Applsci 12 05311 v2

applied
sciences
Article
Numerical Simulation of Non-Stationary Parameter Creep
Large Deformation Mechanism of Deep Soft Rock Tunnel
Jiancong Xu 1, * , Haiyang Wen 1 , Chen Sun 1 , Chengbin Yang 1 and Guorong Rui 2
1 Department of Geotechnical Engineering, Tongji University, Shanghai 200092, China;

1410258@tongji.edu.cn (H.W.); 1832539@tongji.edu.cn (C.S.); 1932507@tongji.edu.cn (C.Y.)
2 China Railway 20th Bureau Group Co., Ltd., Xi’an 710016, China; sycsdxmb@163.com
* Correspondence: xjc1008@tongji.edu.cn
Abstract: The accelerated creep plays an important role in the disasters of soft-rock tunnels under
high stress. However, most of previous studies only involved attenuation creep and uniform creep.
Large deformation disasters of soft rock occurred during the tunneling process in the Qianzhou–
Sanyangchuan Tunnel, Gusu Province, China. In the paper, we developed the nonlinear generalized
Nishihara rheological model with non-stationary parameter creep (NGNRM) to simulate the accel-
erated creep behaviors of soft rocks under high stress, and implemented it in ABAQUS, to reveal
the mechanism of large deformation of soft rock. We proposed the multi-objective back analysis
method of surrounding rock mechanical parameters based on the eXtreme Gradient Boosting and
the non-dominated sorting genetic algorithm-II. In addition, the orthogonal test design method
was used to determine the main parameters affecting the displacement of the tunnel. Using the
proposed method, we can evaluate the large deformation mechanism of deep soft rock tunnels, and
scientifically determine when to reinforce to prevent a large deformation disaster of the tunnel.
Keywords: soft rock; creep; large deformation; numerical simulation; XGBoost; NSGA-II
Citation: Xu, J.; Wen, H.; Sun, C.;
Yang, C.; Rui, G. Numerical
Simulation of Non-Stationary
Parameter Creep Large Deformation
1. Introduction
Mechanism of Deep Soft Rock Tunnel.
Appl. Sci. 2022, 12, 5311. https:// In soft rock strata, the large squeezing deformation of surrounding rock often occurs
doi.org/10.3390/app12115311 during tunneling under high stress [1–6]. Many interesting cases of tunnels, for example,
the section of the F7 fault fracture zone of the Wushaoling Railway Tunnel [7], and the
Academic Editors: Qian Fang and
Saint Martin La Porte access adit of the Lyon–Turin Base Tunnel [2,3], were recorded as
Pengfei Li
representative of large squeezing deformation disasters. These disasters are mostly caused
Received: 25 April 2022 by soft rock rheology [1–4]. Therefore, it is of theoretical and practical interest to investigate
Accepted: 23 May 2022 the large deformation mechanism of deep soft rock tunnels and to explore an efficient
Published: 24 May 2022 numerical modeling method.
Publisher’s Note: MDPI stays neutral
The accelerated creep of soft rock plays an important role in tunnel large deformation
with regard to jurisdictional claims in
disasters under high stress. Many previous works were devoted to investigating the time-
published maps and institutional affil- dependent deformation of surrounding rock around tunnels using viscoelastic models and
iations. viscoplastic models [5]. However, these studies only involved the attenuation creep and
uniform creep of soft rocks [1–6], and seldom involved the accelerated creep.
The identification of rheological curves in the high-stress soft rock test shows that the
viscosity coefficient related to the non-stationary parameter creep is often a function of the
Copyright: © 2022 by the authors. stress level [8], and the previous viscoelastic–plastic models such as the Nishihara model,
Licensee MDPI, Basel, Switzerland. cannot simulate the accelerated rheological characteristics of soft rock. Therefore, only
This article is an open access article the nonlinear rheological model with the non-stationary parameter creep can completely
distributed under the terms and present the accelerated creep characteristics of soft rock under high stress.
conditions of the Creative Commons
In engineering practice, with the variation of the engineering environment and tunnel-
Attribution (CC BY) license (https://
ing process, the mechanical parameters of surrounding rock usually have uncertainty. The
creativecommons.org/licenses/by/
deformation of surrounding rock and the supporting structure is the most direct, obvious
4.0/).
Appl. Sci. 2022, 12, 5311. https://doi.org/10.3390/app12115311 https://www.mdpi.com/journal/applsci

Appl. Sci. 2022, 12, 5311 2 of 23
and comprehensive macroscopic manifestation on the force change of the surrounding

rock support system in the time–space domain of tunneling and support. Therefore, the
mechanical parameters obtained by the displacement back analysis can reflect the influence
of geo-stress, the construction method, and the tunnel structure parameters as a whole. At
present, the back analysis has been widely used to obtain the mechanical parameters of
surrounding rocks by simulating their complicated mechanical behaviors through simple
models combined with measured data [9–15].
Kavanagh and Clough [16] proposed the back analysis theory for the first time, and
they inverted the elastic modulus of elastic solids based on the finite element method.
Since then, some traditional optimization algorithms such as Levenber–Marquard, Gauss–
Newton, Bayesian and other methods were applied to the field of back analysis. Gioda [17]
used the mixed algorithm of the variable rotation method and the simplex method to
invert the mechanics of elastic–plastic rock mass. Later, some scholars proposed the
parameter inverted methods for five commonly used creep models (such as Maxwell,
Kelvin, generalized Kelvin and Burgers models) [18,19].
Although the above methods have made great progress in the field of back analysis of
geotechnical parameters, with the progress of machine learning theory, some scholars have
proposed numerous algorithms with higher computational accuracy and computational
efficiency. Compared with traditional algorithms, these algorithms can better deal with
the complex nonlinear relationship in the geotechnical back analysis, so as to achieve
higher accuracy. The mechanical parameters of surrounding rock may be obtained by
the displacement back analysis using neural networks and machine learning, such as the
genetic algorithm and particle swarm optimization [20–27].
In this study, we propose a nonlinear generalized Nishihara rheological model with the
non-stationary parameter creep (NGNRM), and implement it in ABAQUS. Furthermore, we
propose the multi-objective back analysis method (MBAM) of surrounding rock mechanical
parameters using the eXtreme Gradient Boosting (XGBoost) and the non-dominated sorting
genetic algorithm-II (NSGA-II) for deep soft rock tunnels, to invert the multiple mechanical
parameters of surrounding rock at the same time. Finally, according to the displacement
data of the completed sections or a period of time of the inversion section, we use the
proposed XGBoost–NSGA-II-based MBAM to identify the crucial rheological parameters
of soft rocks, and use the proposed NGNRM method to simulate and analyze the whole
process of large deformation disaster evolution of a deep soft rock tunnel, and reveal its
non-stationary parameter creep large deformation mechanism. These methods have been
successfully applied to the Qianzhou–Sanyangchuan Tunnel located in Gusu Province,
China, with higher precision.
The novelty of this paper is that we propose the NGNRM to fully express the acceler-
ated creep characteristics of soft rock under high stress, and the XGBoost–NSGA-II-based
MBAM, to help in the understanding of numerical simulation rheological models in large
creep evolution during soft rock tunneling.
2. NGNRM with Non-Stationary Parameter Creep

The accelerated creep of soft rock plays an important role in the large deformation
disasters of tunnels. From the introduction, it is known that the previous studies seldom re-
flect the accelerated creep in the whole process of soft rock large deformation. In this paper,
we used the NGNRM (see Figure 1) representing the accelerated creep characteristics of soft
rocks, to investigate the large deformation mechanism of soft rocks. The NGNRM consists
of a Hooke body, a Kelvin body, and a viscoplastic body with the viscosity coefficient as a
function of stress.
 2   
 AT   σs 
where σ is the current stress; σs is the long-term strength of a Bingham body; A
Appl. Sci. 2022, 12, 5311 dimensionless parameter; T 3 of 23 t is l
is the time for the surrounding rock to yield;
time.
Nonstationary
Kelvin body viscoplastic body
Hooke body
E1 σf
E0
σ σ
η1 η2(σ,t)
Figure
Figure 1.1.Rheological
Rheological model
model withwith variable
variable viscoplastic
viscoplastic viscosityviscosity coefficient.
coefficient.
In Figure 1, E0 is the elastic modulus of a Hooke body; E1 and η1 are the elastic
It is noted from Equation (1) that when σ = σ f + σ s body
modulus of a Hooke body and the viscosity coefficient of a Newton
/ A , in
nonlinear creep dege
a Kelvin body,
respectively;
ates σf and ηcreep,
into linear η = stress
2 are the yield E0T /ofAa St., Venant
and strainbody and the viscosity
increases at coefficient
the constant
of a dashpot in the non-stationary viscoplastic body, respectively, and η2 is as a function of
εstress
 = σσs and T , t.that is, steady creep; when σ > σ f + σ s / A , η 2 gradually decrea
/ E0time
withWhen t , reflecting
σ ≥ σf , for theaccelerated
the nonlinear Bingham body, theofexpression
creep process soft rockof
inthe viscosity coeffi-
a high-stress state.
Therefore, when σ ≥ σ f , the one-dimensional creep equation of the nonlinear
cient η2 is denoted by Equation (1) [28].
(σ−σf )t i(2).

coelastic–plastic model is denoted
σ h AEquation
by
 ε = E0 + E0 T +t 1− σ−σ σf A

s
( )
(1)

1 η2 (σ1  h

) = 0 T + t 1 − 1 f · A

E    i2 A σ − σ t
ε (t ) = σ 
E − f
 σ σ
+ 1AT− exp  − σts   +
 0E E  η1    of a Bingham
where σ is the current stress; σs isthe long-term strength
1  σ −σ 
E0 T + t 1 −body; A fis A a 

 yield;t is loadσtime.
dimensionless parameter; T is the time for the surrounding rock to s 
It is noted from Equation (1) that when σ = σf + σs /A, nonlinear creep degenerates
.
into linear
In thiscreep, η=E
paper, 0 T/A,
the and strain
initial strain increases
methodatisthe
usedconstant rate ε = σsthe
to calculate T, that is,
/E0 viscoelastic–pla
steady creep; when σ > σ
constitutive equations fof the + σ /A, η
s NGNRM,2 gradually decreases with t, reflecting the
and its ABAQUS user material subprogram is accelerated
creep process of soft rock in a high-stress state.
veloped by FORTRAN, which is embedded in the ABAQUS. The flowchart of the u
Therefore, when σ ≥ σf, the one-dimensional creep equation of the nonlinear viscoelastic–
defined material
plastic model mechanical
is denoted behavior
by Equation (2). (UMAT) subroutine development for the NGN
is shown in Figure2.
A(σ −by

