Applied Soft Computing 136 (2023) 110066

Contents lists available at ScienceDirect

Applied Soft Computing

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

Review article

Comprehensive review of machine learning in geotechnical reliability

analysis: Algorithms, applications and further challenges
∗
Wengang Zhang a , Xin Gu a , Li Hong a , Liang Han a , Lin Wang b ,
a
School of Civil Engineering, Chongqing University, Chongqing 400045, China
b
School of National Safety and Emergency Management, Beijing Normal University, Zhuhai 519087, China

article info a b s t r a c t

Article history: Geotechnical reliability analysis provides a novel way to rationally take the underlying geotechnical
Received 13 February 2022 uncertainties into account and evaluate the stability of geotechnical structures by failure probability
Received in revised form 12 December 2022 (or equivalently, reliability index) from a probabilistic perspective, which has gained great attention
Accepted 21 January 2023
in the past few decades. With the rapid development of artificial intelligence techniques, various
Available online 2 February 2023
machine learning (ML) algorithms have been successfully applied in geotechnical reliability analysis
Keywords: and the number of relevant papers has been increasing at an accelerating pace. Although significant
Machine learning advances have been made in the past two decades, a systematic summary of this subject is still lacking.
Reliability analysis To better conclude current achievements and further shed light on future research, this paper aims
Geotechnical engineering to provide a state-of-the-art review of ML in geotechnical reliability analysis applications. Through
Uncertainty reviewing the papers published in the period from 2002 to 2022 with the topic of applying ML in the
reliability analysis of slopes, tunneling, and excavations, the pros and cons of the developed methods
are explicitly tabulated. The great achievements that have been made are systematically summarized
from two major aspects. In addition, the four potential challenges and prospective research possibilities
underlying geotechnical reliability analysis are also outlined, including multisensor data fusion, time-
variant reliability analysis, three-dimensional reliability analysis of practical cases, and ML model
selection and optimization.
© 2023 Elsevier B.V. All rights reserved.

Contents

1. Introduction......................................................................................................................................................................................................................... 2
2. Failure probability evaluation ........................................................................................................................................................................................... 2
3. Applications of ML in geotechnical reliability analysis ................................................................................................................................................. 3
3.1. Some typical machine learning algorithms ........................................................................................................................................................ 3
3.1.1. Radial basis function (RBF) ................................................................................................................................................................... 3
3.1.2. Convolutional Neural Network (CNN) ................................................................................................................................................. 4
3.1.3. Multivariate adaptive regression splines (MARS)............................................................................................................................... 4
3.1.4. Extreme learning machine (ELM)......................................................................................................................................................... 5
3.1.5. Extreme gradient boosting (XGBoost) ................................................................................................................................................. 6
3.2. Slopes ...................................................................................................................................................................................................................... 7
3.3. Tunneling ................................................................................................................................................................................................................ 9
3.4. Deep excavations ................................................................................................................................................................................................... 12

Abbreviations: ML, Machine learning; MCS, Monte Carlo simulation; ANN, Artificial neural network; SVM, Support vector machine; RVM, Relevant vector
machine; PSO, Particle swarm optimization; ELM, Extreme learning machine; MARS, Multivariate adaptive regression splines; XGBoost, Extreme gradient boosting;
CNN, Convolutional neural network; AI, Artificial intelligence; BPNN, Back-propagation neural network; 2D, Two-dimensional; 3D, Three-dimensional; RBF, Radial
basis function; GPR, Gaussian process regression; BF, Basic function; GCV, Generalized cross-validation; SVR, Support vector regression; SLFN, Single hidden layer
feed-forward neural network; FS, Safety factor; Pf , Failure probability; RLEM, Random limit equilibrium method; RFEM, Random finite element method; RSM,
Response surface method; RV, Random variable; RF, Random field; LSF, Limit state function; GA, Genetic algorithm; FEM, Finite element method; LS-SVM, Least
squares support vector machine; PCE, Polynomial chaos expansion; ASVM, Adaptive support vector machine; ARVM, Adaptive relevant vector machine; CSRSM,
Collocation-based stochastic response surface method; BCS, Bayesian compressive sampling
∗ Corresponding author.
E-mail address: [email protected] (L. Wang).

https://doi.org/10.1016/j.asoc.2023.110066
1568-4946/© 2023 Elsevier B.V. All rights reserved.
W. Zhang, X. Gu, L. Hong et al. Applied Soft Computing 136 (2023) 110066

4. Insights from previous works ........................................................................................................................................................................................... 13

4.1. Input geotechnical parameter characterization ................................................................................................................................................. 13
4.2. Output failure probability evaluation.................................................................................................................................................................. 13
5. Challenges and future directions ...................................................................................................................................................................................... 14
6. Summary and conclusions................................................................................................................................................................................................. 15
Declaration of competing interest.................................................................................................................................................................................... 16
Data availability .................................................................................................................................................................................................................. 16
Acknowledgments .............................................................................................................................................................................................................. 16
References ........................................................................................................................................................................................................................... 16

engineering problems without being compromised by the pro-

1. Introduction hibitively computational tasks in practical applications. Further-
more, benefiting from the great development of ML, the free and
The safety of geotechnical structures has gained increasing open-source packages of the above-mentioned ML algorithms are
attention in geotechnical engineering practice because their fail- widely available today from the Internet, such as the open-source
ure may cause significant casualties and economic losses, such as packages of the XGBoost and CNN are shared in the well-known
the catastrophic Qianjiangping landslide (e.g., [1]) and Shenzhen GitHub website and can be downloaded for users freely. This
landslide [2]. Evaluating the stability of geotechnical structures open and friendly environment has greatly accelerated the de-
reasonably is an important prerequisite for disaster prevention velopment of ML-based geotechnical reliability analysis methods.
and reduction, and geotechnical reliability analysis provides a Through retrieving from the Web of Science database, a total
novel way to explicitly consider the underlying various geotech- of 306 papers have been published in the period from 2002 to
nical uncertainties (e.g., the inherent variability of geomaterial 2022 concerned with applying ML in the reliability analysis of
properties, transformation uncertainty, and measurement error) slopes, tunneling, and excavations. In the past two decades, more
(e.g., [3,4]) and quantify the safety margin of geotechnical struc- and more geotechnical practitioners have been devoted to the
tures by failure probability (or equivalently, reliability index) application of ML in geotechnical reliability analysis, and this
from a probabilistic perspective. Generally, the evaluation of fail- research topic is expected to have great prospects.
ure probability is not a trivial task in geotechnical reliability Although significant advances have been made in geotechnical
analysis since the complicated implicit performance functions are reliability analysis in the past two decades, a systematic sum-
commonly encountered in engineering practice, leading to the mary on this subject is still lacking. To the best of the authors’
unavailability of analytical solutions. In such a case, the brute- knowledge, this paper is the first comprehensive survey of recent
force Monte Carlo simulation (MCS) method has gained popu- progress in ML-based geotechnical reliability analysis. The main
larity in reliability analysis by virtue of its simplicity, flexibility, contributions of this review can be summarized as follows: (1)
and easy to use for geotechnical engineers. Nonetheless, it suffers It systematically reviews 306 papers with the topic of applying
from a known criticism of extensive computational efforts and ML in the geotechnical reliability analysis of slopes, tunneling,
poor efficiency (e.g., [5–7]). The MCS needs to repeatedly invoke and deep excavations, and the pros and cons of the developed
geotechnical software (e.g., Abaqus, FLAC, and GeoStudio) to per- methods are explicitly tabulated using three tables; (2) It provides
form a large number of deterministic analyses so as to reach the insights into the previous research from two important aspects,
desired accuracy, which may be a computationally expensive task
namely the input geotechnical parameter characterization and
in practical applications.
output failure probability evaluation in geotechnical reliability
With the rapid development of artificial intelligence technolo-
analysis; (3) It outlines four crucial but unsolved problems in
gies, many researchers have contributed to the integration of ML
geotechnical reliability analysis, as well as novel challenges and
and geotechnical reliability analysis for improving computational
promising research directions.
accuracy and efficiency, which gives rise to a lot of successful ap-
This paper aims to present a review of ML in geotechni-
plications (e.g., [5,8–16]). The basic idea of ML in aiding geotech-
cal reliability analysis applications. The remainder of this paper
nical reliability analysis is to reconstruct the high-dimensional
starts with a brief introduction of failure probability evalua-
implicit performance function through learning from the pre-
tion in Section 2. In Section 3, several commonly used ML al-
pared data which mainly contains the input random variables or
gorithms and their applications in the geotechnical reliability
random field samples of geomaterial properties (e.g., cohesion,
analysis are systematically outlined. Besides, the insights from
friction angle, and saturated hydraulic conductivity) as well as the
previous works are also outlined in Section 4. In Section 5, the po-
corresponding quantity of interest (e.g., the factor of safety) that
tential challenges and prospective research possibilities are pre-
is generally calculated from the geotechnical software. As the ML-
sented. Finally, the primary conclusions drawn from this review
based reliability analysis model reaches the desired performance
after sufficient training and proper validation, it can be conve- are summarized in Section 6.
niently used to estimate the failure probability of geotechnical
structures with reasonable accuracy and efficiency. 2. Failure probability evaluation
Until now, various ML algorithms and their variants have
been successfully applied in the geotechnical reliability analysis, In geotechnical reliability analysis, the failure probability Pf (or
such as artificial neural network (ANN) (e.g., [17]), support vec- reliability index β ) is frequently used to measure the safety mar-
tor machine (SVM) (e.g., [18]), relevant vector machine (RVM) gin of geotechnical structures from a probabilistic perspective,
(e.g., [15]), particle swarm optimization (PSO) (e.g., [19]), extreme which can be defined as (e.g., [4]):
learning machine (ELM) (e.g., [20]), multivariate adaptive re-
∫ ∫∫
gression spline (MARS) (e.g., [10,21]), extreme gradient boosting Pf = P [g(Θ ) < 0] = ··· f (Θ )dΘ (1)
g(Θ )<0
(XGBoost) (e.g., [11,22,23]), and convolutional neural network
(CNN) [7,13,14,24,25]. These ML algorithms greatly enhance the where Θ denotes geotechnical parameters that are commonly
computational efficiency in the geotechnical reliability analy- regarded as random variables or random field variables; g(Θ )
sis, allowing engineers and researchers to focus more on the is the performance function concerning geotechnical parameters
2
W. Zhang, X. Gu, L. Hong et al. Applied Soft Computing 136 (2023) 110066

Θ ; P [g(Θ ) < 0] is the possibility of g(Θ ) less than zero; f (Θ )

denotes joint probability density function for Θ . Besides, the
reliability index β of interest can be conveniently obtained by:

