohong L (2021) Reliability analysis considering spatial
considerations for tunnel and cavern design. Rock Mech Rock variability by combining spectral representation method and sup-
Eng 44(4):379–399 port vector machine. Eur J Environ Civ Eng 25(6):1136–1157
Cai M, Kaiser PK, Tasaka Y, Maejima T, Morioka H, Minami M Jia P, Tang CA (2008) Numerical study on failure mechanism of tunnel
(2004) Generalized crack initiation and crack damage stress in jointed rock mass. Tunn Undergr Space Technol 23(5):500–507
thresholds of brittle rock masses near underground excavations. Kavvadas MJ (2003) Monitoring and modeling ground deformations
Int J Rock Mech Min Sci 41(5):833–847 during tunnelling. In: Proceedings of the 11th FIG Symposium
Chehade FH, Shahrour I (2008) Numerical analysis of the interaction on Deformation Measurements. Santorini, Greece, pp 371–390
between twin-tunnels: influence of the relative position and con- Kim H, Major G (1978) Application of Monte Carlo techniques to
struction procedure. Tunn Undergr Space Technol 23(2):210–214 slope stability analysis. In: the 19th US Rock Mechanics Sympo-
Chen D, Xu D, Ren G, Jiang Q, Liu G, Wan L, Li N (2019) Simula- sium: American Rock Mechanics Association. OnePetro. State-
tion of cross-correlated non-Gaussian random fields for layered line, Nevada
rock mass mechanical parameters. Comput Geotech 112:104–119 Kroetz HM, Do NA, Dias D, Beck AT (2018) Reliability of tunnel lin-
Ching J, Hu YG, Yang ZY, Shiau JQ, Chen JC, Li YS (2011) Relia- ing design using the hyperstatic reaction method. Tunn Undergr
bility-based design for allowable bearing capacity of footings on Space Technol 77:59–67
rock masses by considering angle of distortion. Int J Rock Mech Langford JC, Diederichs M (2013) Reliability based approach to tunnel
Min Sci 48(5):728–740 lining design using a modified point estimate method. Int J Rock
Fenton GA (1999) Estimation for stochastic soil models. J Geotech Mech Min Sci 60:263–276
Geoenviron Eng 125(6):470–485 Lü Q, Xiao Z, Zheng J, Shang Y (2018) Probabilistic assessment of
Fenton GA, Griffiths DV (2008) Risk assessment in geotechnical engi- tunnel convergence considering spatial variability in rock mass
neering. Wiley properties using interpolated autocorrelation and response surface
Fortsakis P, Litsas D, Kavvadas M, Trezos K (2011) Reliability analy- method. Geosci Front 9(6):1619–1629
sis of tunnel final lining. Geotech Safe Risk: ISGSR;409–418 Maazallahi V, Majdi A (2020) Numerical appraisal of rock mass ani-
Funatsu T, Hoshino T, Sawae H, Shimizu N (2008) Numerical analy- sotropy effect on elastic deformations of a circular tunnel. Arab
sis to better understand the mechanism of the effects of ground J Geosci 13(13):1–17
supports and reinforcements on the stability of tunnels using Mazraehli M, Zare S (2020) An application of uncertainty analysis to
the distinct element method. Tunn Undergr Space Technol rock mass properties characterization at porphyry copper mines.
23(5):561–573 Bul Eng Geolog Environ 79(7):3721–3739
Geodata (2016) Detailed design of Alborz main tunnel. Retrieved from Napa-García GF, Beck AT, Celestino TB (2017) Reliability analyses
Tehran-Shomal freeway project. Turin of underground openings with the point estimate method. Tunn
Haldar S, Babu GS (2008) Effect of soil spatial variability on the Undergr Space Technol 64:154–163
response of laterally loaded pile in undrained clay. Comp Geo- Oreste P (2005) A probabilistic design approach for tunnel supports.
tech 35(4):537–547 Comput Geotech 32(7):520–534
Han W, Jiang Y, Li N, Koga D, Sakaguchi O, Chen H (2021) Safety Pandit B, Babu GS (2021) Probabilistic stability assessment of tunnel-
evaluation and failure behavior of degraded tunnel structure with support system considering spatial variability in weak rock mass.
compound diseases of voids and lining defects. Arab J Geosci Comp Geotech 137:104242
14(15):1–10 Pandit B, Tiwari G, Latha GM, Sivakumar Babu GL (2018) Stability
Heidarzadeh S, Saeidi A, Rouleau A (2020) Use of probabilistic analysis of a large gold mine open-pit slope using advanced proba-
numerical modeling to evaluate the effect of geomechanical bilistic method. Rock Mech Rock Eng 51(7):2153–2174
parameter variability on the probability of open-stope failure: a Panet M (1995) Le calcul des tunnels par la méthode convergence-
case study of the Niobec Mine, Quebec (Canada). Rock Mech confinement. Presses de l’Ecole Nationale des Ponts et Chausses,
Rock Eng 53(3):1411–1431 Paris
Hoek E (1998) Reliability of Hoek-Brown estimates of rock mass Pelizza S, Oreste P, Peila D, Oggeri C (2000) Stability analysis of a
properties and their impact on design. Int J Rock Mech Min Sci large cavern in Italy for quarrying exploitation of a pink marble.
35(1):63–68 Tunn Undergr Space Technol 15(4):421–435
Hsu SC, Nelson PP (2006) Material spatial variability and slope Phoon KK, Kulhawy FH (1999) Characterization of geotechnical vari-
stability for weak rock masses. J Geotech Geoenviron Eng ability. Can Geotech J 36(4):612–624
132(2):183–193 Riedmüller G, Schubert W (1999) Rock mass modeling in tunnelling
Idris MA, Nordlund E (2019) Probabilistic-based stope design method- versus rock mass classification using rating methods. In: Vail
ology for complex ore body with rock mass property variability. Rocks 1999, The 37th US Symposium on Rock Mechanics. One-
J Min Sci 55:743–750 Petro. Vail, Colorado, 601–605
Idris MA, Saiang D, Nordlund E (2011) Numerical analyses of the Sari M (2009) The stochastic assessment of strength and deformability
effects of rock mass property variability on open stope stabil- characteristics for a pyroclastic rock mass. Int J Rock Mech Min
ity. In: 45th US Rock Mechanics/Geomechanics Symposium: Sci 46:613–626
American Rock Mechanics Association. OnePetro. San Francisco, Sari M, Karpuz C (2006) Rock variability and establishing confining
California pressure levels for triaxial tests on rocks. Int J Rock Mech Min
Idris MA, Saiang D, Nordlund E (2012) Consideration of the rock Sci 43:328–335
mass property variability in numerical modeling of open stope Seguini M, Nedjar D (2016) Modeling of soil-structure interaction
stability. In: Bergmekanikdag: Stiftelsen bergteknisk forskning- behaviour: geometric nonlinearity of buried structures combined
Befo, Stockholm, Sweden, 111–123 to spatial variability. Eur J Environ Civ Eng 10(2):134–141
Idris MA, Nordlund E, Saiang D (2016) Comparison of different proba- Song KI, Cho GC, Lee SW (2011) Effects of spatially variable
bilistic methods for analyzing stability of underground rock exca- weathered rock properties on tunnel behavior. Prob Eng Mech
vations. Electron J Geotech Eng 21(21):6555–6585 26(3):413–426