1
UMATε(tis) =a σFORTRAN 1 program Einterface t
σf )ABAQUS
+ 1 − exp − t 1
+ provided
h i to users
(2) to de
E E η σ − σf
their own material properties, enabling users to use
0 1 1 E 0 T + t 1 −
material σmodels
s
A that are not defi
In this paper, the initial strain method is used to calculate the viscoelastic–plastic
constitutive equations of the NGNRM, and its ABAQUS user material subprogram is
developed by FORTRAN, which is embedded in the ABAQUS. The flowchart of the user-
defined material mechanical behavior (UMAT) subroutine development for the NGNRM is
shown in Figure 2.
UMAT is a FORTRAN program interface provided by ABAQUS to users to define
their own material properties, enabling users to use material models that are not defined
in the ABAQUS material library. The UMAT realizes the data exchange with ABAQUS
through the interface with the ABAQUS main solver.
Appl. Sci. 2022, 12, x FOR PEER REVIEW 4 of 25
in the ABAQUS material library. The UMAT realizes the data exchange with ABAQUS
Appl. Sci. 2022, 12, 5311 4 of 23
through the interface with the ABAQUS main solver.
Drucker–Prager yield criterion is adopted in this paper, denoted by Equation (3).
Incremental step start
Viscoelasto-plastic
calculation Calculate current nonstationary
Call UMAT and get state variables
coefficient of viscosity
Calculate viscoelastic strain increment Calculate viscoplastic strain rate
Calculate the current Calculate the sum of viscoelastic strain

stress level increment and viscoplastic strain increment
Judge whether Calculate Jacobian viscoelasto-plastic

No
F<0? stiffness matrix
Yes
Calculate Jacobian elastic stiffness

matrix on viscoplastic strain rate = 0
Calculate stress increment using initial strain method
No Has the calculation of main program

balance iteration has converged?
Yes
Output the state variables of the current incremental step and go to the next
viscoelasto-plastic analysis step
Figure 2. 2.
Figure UMAT flowchart
UMAT ofof
flowchart nonlinear generalized
nonlinear Nishihara
generalized rheological
Nishihara analysis.
rheological analysis.
Drucker–Prager yield criterion is adopted in this paper, denoted by Equation (3).


 F(σ, c) = p J + η I − ξc
 F ( σ , c )√= J22 +η I11 − ξ c


 η = √ 3 sin ϕ




 sin2ϕ
3 (33+sin ϕ) (3)
η= √ (3)
sinϕ22 ϕ )
√3+·cos

 ξ =3√ (3




 3 (3+sin ϕ)
3 ⋅ cos ϕ
where c and ϕ are the internal cohesion
ξ= and internal friction angle, respectively; I1 is the
 2
(3 + sinof
first invariant of stress; J2 is thesecond3invariant ϕdeviatoric
) stress.
The UMAT subroutine of the NGNRM are implemented by following steps:
Step 1 Call the UMAT subroutine. ABAQUS provides state variables such as stress
and strain at the previous moment and the increment of strain at the current moment.
Step 2 Calculate the viscoelastic strain increments of elements.
Appl. Sci. 2022, 12, 5311 5 of 23
Step 3 Calculate the current stress level and the yield function of the viscoplastic
body, and judge whether <0. If yes, it will enter the next step, and conduct the viscoelastic
calculation. Otherwise, it will enter Step 5.
Step 4 Calculate the Jacobian elastic stiffness matrix when the viscoplastic strain rate = 0.
Step 5 Conduct the viscoelastic–plastic calculation. Calculate the nonstationary coeffi-
cient of the viscosity and viscoplastic strain rate of elements.
Step 6 Calculate the sum of viscoelastic strain increment and viscoplastic strain incre-
ment of elements.
Step 7 Calculate the Jacobian viscoelastic–plastic stiffness matrix.
Step 8 Calculating stress increment using the initial strain method.
Step 9 Determine whether the balance iteration calculation of the main program has
converged. If yes, it will enter the next step. Otherwise, it will return to Step 1.
Step 10 Output the state variables of the current incremental step and go to the next
viscoelastic–plastic analysis step.
3. XGBoost
XGBoost [29,30] takes into account the second order term of Taylor expansion and
introduces richer derivative information, so that it can better capture nonlinear information.
Besides, XGBoost can reduce the over-fitting effect that may be caused by the gradient
boosting algorithm by introducing regular terms.
XGBoost is an improvement on the gradient boosting algorithm, developed by Chen
and Guestrin [29]. It uses Newton’s method to solve the extreme value of the loss function,
and adds a regularization term to the loss function. Its loss function is denoted by the
second-order Taylor expansion. On training, the objective function is composed of two
parts: the first part is the loss of the gradient boosting algorithm, and the second part is the
regularization term.
Given a training sample data set T = {(x1 , y1 ), (x2 , y2 ), · · · , (x N , y N )}, xi ∈ χ ∈ Rn ,
χ is the input space of samples, yi ∈ Y ∈ R, Y is the output space, n is the dimension of
input space. If the input space χ is divided into non-intersecting regions R1 , R2 , · · · , R J ,
and the constant output c j is determined on each region, then the tree can be denoted by
Equation (4).
J
T (x; Θk ) = ∑ c j I x ∈ R j

(4)
j =1
where Θ = ( R1 , c1 ), ( R2 , c2 ), · · · , R J , c J represents the region division of the tree and

the constants on each region; J is the complexity of the regression tree, that is, the number
of leaf nodes.
XGBoost’s boosting tree model can be expressed by Equation (5) as an additive model
of decision trees.
K
f K (x) = ∑ T (x; Θk ) (5)
k =1
where T (x; Θk ) is the decision tree; Θk is parameters of the decision tree; K is the number
of trees.
The boosting tree algorithm in XGBoost uses a forward step-by-step algorithm as follows:
Firstly, the initial boosting tree f 0 (x) = 0 is determined.
Secondly, the model at the k-th step is denoted by Equation (6).
f k (x) = f k−1 (x)+ T (x; Θk ) , k = 1, 2, · · · , K (6)
where f k−1 (x) is the current model; Θk is the parameters of the decision tree, determined
by the minimization of empirical risk, denoted by Equation (7).
N
Θ̂k = argmin ∑ L(yi , f k−1 (xi ) + T (xi ; Θk )) (7)
Θk
i =1
Appl. Sci. 2022, 12, 5311 6 of 23
where Θ̂k is parameters of the k-th tree; L(yi , f k−1 (xi ) + T (xi ; Θk )) is the loss value between
the observed value y and the predicted value f k (x) for the i-th sample.
The loss function may be defined as L(y, f (x)) = (y − f (x))2 using the least square
error criterion, denoted by Equation (8).
l y, y0 = L(y, f k−1 (x) + T (x; Θk )) = [y − f k−1 (x) − T (x; Θk )]2 = [r − T (x; Θk )]2

(8)
where r = y − f k−1 (x) is the fitting residual of the current model; f (x) is the prediction
value obtained by the fitting model, y0 = f (x); f k−1 (x) is the fitting model obtained at the
k − 1-th step.
Thirdly, calculate the residual by Equation (9).
rki = yi − f k−1 (xi ), i = 1, 2, · · · , N (9)
Learn a regression tree based on the fitting residual rki , and obtain T (x; Θk ).
Fourthly, update f k (x) = f k−1 (x)+ T (x; Θk ) .
K
Fifthly, obtain boosting tree of regression problem f K (x) = ∑ T (x; Θk ).
k =1
The loss function for XGBoost is defined as L(φ), denoted by Equation (10).
n
∑ l ( yi , y 0 i ) + ∑ Ω ( f k ) = l y, y0 + ∑ Ω( f k ) =[r − T (x; Θk )]2 + ∑ Ω( f k )

L(φ) = (10)
i =1 k k k
n
where n is the number of training samples; ∑ l (yi , y0 i ) is the loss value caused using the
i =1
gradient boosting algorithm, calculated by Equation (8), assuming it is a differentiable
convex function; Ω( f k ) is the regularization term; f k is the weak learner function.
The complexity of the XGBoost model is defined by the regularization term Ω( f k ),
denoted by Equation (11).
1 λ J
Ω( f k ) = γTln + λkωk2 = γTln + ∑ ωtj2 (11)
2 2 j =1
where γ and λ are two artificially set coefficients; ω is the vector formed by the values of
all leaf nodes of the decision tree; Tln is the number of leaf nodes; ωtj is the weight of leaf j.
The ultimate goal of XGBoost is to minimize the objective function in Equation (10).
The objective function in Equation (10) is approximately solved by the Newton’s method,
and denoted by the second-order Taylor expansion at the point yi,t 0
−1 and Equation (12).

n
∂L(yi ,yi0 )

∂2 L(y ,y0 )

0
Lt (φ) ≈ ∑ l yi , yi,t f t ( xi ) + 21 ∂y0i2 i y0 =y0 f 2 ( xi )
−1 + ∂yi0 y0 =y0
i i,t−1 i i i,t−1 t −1
i =1
(12)
T
+γTln + λ
2 ∑ ωtj2
j =1
where f t ( xi ) is the current weak learner; xi is the i-th training sample.