β = Φ −1 (1 − Pf ) (2)
where Φ −1 denotes the inverse of the cumulative distribution
function of the standard normal distribution.
Generally, it is difficult to solve the Pf analytically by inte-
grating the performance function over the failure zone. This is
because the performance function is usually implicit and highly
nonlinear in geotechnical engineering. In such a case, it is prefer-
able to calculate the Pf approximately by MCS due to its concep-
tual simplicity and easy implementation for geotechnical engi-
neers. Taking the slope reliability analysis for example, the Pf can
be calculated by:
NMCS
1 ∑
Pf = I [FSi < 1] (3)
NMCS
i=1
Fig. 1. Proportion of ML in the reliability analysis of geotechnical structures.
where NMCS is the total number of MCS; FSi denotes the safety
factor (FS) based on the ith realization of random variables or
random fields; I [·] is an indicator function for judging whether The key principle of ML is to learn from historic experience
the slope failure occurs or not. If FSi is less than the threshold
automatically and rationally through computer algorithms, and
value (i.e., 1), I [·] = 1; otherwise, I [·] = 0. As indicated by Eq. (3),
then apply what it has learned to new problems, as known as
it is a cumbersome task to calculate Pf which involves a large
the generalization ability. Until now, ML has been widely used
number of deterministic slope stability analyses. In such a case,
in the realm of geotechnical engineering because it provides
an increasing number of ML algorithms have been integrated
a versatile tool to reveal and handle uncertainty and random-
into geotechnical reliability analysis to improve computational
ness (e.g., [26–33]), which are frequently faced by geotechnical
efficiency, as discussed in the next section.
engineers and researchers. This subsection will provide a brief
introduction of several presentative ML algorithms popularized
3. Applications of ML in geotechnical reliability analysis
in the geotechnical reliability analysis.
From the statistical data retrieved from the Web of Science
3.1.1. Radial basis function (RBF)
database, the main applications of ML in the reliability anal-
The RBF is an important kind of feed-forward neural network,
ysis of geotechnical structures focus on slopes, tunneling, and
which is designed based on localized basis functions and iterative
excavations. Fig. 1 depicts the proportion of ML applications in
function approximation networks. The RBF neural network con-
these three fields. It is obvious that most ML applications are
sists of three layers, i.e., input, hidden, and output layers (Fig. 3).
devoted to the slope (i.e., 48%) in the past two decades, followed
Like other ANNs, the input layer takes input and the output layer
by excavation, and tunneling accounts for the smallest propor-
provides the target value. The main feature of RBF lies in its
tion. If ‘machine learning’, ‘slope’, and ‘reliability’ are used as
hidden layer, where neurons process information with the appli-
keywords in a search engine of the Web of Science database,
cation of the nonlinear transformation from the input space to the
the annual number of published papers for ‘slope’ relating to
output space. More specifically, each neuron in the hidden layer
‘machine learning’ and ‘reliability’ can be obtained, as shown
in Fig. 2(a). It can be seen that the application of ML in slope uses the radial basis function as a non-linear activation function
reliability analysis shows a dramatic increase in 2019 and 2021. whose outputs are inversely proportional to the distance from
Similarly, the number of publications in tunneling and excavation the center of the neuron [34]. Therefore, the neuron center, the
can also be obtained as shown in Fig. 2(b) and (c), respectively. distance scale, and the shape of the radial function are the model
As expected, the ML application in the reliability analysis of parameters. For jth output, yj (i) can be expressed as follows:
these two geotechnical structures has an increasing trend. Thus, K
∑
this section will focus on previous studies related to reliability yj (i) = wjk Φ [x(i), ck , σk ], j = 1, 2, . . . , n i = 1, 2, . . . , N (4)
analysis of the three main geotechnical structures (i.e., slopes, k=1
tunneling, and deep excavations) based on ML techniques.
where wjk denotes connection weight, Φ [x(i), ck , σk ] is the spe-
cific RBF used, σk ∈ Rm and ck ∈ Rm are the width value vector
3.1. Some typical machine learning algorithms
and the center value vector of RBF, respectively, and K denotes
the number of RBFs used. ck = [ck1 , ck2 , . . . , ckm ]T ∈ Rm , k =
Artificial intelligence (AI) tends to build intelligent systems
1, 2, . . . , K and σk = [σk1 , σk2 , . . . , σkm ]T ∈ Rm , k = 1, 2, . . . , K .
that learn, adapt, mimic and even exceed human intelligence.
The most commonly used function Φ [·] is the Gaussian form,
In the past few decades, AI methods have been commonly used
which can be expressed as follows:
for solving scientific, engineering, economic and medical prob-
lems. These methods are considered as useful tools to address −(xi − cki )2
practical problems, as they can exploit and capture complex, ϕ[xi , cki , σki ] = exp( ) (5)
2σki2
dynamic, and non-linear relationships hidden in data. Currently,
AI is used in a wide range of applications, including data mining, Therefore, the jth output in Eq. (4) can be calculated as follows:
pattern recognition, natural language processing, self-driving, and K
{ m
}
∑ ∑
so on. ML is seen as a primary component of AI and it is pro- yj (i) = wjk exp − [(xi − cki ) /2σ ]
2 2
(6)
ki
posed to solve problems that cannot be explicitly programmed. k=1 i=1

3
W. Zhang, X. Gu, L. Hong et al. Applied Soft Computing 136 (2023) 110066

Compared with traditional methods, CNN has proven its supe-

riority in both accuracy and efficiency. The core feature of CNN
lies in the building block which is composed of many layers that
are interconnected with each other using a set of shared weights
and biases. The architecture of a typical CNN (Fig. 4) consists of
convolutional, pooling, and fully connected layers.
Convolution is a powerful tool used for extracting features.
For an input image, a square convolutional kernel with weights
is usually used to traverse each pixel. Each input pixel and the
corresponding area of the convolution kernel are multiplied by
the corresponding weights of the convolution kernel, which is
also called the filter bank. The result of this product is then
summed again, adding the bias of the convolution kernel. Finally,
this process results in a pixel in the output image, which can be
expressed as follows:
∑
yi = xi × wij + bi (8)
i

where xi denotes the input of each channel, wij is the corre-

sponding weight of each convolutional kernel, bi is the bias of
the convolutional kernel, and yi represents the final output. Gen-
erally, the results of this procedure are then passed through a
nonlinear function, such as Relu and tanh. This approach not only
effectively solves the disappearing gradient problem, but also has
high execution speed and precision.
Pooling is a down-sampling operation, which selects spatial
invariance by reducing the resolution of the feature map. Usually,
the unit in a pooling layer receives the maximum of a local patch
of units in one feature map, and there are two main types of pool-
ing operations. The max-pooling operator selects the maximum
value from an image array’s subarrays. By contrast, the average
pooling operator selects the mean value.
Fully connected layers are followed behind the convolution
and pooling procedure. They are the traditional multi-layer per-
ceptron neural network, which is one of the most popular ANNs.
The term ‘fully connected’ implies that all units in the feature
maps are concatenated together into a form of a vector.

3.1.3. Multivariate adaptive regression splines (MARS)

The MARS proposed by Friedman [38] is a nonlinear, nonpara-
metric data-driven method to capture the underlying relationship
between dependent variables and input variables without any
specific assumption about the functional form. The word ‘splines’
in MARS means these splines are smoothly connected. According
to Friedman [38], these curves are named as basic functions (BFs),
Fig. 2. Annual number of published papers relating to the ML application where the interface points to connect any two adjacent pieces
in reliability analysis of different geotechnical structures (Retrieved in August which are named as knots. These knots symbolize the beginning
2022).
of some data in one specific region and the end of the data
from another region, where the candidate knots are randomly set.
Generally, the procedure to construct a MARS model contains two
where cki is the center of the Gaussian function and σki denotes parts, named as the forward phase and the backward phase. To
the Gaussian function spread. One of the selection methods for be specific, the forward phase is to add functions and determine
σk is expressed as follows: the underlying knots to improve the prediction accuracy, and the
1 ∑ least effective terms for MARS will be deleted in the backward
σk2 = ∥x − ck ∥ (7)
phase.
mk
x∈θk
MARS builds up the implicit functional relationship via BFs,
where mk is the number of sample data in the kth cluster, and θk including piecewise linear and piecewise cubic functions. For
is the kth cluster of the training set. simplicity, only the piecewise linear function is demonstrated.
Piecewise linear functions are in the form of max(0, x − t) corre-
sponding to a knot at the value of t, which means only the positive
3.1.2. Convolutional Neural Network (CNN)
part will be adopted, otherwise it will be given a value of zero:
The CNN was first proposed by Fukushima and Miyake [36],
x − t , if x ≥ t
{
and it has been successfully used for the classification, segmen-
max(0, x − t) = (9)
tation, and recognition of images over the past two decades. 0, other w ise
4
W. Zhang, X. Gu, L. Hong et al. Applied Soft Computing 136 (2023) 110066

Fig. 3. A typical structure of RBF neural network.

Source: Adapted from Qasem et al. [35].

Fig. 4. A typical structure of CNN.

Source: Adapted from Li et al. [37].

Then, the MARS model will be built up of a linear combination subsets are compared using the less computation-consuming
of BFs and their interactions as follows: method of Generalized Cross-Validation (GCV) to avoid overfit-
m
∑ ting. Interesting readers can refer to Zhang and Goh [39,40] for
f (X) = β0 + βi λi (X) (10) more details on the calculation of the GCV. Generally, MARS is an
i=1
adaptive procedure because the selection of BFs and the variable
knot locations are data-based and dependent on the specific
in which each λi (x) is a basis function in the form of a spline problem.
function, or the product of two or more spline functions already
contained in the model. The constant coefficients β can be es- 3.1.4. Extreme learning machine (ELM)
timated via the least-squares method. In order to obtain the The ELM is a computer algorithm proposed by Huang et al.
model in Eq. (10), a forward selection procedure is first conducted [41] and it can be described as a least square-based single hidden
with the training data. Considering a current model with m basis layer feed-forward neural network (SLFN). Fig. 5 shows a typical
functions, the next pair is added to the model in the form: architecture of ELM. For a single hidden layer ELM model with
β̂m+1 λi (X ) max(0, Xj − t) + β̂m+2 λi (X ) max(0, t − Xj ) (11) N samples {xi , yi }, where xi = [xi1 , xi2 , . . . , xin ]T ∈ Rn and yi =
[yi1 , yi2 , . . . , yik ]T ∈ Rk , the standard SLFN with Ñ hidden neurons
Similarly, each β is evaluated via the least square method. When and the activation function g(x) can be expressed as follows:
a basic function is added to the MARS model, the interactions Ñ
between any two pre-existing BFs in the model have been consid- ∑
ered. BFs will be continually considered until the optimal model
βi g(wi xj + bi ) = oj , j = 1, 2, . . . , N (12)
i=1
is obtained without underfitting or overfitting.
To reduce the number of terms, a procedure named backward where wi = [wi1 , wi2 , . . . , win ]T is the input weight of ith hidden
deletion follows. This procedure aims to find a close to optimal layer neuron; βi = [βi1 , βi2 , . . . , βik ]T is the output weight of ith
model by removing the superfluous variables. In this phase, the hidden layer neuron; bi is the bias of ith hidden layer neuron;
algorithm prunes the model by deleting the BFs with the lowest wi xj is the inner product of wi and xj ; oj = [oj1 , oj2 , . . . , ojk ]T is
contribution to the model until it finds the best sub-model. Model the output parameter of the jth input parameter.
5
W. Zhang, X. Gu, L. Hong et al. Applied Soft Computing 136 (2023) 110066

Fig. 5. A typical structure of ELM.