13
1490 Page 18 of 18 Arab J Geosci (2022) 15: 1490

Srivastava A (2012) Spatial variability modeling of geotechnical Xing Y, Kulatilake PH, Sandbak LA (2018) Effect of rock mass and
parameters and stability of highly weathered rock slope. Indian discontinuity mechanical properties and delayed rock supporting
Geotech J 42(3):179–185 on tunnel stability in an underground mine. Eng Geolog 238:62–75
Taghizadeh H, Zare S, Mazraehli M (2020) Analysis of rock load for Yu X, Cheng J, Cao C, Li E, Feng J (2019) Probabilistic analysis of
tunnel lining design. Geotech Geolog Eng 38(3):2989–3005 tunnel liner performance using random field theory. Adv Civ Eng
Timoshenko SP (1976) Strength of materials, 3rd edn. Krieger Publish- 2019
ing Company, Florida Zhang W, Han L, Gu X, Wang L, Chen F, Liu H (2022) Tunneling and
Tiwari G, Pandit B, Gali ML, Babu GS (2017) Probabilistic analysis deep excavations in spatially variable soil and rock masses: a short
of tunnels considering uncertainty in peak and post-peak strength review. Undergr Space 7(3):380–407
parameters. Tunn Undergr Space Technol 70:375–387 Zhu W, Li S, Li S, Chen W, Lee C (2003) Systematic numerical simu-
Tiwari G, Pandit B, Gali ML, Babu GS (2018) Analysis of tunnel sup- lation of rock tunnel stability considering different rock condi-
port requirements using deterministic and probabilistic approaches tions and construction effects. Tunn Undergr Space Technol
in average quality rock mass. Int J Geomech 18(4):1–20 18(5):531–536
Vanmarcke EH (1977) Probabilistic modeling of soil profiles. J Geo-
tech Eng Div 103(11):1227–1246 Springer Nature or its licensor holds exclusive rights to this article under
Vanmarcke E (1983) Random fields.  The MIT Press, Cambridge, a publishing agreement with the author(s) or other rightsholder(s);
Massachusetts author self-archiving of the accepted manuscript version of this article
is solely governed by the terms of such publishing agreement and
applicable law.

13

You might also like