To simplify the expression of Equation (12), the first-order negative gradient and the
second-order negative gradient are recorded as gti and hti denoted by Equation (13).
 0 2 0
 gti = ∂L(yi0,yi ) y0 =y0 , hti = ∂ L(y0i2,yi ) y0 =y0

∂yi i i,t−1 ∂yi i i,t−1
(13)
 Gt = ∑i∈ I gti , Ht = ∑i∈ I hti

j j
Appl. Sci. 2022, 12, 5311 7 of 23
Delete the constant term of Equation (12) that is independent of f t ( xi ). The form of the
final loss function can be simplified to Equation (14) when Gt = ∑i∈ Ij gti and Ht = ∑i∈ Ij hti .
T
1
Lt = ∑ Gt ω j + ( Ht + λ)ω 2j + γTln
2
(14)
i =1
4. NSGA-II –XGBoost Based Multi-Objective Back Analysis

A numerical model and the optimization method are two key components of the
back analysis. For deep soft rock tunnels, the numerical modeling is time consuming. In
this paper, XGBoost was used to represent numerical model and to map the relationship
between the mechanical parameters of surrounding rock and the displacement of each
measuring point, to reduce the computing time greatly. NSGA-II was adapted to search
multi-objectively the geomechanical parameters as an optimization method.
4.1. XGBoost-Based Relationship between Displacement and Geomechanical Parameters

We used XGBoost to map the nonlinear relationship between geomechanical parame-
ters such as elastic modulus, internal friction angle, cohesion, geo-stress coefficients, and
monitored displacements, where the surrogate model is built to replace the numerical
simulation model. The mathematical model of the multi-output XGBoost, XGBoost(X), is
defined as:
XGBoost(X): R N → R M

(15)
Y = XGBoost(X)
where X = (x1 , x2 , · · · , x N ), xi (i = 1, 2, · · · , N ) is a vector of geomechanical parameters
such as elastic modulus, internal friction angle, and cohesive force, etc.; N is the number of
mechanical parameters; Y = (y1 , y2 , · · · , y M ) is the M dimensional vector of the monitoring
data, such as displacement. In this case, the observable output is the displacement in the
back analysis, correspondingly, M represents the dimension of displacement data.
To obtain XGBoost(X), a training process based on a known dataset is needed. The
necessary training samples were created for this work by combining the FEM numerical
analysis and test design, which is used to obtain the displacements of the surrounding rock
initial support system according to a given set of mechanical parameters of the surrounding
rock. For this study, the sampling by the uniform test design method was adopted to
build samples for mechanical parameters of surrounding rock, and these geomechanical
parameters were defined as the input of XGBoost. The displacement was defined as the
output of XGBoost.
4.2. XGBoost–NSGA-II-Based Back Analysis

We propose the XGBoost–NSGA-II-based back analysis method for the deep soft
rock tunnel engineering. The monitored displacements at different monitoring points
are denoted herein as Ymonitor1 , Ymonitor2 , · · · , Ymonitord ; the predicted displacement by the
XGBoost-based surrogate model at different monitoring points are denoted as y1 , y2 , · · · , yd .
The non-dominated sorting genetic algorithm-II (NSGA-II) is a fast non-dominated
multi-objective optimization algorithm based on the Pareto optimal solution with an elite
retention strategy [31,32].
The objective function of multi-objective optimization is a set of functions containing
multiple members, and its optimization form is denoted by Equation (16).
θ∗ = argmin L(θ) = argmin{`1 (θ), `2 (θ), · · · , `d (θ)} (16)

θ θ
where θ is the geomechanical parameters to be inverted; “argmin” represents the variable

θ when the objective function L(θ) is minimized, i.e., θ∗ ; d is the number of monitoring
points for each section of tunnel.
Appl. Sci. 2022, 12, 5311 8 of 23
An objective function is set for each monitoring point, and each objective function is
represented by the mean square error (MSE). Finally, the multi-objective function is denoted
by Equation (17). 

 min`1 = (Ymonitor1 − y1 )2
min`2 = (Ymonitor2 − y2 )2


minL(θ) = (17)

 ···
min`d = (Ymonitord − yd )2


where `i (i = 1, 2, · · · , d) represents the MSE of the i-th monitoring point.

NSGA-II is used to solve the model parameters according to the following steps:
Step 1 Structural representation of solutions. The solution of the model optimization
problem may be expressed by a floating point vector or binary vector. The binary vectors
are used as chromosomes to represent the true value of the solution space, and the length of
the vector is determined by the accuracy of solving the problem. The chromosome number
of the initial population is pop_size.
Step 2 Generation of the initial population of feasible solutions. If there is no experience
and prior knowledge, the initial population can be randomly generated.
Step 3 The size of the initial population pop_size. The number of chromosomes pop_size
is 50~100. By random sampling for pop_size times, the initial sample of the population at
t = 0 may be obtained.
Step 4 Evaluation function. Supposing the chromosomes are V1 , V2 , · · · , Vpop_size , an
order-based evaluation function is defined by Equation (18).
eval(Vi ) = α(1 − α)i−1 (i = 1, 2, · · · , pop_size, α = 0.5) (18)
The pop_size chromosomes are decoded, and the adaptability of each chromosome is
evaluated. If the optimal solution remains unchanged for no less than three consecutive
generations, the solution is the optimal solution, therefore proceed to Step 9 to obtain the
optimal solution. Otherwise, continue to Step 5 to select chromosomes.
Step 5 Chromosome selection. The selection process of chromosomes is based on
spinning the wheel pop_size times. Based on the fitness of each chromosome, each rotation
selects a chromosome for the new population. The selection process is as follows:
First, for each chromosome Vi , calculate the cumulative probability qi by Equation (19).

 q0 = 0

i (19)
 q i = ∑ eval(Vj )(i = 1, 2, · · · , pop_size)
j =1


Second, generate a random number r from 0, q pop_size . If qi−1 < r ≤ qi , select the i-th
chromosome Vi (1 ≤ i ≤ pop_size).
Third, repeat the above two steps for the pop_size times to get pop_size duplicated
chromosomes.
Step 6 Conduct the hybridization operation to generate pop_size new chromosomes
with an appropriate crossover probability Pc = 0.4~0.9.
Step 7 Select Pm × pop_size chromosomes from pop_size chromosomes for mutation
operation with an appropriate probability Pm = 0.001~0.1.
Step 8 Let t = t + 1, and go to Step 4.
Step 9 Solving the model is terminated, and the optimal solution θ∗ is obtained.
4.3. Computation Procedure for the XGBoost–NSGA-II-Based Back Analysis

The algorithm for the XGBoost–NSGA-II-based back analysis uses the NSGA-II ap-
proach to calculate the optimal combination of geomechanical parameters in the deep
soft rock tunnel engineering. XGBoost is adopted to surrogate this relationship instead of
numerical models used in traditional back analysis methods. The back analysis procedure
is shown in Figure 3 and explained below:
Step 5 Based on the sets of samples built in Step 4, the XGBoost model can be obtained
by solving Equation (15).
Appl. Sci. 2022, 12, 5311 9 of 23
Step 6 Build the NSGA-II model based on Equations (16) and (17) to back-calculate
the geomechanical parameters.
Start
Step 1 Collect engineering information Build samples for XGBoost
Build test parameter sets using Build numerical model

Step 2
uniform test design method using ABAQUS
Step 3 Calculate the displacement of each sample by numerical model
Step 4 XGBoost algorithm Build training samples for XGBoost
Build XGBoost model to replace

Step 5 Build XGBoost model
the numerical model
NSGA-II method
Step 6 Back- calculate geo-mechanical parameters
End
Figure3.3. Flowchart
Figure Flowchart of
ofXGBoost–NSGA-II-based
XGBoost–NSGA-II-basedback
backanalysis.
analysis.
StepStudy
5. Case 1 Collect the engineering information, such as geological conditions, field test
and monitoring data, indoor
5.1. Problem Description and Itsexperimental
Parameters data, and in situ geostress, etc.
Step 2 According to the information collected in Step 1, determine the range of me-
In the section, the Qinzhou–Sanyangchuan soft rock tunnel located in Gansu Prov-
chanical parameters of surrounding rock that need to be recognized in the training set, and
ince, People’s Republic of China, is investigated as an illustrative example. Surrounding
build their sample sets by the uniform test design method and a numerical model using
rock of Section K3 + 038~100 of the tunnel is strongly weathered granite gneiss. The buried
finite element methods (FEM) or finite difference method, such as ABAQUS and ANSYS or
depth of Section K3 + 038~100 of the tunnel is 135~152 m. The aggregate score of the rock
FLAC3D.
massStep
rating (RMR) is the
3 Calculate 33, and its rock mass
displacement classification
values with RMR rock
of the surrounding is determined to be IV.
primary support
The tunnel has a height of 10.23 m and a span
system corresponding to each mechanical parameter sample. of 12.46 m, shown in Figure 4. Design pa-
rameters of the initial support of the tunnel: length of the rock bolt φ 25-5
Step 4 Build the sample set for XGBoost. The mechanical parameter sets of the sur- = 3.5 m; spacing
rounding rock
of the grille canframe
steel φ 25
be built. The= displacement
0.6 m; thicknessvalues of each sample
of sprayed concreteset, calculated
= 25 cm; steelby usingφ
mesh
FEM or other numerical methods, are used to define the XGBoost model. Sample
8 = 20 × 20 cm. Design parameters of the second lining of the tunnel: thickness of reinforce- sets are
composed of the geomechanical parameters of monitoring points and their
ment = 33~45 cm. The buried depth of the tunnel is 148 m. Horizontal geostress in the areacorresponding
displacement
of the tunnel values.
is 4.125~4.411 MPa. The design parameters of initial support and second
Step
lining are5shown
Based on the sets
in Table 1. of samples built in Step 4, the XGBoost model can be obtained
by solving Equation (15).
The initial support of Section K3 + 038~050 of the tunnel was completed on 9 October
Step 6 Build the NSGA-II model based on Equations (16) and (17) to back-calculate
2019. Since the completion of the initial support, the deformation of tunnel initial support
the geomechanical parameters.
structure had been increasing, leading to the collapse of Section K3 + 038~050 of this tunnel
onCase
5. 12 November
Study 2019, show in Figure 5. The tunnel was monitored by the infrared range-
finder during the
5.1. Problem Descriptionwholeand
period before the disaster. On the collapse of these sections of the
Its Parameters
tunnel, the maximum value of the vault settlement and the horizontal displacement of one
In the section, the Qinzhou–Sanyangchuan soft rock tunnel located in Gansu Province,
of side walls are 46.2 cm and 40.0 cm, respectively.
People’s Republic of China, is investigated as an illustrative example. Surrounding rock
of Section K3 + 038~100 of the tunnel is strongly weathered granite gneiss. The buried
depth of Section K3 + 038~100 of the tunnel is 135~152 m. The aggregate score of the
rock mass rating (RMR) is 33, and its rock mass classification with RMR is determined
to be IV. The tunnel has a height of 10.23 m and a span of 12.46 m, shown in Figure 4.
Appl. Sci. 2022, 12, 5311 10 of 23
Design parameters of the initial support of the tunnel: length of the rock bolt φ25-5 = 3.5 m;
spacing of the grille steel frame φ25 = 0.6 m; thickness of sprayed concrete = 25 cm; steel
mesh φ8 = 20 × 20 cm. Design parameters of the second lining of the tunnel: thickness of
reinforcement = 33~45 cm. The buried depth of the tunnel is 148 m. Horizontal geostress
Appl. Sci. 2022, 12, x FOR PEER REVIEW
in the area of the tunnel is 4.125~4.411 MPa. The design parameters of initial support11 of 25
and
second lining are shown in Table 1.
Figure 4.
Figure 4. Initial
Initial support and second
support and second lining
lining of
of tunnel
tunnel (unit:
(unit: cm).
cm).
Table 1. Tunnel support mechanical parameters.
Initial Support Concrete Lining Rock Bolt