The model can be continuously improved by minimizing the gradient boosting framework and is also a decision-tree-based
errors between output and actual value, which can be expressed ensemble method. The core principle of this method is that it
as: builds classification or regression trees one by one, and then the
N
∑ residuals of the previous tree are used to train the subsequent
model. In the training process, this method can integrate values
 
oj − yj  = 0 (13)
j=1
of the previously trained trees to achieve a better outcome. It can
use the pruning procedure to avoid over-fitting, which reduces
Eq. (12) is known as the cost function and it can be trans- the size of a decision tree by removing decision nodes that con-
formed into: tribute little to target values. Fig. 6 shows the schematic diagram
Ñ
∑ of the computational process of XGBoost and the prediction is
βi g(wi xj + bi ) = yj , j = 1, 2, . . . , N (14) calculated as:
i=1 t
(t )
∑ (t −1)
And it can be expressed in a simplified way: ŷi = fk (xi ) = ŷi + ft (xi ) (17)
k=1
Hβ = Y (15) (t ) (t −1)
where ŷi denotes the final tree model; ŷi is the previous tree
g(w1 x1 + b1 ) g(wÑ x1 + bÑ )
⎡ ⎤
··· model; ft (xi ) is the newly generated tree model, and t is the total
where H
.. .. .. number of base tree models.
= ⎣ . is the
⎢ ⎥
. . ⎦
To achieve optimal model performance, it is important to
g(w1 xN + b1 ) ··· g(wÑ xN + bÑ ) N ×Ñ select appropriate values of depth and number of trees. The
β1T
⎤ ⎡
objective function is defined as:
⎢ . ⎥
output matrix of the hidden layer, β = ⎣ .. ⎦ is the output t t
(t )
Obj(t ) =
∑ ∑
βÑT L(yi , ŷi ) + Ω (fi ) (18)
⎡ ⎤ Ñ ×k
Y1T i=1 i=1

⎢.⎥ (t )
where yi is the actual value; ŷi is the predicted value; L(yi , ŷi )
(t )
weight, and Y = ⎣ .. ⎦ is the target matrix. H is usually
T is the loss function and Ω (fi ) is the regularization term.
YN N ×k
Substituting Eq. (17) into Eq. (18) and then following some
obtained by the gradient descent method. The input weights wi deduction steps (seen in [42]), Eq. (18) can be obtained as:
and hidden bias bi can be randomly generated at the beginning of
t
the training stage, and can be fixed during the training procedure.
Obj(t ) =
∑ (t −1)
The output weight βi can be determined by solving the least L(yi , ŷi + ft (xi )) + Ω (ft ) + constant (19)
square solution of the linear system H β = Y : i=1

Following Chen and Guestrin [42], the objective function can

β̂ = H † Y (16)
be transformed into the form of the following equation:
where H † is the Moore–Penrose generalized inverse of the matrix t [ ]
H. (t )
∑ 1
Obj = gi ft (xi ) + hi ft2 (xi ) + Ω (ft ) (20)
2
i=1
3.1.5. Extreme gradient boosting (XGBoost)
(t −1) (t −1) (t −1) (t −1)
The XGBoost was proposed as an alternative method to pre- where gi = ∂ yi L(yi , yi ) and hi = ∂ 2 yi L(yi , yi ) are the
dict a response variable given certain covariates [42]. It uses a first and second-order gradient statistics on the loss function.
6
W. Zhang, X. Gu, L. Hong et al. Applied Soft Computing 136 (2023) 110066

Fig. 6. The schematic diagram of XGBoost algorithm.

Source: Adapted from Mo et al. [43].

The regularization term Ω (ft ) is expressed by Eq. (21) and parts: (1) A number of random variables or random field samples
it is used to reduce model complexity and avoid over-fitting, of geotechnical soil properties are firstly generated via MCS;
enhancing the generalization ability. (2) RFEM/RLEM is subsequently performed to calculate all the
1 corresponding FSs; (3) Consider these random variable samples or
Ω (ft ) = γ T + λ ∥w∥2 (21) random field samples as input, and the calculated FSs as the out-
2
put. After the procedure of training and validation with a specific
where T is the number of leaves; w is the corresponding weights ML algorithm, the target model is established to capture the po-
of the leaf; λ and γ are coefficients, and the default values are 1 tential relationship between the input random variable/random
and 0, respectively. field samples and the output FSs; (4) Through the obtained ML-
based model, the FSs corresponding to another large number of
3.2. Slopes random field samples can be calculated efficiently and accurately.
To enhance the understanding of the ML-aided method, Fig. 7 de-
With the increasing population, mountain areas are gradually scribes the differences between the traditional reliability analysis
being used, and more and more infrastructures are built near
and the ML-based reliability analysis of slope stability, where the
slopes to alleviate the pressure of the limited land resources [44].
uncertainty of soil parameters is characterized under the random
In these areas, landslide hazards are widespread, thus fatalities
field theory framework. It is noted that the abbreviation FOS in
and economic losses often occur. For disaster prevention and
Fig. 7 is the same as the FS mentioned in this paper.
mitigation, slope stability analysis is often performed. In the tra-
The remaining part of this subsection is to illustrate the appli-
ditional deterministic analysis, the FS measures whether a slope
cations of some typical ML algorithms in the reliability analysis of
is safe, which is calculated via the limit equilibrium method, finite
slope stability, especially for high dimensional or highly nonlinear
element limit analysis, or strength reduction method. Actually,
response surface method (RSM) (e.g., [14,44,45]).
the geotechnical parameters are rarely homogeneous by nature,
Zhang et al. [46] proposed a Kriging-based RSM to conduct the
e.g., soil properties in one site are different from those in another
system reliability analysis for soil slopes, where the undrained
site due to various deposition conditions or loading histories
shear strength was modeled as random variables. Compared with
[3]. The spatial variability of soil properties will be ignored by
the traditional RSM method, the Kriging-based RSM performs
the aforementioned deterministic methods when assessing the
slope stability and it is assumed that the soil parameters from the global approximation, which can evaluate the system relia-
different soil layers are all constant. To improve the accuracy of bility of soil slopes more accurately. Kang et al. [47] adopted a
the slope stability assessment, the spatial variability of soil prop- novel machine learning method called v-support vector machine
erties is considered and reliability analysis methods are adopted to construct the surrogate model for FS prediction, which can
to estimate the slope stability with a widely-used indicator of achieve an accurate evaluation of the system failure probability.
Pf via the brute-force MCS method, which may be seriously Ji et al. [48] performed the slope system reliability analysis via
computation-consuming. To be specific, the spatial variability of surrogate models, which were based on the least-squares support
soil parameters is characterized via the random field theory, vector machine regression. Results showed that the developed
then each FS corresponding to a specific random field will be model could guarantee robust and accurate estimations of the
calculated. Based on the obtained FSs via the MCS, the Pf can stability for spatially variable slopes. Chivata Cardenas [49] pro-
be calculated. In the field of risk assessment for slopes in the posed a meta-modeling approach based on Bayesian networks
geotechnical engineering, the aforementioned method to calcu- to explore the influences of uncertainties on slope stability. Liu
late the Pf is named as random limit equilibrium method (RLEM) et al. [8] further extended the MARS-based RSM to evaluate the
and random finite element method (RFEM) correspondingly. reliability of slopes in spatially variable soils. Based on their study,
In order to improve the computation efficiency, some efficient it could be found that the proposed method could evaluate the Pf
methods named response surface method or surrogate models of the slopes considering the spatial variability of soil properties
are proposed, i.e., the quadratic polynomial function. With the de- with enough efficiency and accuracy. He et al. [44] employed
velopment of artificial intelligence technologies, a more efficient neural networks and SVM as surrogate models to accelerate com-
method named ML-aided reliability analysis is gradually used in putational efficiency. Results showed that the required time for
the reliability analysis of slope stability considering the spatial calculating the failure probability with predefined accuracy can
variability of soil properties. Such an ML-aided method is similar be significantly reduced. Wang et al. [11] carried out the reliabil-
to the ‘surrogate models’ method. The main procedures for the ity analysis of the earth dam slope with the aid of XGBoost. It was
establishment of the ML-aided method contain the following found that this novel method could evaluate the Pf rationally and
7
W. Zhang, X. Gu, L. Hong et al. Applied Soft Computing 136 (2023) 110066

Fig. 7. Illustration of main procedures for traditional/ML-aided reliability analysis.

Source: From He et al. [44].

efficiently. Moreover, the coefficient of variation, the scale of fluc- The researches mentioned above are all two-dimensional (2D)
tuation, and the auto-correlation function may significantly affect problems to simplify the calculation. In nature, most landslides
the reliability analysis results. Wang and Goh [14] developed a are in three-dimensional (3D) space. Thus, 3D slope reliability
novel method based on CNNs to accelerate the process of reliabil- analysis can more accurately evaluate the true slope stability.
ity analysis. They took a two-layer slope as an example and found 2D reliability analysis is inherently time-consuming, therefore
3D reliability analysis may consume much more time. In order
that CNNs combined with MCSs could deal with uncertain events
to successfully perform the 3D slope reliability analysis, it is
with low failure probability. Compared with other metamodel-
necessary to extend the idea of 2D ML-based RSM to 3D problems.
based methods, i.e., single/multiple stochastic response surface
Recently, Song et al. [45] adopted the RBF to construct the RSM
method, CNNs-based RSM are more efficient. Recently, Wang and and then performed the 3D slope reliability analysis. In their
Goh [7] further combined the maximum entropy method with study, the strength parameters (i.e., cohesion and friction angle)
CNN-based surrogate model to estimate a multi-layered slope were considered as random variables. It is found that the 2D slope
stability, where the accuracy and efficiency are guaranteed at the reliability analysis is too conservative and may overestimate the
same time. Pf , and the RBF-based RSM can guarantee accuracy, feasibility, and
8
W. Zhang, X. Gu, L. Hong et al. Applied Soft Computing 136 (2023) 110066