EIS /GPa Poisson’s Ratio ν ECL /MPa Poisson’s Ratio νCL Ebolt /MPa νbolt
26.311 0.22 28,000 0.27 210,000 0.3
The initial support of Section K3 + 038~050 of the tunnel was completed on 9 October
2019. Since the completion of the initial support, the deformation of tunnel initial support
structure had been increasing, leading to the collapse of Section K3 + 038~050 of this
tunnel on 12 November 2019, show in Figure 5. The tunnel was monitored by the infrared
rangefinder during the whole period before the disaster. On the collapse of these sections
of the tunnel, the maximum value of the vault settlement and the horizontal displacement
of one of side walls are 46.2 cm and 40.0 cm, respectively.
Appl. Sci. 2022, 12, 5311 11 of 23
Figure 4. Initial support and second lining of tunnel (unit: cm).
Figure5.5.Collapse
Figure Collapseof
oftunnel
tunnelprimary
primarysupport
supportstructure.
structure.
5.2. Modelling Approach

ABAQUS is a powerful finite element software for engineering simulation owned
by Dassault SIMULIA. In this paper, we develop the NGNRM and implement it in
ABAQUS_2020. The NGNRM is used to simulate the evolution process of the non-stationary
parameter creep large deformation of the tunnel’s surrounding rock primary support sys-
tem, revealing its disaster mechanism.
5.3. Equivalence of Initial Support Parameters

In order to facilitate the numerical simulation, the design parameters of the tunnel’s
initial support structure need to be equivalent, as follows:
(1) The spacing of the grille steel frame is 0.6 m, and the initial support thickness is
0.25 m. Take the initial support of a rectangle with width b = 1.8 m and height h = 0.25 m as
the model. The stiffness of the model is calculated by the equivalent stiffness method.
The equivalent concrete elastic modulus of the shotcrete supported by the grillage
steel frame E10 can be calculated by Equation (20).
E10 I1 = E1 I1 + E2 ∑ I2 (20)
where E1 is the elastic modulus of the concrete; I1 is the moment of concrete inertia for the
calculation model; E2 is elastic modulus of the steel bar; I2 is the moment of inertia of the
virtual axis of each bar in the grille steel frame, denoted by Equation (21). In the study,
E1 = 23 GPa, E2 = 210 GPa.
πd4
I2 = + Ab12 (21)
64
where d is the circle diameter of the section of steel bar; A is the circle area of the bar section;
b1 = the half of the thickness of the initial support structure minus the thickness of the
protective layer and a steel bar radius.
In this illustrative example, E10 = 26.311 GPa, obtained by Equation (20).
Appl. Sci. 2022, 12, 5311 12 of 23
(2) The reinforcement effect of grouting a rock bolt on surrounding rock can be equivalent
to improving the cohesion of surrounding rock, which may be obtained by Equation (22).
τs Ss
c1 = c0 + η (22)
ab
where c0 , c1 are, respectively, the cohesion of surrounding rock before and after reinforce-
σbs
ment; τs = √ , σ is the yield strength of the rock bolt; Ss is the cross-sectional area of the
3 bs
rock bolt; a and b are, respectively, longitudinal spacing and transverse spacing; η is an
experience coefficient. In this paper, η = 0.963; σs = 335 MPa; a = 0.6 m; b = 1.2 m.
In this illustrative example, c1 = c0 + 127.05 kPa, obtained by Equation (22).
5.4.
5.4.Establishment
Establishmentofofthe
theNumerical
NumericalModel
Model
AAtunnel
tunnelmodel
modelisisestablished
establishedas asshown
shownininFigure
Figure6.6.The
Theground
groundstress
stresswas
wasobtained
obtained
from field investigation. The size of the model is 80 m ×
from the field investigation. The size of the model is 80 m × 75 m (length × height).The
the 75 m (length × height). The
displacement
displacementconstraints
constraintsareareapplied
appliedon onthe
thebottom
bottomandandsides
sidesofofthe
themodel,
model,horizontal
horizontal
ground
groundstress
stressisisapplied
appliedononboth
bothsides of the
sides model,
of the model,andand
self-weight stress
self-weight is applied
stress to theto
is applied
entire model. Considering the symmetry of the model, we arranged
the entire model. Considering the symmetry of the model, we arranged six monitoringsix monitoring points
D1~D6
points on the right
D1~D6 side
on the of the
right sideinitial
of thesupport structure
initial support of the tunnel
structure of the(Figure
tunnel 6).
(Figure 6).
Figure6.6.Tunnel
Figure Tunnelmodel.
model.
5.5.Determination
5.5. DeterminationofofNumerical
NumericalSimulation
SimulationParameters
Parameters
Weuse
We use the
the proposed
proposed XGBoost–NSGA-II-based
XGBoost–NSGA-II-basedback backanalysis
analysismethod
method to to
inversely an-
inversely
alyze the
analyze mechanical
the mechanical parameters
parameters of surrounding
of surroundingrockrock
according to the
according todisplacement mon-
the displacement
itoring datadata
monitoring of D3, D4 D4
of D3, andand
D6 D6of tunnel Section
of tunnel K3 K3
Section + 060, andand
+ 060, obtain thethe
obtain mechanical pa-
mechanical
rameters ininTable
parameters Table2.2.InInTable ρ
Table2,2, ρ isisthe
thedensity
densityofofthe
therock mass;E0Eis0 Young’s
rockmass; is Young’s modu-
modulus
oflus the rock mass; ν is Poisson’s ratio of the rock mass; E1 is the elastic modulus of
theofrock mass; ν is Poisson’s ratio of the rock mass; E 1 is the elastic modulus of a Hooke
body in a Kelvin body; η1 is the viscosity coefficient of a Newtonian body in a Kelvin body;
ca0 is
Hooke body in of
the cohesion a Kelvin
surrounding η1 is ϕthe
body; rock; viscosity
is the coefficient
internal of a Newtonian
friction angle body in a
of the surrounding
rock;
Kelvin λ isbody; c0 is theof
the coefficient confinement
cohesion pressure; σrock;
of surrounding f and σ ϕs are, respectively,
is the the long-term
internal friction angle of
strength and instantaneous strength of a Bingham body; E10 is the equivalent concrete
the surrounding rock; λ is the coefficient of confinement pressure; σ f and σ s are, re-
elastic modulus of the shotcrete supported by the grillage steel frame; A is a dimensionless
spectively,Tthe
parameter; long-term
is the time forstrength and instantaneous
the surrounding strength of a Bingham body; E1′ is
rock to yield.
the equivalent concrete elastic modulus of the shotcrete supported by the grillage steel
frame; A is a dimensionless parameter; T is the time for the surrounding rock to yield.
Table 2. Numerical simulation parameters of the model.
ρ /kg/m3 E 0 /GPa ν E1 /GPa η1 /GPa.h c0 /MPa ϕ /°

Appl. Sci. 2022, 12, 5311 13 of 23
Table 2. Numerical simulation parameters of the model.
ρ/kg/m3 E0 /GPa ν E1 /GPa η1 /GPa.h c0 /MPa ϕ/◦
Appl. Sci. 2022, 12, x FOR PEER REVIEW

2370 2.3 0.33 8.01 12,050 0.21 33 15 of 25
0
σf /MPa A T λ σs /MPa E1 /GPa
8.0 4.2 1650 1.188 35.62 26.311
Then, the obtained mechanical parameters are substituted into the model for calcu-
Then,
lation. the obtained
In addition, themechanical parameters
relative errors (REs) are substituted
between into the model
the calculated for calcula-
values and the meas-
tion. In addition, the relative errors (REs) between the calculated values and the measured
ured values of displacements are 0.136~4.654%, 0.313~4.737%, and 0.25~3.959% for the
values of displacements are 0.136~4.654%, 0.313~4.737%, and 0.25~3.959% for the monitor-
monitoring points D3, D4, and D6, respectively. The comparison between the calculated
ing points D3, D4, and D6, respectively. The comparison between the calculated values and
values and the
the measured measured
values values
of D3, D4, and of
D6D3, D4, and D6
displacements is displacements
shown in Figureis 7. shown in Figure 7
Combining
Combining with Figure 7 and the REs of the displacements of each measuring
with Figure 7 and the REs of the displacements of each measuring point, we can note that point, we
can note thatofthe
the accuracy the accuracy
surroundingof the
rocksurrounding rock mechanical
mechanical parameters obtainedparameters obtained by
by the proposed
the
backproposed back analysis
analysis method can meetmethod can
the actual meet the actual
engineering engineering requirements.
requirements.
D3_Calculated value
5 D3_Measured value
D4_Calculated value
D4_Measured value
D6_Calculated value
4 D6_Measured value
Displacement /cm
0
5 10 15 20 25 30 35
Time /d
Figure 7.7.Comparison
Figure Comparison of of
measured andand
measured calculated values.
calculated values.
5.6. Parametric Study

5.6. Parametric Study
In the stability analysis of the tunnel using the NGNRM, thirteen parameters of
In the stability
non-stationary analysis of the
viscoelastic–plastic tunnel
model using the
are shown NGNRM,
in Table 2. thirteen parameters of non-
stationary
In this viscoelastic–plastic
paper, the orthogonalmodel are shown
experimental design in method
Table 2.(OEDM) is used to study
the sensitivity of nine parameters of NGNRM (E0 , ν, c0 , ϕ, Emethod
In this paper, the orthogonal experimental design 1 , η1 , σf , (OEDM) is used
λ, E10 ) on the vaultto study
sensitivity of nine parameters of NGNRM ( E0 ,ν ,tunnel.
settlement and horizontal displacement of the side wall of the
the c , ϕ The σ f , λ, E1′ ) on the
, η1 ,parameters
, E1nine
are selected as test factors. The values of these nine parameters in0Table 2 and the values of
vault settlement
their upper and20%
and lower horizontal
are takendisplacement of the
as the level of the test side wall
factors. In of the tunnel.
addition, The nine pa-
four empty
rameters
columns are are selected
added asthe
to test test factors. The
significance values
of the of these
test factors. nine parameters
Therefore, in Table 2 and
L27 (313 ) is selected
the values of their upper and lower 20% are taken as the level of the test factors. In addi-
as the orthogonal design test table.
tion,For
four vault settlement
empty columns andare
horizontal
addeddisplacement of the side wall,
to test the significance of the the test
variances of the
factors. Therefore
nine parameters
13 are calculated, respectively, and the statistics F is constructed.
L27 (3 ) is selected as the orthogonal design test table.
Variance of column j in the orthogonal table can be expressed by Equation (23).
For vault settlement and horizontal displacement of the side wall, the variances of
SS j
the nine parameters are calculated, respectively,
MS j =
and the statistics F is constructed.
(23)
Variance of column j in the orthogonald f jtable can be expressed by Equation (23).
where SS j is the sum of squares of deviation; d f j is theSS

degrees of freedom (DOF) of column
j
j in the orthogonal table. MS j = (23)
df j
where SS j is the sum of squares of deviation; df j is the degrees of freedom (DOF) of
column j in the orthogonal table.
2, 12, x FOR PEER REVIEW 16 of 25
Appl. Sci. 2022, 12, 5311 14 of 23
As thus, F of factor j can be calculated by Equation (25).

Variance of the error term in the orthogonal table may be expressed by Equation (24).
MS
Fj == MSerrorj = SSerror
(25)(24)
MSerror d f error
where SSerror is the sum of squares of deviation of error term; d f error is the sum of DOF of
In order to improve
error theorthogonal
term in the sensitivity of the F-test, all factors of MS j ≤ 2 MS error to-
table.
Δ
As thus,
gether with four empty F of factor
columns arej can be calculated
included in theby Equation
error expressed by e ; Fα is
term,(25).
the critical value of F ; e is the error term considering MSonly
j two empty columns; SOV is
Fj = (25)
source of variation; Sig. denotes the significance of the MSinfluencing
error factor.
Finally, the significance test of the
In order to improve the effect of these
sensitivity of thetest factors
F-test, is quantitatively
all factors of MS j ≤ 2MSassessed
error together
by their variances. ∆
with four empty columns are included in the error term, expressed by e ; Fα is the critical
Based on value
the significance
of F; e is the(Sig.) of theconsidering
error term influencing onlyfactors on the
two empty vault settlement
columns; SOV is source of of
the tunnel, we obtained
variation; the sensitivity
Sig. denotes ranking
the significance of the influencing
of the influencing factor. factors when
F j > F0.01（8,12） as follows: ν > ϕ > E0 > λ . Based on the significance of the influencing
Finally, the significance test of the effect of these test factors is quantitatively assessed
by their variances.
factors on the sidewall
Basedhorizontal displacement
on the significance of influencing
(Sig.) of the the tunnel,factors
we obtained thesettlement
on the vault sensitivity of the
ranking of the influencing factors when F j > F0.01（8,14） as follows: ν > ϕ > λ > E
tunnel, we obtained the sensitivity ranking of the influencing factors when F > F
j 0 . 0.01(8,12)
as follows: ν > ϕ > E0 > λ. Based on the significance of the influencing factors on the
ν , ϕ ,horizontal
Therefore,sidewall E0 , λ may be determined
displacement as thewe
of the tunnel, main parameters
obtained of the ranking
the sensitivity NGNRM of the
influencing factors
by the above significance when
analysis j > influencing
ofFthe F0.01(8,14) as follows: ν > ϕ > λ > E0 .
factors.
Therefore, ν, ϕ, E0 , λ may be determined as the main parameters of the NGNRM by
the above significance analysis of the influencing factors.
5.7. Results and Discussion
5.7. Results
(1) Influence and Discussion
of surrounding rock Poisson’s ratio
From Figure 8, (1)we note that
Influence with the increase
of surrounding in the
rock Poisson’s reduction rate of surrounding
ratio
From Figure 8, we note that with
rock Poisson’s ratio, tunnel perimeter displacements increasethe increase in with
the reduction
time, butrate of surrounding
they are only
rock Poisson’s ratio, tunnel perimeter displacements increase with time, but they are only
slightly changed. In general, the curvature of the curve of the vault settlement versus time
slightly changed. In general, the curvature of the curve of the vault settlement versus time
(Figure 8b) is larger
(Figurethan
8b) isthat ofthan
larger the displacement of the side
that of the displacement wall
of the sideversus time time
wall versus (Figure 8a).8a).
(Figure
Poisson's ratio reduction rate 0

Poisson's ratio reduction rate 10%
0.20
Displacement /m
0.15 Poisson's ratio reduction rate 60%

0.10
0.05
0.00
0 200 400 600 800 1000 1200 1400 1600 1800
Time /h
(a)
Figure 8. Cont.
0.25
, 12, x FOR PEER REVIEW 17 of 25
Appl. Sci. 2022, 12, 5311 15 of 23
Poisson's ratio reduction rate 0
0.20
0.25 Poisson's ratio reduction rate 30%
/m
Poisson's ratio reduction
raterate
0 50%
Displacement
0.15 Poisson's ratio reduction
Poisson's
Poisson's ratio
ratio reduction
reduction raterate
10%60%
0.20 Poisson's ratio reduction
Poisson's ratio reduction raterate
20%70%
Poisson's
Poisson's ratio
ratio reduction
reduction raterate
30%80%
Poisson's ratio reduction
raterate 90%
/m
0.10 Poisson's ratio reduction 40%
Displacement Poisson's ratio reduction rate 50%
0.15
0.05
0.10
0.00
0.05 0 200 400 600 800 1000 1200 1400 1600 1800
Time /h
0.00
(b)
0 200 400 600 800 1000 1200 1400 1600 1800
Figure 8. Variation of perimeter displacement with Poisson’s ratio. (a) D3 monitoring point. (b) D6
Time /h
monitoring point.
(b)
(2)8.Influence
Figure Variation ofofperimeter
Figure 8.surrounding rock displacement
Variationdisplacement
of perimeter internal friction
with Poisson’s angle
ratio. (a)ratio.
with Poisson’s D3 monitoring point.point.
(a) D3 monitoring (b) D6(b) D6
monitoring point.
monitoring point.
From Figure 9, we note that with the increase in the reduction rate of the surrounding
rock internal friction angle, tunnel
(2) Influence perimeter
of surrounding rockdisplacements increase with time, but there
internal friction angle
(2) Influence
is a big change. With of surrounding rock internal friction angle
Fromthe increase
Figure 9, we notein thethat reduction rate in
with the increase ofthe
thereduction
surrounding rock
rate of the internal
surrounding
From Figurerock9,internal
we note that with the increase
friction angle, both the curvature
friction of the
angle, curve
tunnel of in
thethe
perimeter reduction rate
displacements
displacements ofofthe
increasethewith
surrounding
side wallbut
time, versus
there
rock internal friction
is a big angle,
change. tunnel
With theperimeter
increase displacements
in the reduction increase
rate of the with time, but
surrounding rockthere
internal
time (Figure 9a) and that of the vault settlement versus time (Figure 9b) decrease.
is a big change.friction
With angle,
the increase
both the in the reduction
curvature of the curve rateofof
thethe surrounding
displacements rock
of the sideinternal
wall versus
(3) Influence of the
time (Figure
surrounding
9a) and that
rock
of the
elastic modulus
vaultofsettlement versus time of(Figure 9b) decrease.
friction angle, both the curvature of the curve the displacements the side wall versus
time (Figure 9a) and that of the vault settlement versus time (Figure 9b) decrease.
Internal friction angle reduction rate 0
(3) Influence of the surrounding rock elastic modulus
Internal friction angle reduction rate 10%
0.20
0.20 Internal friction angle reduction rate 50%
/m

Displacement

Internalfriction
Internal friction angle
angle reduction
reduction raterate
50% 80%
/m

Displacement

0.10
0.05
0.05
0.00
0 200 400 600 800 1000 1200 1400 1600 1800
0.00 Time /h
0 200 400 600 800
(a) 1000 1200 1400 1600 1800
Time /h
Figure 9. Cont.
(a)
12, x FOR PEER REVIEW 18 of 25
Appl. Sci. 2022, 12, 5311 16 of 23

Displacement /m
0.10
0.05
0.00
0 200 400 600 800 1000 1200 1400 1600 1800
Time /h
(b)
Figure 9. Variation of 9.perimeter
Figure displacement
Variation of with internal
perimeter displacement friction
with internal angle.
friction (a) D3
angle. monitoring
(a) D3 monitoring point.
point.
(b) D6 monitoring
(b) D6 monitoring point. point.
(3) Influence of the surrounding rock elastic modulus