Table 1
Summary of ML-based reliability analysis of slopes.
Dimension Parameter ML Advantages Limitations References
characterization algorithms
2D RV ANN High accuracy; Requiring a large number of initial parameters; Cho [17]
Parallel distributed processing; Consuming a long time during model training
Strong robustness
2D RV Kriging Predominant fitting Only valid within the range of calibration data Zhang et al. [46]
performance in highly
nonlinear problems
2D RV v-SVM Good generalization capability; Unable to provide sequential prediction; Kang et al. [47]
Adaptable for high-dimensional Single prediction outputs
problems;
2D RV LS-SVM Computationally efficient; Unable to provide sequential prediction; Ji et al. [48]
Good generalization capability; Single prediction outputs
Adaptable for high-dimensional
problems
2D RV Bayesian Graphical representation; Requiring the known prior information; Chivata Cardenas
networks Use of prior knowledge; Sensitive to the format of the input data [49]
Relevant variables
identification;
Propagation of uncertainty;
Inference enabled;
Use of incomplete observations
2D RF MARS No need to deal with a large Unable to provide sequential prediction; Liu et al. [8]
amount of data and high Single prediction outputs
dimension data
2D RF ANN, SVM Good generalization capability; Unable to provide sequential prediction; He et al. [44]
Adaptable for high-dimensional Single prediction outputs
problems
2D RF XGBoost High accuracy; Requiring large memory Wang et al. [11]
Parallel distributed processing;
Strong robustness
2D RF CNN Adaptable for high-dimensional Large number of samples required Wang and Goh
problems; [7,14]
High performance in feature
classification
3D RF RBF Better nonlinear fitting ability Unexplainable Song et al. [45]
3D RV RVM Adaptable for high-dimensional Time-consuming Li et al. [15]
non-linear problems;
Accurate and efficient
2D RV GWO- Strong fitting and Poor readability and interpretability; Ling et al. [5]
MKELM generalization No sequential prediction ability;
Single output prediction

superiority. Li et al. [15] developed active learning relevant vector face. In addition, due to the complexity of geological conditions
machine combined with MCS to perform the reliability analysis and the limitation of investigation technique, the geological con-
for rock slopes. It is found that the proposed method could avoid ditions in front of the cutter head are often unknown and change
overfitting in the period of model training and ensure enough cal- with the advanced process, which also will pose great threats
culation efficiency. Furthermore, it is verified that this combined to the safety of equipment and personnel life. Therefore, how to
method slightly outperforms some previously online methods, es- overcome these problems to improve the safety, efficiency, and
pecially for high non-linear problems. Ling et al. [5] combined the construction quality of shield tunnels is necessary.
strength reduction method, multi-kernel-based extreme learning In the past, deterministic analysis was the main analysis mea-
machine, and multi-objection gray wolf optimization to form a sure, while it did not take the uncertainty of parameters into
novel reliability analysis method. In their study, the cohesion consideration, which might not be realistic. Therefore, nowadays,
and friction angle were considered as random variables. Results reliability-based design is becoming more and more prevalent
showed that the developed method could evaluate slope stability
in the geotechnical engineering community. In addition, conven-
with sufficient accuracy and efficiency. Moreover, it is found
tional reliability is mainly based on intensive numerical simu-
that this proposed method has the potential to calculate Pf for
lation, which requires expensive computational cost, especially
slopes with the consideration of the coefficients of variation and
for complex geotechnical structure systems. Limit state function
correlation of different soil properties. More details about the
applications of ML in the reliability analysis of deep excavations (LSF) describes the engineering behaviors, while it generally can-
are presented in Table 1. not be known explicitly, which brings lots of difficulties for relia-
bility analysis. As ML is good at capturing the complex nonlinear
3.3. Tunneling and multivariate relationship between the target and features,
ML is gradually incorporated into reliability analysis to improve
Nowadays, with the development of the city, the tunnel plays computational efficiency and accuracy.
an increasingly important role in linking different places. During Due to excavation, the stress field around a tunnel will change
construction, several problems will be faced, such as surface set- so that the squeezing deformation may happen, which can result
tlement, shield machine performance, and stability of excavation in instability of the tunnel and pose a great challenge to later
9
W. Zhang, X. Gu, L. Hong et al. Applied Soft Computing 136 (2023) 110066

maintenance. Therefore, it is necessary to evaluate the reliability whose results indicate that this presented approach can effec-
of surrounding rock deformation. Qi et al. [50] first established tively capture the nonlinear and high dimension characteristic of
a reliability evaluation model by combining the commonly used LSF.
numerical simulation technique and reliability theory. This re- Actually, besides RSM, the variance reduction technique is an-
liability evaluation model is a constraint solving problem and other method to reduce computational cost, such as importance
beyond the capability of the conventional analytical method, and sampling, directional simulation, Latin hypercube sampling, and
hence, genetic algorithm (GA) is incorporated into it as GA is good subset simulation. It is found that these two methods can be
at finding global optimization. combined to perform the reliability analysis with high efficiency
Currently, the reliability-based method is being paid more and accuracy [57].
and more attention in geotechnical engineering, such as tunnel Based on previous research, Cui and Ghosn [58] proposed
engineering. In engineering practice, the LSF of an engineering a subset simulation–Kriging & K-means method, in which the
structure is not known explicitly but can be known implicitly Kriging algorithm is employed to build surrogate models for each
with aid of numerical simulation technique, which means that cluster of data points representing a potential failure mode as
conventional reliability analysis needs a large number of itera- identified via the K-means algorithm. Therefore, besides the high
tive calculations [51]. The RSM is a practical method in relia- efficiency, this method can also identify different failure modes
bility analysis in geotechnical engineering, because it makes use in the structure system, which can provide more reference to
of the merit of the finite element method (FEM) and can re- engineers. If this method is devoted to tunnel engineering, it will
duce the number of numerical simulations. However, traditional produce some promising works.
polynomial-based RSMs are frequently unable to approximate the In view of building metamodel for reliability analysis, Samui
true LSF. To overcome this problem, ANN-based and SVM-based et al. [59] have studied the application of RVM in reliability anal-
RSM were presented [52]. ANN has several shortcomings: a low ysis, which indicates that can predict the implicit LSF. Therefore,
convergence rate, easily falling into local minimum, and overfit- in terms of the metamodel technique, Li et al. [15] presented a
ting problem. In view of those problems, SVM-based RSM has new reliability analysis method combining an adaptive relevant
a better performance in practice. Inspired by pioneering works, vector machine (ARVM) and MCS, in which the stability of the
SVM has been devoted to multiple reliability analysis and is surrounding rock of a circular tunnel is taken into consideration
proven that it has a satisfying performance (e.g., [53]). for validation of the presented method. The results show that the
simulated limit state surface by ARVM-MCS is rather close to the
As an ML algorithm, SVM involves several hyper-parameters,
true limit state surface, indicating that the presented method has
and generally, the optimal hyper-parameters can build the best
good performance.
SVM model for a specific problem. As a result, sequential mini-
To study the settlement and bending moment of a shallow
mum optimization-based SVM algorithm was previously utilized
buried circular tunnel in a soil mass, Hamrouni et al. [60] adopted
to construct the SVM model. Considering that the calculation
RSM to analyze the serviceability of this tunnel in view of re-
process is time-consuming, Zhao et al. [51] proposed least squares
liability. Considering the expensive computational time, GA is
support vector machine (LS-SVM) and further integrate it with
introduced to find the optimal solutions, and the results show
RSM to build an LS-SVM-based RSM. Compared with traditional
that GA eliminates the successive iterative method utilized by
RSM and other methods, LS-SVM-based RSM can reach the final
the classical RSM method and saves much computation time. It
target using fewer iterations. Li et al. [54] developed a hybrid
is easy to understand that several types of geological conditions
model by integrating an experimental design called uniform de-
may exist in one section of the tunnel, and hence, Zhou et al.
sign with SVM, where the uniform design is utilized to generate
[61] investigated the effect of a weak interlayer on the stability
sampling points and SVM is applied to build a response surface
of the tunnel face, where an active learning method combining
approximating the true implicit LSF. Through three examples, the Kriging and MCS was employed for the reliability analysis, whose
presented method is proven that it has a powerful potential for flowchart is illustrated in Fig. 8. At the same time, the relative im-
probabilistic assessment of tunnel stability involving an inexplicit portance of each soil property is evaluated by integrating global
LSF. sensitivity analysis. Compared with conventional approaches, the
For reliability analysis, Bai et al. [55] found that RBF is capable proposed method is capable of providing multiple interesting
of constructing LSF with high accuracy and good performance. results, including the probability of failure, model response statis-
Therefore, Wang et al. [56] applied three RBFs to construct meta- tics and sensitivity index with a relatively lower computational
models of LSF. Two tunnel examples were taken into considera- cost.
tion for validation of three RBFs in tunnel reliability analysis, and Nowadays, the numerical simulation technique is a useful
the results prove that the metamodel based on RBFs can effec- tool in geotechnical engineering. Proper selection of geomechan-
tively and efficiently perform reliability for complex engineering ical parameters is a key component in geotechnical numerical
systems with lower computational cost than traditional RSM. simulation. To this end, parameter inversion is adopted. In the
Metamodel-based Monte Carlo method is also a useful tool in early period, displacement back-analysis is often used to identify
the reliability analysis community, several strategies have been the mechanical parameters of the surrounding rock mass of a
put forward to build a suitable metamodel for a specific problem, tunnel [62]. Later, RSF or modified RSF by ANN has been put
including polynomial chaos expansion (PCE), Kriging model, and forward to perform mechanical parameter inversion. In light of
ANN. PCE is easily subjected to the ‘curse of dimensionality’, the drawbacks of ANN, SVR is gradually incorporated into RSF. To
and the drawbacks of ANN have been discussed above. As pre- obtain the best SVR model, two optimization algorithms (i.e., GA
mentioned, SVM has several merits, and for example, it can avoid and PSO) can be utilized to achieve good results. However, GA
overfitting problems and is capable of high dimension and highly often has a low convergence rate, and PSO easily falls into local
nonlinear problems. Pan and Dias [57] developed a reliability minimum and also has relatively lower accuracy. As a result,
system by combining adaptive SVM (ASVM) and MCS, in which Zhuang et al. [63] adopted a new multi-strategy artificial fish
system ASVM adopts an active learning strategy. Four examples, swarm algorithm to find the best SVR model.
including nonlinearity, high dimension, and implicit limit state For the tunneling work, the tunnel roof is also a key part that
function problems (i.e. stability analysis of a tunnel face), are should be paid attention to. If the tunnel is shallowly buried, the
used to validate the applicability of the proposed ASVM-MCS, significant tunnel roof defection can lead to ground settlement
10
W. Zhang, X. Gu, L. Hong et al. Applied Soft Computing 136 (2023) 110066

Fig. 8. Flowchart of an active learning method based on Kriging and MCS.