From Figure 10, From we note10,
Figure thatwe with thewith
note that increase in the
the increase in reduction
the reductionraterate of thesurrounding
of the surround-
ing rock elasticrock
modulus, tunneltunnel
elastic modulus, perimeter
perimeterdisplacements
displacements increase with
increase with time,
time, but but
therethere
is a bigis
change. With the increase in the reduction rate of the surrounding
a big change. With the increase in the reduction rate of the surrounding rock elastic mod- rock elastic modulus,
ulus, both theboth the curvature
curvature of theof the curveof
curve of the
thedisplacements
displacements of theof
side wall
the versus
side walltime (Figuretime
versus 10a)
and that of the vault settlement versus time (Figure 10b) increase.
(Figure 10a) and that of the vault
(4) Influence of thesettlement
lateral pressureversus time (Figure 10b) increase.
coefficient
(4) Influence ofFromtheFigure
lateral 11,pressure
we note that coefficient
with the increase in the reduction rate of the lateral
From Figure 11, coefficient,
pressure we note that tunnelwith the increase
perimeter in the
displacements reduction
increase with time.rateHowever,
of the lateral
for the
reduction rate of 10% among the lateral pressure coefficient,
pressure coefficient, tunnel perimeter displacements increase with time. However, for the curvature of the curve
the
of the perimeter displacements and lateral pressure coefficient increases with time and
reduction rate of 10% among the lateral pressure coefficient, the curvature of the curve of
its variation is the largest. In general, for the reduction rate of 10~30% among the lateral
the perimeter pressure
displacements and
coefficient, the lateral
curvature pressure coefficient
of the curve increases
of the perimeter with time
displacement andand its
lateral
variation is the largest.
pressure In general,
coefficient for with
increases the time.
reduction rate of of10%~30%
The curvature among the lateral
the curve corresponding to the
other reduction
pressure coefficient, rates are of
the curvature onlytheslightly
curvechanged.
of the perimeter displacement and lateral
(5) Influence of the elastic modulus of primary support
pressure coefficient increases with time. The curvature of the curve corresponding to the
From Figure 12, we note that with the increase in the reduction rate of the elastic
other reduction rates are
modulus only slightly
of primary support, changed.
tunnel perimeter displacements increase with time. In
(5) Influence of the
general, the elastic
curvature modulus
variationof of primary
the curve of support
the displacements of the side wall versus
From Figure 12, we note that with the increase in the
time is only slightly changed (Figure 12a). However, withreduction
the increase rate
in theof the elastic
reduction rate
of the elastic
modulus of primary modulus
support, of primary
tunnel support,displacements
perimeter the curvature of the curve ofwith
increase the vault
time. settlement
In gen-
versus time (Figure 12b) has a relatively large change. That is to say, the initial support
eral, the curvature variation of the curve of the displacements of the side wall
stiffness has a greater influence on the vault settlement than the side wall displacement.
versus time
is only slightly changed (Figure 12a).
(6) Non-stationary However,
parameter withdeformation
creep large the increase in the reduction rate of
mechanism
the elastic modulus In of
theprimary
following,support,
we use the the curvature
proposed of the
method curve of the
to investigate the vault settlement
large deformation
mechanism of Section K3 + 038~040 of the tunnel.
versus time (Figure 12b) has a relatively large change. That is to say, the initial support
stiffness has a greater influence on the vault settlement than the side wall displacement.
ci. 2022, 12, x FOR PEER REVIEW 19 of 25
Sci. 2022, 12, x FOR

Appl. Sci.PEER
2022,REVIEW
12, 5311 19 of 25 17 of 23
1.8
Elastic modulus reduction rate 0
Elastic modulus reduction rate 10%
1.6
1.8 Elastic modulus reduction
Elastic modulus reduction raterate
0 20%
Elastic
Elastic modulus
modulus reduction
reduction raterate
10%30%
1.4
1.6 Elastic
Elastic modulus
modulus reduction
reduction raterate
20%40%
Elastic
Elastic modulus
modulus reduction
reduction raterate
30%50%
/m /m
1.4
1.2 Elastic
Elastic modulus
modulus reduction
reduction raterate
40%60%
Elastic
Elastic modulus
modulus reduction
reduction raterate
50%70%
Displacement
1.2
1.0 Elastic
Elastic modulus
modulus reduction
reduction raterate
60%80%
Elastic
Elastic modulus
modulus reduction
reduction raterate
70%90%
Displacement 1.0 Elastic modulus reduction rate 80%

0.8 Elastic modulus reduction rate 90%
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0.0
0.0 0 200 400 600 800 1000 1200 1400 1600 1800
0 200 400 600 800 1000/h
Time 1200 1400 1600 1800
Time
(a) /h
2.2
(a)
2.2
2.0 Elastic
Elastic modulus
modulus reduction
reduction raterate
0 10%
2.0 Elastic
Elastic modulus
modulus reduction
reduction raterate
10%20%
1.8 Elastic
Elastic modulus
modulus reduction
reduction raterate
20%30%
1.8 Elastic
Elasticmodulus
modulusreduction raterate
reduction 30%40%
1.6 Elastic
Elasticmodulus
reduction 40%50%
1.6 Elastic modulus reduction raterate
50%60%
/m/m
Elastic modulus reduction

1.4 Elastic
Elasticmodulus
reduction 60%70%
Displacement

70%80%
Displacement

80%90%
1.2 Elastic modulus reduction
1.0
1.0
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0.0
0.0
00 200
200 400
400 600600 800800 1000
1000 1200120014001400160016001800 1800
Time
Time /h /h
(b)(b)
Figure 10.
Figure 10. Variation
Variation of
Figureofperimeter
perimeter
10. displacement
displacement
Variation with
of perimeterwithelastic modulus.
elastic
displacement (a) (a)
modulus. D3 D3
with elastic monitoring point.
monitoring
modulus. (a) (b) (b) point. (b) D6
D3point.
monitoring
D6 monitoring point.
monitoring point.
0.22
0.22 Lateral pressure coefficient reduction rate 0
Lateral pressure coefficient reduction rate 0
0.20 Lateral pressure coefficient reduction rate 10%
Lateral pressure coefficient reduction rate 20%
0.18
0.18 Lateral
Lateral pressure
pressure coefficient
coefficient reduction
reduction raterate
40%30%
0.16 Lateral
Lateral pressure
reduction raterate
50%40%
0.16 Lateral pressure coefficient reduction
raterate
60%50%
/m/m
Lateral pressure coefficient reduction

0.14 Lateral
Lateral pressure
reduction raterate
70%60%
Displacement

Displacement
0.12
0.12
0.10
0.10
0.08
0.08
0.06
0.06
0.04
0.04
0.02
0.02
0.00
0 200 400 600 800 1000 1200 1400 1600 1800
0.00
0 200 400 600 Time
800 /h 1000 1200 1400 1600 1800
(a) Time /h
(a)
Figure 11. Cont.
i. 2022, 12, Appl.
x FORSci.
PEER REVIEW
2022, 12, 5311 20 of 25 18 of 23
Sci. 2022, 12, x FOR PEER REVIEW 20 of 25
0.25
Lateral pressure coefficient reduction rate 0
0.25 Lateral pressure coefficient
coefficientreduction
reductionrate
rate0 10%
Lateral pressure
Lateral pressure coefficient
Lateral pressure coefficientreduction
reductionrate
rate10%
20%
reductionrate
rate20%
30%
0.20
reductionrate
rate30%
40%
0.20 Lateral pressure coefficient
reductionrate
rate40%
50%
Lateral pressure
reductionrate
rate50%
60%
/m/m
Lateral pressure
reductionrate
rate60%
70%
Displacement
0.15
Displacement 0.15
0.10
0.10
0.05
0.05
0.00
0.00
0 200 400 600 800 1000 1200 1400 1600
0 200 400 600 800 1000 1200 1400 1600
Time/h/h
Time
(b)
(b)
Figure
Figure11.
11.Variation
Variationof
ofperimeter
Figure perimeter
11. displacement
displacement
Variation with
with
of perimeter lateralpressure
lateral pressure
displacement coefficient.
withcoefficient. (a)(a)
lateral pressure D3D3 monitoring
monitoring
coefficient. (a) D3 monitoring
point.
point.(b)
(b)D6
D6monitoring
point. (b)point.
monitoring point.
0.25
0.25
Elastic modulus reductionrate
rate 00
Elastic modulus reductionrate
rate 10%
10%
Elastic modulus
reductionrate
rate 20%
20%
0.20
0.20 Elastic modulus reduction rate 40%
Elastic modulus
reductionrate
rate 50%
50%
/m
Displacement/m
0.15
Displacement
0.15
0.10
0.10
0.05
0.05
0.00
0.00 0 200 400 600 800 1000 1200 1400 1600 1800
0 200 400 600 800 1000 1200 1400 1600 1800
Time /h
Time /h
(a)
(a)
0.25 Elastic modulus
modulus reduction
reductionrate
rate 10%
0
modulus reduction
reductionrate
rate 20%
10%
Elastic modulus
modulus reduction
reductionrate
rate 30%
20%
Elastic modulus
modulus reduction
reductionrate
rate 40%
30%
reductionrate
rate 50%
40%
/m /m

reductionrate
rate 60%
50%
Displacement
0.15
Displacement
0.15
0.10
0.10
0.05
0.05
0.00
0 200 400 600 800 1000 1200 1400 1600
0.00
0 200 400 600 Time
800 /h 1000 1200 1400 1600
(b)Time /h
Figure of12.
perimeter displacement
Variation (b)
with
of perimeter elastic modulus
displacement withofelastic
primary support.
modulus of (a) D3 support. (a) D3
primary
monitoring point. (b) D6 monitoring point.
monitoring
Figure 12. Variation point. displacement
of perimeter (b) D6 monitoring
withpoint.
elastic modulus of primary support. (a) D3
monitoring point. (b) D6 monitoring point.
the vault have the largest among them.
From Figures 14–16, and Table 3, we note that the displacements of the monitoring
points from the vault to the side wall basically present a sequence from large to small,
which is consistent with the collapse of the tunnel primary support structure shown in
Figure 5.
Appl. Sci. 2022, 12, 5311 From Figure 16, we also note that the large deformation of the initial support 19 ofofthe
23
tunnel has undergone the following five stages of evolution:

(1) At the first stage, the displacement rate is very small, for example, its average
displacement
We use the rate of monitoring
proposed point NO.1
back analysis methodis 0.226 cm/d from
to inversely 0 h tothe
analyze 200mechanical
h of rheologypa-
time.
rameters of surrounding rock according to the displacement monitoring data of monitoring
points(2)D4Theand D6 of Section
average K3 + 040
displacement ratefrom 1 to
at the 24 h ofstage
second rheology time.
is much Then,
larger thethat
than obtained
at the
mechanical
first stage. Forparameters
example,are substituted
the into the model
average displacement forof
rate calculation.
monitoring Inpoint
addition,
NO.1theatREs
the
between
second stagethe calculated values
is 0.915 cm/d fromand
200the measured
h to values of time.
400 h of rheology displacements are 0.116~4.661
and 0.19~4.015 for thedisplacement
(3) The average monitoring points D4the
rate at andthird
D6, respectively.
stage is much The comparison
larger than thatbetween
at the
the calculated
second stage. Forvalues and the
example, themeasured values of D4rate
average displacement andofD6 displacements
monitoring point is shown
NO.1 in
at the
Figure 13. Combining
third stage is 1.390 cm/dwith Figure
from 400 13 and
h to the hREs
542.5 of the displacements
of rheology of each measuring
time. The displacement of the
point,
primary wesupport
can note that theof
structure accuracy of the
the tunnel surrounding
reaches the reservedrockdeformation
mechanical design
parameters
valueob-
20
tained by the proposed back analysis method can
cm when the surrounding rock force rheology time is 542.5 h.meet the actual engineering requirements.
25
D4_Calculated value
D4_Measured value
D6_Calculated value
20 D6_Measured value
Displacement /cm
15
10
0
0 5 10 15 20 25
Time /d
Variation of
Figure 13. Variation of perimeter
perimeter displacement
displacement with
with elastic
elastic modulus
modulus of
of primary
primarysupport.
support.
From Figures 14–16, we note that the displacements of monitoring points No. 1
(Node 8), No. 2 (Node 69), No. 3 (Node 13), No. 4 (Node 70), No. 5 (Node 17), and No. 6
(Node 20) on tunnel primary support structure present an approximate upward concave
parabolic change with rheological time. The displacements of the monitoring22point
Appl. Sci. 2022, 12, x FOR PEER REVIEW of 25 No. 1
at the vault have the largest among them.
Figure 14.
Figure 14. Contour
Contourmap
mapofoftunnel
tunneldisplacement at rheological
displacement timetime
at rheological 721 h.
721 h.
Appl. Sci. 2022, 12, 5311 20 of 23
Figure 14. Contour map of tunnel displacement at rheological time 721 h.

Figure15.
Figure 15.Contour
Contourmap
mapof
oftunnel
tunneldisplacement
displacementatatrheological
rheologicaltime
time837
837h.h.
(4) The average

Node 8
displacement rate at the fourth stage is very much larger than that at
the third
stage.
NodeFor
0.6 13 example, the average displacement rate of monitoring point NO.1 at
Node 17 Fifth stage
the fourth stage
Node 20 2.336 cm/d from 542.5 h to 800 h of rheology time.
is
0.5 Node 69
(5) The average displacement rate at the fifth stage is very much larger than that at
Node 70
the fourth stage. For example, the average displacement rate of monitoring point NO.1 at
Displacement /m
0.4
the fifth stage is 3.695 cm/d from 800 h to 850 h of rheology time.
Fourth stage
0.3
First stage Second stage Third stage
Reserved deformation
0.2
design value
0.1
0.0
0 200 400 600 800 1000
Time /h
Figure 16.
Variation curve
curve of
ofthe
thedisplacements
displacementsof
oftunnel
tunnelmeasuring
measuringpoints
pointswith
withtime.
time.
From
From Figures 14–16,
Figure 17, andnote
we also Table 3,the
that wedisplacement
note that the velocities
displacements
of the of the monitoring
monitoring points
points from
increase the vault to linearly
approximately the sidefrom
wall the
basically present
sixth day a sequence
to the twentiethfromday large to small,
after rheology
which
begins,isand
consistent with the rapidly,
then it increases collapseespecially
of the tunnel
afterprimary
28 days support structure
of rheology. That isshown
to say,init
Figure 5.
is most scientific and reasonable to reinforce and dispose of the large deformation disaster
of the tunnel within 20 days after the initial support.
Table 3. Displacement and velocity of monitoring points around the tunnel.
Table 3. Displacement and velocity of monitoring points around the tunnel.
Monitoring Point Displacement/m Displacement
Number Displacement/m Time 30 d Time 34.833 d Average Rate/cm/d
Monitoring Point Number Displacement Average Rate/cm/d
Time
NO.130 d Time
0.35334.833 d 0.505 3.145
NO.1 0.353
NO.2 0.505
0.346 0.495 3.145 3.083
NO.2 NO.3
0.346 0.325
0.495 0.468 3.083 2.959
NO.3 NO.4
0.325 0.308
0.468 0.447 2.959 2.876
NO.5 0.301 0.440 2.876
NO.4 0.308
NO.6 0.447
0.306 0.441 2.876 2.793
NO.5 0.301 0.440 2.876
NO.6 0.306 0.441 2.793
NO.1
4.0
NO.2
NO.3
0.5 Node 69
Node 70
Displacement /m
0.4
Fourth stage
0.3
First stage Second stage Third stage
Appl. Sci. 2022, 12, 5311 21 of 23
Reserved deformation
0.2
design value
0.1
From Figure 16, we also note that the large deformation of the initial support of the
tunnel has undergone the following five stages of evolution:
0.0
(1) At the first stage, the displacement rate is very small, for example, its average
0 200 400 600 800 1000
displacement rate of monitoring point NO.1 is 0.226 cm/d from 0 h to 200 h of rheology time.
Time /h
(2) The average displacement rate at the second stage is much larger than that at the
first stage. curve ofthe
For example, theaverage
displacements of tunnel measuring
displacement points withpoint
rate of monitoring time.NO.1 at the
second stage is 0.915 cm/d from 200 h to 400 h of rheology time.
From
(3) The Figure 17, we
average also note that
displacement ratetheatdisplacement
the third stage velocities
is much oflarger
the monitoring
than thatpoints
at the
increase approximately linearly from the sixth day to the
second stage. For example, the average displacement rate of monitoring point twentieth day after
NO.1rheology
at the
begins,
third stageandisthen it increases
1.390 cm/d from rapidly,
400 h toespecially
542.5 h of after 28 daystime.
rheology of rheology. That is to say,
The displacement it
of the
is most scientific
primary supportand reasonable
structure of thetotunnel
reinforce and dispose
reaches of the large
the reserved deformation
deformation designdisaster
value
of
20 the
cm tunnel
when the within 20 days after
surrounding rockthe
forceinitial support.
rheology time is 542.5 h.
(4) The average displacement rate at the fourth stage is very much larger than that at
Table 3. Displacement
the third and velocity
stage. For example, theofaverage
monitoring points around
displacement ratetheoftunnel.
monitoring point NO.1 at
the fourth stage is 2.336 cm/d
Displacement/m from 542.5 h to 800 h of rheology time.
Monitoring Point Number (5) The average displacement rate at the fifth stage Displacement
is very muchAverage Rate/cm/d
larger than that at the
Time 30 d Time 34.833 d
fourth stage. For example, the average displacement rate of monitoring point NO.1 at the
NO.1 0.353 0.505 3.145
fifth stage is 3.695 cm/d from 800 h to 850 h of rheology time.
NO.2 From Figure 17, we also note that the displacement velocities of3.083
0.346 0.495 the monitoring points
NO.3 0.325
increase approximately 0.468the sixth day to the twentieth
linearly from 2.959day after rheology
NO.4 begins, and 0.308
then it increases rapidly, 0.447
especially after 28 days of rheology. 2.876 That is to say, it is
NO.5 0.301and reasonable to reinforce
most scientific 0.440 and dispose of the large deformation 2.876 disaster of
NO.6 the tunnel0.306within 20 days after the0.441 initial support. 2.793
NO.1
4.0
NO.2
NO.3
3.5
NO.4
NO.5
Displacement velocity /cm/d
3.0 NO.6
2.5
2.0
1.5
1.0
0.5
0.0
-0.5
0 5 10 15 20 25 30 35 40
Time /d
Figure 17. Curve of displacement velocities and time of measuring points.
6. Conclusions
1. The NGNRM with non-stationary parameter creep developed in this paper can
effectively evaluate the large deformation mechanism of soft rock tunnels.
2. The proposed XGBoost–NSGA-II-based back analysis can simultaneously obtain
multiple mechanical parameters of the surrounding rock using machine learning algorithms
to construct surrogate models of numerical simulation, and optimize parameters of the
surrogate model.
3. Through the analysis of the significance of the influencing factors, it is shown that ν,
ϕ, E0 , and λ are four main parameters of the NGNRM affecting the large deformation of
soft rock tunnels.
4. The large deformation of the initial support of the tunnel generally undergoes the
five stages of evolution from a very small to a very high displacement rate.
Appl. Sci. 2022, 12, 5311 22 of 23
5. For soft rock tunnels, the displacement velocities of monitoring points increase
approximately linearly within a certain period of time after rheology begins, and then it
increases rapidly, especially with large acceleration, in a very short period of time before
the collapse of the surrounding rock primary support system.
6. It is most scientific and reasonable to reinforce and dispose of the large deformation
disaster of a tunnel before the displacement rate of monitoring points changes from uniform
variation to accelerated development.
Author Contributions: J.X.: conceptualization and constitutive model development; H.W.: numerical
simulation and analysis; C.S.: Back analysis method development; C.Y.: software; writing—review
and editing; G.R.: data curation, investigation. All authors have read and agreed to the published
version of the manuscript.
Funding: This research was funded by Science and Technology Project of China Railway 20th Bureau
Group Co., Ltd., and China Postdoctoral Science Foundation, grant number 20060390165.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Data are contained within this article.
Conflicts of Interest: The authors declare no conflict of interest.
References
1. Debernardi, D.; Barla, G. New Viscoplastic model for design analysis of tunnels in squeezing conditions. Rock Mech. Rock Eng.
2009, 42, 259–288. [CrossRef]
2. Bonini, M.; Barla, G. The Saint Martin La Porte access adit (Lyon-Turin Base Tunnel) revisited. Tunn. Undergr. Space Technol. 2012,
30, 38–54. [CrossRef]
3. Barla, G.; Debernardi, D.; Sterpi, D. Time-dependent modeling of tunnels in squeezing conditions. Int. J. Geomech. 2012, 12,
697–710. [CrossRef]
4. Barla, G.; Barla, M.; Bonini, M.; Debernardi, D. Guidelines for TBM tunnelling in squeezing conditions—A case study. Geotech.
Lett. 2014, 4, 83–87. [CrossRef]
5. Zhang, J.Z.; Zhou, X.P.; Yin, P. Visco-plastic deformation analysis of rock tunnels based on fractional derivatives. Tunn. Undergr.
Space Technol. 2019, 85, 209–219. [CrossRef]
6. Do, D.P.; Tran, N.T.; Mai, V.T.; Hoxha, D.; Vu, M.N. Time-dependent reliability analysis of deep tunnel in the viscoelastic Burger
rock with sequential installation of liners. Rock Mech. Rock Eng. 2020, 53, 1259–1285. [CrossRef]
7. Liu, Z.C.; Li, W.J.; Zhang, S.M.; Zhu, Y.Q. Synthetical analysis on monitoring of Wushaoling railway tunnel. Tunn. Undergr. Space
Technol. 2006, 21, 363–364. [CrossRef]
8. Sun, J. Rock rheological mechanics and its advance in engineering applications. Chin. J. Rock Mech. Eng. 2007, 26, 1081–1106.
9. Sakurai, S.; Takeuchi, K. Back analysis of measured displacements of tunnels. Rock Mech. Rock Eng. 1983, 16, 173–180. [CrossRef]
10. Gioda, G.; Sakurai, S. Back analysis procedures for the interpretation of field measurements in geomechanics. Int. J. Numer. Anal.
Met. 1987, 11, 555–583. [CrossRef]
11. Kaiser, P.; Zou, K.D.; Lang, P.A. Stress determination by back-analysis of excavation-induced stress changes—A case study. Rock
Mech. Rock Eng. 1990, 23, 185–200. [CrossRef]
12. Cai, M.; Morioka, H.; Kaiser, P.K.; Tasaka, Y.; Kurose, H.; Minami, M.; Maejima, T. Back-analysis of rock mass strength parameters
using AE monitoring data. Int. J. Rock Mech. Min. 2007, 44, 538–549. [CrossRef]
13. Ghorbani, M.; Sharifzadeh, M. Long term stability assessment of Siah Bisheh powerhouse cavern based on displacement back
analysis method. Tunn. Undergr. Space Technol. 2009, 24, 574–583. [CrossRef]
14. Hisatake, M.; Hieda, Y. Three-dimensional back-analysis method for the mechanical parameters of the new ground ahead of a
tunnel face. Tunn. Undergr. Space Technol. 2008, 23, 373–380. [CrossRef]
15. Miranda, T.; Dias, D.; Eclaircy-Caudron, S.; Correia, A.G.; Costa, L. Back analysis of geomechanical parameters by optimisation of
a 3D model of an underground structure. Tunn. Undergr. Space Technol. 2011, 26, 659–673. [CrossRef]
16. Karangh, K.T.; Clough, R.W. Finite element application in the characterization of elastic solids. Int. J. Solids Struct. 1972, 7, 11–23.
17. Gioda, G.; Jurina, L. Numerical identification of soil-structure interaction method in pressures. Int. J. Numer. Anal. Met. 1981, 5,
33–56. [CrossRef]
18. Wang, Z.; Li, Y. Back analysis of measured rheologic displacements of underground openings. In Proceedings of the 6th
Conference on Num. Meth. in Geom., Innsbruck, Austria, 11–15 April 1988; pp. 2291–2297.
19. Wang, Z.; Li, Y. Back analysis of viscoparameters and strata stress in underground opening. In Proceedings of the International
Symposium on Underground Engineering, New Delhi, India, 14–17 April 1988; pp. 181–186.
Appl. Sci. 2022, 12, 5311 23 of 23
20. Feng, X.; Zhang, Z.; Yang, C.; Lin, Y. Research on evolutionary neural network method for displacement back analysis. Chin. J.
Rock Mech. Eng. 1999, 18, 497–502.
21. Shang, Y.J.; Cai, J.G.; Hao, W.D.; Wu, X.Y.; Li, S.H. Intelligent back analysis of displacements using precedent type analysis for
tunneling. Tunn. Undergr. Space Technol. 2002, 17, 381–389. [CrossRef]
22. Feng, X.; Zhao, H.; Li, S. A new displacement back analysis to identify mechanical geo-material parameters based on hybrid
intelligent methodology. Int. J. Numer. Anal. Met. 2004, 28, 1141–1165. [CrossRef]
23. Zhao, H.B.; Yin, S. Geomechanical parameters identification by particle swarm optimization and support vector machine. Appl.
Math. Model. 2009, 33, 3997–4012. [CrossRef]
24. Zhao, H.; Ru, Z.; Yin, S. A practical indirect back analysis approach for geomechanical parameters identification. Mar. Georesour.
Geotechnol. 2015, 33, 212–221. [CrossRef]
25. Rechea, C.; Levasseur, S.; Finno, R. Inverse analysis techniques for parameter identification in simulation of excavation support
systems. Comput. Geotech. 2008, 35, 331–345. [CrossRef]
26. Vardakos, S.; Gutierrez, M.; Xia, C. Parameter identification in numerical modeling of tunneling using the differential evolution
genetic algorithm (DEGA). Tunn. Undergr. Space Technol. 2012, 28, 109–123. [CrossRef]
27. Pichler, B.; Lackner, R.; Mang, H.A. Back analysis of model parameters in geotechnical engineering by means of soft computing.
Int. J. Numer. Meth. Eng. 2003, 57, 1943–1978. [CrossRef]
28. Sun, J. Geotechnical Material Rheology and Its Engineering Application; China Construction Industry Press: Beijing, China, 1999; p. 545.
29. Chen, T.; Guestrin, C. XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, 13–17 August 2016; pp. 785–794.
30. Zhou, J.; Qiu, Y.G.; Zhu, S.L.; Armaghani, D.J.; Khandelwal, M.; Mohamad, E.T. Estimation of the TBM advance rate under hard
rock conditions using XGBoost and Bayesian optimization. Undergr. Space 2021, 6, 506–515. [CrossRef]
31. Jiang, P.; Wang, C.C.; Zhou, Q.; Shao, X.Y.; Shu, L.S.; Li, X.B. Optimization of laser welding process parameters of stainless steel
316L using FEM, Kriging and NSGA-II. Adv. Eng. Softw. 2016, 99, 147–160. [CrossRef]
32. Ardakan, M.A.; Rezvan, M.T. Multi-objective optimization of reliability-redundancy allocation problem with cold-standby
strategy using NSGA-II. Reliab. Eng. Syst. Saf. 2018, 172, 225–238. [CrossRef]

Applsci 12 05311 v2

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Applsci 12 05311 v2

Uploaded by

Copyright:

Available Formats

applied

1 Department of Geotechnical Engineering, Tongji University, Shanghai 200092, China;

Appl. Sci. 2022, 12, 5311. https://doi.org/10.3390/app12115311 https://www.mdpi.com/journal/applsci

and comprehensive macroscopic manifestation on the force change of the surrounding

2. NGNRM with Non-Stationary Parameter Creep

Incremental step start

Calculate viscoelastic strain increment Calculate viscoplastic strain rate

Calculate the current Calculate the sum of viscoelastic strain

Judge whether Calculate Jacobian viscoelasto-plastic

Calculate Jacobian elastic stiffness

Calculate stress increment using initial strain method

No Has the calculation of main program

Drucker–Prager yield criterion is adopted in this paper, denoted by Equation (3).

where Θ = ( R1 , c1 ), ( R2 , c2 ), · · · , R J , c J represents the region division of the tree and

f k (x) = f k−1 (x)+ T (x; Θk ) , k = 1, 2, · · · , K (6)

rki = yi − f k−1 (xi ), i = 1, 2, · · · , N (9)

where f t ( xi ) is the current weak learner; xi is the i-th training sample.

4. NSGA-II –XGBoost Based Multi-Objective Back Analysis

4.1. XGBoost-Based Relationship between Displacement and Geomechanical Parameters

4.2. XGBoost–NSGA-II-Based Back Analysis

θ∗ = argmin L(θ) = argmin{`1 (θ), `2 (θ), · · · , `d (θ)} (16)

where θ is the geomechanical parameters to be inverted; “argmin” represents the variable

where `i (i = 1, 2, · · · , d) represents the MSE of the i-th monitoring point.

eval(Vi ) = α(1 − α)i−1 (i = 1, 2, · · · , pop_size, α = 0.5) (18)

4.3. Computation Procedure for the XGBoost–NSGA-II-Based Back Analysis

Step 1 Collect engineering information Build samples for XGBoost

Build test parameter sets using Build numerical model

Step 3 Calculate the displacement of each sample by numerical model

Step 4 XGBoost algorithm Build training samples for XGBoost

Build XGBoost model to replace

Step 6 Back- calculate geo-mechanical parameters

Table 1. Tunnel support mechanical parameters.

Initial Support Concrete Lining Rock Bolt

Figure 4. Initial support and second lining of tunnel (unit: cm).

5.2. Modelling Approach

5.3. Equivalence of Initial Support Parameters

Table 2. Numerical simulation parameters of the model.

ρ /kg/m3 E 0 /GPa ν E1 /GPa η1 /GPa.h c0 /MPa ϕ /°

Table 2. Numerical simulation parameters of the model.

ρ/kg/m3 E0 /GPa ν E1 /GPa η1 /GPa.h c0 /MPa ϕ/◦

Appl. Sci. 2022, 12, x FOR PEER REVIEW

5.6. Parametric Study

where SS j is the sum of squares of deviation; d f j is theSS

As thus, F of factor j can be calculated by Equation (25).

Poisson's ratio reduction rate 0

0.15 Poisson's ratio reduction rate 60%

0.15 Internal friction angle reduction rate 60%

Internal friction angle reduction rate 40%

0.15 Internal friction angle reduction rate 60%

Internal friction angle reduction rate 80%

Internal friction angle reduction rate 0

(3) Influence of the surrounding rock elastic modulus

Sci. 2022, 12, x FOR

Displacement 1.0 Elastic modulus reduction rate 80%

Elastic modulus reduction

1.4 Elastic modulus reduction raterate

1.2 Elastic modulus reduction raterate

Lateral pressure coefficient reduction

0.14 Lateral pressure coefficient reduction rate 70%

0.20 Elastic modulus

tunnel has undergone the following five stages of evolution:

Figure 14. Contour map of tunnel displacement at rheological time 721 h.

Appl. Sci. 2022, 12, x FOR PEER REVIEW 23 of 25