Source: From Zhou et al. [61].

or even ground subsidence. While the conventional 3D numerical maximum tunnel roof defection, and ANN is used to establish the
simulation can be used to assess the reliability of tunnel roof, the surrogate model about the maximum tunnel roof defection, and
estimation cannot be in real-time due to its expensive computa- at last, a simplified approach is further proposed to estimate the
tional cost. To this end, Zhou et al. [64] did a study to combine the exceedance probability of tunnel roof defection. Similarly, Verma
conventional numerical simulator and ML algorithms. In the work et al. [65] also established three surrogate models to represent
of Zhou et al. [64], according to the actual project information, the limit state function of tunneling using three ML algorithms
FLAC 3D is employed to generate the synthetic database about including collocation-based stochastic response surface method
11
W. Zhang, X. Gu, L. Hong et al. Applied Soft Computing 136 (2023) 110066

Table 2
Summary of ML-based reliability analysis of tunneling.
Dimension Parameter ML Advantages Limitations References
characterization algorithms
2D RV SVM (1) The current approach combining uniform (1) It is difficult to use the analysis Li et al. [54]
design and SVM can be used successfully for of variance for data analysis; (2) The
their reliability analyses involving the implicit spatial variability of soils cannot be
LSF; (2) Suitable for small number of sampling considered.
points.
2D RV RBF (1) Accurate estimations of the failure (1) A low convergence rate; (2) easily Wang et al. [56]
probability; (2) Less computational effort. falling into local optimum; (3) The
spatial variability of soils cannot be
considered.
2D RV ASVM (1) Be capable of handling issues involving (1) The proposed method is only Pan and Dias [57]
nonlinearity, high-dimension, and implicit applicable for the assessment of
limit state function; (2) The errors of the failure probability, and hence, it
estimated failure probabilities by the proposed cannot provide other results related
ASVM-MCS are under control and rather small. to probabilistic analysis; (2) The
current method cannot deal with the
issue of small failure probability; (3)
The spatial variability of soils cannot
be considered.
2D RV Kriging & (1) The proposed method can identify the (1) The K value in K-means should Cui and Ghosn [58]
K-means important failure modes; (2) With the aid of be specified before; (2) K-means is
Subset Simulation, it can estimate the system sensitive to the center of cluster and
reliability and solve nonlinear problems with outliers; (3) K-means is not suitable
low probabilities of failure. for classification of non-convex
shapes; (4) The spatial variability of
soils cannot be considered.
2D RV GA The unknown parameters in the performance The spatial variability of soils cannot Hamrouni et al. [60]
function can be optimized by GA. be considered.
3D RV Kriging Good accuracy and less computational cost. (1) The spatial variability of soils Zhou et al. [61]
cannot be considered so that the
Pf may be over-estimated slightly; (2)
The positive correlation between the
weak layer soils and the nearby soils.
2D RV ARVM (1) Avoid the over-fitting problems; (2) The spatial variability of soils cannot Li et al. [15]
Suitable for high-dimensional and non-linear be considered.
problems with less computational costs.
3D RV ANN It is computational efficiency and makes the (1) The correlations among the input Zhou et al. [64]
probabilistic analysis computationally viable. random variables are not considered
well, especially for the site-specific
database; (2) ANN has several
shortcomings: a low convergence
rate, easily falling into local
minimum, and overfitting problem.
2D RF CNN (1) Less computational costs; (2) The nonlinear The determination of Zhang et al. [66]
pattern between the random field of soil hyper-parameters in CNN is complex.
property input and the factor of safety of
tunnel target can be deserted; (3) The
confidence interval of predicted results can be
produced.
2D RV CSRSM, (1) Less computational costs; (2) The epistemic The results from MARS fall on the Verma et al. [65]
MARS uncertainty can be considered; (3) CSRSM and conservative side as compared to the
MARS both are found appropriate for both the CSRSM and it produces more
normal and non-normal random variables. unexpected scenarios which need to
be considered.

(CSRSM), multi-gene genetic programming, and MARS, and in 3.4. Deep excavations
conjunction with the MCS, the failure probability of tunnel is
also estimated accurately and efficiently. The results also show In recent years, many researchers majored in applications of
that the computational cost is evidently reduced, and the CSRSM ML in evaluating key parameters during deep excavation, such as
and MARS can incorporate the epistemic uncertainty very well. determining the deflection of support wall and ground settlement
Compared with regular ML algorithms, Zhang et al. [66] applied or basal heave displacement via polynomial regression, MARS,
the CNN to predict the tunnel performance, in which the basic ANN (e.g., [67–69]). Additionally, more and more studies also
framework is similar, that is, numerical simulation considering concentrated on geotechnical reliability analysis via random field
the spatial variability of soils is used to produce the synthetic theory. Different from deterministic analysis, reliability analysis
database for training and testing, and then, a surrogate model usually needs more computational effort since the complicated
about tunnel performance is built by CNN architecture. The nov- implicit performance functions are commonly encountered in
elty in the work of Zhang et al. [66] is that the presented method engineering practice.
can obtain the results with the confidence interval. Addition- Huang and Wang [70] applied ANN in the reliability analysis
ally, more details about the applications of ML in the reliability of braced excavation. The mechanical and physical properties
analysis of tunneling are tabulated in Table 2. of soil were regarded as input layers and the deformation and
12
W. Zhang, X. Gu, L. Hong et al. Applied Soft Computing 136 (2023) 110066

safety factors were regarded as output layers. ANN model was limited, which poses great difficulty in rationally characterizing
adopted to determine the limit state functions so that the reli- the geotechnical parameters due to the unavailability of statistical
ability index can be obtained. Cao et al. [71] applied the SVM information (e.g., means, standard deviations, and probability
in the system reliability analysis of foundation excavation. The distribution). Such statistical information obtained from classical
SVM model was adopted to establish response surface equations statistical analyses generally requires a considerable amount of
and the Pearson correlations analysis was adopted to evaluate the test data so as to generate statistically meaningful results. To
correlations of multiple failure modes. Zhang et al. [68] combined address this issue, machine learning provides a rational way to
the Back-propagation neural network (BPNN) and the optimized portray the spatially variable geotechnical parameters based on
gray discrete Verhulst model to predict the foundation pit set- limited site-specific data. Wang et al. [76] proposed a Bayesian
tlement. The results showed that the proposed model inherited machine learning method called Bayesian compressive sampling
the advantages of the two models, which could satisfactorily de- (BCS) method to directly simulate the random field samples from
scribe the settlement monitoring projects. Kumar and Samui [20] sparsely measured data. Hu et al. [77] further applied the BCS
proposed the ELM-based and MARS-based first-order second- method for generating 2D isotropic or anisotropic random field
moment methods for the reliability analysis of pile foundations, samples from sparse measurements. Wang et al. [78] optimized
which significantly reduces the computational pains and memory the BCS for simulating random field samples with a linear or
requirements in the reliability analysis. It was found that the nonlinear trend directly from sparse measurements without de-
MARS model performed much better than the ELM model in their trending. Thus, it can be seen that the newly developed BCS
research. He et al. [72] evaluated the failure probability in braced method provides a versatile tool for generating the random field
excavation using Bayesian networks with integrated model up- samples directly from the limited measurement data, which is
dating, which overcomes the limitation of the complex likelihood able to simulate the non-stationary non-Gaussian random field
function and enables failure probabilities to be determined with and random field with unknown trend function. Besides, inspired
real-time result updating. Ray et al. [9] adopted three different by the powerful ability of CNN in the field of image recognition,
ML algorithms to study the shallow foundation reliability based Zhang et al. [79,80] developed an efficient CNN-based approach
on settlement criteria, in which the soil parameters were taken as for estimating the horizontal scale of fluctuation and vertical
input variables while the settlement of shallow foundations was scale of fluctuation based on the limited cone penetration test
regarded as output. Wang et al. [13,24] proposed a new method data. Compared with the BCS method that generates random
that can decrease computing efforts underlying the reliability field samples directly from sparse measurements, the CNN-based
analysis of braced excavations and shallow foundations. In their approach first quantifies the critical parameters required in the
study, the random fields were treated as images in the CNN, autocorrelation functions (e.g., the single exponential autocorre-
and the predictions of FEM can be regarded as output layers lation) and then uses them for random field characterization. Both
that contain information about the random variabilities in both the BCS method and CNN-based method offer a rational approach
spatial distribution and intensity, as described in Fig. 9. Similarly, for probabilistically characterizing the spatially variable geotech-
Wu et al. [25] applied the CNN to predict the wall deflection nical parameters, addressing the bottleneck of sparse and limited
of braced excavation considering the spatial variability of soil site-specific data that are frequently encountered in geotechnical
parameters, and compared its predictive performance with the engineering practice.
XGBoost model.
On the other hand, ML methods also played an important 4.2. Output failure probability evaluation
role in taking good use of the information of in-site investigation
data for soil parameter characterization. For example, in condi- The evaluation of failure probability (or equivalently, reliabil-
tional random field modeling, the assumptions based on historical ity index) is a major concern in geotechnical reliability analysis,
statistical information were regarded as prior information, and which quantifies the safety margin of geotechnical structures
the in-site investigation information will update the prior model from a probabilistic perspective. Although the MCS method has
and obtain the posterior model. For example, Juang et al. [73] been widely used in reliability analysis due to the advantages
conducted probabilistic inverse analysis of wall deformations and of simplicity and flexibility, it is known that the MCS method
ground settlement induced by excavations via the Bayesian up- generally requires extensive computational efforts and is often
dating method. Luo and Hu [74] expanded the Bayesian method criticized for poor efficiency. As a promising alternative, various
to update multiple soil parameters using field data on wall de- ML algorithms and their variants have been successfully applied
flections and ground surface settlement from braced excavations. in the past two decades for facilitating the evaluation of failure
Leung et al. [75] proposed a novel conditional random field mod- probability, such as MARS, XGBoost, CNN, and so on. These ML
eling method and discussed the optimal arrangement of boring algorithms significantly improve computational efficiency and
holes in project sites. More details about the applications of ML boost the application of geotechnical reliability analysis methods
in the reliability analysis of deep excavations are listed in Table 3. in solving practical engineering problems (e.g., reliability-based
design), allowing engineers and researchers to concentrate more
4. Insights from previous works on practical engineering problems without being compromised
by the prohibitively computational tasks in practical applications.
After reviewing the research contributed by previous Actually, the key idea of ML in aiding geotechnical reliability
researchers, great achievements have been made in geotechnical analysis is to approximate the high-dimensional implicit perfor-
reliability analysis with the aid of ML, which are summarized mance function through learning from a preparatory database,
from two important aspects in this section. which contains the input geotechnical parameters (e.g., shear
strength parameters and hydraulic parameters) and the output
4.1. Input geotechnical parameter characterization quantity of interest (e.g., the FS for slopes). As the established
ML models achieve a desired performance after sufficient training
Probabilistic characterization of the spatially variable geotech- and rational validation, they can be conveniently used to estimate
nical parameters is a significant prerequisite in geotechnical relia- the failure probability of geotechnical infrastructures.
bility analysis. It is known that the measured data obtained from It can be seen that the calibration of ML models plays a signif-
various in-situ and/or laboratory tests are relatively sparse and icant role in ML-aided geotechnical reliability analysis methods,
13
W. Zhang, X. Gu, L. Hong et al. Applied Soft Computing 136 (2023) 110066

Fig. 9. Framework of CNN model.

Source: From Wang et al. [13].

Table 3
Summary of ML-based reliability analysis of deep excavations.
Dimension Parameter ML Advantages Limitations References
characterization algorithms
2D RV MARS Simple and explicit expression Single output prediction; Zhang and Goh
Number of basic functions is [40], Kumar and
hard to determine Samui [20]
2D RV BPNN Multiple output prediction; Poor readability Zhang et al. [68]
Low computational cost
2D RV SVM Simple and explicit expression Poor nonlinear mapping ability Cao et al. [71]
2D RV ANN Simple and explicit expression Single output prediction; Poor Huang and Wang
nonlinear mapping ability [70]
2D RV ELM Higher learning efficiency; More computational cost; More Kumar and Samui
Higher generalization ability nodes will lead to overfitting [20]
2D RV Bayesian Available for sparse data; Single output prediction Juang et al. [73],
updating Available for multidimensional Luo and Hu [74]
monitoring data
2D RV PSO Low computational cost Poor generalization ability Ray et al. [9]
2D RF CNN Low computational cost; Only available for input data Wang et al. [13],
Available for multiple input with matrix type Wang et al. [24],
data Wu et al. [25]
3D RV Bayesian The real-time updated failure Prior distribution must be He et al. [72]
network probabilities in the excavation defined reasonably
process are available
3D RF Bayesian Available for sparse data; Posterior model also depends Leung et al. [75]
updating Considering both prior on prior information
information and in-site
investigation data

which highly depend on the preparatory database. Compared analysis, the computational efforts may increase significantly.
with the computational efforts in the preparation of a database Thus, preparing a training database becomes a computationally
that generally necessitates a large number of deterministic anal- expensive and time-consuming task in geotechnical reliability
yses through repeatedly invoking geotechnical software, the com- analysis, which emphasizes the necessity of developing new ML-
putational cost of the established ML models in the estimation of aided geotechnical reliability analysis methods that require fewer
failure probability can be negligible. For example, Wang et al. [11] training samples.
took about 3 h and 20 min on a personal desktop computer to
prepare a database with a total of 2198 samples, while just about 5. Challenges and future directions
15 s were consumed in predicting the failure probability of 10,000
testing samples when using the established XGBoost model. If we Benefited from the great development of AI technologies, more
extend the 2D model analyzed in Wang et al. [11] to 3D reliability and more ML algorithms including several commonly used ML
14
W. Zhang, X. Gu, L. Hong et al. Applied Soft Computing 136 (2023) 110066

algorithms and some latest advanced ML methods (e.g., XGBoost Finally, the monitoring information was rarely incorporated
and CNN) have been applied to the geotechnical reliability anal- into the geotechnical parameter characterization when compared
ysis in the past decades, which gives rise to a lot of ML-based with the measured data obtained from various in-situ and/or lab-
reliability analysis methods and greatly facilitates the implemen- oratory tests. In the past few decades, previous researchers have
tation of reliability analysis in geotechnical engineering practice. contributed a lot to probabilistically characterize the spatially
Although great achievements have been obtained, there still ex- variable geotechnical parameters based on sparse and limited
ist many remaining issues needed to be further explored, as site-specific data. It is well known that more and more advanced
summarized in Fig. 10. sensors are installed onto geotechnical infrastructures. MegaByte
Firstly, most of the available applications were concentrated or even GigaByte of monitoring data are produced each hour,
on time-invariant reliability analysis, and the time-variant reli- which would be transferred by the fifth-generation (5G) mobile
ability analysis has been rarely reported. The time-variant re- network. Therefore, the requirement for making the best use of
liability analysis generally focuses on the evaluation of failure these valuable monitoring data in geotechnical reliability analysis
probability during a given period, in contrast, the time-invariant is even more demanding.
reliability analysis pays attention to the failure probability at a
specified time point of interest which is unable to consider the ef- 6. Summary and conclusions
fects of geomaterial properties and external loading may be vary-
ing with time. For example, it is well recognized that the reservoir This paper reviewed previous studies on the applications of ML
slope stability is significantly affected by the periodic reservoir in different geotechnical reliability analysis problems in the past
water level fluctuation and seasonal rainfall (e.g., [81,82]), leading two decades. The evaluation of failure probability is a primary
to the reservoir slope reliability varying with the external envi- concern in the geotechnical reliability analysis and the direct
ronment. Besides, as a key component of tunneling engineering, MCS method usually requires extensive computational efforts to
the lining structure is frequently subjected to groundwater seep- ensure the desired accuracy. With the great development of ML
age which may cause material deterioration with service time, in the past few decades, many researchers have contributed to
potentially threatening the serviceability and safety of the tunnel- the evaluation of failure probability by incorporating ML algo-
ing [83]. Thus, the time-variant reliability analysis of geotechnical rithms, such as MARS, XGBoost, CNN, and so on. These ML-based
structures with the aid of ML is recommended in future research. reliability analysis methods provide a powerful tool to calculate
Secondly, we discuss how to select an appropriate ML al- the failure probability with higher accuracy and efficiency, allow-
gorithm among numerous candidates and rationally determine ing geotechnical engineers to concentrate more on engineering
the associated hyper-parameters. Until now, many open-source problems rather than the prohibitively computational tasks in
packages of ML algorithms are available on the Internet and can practical applications.
be downloaded freely for scientific research, such as the packages By reviewing the published research in the field of geotechni-
of the XGBoost and CNN which are shared in the well-known cal reliability analysis, it is evident that ML algorithms and their
open-source website called GitHub. This open and friendly envi- variants have been successfully applied in the reliability analysis
ronment allows researchers and engineers to have more choices of geotechnical structures including slopes, tunneling, and deep
when they want to perform geotechnical reliability analysis with excavations. Each ML algorithm has its own merits and shortages,
the aid of the ML algorithm, enabling the possibility of addressing and thus it is advisable to select an appropriate ML algorithm
the same engineering problem using different ML algorithms. The according to the computational efficiency, memory consumption,
associated hyper-parameters can be determined through opti- and prediction performance in practical applications since no
mization strategies [84]. It is undeniable that each ML algorithm single or particular model can be always regarded as the most
has its own merits and shortages, and thus no single or particular appropriate one in solving different geotechnical problems.
model can be always regarded as the most appropriate one in Although remarkable progress has been achieved in the appli-
solving geotechnical problems. Thus, it is advisable to choose cation of ML in geotechnical reliability analysis, there still exist
an appropriate ML algorithm according to the computational many challenges that need further attention in future studies.
efficiency, memory consumption, and prediction performance in Most of the available studies were concentrated on time-invariant
practical applications. reliability analysis which pays attention to the failure probability
Furthermore, previous research mainly focused on simpli- at a specific time point of interest and is unable to consider the
fied models and lacked realistic geotechnical case applications. effects of geomaterial deterioration and external loading variation
Through reviewing the published research on slope reliability with time. In addition, previous research mainly focused on sim-
analysis using ML algorithms, the idealized simple slopes have plified models or two-dimensional reliability analysis, leading to
been widely studied, which may be far away from the realistic the lack of realistic geotechnical case applications. Furthermore,
situation. Compared with the commonly used 2D slope model, the computational efficiency needs to be further improved be-
the reliability analysis of slope stability in 3D space not only cause preparing a training dataset is generally computationally
can reflect the stability state accurately, but also reveals the expensive and time-consuming in geotechnical reliability analysis
deformation and mechanical behaviors from a more practical which needs to repeatedly conduct a large number of determinis-
perspective. Therefore, it is necessary to extend the ML-based tic analyses. Thus, future research is recommended to introduce
reliability analysis methods to 3D cases. Similarly, many works some advanced ML algorithms or ensemble/hybrid models that
related to tunneling and deep excavations simply consider a sin- require fewer training samples, so as to reduce the computa-
gle circular tunnel or a simple deep excavation model in 2D space tional burden in geotechnical practical applications. In addition,
without any adjacent building, tunnel, or deep excavation. Under the future study reviewing of machine learning methods for 5G
the current complex building environment, especially in urban data and their application in geotechnical reliability analysis, is
areas with high densities of buildings, the existing infrastructures desirable.
may be affected by adjacent construction, such as the existing Benefiting from the rapid development of ML, a large number
piles which can be influenced by adjacent deep excavations, and of open-source packages are available on the Internet and can
the existing tunnels which can be affected by a new tunnel. These be downloaded freely for scientific research, allowing researchers
works are far away from the true nature of the current city. Thus, and engineers to address practical engineering problems using
future research should focus more on the reliability analysis of different algorithms of interest. This open and friendly environ-
realistic geotechnical cases with the aid of ML. ment can not only stimulate the growing tendency of applying
15
W. Zhang, X. Gu, L. Hong et al. Applied Soft Computing 136 (2023) 110066

Fig. 10. Research progress, gaps, and future directions.

ML in geotechnical engineering, but also accelerate the devel- [2] L.T. Zhan, Z. Zhang, Y.M. Chen, R. Chen, S. Zhang, J. Liu, A.G. Li, The
opment of ML-based reliability analysis methods. In the near 2015 Shenzhen catastrophic landslide in a construction waste dump :
Reconstitution of dump structure and failure mechanisms via geotechnical
future, the above-mentioned challenges including time-variant
investigations, Eng. Geol. 238 (2018) 15–26, http://dx.doi.org/10.1016/j.
reliability analysis, three-dimensional reliability analysis of re- enggeo.2018.02.019.
alistic geotechnical cases, and more efficient reliability analysis [3] K. Phoon, Z. Cao, J. Ji, Y.F. Leung, S. Najjar, T. Shuku, C. Tang, Z.Y. Yin,
approaches, are expected to be successfully conquered with the Y. Ikumasa, J. Ching, Geotechnical uncertainty, modeling, and decision
making, Soils Found 62 (5) (2022) 101189, http://dx.doi.org/10.1016/j.
aid of ML.
sandf.2022.101189.
[4] S.H. Jiang, J. Huang, D.V. Griffiths, Z.P. Deng, Advances in reliability and
Declaration of competing interest risk analyses of slopes in spatially variable soils: A state-of-the-art review,
Comput. Geotech. 141 (2022) http://dx.doi.org/10.1016/j.compgeo.2021.
104498.
The authors declare that they have no known competing finan- [5] Q. Ling, Q. Zhang, Y. Wei, L. Kong, L. Zhu, Slope reliability evaluation
cial interests or personal relationships that could have appeared based on multi-objective grey wolf optimization-multi-kernel-based ex-
to influence the work reported in this paper. treme learning machine agent model, Bull. Eng. Geol. Environ. 80 (2021)
2011–2024, http://dx.doi.org/10.1007/s10064-020-02090-5.
[6] J.Z. Li, S.H. Zhang, L.L. Liu, L. Huang, Y.M. Cheng, D. Dias, Probabilistic
Data availability analysis of pile-reinforced slopes in spatially variable soils with rotated
anisotropy, Comput. Geotech. 146 (2022) 104744, http://dx.doi.org/10.
1016/j.compgeo.2022.104744.
Data will be made available on request. [7] Z.Z. Wang, S.H. Goh, A maximum entropy method using fractional mo-
ments and deep learning for geotechnical reliability analysis, Acta Geotech.
17 (4) (2022) 1147–1166, http://dx.doi.org/10.1007/s11440-021-01326-2.
Acknowledgments [8] L. Liu, S. Zhang, Y. Cheng, L. Liang, Advanced reliability analysis of slopes
in spatially variable soils using multivariate adaptive regression splines,
The authors are grateful to the financial supports from Na- Geosci. Front. 10 (2) (2019) 671–682, http://dx.doi.org/10.1016/j.gsf.2018.
03.013.
tional Natural Science Foundation of China (52008058 and
[9] R. Ray, D. Kumar, P. Samui, L.B. Roy, A.T.C. Goh, W. Zhang, Application of
52078086), Program of Distinguished Young Scholars, Natural Sci- soft computing techniques for shallow foundation reliability in geotechni-
ence Foundation of Chongqing, China (cstc2020jcyj-jq0087), and cal engineering, Geosci. Front. 12 (1) (2021) 375–383, http://dx.doi.org/10.
High-end Foreign Expert Introduction program, China 1016/j.gsf.2020.05.003.
(DL2021165001L, G20200022005, and G2022165004L). Special [10] L. Wang, C. Wu, X. Gu, H. Liu, G. Mei, W. Zhang, Probabilistic stability
analysis of earth dam slope under transient seepage using multivari-
thanks are given to the anonymous reviewers for their valuable ate adaptive regression splines, Bull. Eng. Geol. Environ. 79 (6) (2020)
comments and constructive suggestions. 2763–2775, http://dx.doi.org/10.1007/s10064-020-01730-0.
[11] L. Wang, C. Wu, L. Tang, W. Zhang, S. Lacasse, H. Liu, L. Gao, Efficient
reliability analysis of earth dam slope stability using extreme gradient
References boosting method, Acta Geotech. 15 (11) (2020) 3135–3150, http://dx.doi.
org/10.1007/s11440-020-00962-4.
[1] F.W. Wang, Y.M. Zhang, Z.T. Huo, X.M. Peng, S.M. Wang, S. Yamasaki, [12] H. Wang, L. Zhang, K. Yin, H. Luo, J. Li, Landslide identification using
Mechanism for the rapid motion of the Qianjiangping landslide during machine learning, Geosci. Front. 12 (1) (2021) 351–364, http://dx.doi.org/
reactivation by the first impoundment of the Three Gorges dam reservoir, 10.1016/j.gsf.2020.02.012.
China. Landslides 5 (4) (2008) 379–386, http://dx.doi.org/10.1007/s10346- [13] Z.Z. Wang, C. Xiao, S.H. Goh, M.X. Deng, Metamodel-based reliability
008-0130-7. analysis in spatially variable soils using convolutional neural networks, J.

16
W. Zhang, X. Gu, L. Hong et al. Applied Soft Computing 136 (2023) 110066

Geotech. Geoenviron. Eng. 147 (3) (2021) 04021003, http://dx.doi.org/10. [36] K. Fukushima, S. Miyake, Neocognitron: A Self-Organizing Neural Network
1061/(asce)gt.1943-5606.0002486. Model for a Mechanism of Visual Pattern Recognition, Competition and
[14] Z.Z. Wang, S.H. Goh, Novel approach to efficient slope reliability analysis Cooperation in Neural Nets, Springer, 1982, pp. 267–285, http://dx.doi.org/
in spatially variable soils, Eng. Geol. 281 (1) (2021) 105989, http://dx.doi. 10.1007/978-3-642-46466-9_18.
org/10.1016/j.enggeo.2020.105989. [37] B. Li, K.C.P. Wang, A. Zhang, E. Yang, G. Wang, Automatic classification
[15] T.Z. Li, Q. Pan, D. Dias, Active learning relevant vector machine for of pavement crack using deep convolutional neural network, Int. J. Pave-
reliability analysis, Appl. Math. Model. 89 (2021) 381–399, http://dx.doi. ment Eng. 21 (2020) 457–463, http://dx.doi.org/10.1080/10298436.2018.
org/10.1016/j.apm.2020.07.034. 1485917.
[16] Q.J. Pan, Y.F. Leung, S.C. Hsu, Stochastic seismic slope stability assessment [38] J.H. Friedman, Multivariate adaptive regression splines, Ann. Statist. 19
using polynomial chaos expansions combined with relevance vector ma- (1991) 1–67, http://dx.doi.org/10.1214/aos/1176347963.
chine, Geosci. Front. 12 (1) (2021) 405–414, http://dx.doi.org/10.1016/j.gsf.
[39] W.G. Zhang, A.T.C. Goh, Multivariate adaptive regression splines for anal-
2020.03.016.
ysis of geotechnical engineering systems, Comput. Geotech. 48 (2013)
[17] S.E. Cho, Probabilistic stability analyses of slopes using the ANN-based
82–95, http://dx.doi.org/10.1016/j.compgeo.2012.09.016.
response surface, Comput. Geotech. 36 (2009) 787–797, http://dx.doi.org/
[40] W.G. Zhang, A.T.C. Goh, Multivariate adaptive regression splines and neural
10.1016/j.compgeo.2009.01.003.
network models for prediction of pile drivability, Geosci. Front. 7 (1) (2016)
[18] S. Li, H.B. Zhao, Z. Ru, Slope reliability analysis by updated support vector
45–52, http://dx.doi.org/10.1016/j.gsf.2014.10.003.
machine and Monte Carlo simulation, Nat. Hazards 65 (1) (2013) 707–722,
http://dx.doi.org/10.1007/s11069-012-0396-x. [41] G.B. Huang, Q.Y. Zhu, C.K. Siew, Extreme learning machine: A new learning
[19] P. Yi, K. Wei, X. Kong, Z. Zhu, Cumulative PSO-Kriging model for slope scheme of feedforward neural networks, in: 2004 IEEE International Joint
reliability analysis, Probab. Eng. Mech. 39 (2015) 39–45, http://dx.doi.org/ Conference on Neural Networks (IEEE Cat. No. 04CH37541), Ieee, 2004, pp.
10.1016/j.probengmech.2014.12.001. 985–990, http://dx.doi.org/10.1109/IJCNN.2004.1380068.
[20] M. Kumar, P. Samui, Reliability analysis of pile foundation using ELM and [42] T. Chen, C. Guestrin, Xgboost: A scalable tree boosting system, in: Pro-
MARS, Geotech. Geol. Eng. 37 (4) (2019) 3447–3457, http://dx.doi.org/10. ceedings of the 22nd Acm Sigkdd International Conference on Knowledge
1007/s10706-018-00777-x. Discovery and Data Mining, 2016, pp. 785–794, http://dx.doi.org/10.1145/
[21] H. Shen, J. Li, S. Wang, Z. Xie, Prediction of load–displacement performance 2939672.2939785.
of grouted anchors in weathered granites using FastICA-MARS as a novel [43] H. Mo, H. Sun, J. Liu, S. Wei, Developing window behavior models for
model, Geosci. Front. 12 (1) (2021) 415–423, http://dx.doi.org/10.1016/j. residential buildings using XGBoost algorithm, Energy Build. 205 (2019)
gsf.2020.05.004. 109564, http://dx.doi.org/10.1016/j.enbuild.2019.109564.
[22] M.X. Wang, D.R. Huang, G. Wang, D.Q. Li, SS-XGBoost: A machine learning [44] X. He, H. Xu, H. Sabetamal, D. Sheng, Machine learning aided stochastic
framework for predicting newmark sliding displacements of slopes, J. reliability analysis of spatially variable slopes, Comput. Geotech. 126 (2020)
Geotech. Geoenviron. Eng. 146 (9) (2020) 04020074, http://dx.doi.org/10. http://dx.doi.org/10.1016/j.compgeo.2020.103711.
1061/(asce)gt.1943-5606.0002297. [45] L. Song, X. Yu, B. Xu, R. Pang, Z. Zhang, 3D slope reliability analysis based
[23] W.G. Zhang, C.Z. Wu, L.B. Tang, X. Gu, L. Wang, Efficient time-variant on the intelligent response surface methodology, Bull. Eng. Geol. Environ.
reliability analysis of Bazimen landslide in the Three Gorges Reservoir 80 (2021) 735–749, http://dx.doi.org/10.1007/s10064-020-01940-6.
Area using XGBoost and LightGBM algorithms, Gondwana Res. (2022) [46] J. Zhang, H.W. Huang, K.K. Phoon, Application of the kriging-based re-
http://dx.doi.org/10.1016/j.gr.2022.10.004. sponse surface method to the system reliability of soil slopes, J. Geotech.
[24] Z.Z. Wang, S.H. Goh, W. Zhang, Reliability-based design in spatially variable Geoenviron. Eng. 139 (2013) 651–655, http://dx.doi.org/10.1061/(asce)gt.
soils using deep learning: An illustration using shallow foundation, Georisk 1943-5606.0000801.
1 (2022) 1–15, http://dx.doi.org/10.1080/17499518.2022.2083178.
[47] F. Kang, Q. Xu, J. Li, Slope reliability analysis using surrogate models via
[25] C. Wu, L. Hong, L. Wang, R. Zhang, S. Pijush, W. Zhang, Prediction of
new support vector machines with swarm intelligence, Appl. Math. Model.
wall deflection induced by braced excavation in spatially variable soils
40 (2016) 6105–6120, http://dx.doi.org/10.1016/j.apm.2016.01.050.
via convolutional neural network, Gondwana Res. (2022) http://dx.doi.org/
[48] J. Ji, C. Zhang, Y. Gui, Q. Lue, J. Kodikara, New observations on the
10.1016/j.gr.2022.06.011.
application of LS-SVM in slope system reliability analysis, J. Comput. Civ.
[26] T. Zhao, Y. Wang, Interpolation and stratification of multilayer soil property
Eng. 31 (2017) http://dx.doi.org/10.1061/(asce)cp.1943-5487.0000620.
profile from sparse measurements using machine learning methods, Eng.
Geol. 265 (2020) 105430, http://dx.doi.org/10.1016/j.enggeo.2019.105430. [49] I. Chivata Cardenas, On the use of Bayesian networks as a meta-modelling
[27] J. Li, P. Li, D. Guo, X. Li, Z. Chen, Advanced prediction of tunnel boring approach to analyse uncertainties in slope stability analysis, Georisk-
machine performance based on big data, Geosci. Front. 12 (1) (2021) Assess. Manag. Risk Eng. Syst. Geohazards 13 (2019) 53–65, http://dx.doi.
331–338, http://dx.doi.org/10.1016/j.gsf.2020.02.011. org/10.1080/17499518.2018.1498524.
[28] Z. Liu, G. Gilbert, J.M. Cepeda, A.O.K. Lysdahl, L. Piciullo, H. Hefre, S. Lacasse, [50] C. Qi, R. Xu, J. Wu, J. Yu, Reliability analysis of rock mass deformation in
Modelling of shallow landslides with machine learning algorithms, Geosci. tunnel excavation based on genetic algorithm, J. Eng. Geol. 16 (2) (2008)
Front. 12 (1) (2021) 385–393, http://dx.doi.org/10.1016/j.gsf.2020.04.014. 258–262, (In Chinese).
[29] P. Zhang, Z.Y. Yin, Y.F. Jin, T.H.T. Chan, F.P. Gao, Intelligent modelling [51] H. Zhao, Z. Ru, X. Chang, S.D. Yin, S.J. Li, Reliability analysis of tunnel
of clay compressibility using hybrid meta-heuristic and machine learning using least square support vector machine, Tunn. Undergr. Space Technol.
algorithms, Geosci. Front. 12 (1) (2021) 441–452, http://dx.doi.org/10.1016/ 41 (2014) 14–23, http://dx.doi.org/10.1016/j.tust.2013.11.004.
j.gsf.2020.02.014. [52] C. Bucher, T. Most, A comparison of approximate response functions in
[30] R.H. Zhang, C.Z. Wu, A.T.C. Goh, T. Bohlke, W.G. Zhang, Estimation of structural reliability analysis, Probab. Eng. Mech. 23 (2–3) (2008) 154–163,
diaphragm wall deflections for deep braced excavation in anisotropic http://dx.doi.org/10.1016/j.probengmech.2007.12.022.
clays using ensemble learning, Geosci. Front. 12 (1) (2021) 365–373, [53] A. Roy, S. Chakraborty, Support vector regression based metamodel by
http://dx.doi.org/10.1016/igsf.2020.03.003. sequential adaptive sampling for reliability analysis of structures, Reliab.
[31] W. Zhang, H. Li, Y. Li, H. Liu, Y. Chen, Application of deep learning Eng. Syst. Saf. 200 (2020) http://dx.doi.org/10.1016/j.ress.2020.106948.
algorithms in geotechnical engineering: A short critical review, Artif. [54] X. Li, X. Li, Y. Su, A hybrid approach combining uniform design and support
Intell. Rev. 54 (8) (2021) 5633–5673, http://dx.doi.org/10.1007/s10462- vector machine to probabilistic tunnel stability assessment, Struct. Saf. 61
021-09967-1. (2016) 22–42, http://dx.doi.org/10.1016/j.strusafe.2016.03.001.
[32] S. Zheng, Y.X. Zhu, D.Q. Li, Z.J. Cao, Q.X. Deng, K.K. Phoon, Probabilistic
[55] Y.C. Bai, X. Han, C. Jiang, J. Liu, Comparative study of metamodeling
outlier detection for sparse multivariate geotechnical site investigation
techniques for reliability analysis using evidence theory, Adv. Eng. Softw.
data using Bayesian learning, Geosci. Front. 12 (1) (2021) 425–439, http:
53 (2012) 61–71, http://dx.doi.org/10.1016/j.advengsoft.2012.07.007.
//dx.doi.org/10.1016/j.gsf.2020.03.017.
[33] J. Ching, I. Yoshida, K. Phoon, Comparison of trend models for geotech- [56] Q. Wang, H. Fang, L. Shen, Reliability analysis of tunnels using a metamod-
nical spatial variability: Sparse Bayesian learning vs. Gaussian process eling technique based on augmented radial basis functions, Tunn. Undergr.
regression, Gondwana Res. (2022) http://dx.doi.org/10.1016/j.gr.2022.07. Space Technol. 56 (2016) 45–53, http://dx.doi.org/10.1016/j.tust.2016.02.
011. 007.
[34] A. Karkevandi-Talkhooncheh, A. Rostami, A. Hemmati-Sarapardeh, M. Ah- [57] Q. Pan, D. Dias, An efficient reliability method combining adaptive support
madi, M.M. Husein, B. Dabir, Modeling minimum miscibility pressure vector machine and Monte Carlo simulation, Struct. Saf. 67 (2017) 85–95,
during pure and impure CO2 flooding using hybrid of radial basis function http://dx.doi.org/10.1016/j.strusafe.2017.04.006.
neural network and evolutionary techniques, Fuel 220 (2018) 270–282, [58] F. Cui, M. Ghosn, Implementation of machine learning techniques into the
http://dx.doi.org/10.1016/j.fuel.2018.01.101. subset simulation method, Struct. Saf. 79 (2019) 12–25, http://dx.doi.org/
[35] S.N. Qasem, I. Ebtehaj, H. Bonakdari, Potential of radial basis function 10.1016/j.strusafe.2019.02.002.
network with particle swarm optimization for prediction of sediment [59] P. Samui, T. Lansivaara, D. Kim, Utilization relevance vector machine for
transport at the limit of deposition in a clean pipe, Sustain. Water Resour. slope reliability analysis, Appl. Soft Comput. 11 (5) (2011) 4036–4040,
Manag. 3 (2017) 391–401, http://dx.doi.org/10.1007/s40899-017-0104-9. http://dx.doi.org/10.1016/j.asoc.2011.03.009.

17
W. Zhang, X. Gu, L. Hong et al. Applied Soft Computing 136 (2023) 110066

[60] A. Hamrouni, D. Dias, B. Sbartai, Probability analysis of shallow circular [73] C.H. Juang, Z. Luo, S. Atamturktur, H.W. Huang, Bayesian updating of soil
tunnels in homogeneous soil using the surface response methodology parameters for braced excavations using field observations, J. Geotech.
optimized by a genetic algorithm, Tunn. Undergr. Space Technol. 86 (2019) Geoenviron. Eng. 139 (3) (2013) 395–406, http://dx.doi.org/10.1061/(ASCE)
22–33, http://dx.doi.org/10.1016/j.tust.2019.01.008. GT.1943-5606.0000782.
[61] S. Zhou, X. Guo, Q. Zhang, D. Dias, Q.J. Pan, Influence of a weak layer on the [74] Z. Luo, B. Hu, Bayesian model and parameter calibration for braced
tunnel face stability - Reliability and sensitivity analysis, Comput. Geotech. excavations in soft clays, Mar. Georesour. Geotechnol. 38 (10) (2020)
122 (2020) http://dx.doi.org/10.1016/j.compgeo.2020.103507. 1235–1244, http://dx.doi.org/10.1080/1064119X.2019.1673855.
[62] N. Moreira, T. Miranda, M. Pinheiro, P. Fernandes, D. Dias, L. Costa, J. Sena-
[75] Y.F. Leung, W. Liu, Y. Lei, S.C. Hsu, Quantifying cost-effectiveness of
Cruz, Back analysis of geomechanical parameters in underground works
subsurface strata exploration in excavation projects through geostatistics
using an evolution strategy algorithm, Tunn. Undergr. Space Technol. 33
and spatial tessellation, Autom. Constr. 90 (2018) 243–252, http://dx.doi.
(2013) 143–158, http://dx.doi.org/10.1016/j.tust.2012.08.011.
org/10.1016/j.autcon.2018.02.032.
[63] D.Y. Zhuang, K. Ma, C.A. Tang, Z.Z. Liang, K.K. Wang, Z.W. Wang, Mechanical
parameter inversion in tunnel engineering using support vector regression [76] Y. Wang, T.Y. Zhao, K.K. Phoon, Direct simulation of random field samples
optimized by multi-strategy artificial fish swarm algorithm, Tunn. Undergr. from sparsely measured geotechnical data with consideration of uncer-
Space Technol. 83 (2019) 425–436, http://dx.doi.org/10.1016/j.tust.2018.09. tainty in interpretation, Can. Geotech. J. 55 (6) (2018) 862–880, http:
027. //dx.doi.org/10.1139/cgj-2017-0254.
[64] M. Zhou, M. Shadabfar, Y. Xue, Y. Zhang, H. Huang, Probabilistic analysis [77] Y. Hu, T.Y. Zhao, Y. Wang, C. Choi, C.W.W. Ng, Direct simulation of
of tunnel roof deflection under sequential excavation using ANN-based two-dimensional isotropic or anisotropic random field from sparse mea-
Monte Carlo simulation and simplified reliability approach, ASCE-ASME J. surement using Bayesian compressive sampling, Stoch. Environ. Res. Risk
Risk Uncertain. Eng. Syst. A 7 (4) (2021) 1–14, http://dx.doi.org/10.1061/ Assess. 33 (8–9) (2019) 1477–1496, http://dx.doi.org/10.1007/s00477-019-
ajrua6.0001170. 01718-7.
[65] A.K. Verma, A. Pain, E. Agarwal, D. Pradhan, Reliability assessment of [78] Y. Wang, T.Y. Zhao, Y. Hu, K.K. Phoon, Simulation of random fields with
tunnels using machine learning algorithms, Indian Geotech. J. 52 (2022) trend from sparse measurements without detrending, J. Eng. Mech. 145
780–798, http://dx.doi.org/10.1007/s40098-022-00610-6. (2) (2019) 1–12, http://dx.doi.org/10.1061/(ASCE)EM.1943-7889.0001560.
[66] J. Zhang, K.K. Phoon, D. Zhang, H. Huang, C. Tang, Deep learning-based [79] J.Z. Zhang, K.K. Phoon, D.M. Zhang, H.W. Huang, C. Tang, Novel approach to
evaluation of factor of safety with confidence interval for tunnel defor-
estimate vertical scale of fluctuation based on CPT data using convolutional
mation in spatially variable soil, J. Rock Mech. Geotech. Eng. 13 (2021)
neural networks, Eng. Geol. 294 (2021) 106342, http://dx.doi.org/10.1016/
1358–1367, http://dx.doi.org/10.1016/j.jrmge.2021.09.001.
j.enggeo.2021.106342.
[67] Y. Zhou, S. Li, C. Zhou, H. Luo, Intelligent approach based on random forest
for safety risk prediction of deep foundation pit in subway stations, J. [80] J.Z. Zhang, D.M. Zhang, H.W. Huang, K.K. Phoon, C. Tang, G. Li, Hybrid
Comput. Civ. Eng. 33 (1) (2019) http://dx.doi.org/10.1061/(asce)cp.1943- machine learning model with random field and limited CPT data to
5487.0000796. quantify horizontal scale of fluctuation of soil spatial variability, Acta
[68] C. Zhang, J.Z. Li, Y. He, Application of optimized grey discrete Verhulst-BP Geotech. (2021) http://dx.doi.org/10.1007/s11440-021-01360-0.
neural network model in settlement prediction of foundation pit, Environ. [81] D. Huang, D.M. Gu, Y.X. Song, D.F. Cen, B. Zeng, Towards a complete
Earth Sci. 78 (15) (2019) 15, http://dx.doi.org/10.1007/s12665-019-8458-y. understanding of the triggering mechanism of a large reactivated landslide
[69] W. Zhang, R. Zhang, C. Wu, A.T.C. Goh, S. Lacasse, Z. Liu, H. Liu, State-of- in the Three Gorges reservoir, Eng. Geol. 238 (2018) 36–51, http://dx.doi.
the-art review of soft computing applications in underground excavations, org/10.1016/j.enggeo.2018.03.008.
Geosci. Front. 11 (4) (2020) 1095–1106, http://dx.doi.org/10.1016/j.gsf. [82] W. Zhang, L. Tang, H. Li, L. Wang, L. Cheng, T. Zhou, X. Chen, Probabilistic
2019.12.003. stability analysis of Bazimen landslide with monitored rainfall data and
[70] F.K. Huang, G.S. Wang, ANN-based reliability analysis for deep excavation, water level fluctuations in Three Gorges reservoir, China, Front. Struct.
in: EUROCON 2007 - The International Conference on Computer as a Tool,
Civ. Eng. (2020) http://dx.doi.org/10.1007/s11709-020-0655-y.
2007, pp. 2039–2046, http://dx.doi.org/10.1109/EURCON.2007.4400328.
[83] C. Wang, Reliability-based design of lining structures for underground
[71] J. Cao, C.N. Sun, H.M. Liu, Application of SVM to reliability analysis of
space against water seepage, Undergr. Space (2020) http://dx.doi.org/10.
foundation excavations system, Adv. Mater. Res. 859 (2014) 315–321,
1016/j.undsp.2020.03.004.
http://dx.doi.org/10.4028/www.scientific.net/AMR.859.315.
[72] L.X. He, Y. Liu, S.F. Bi, L. Wang, M. Broggi, M. Beer, Estimation of failure [84] W.G. Zhang, C.Z. Wu, H.Y. Zhong, Y.Q. Li, L. Wang, Prediction of undrained
probability in braced excavation using Bayesian networks with integrated shear strength using extreme gradient boosting and random forest based
model updating, Undergr. Space (2020) http://dx.doi.org/10.1016/j.undsp. on Bayesian optimization, Geosci. Front. (2021) http://dx.doi.org/10.1016/
2019.07.001. j.gsf.2020.03.007.

18