1 s2.0 S0886779821001371 Main

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 23

Tunnelling and Underground Space Technology 113 (2021) 103946

Contents lists available at ScienceDirect

Tunnelling and Underground Space Technology


incorporating Trenchless Technology Research
journal homepage: www.elsevier.com/locate/tust

State-of-the-art review of geotechnical-driven artificial intelligence


techniques in underground soil-structure interaction
S.C. Jong a, D.E.L. Ong a, *, E. Oh b
a
Griffith University, School of Engineering and Built Environment, 170 Kessels Road, Nathan, Queensland 4111, Australia
b
Griffith University, School of Engineering and Built Environment, 1 Parklands Drive, Gold Coast Campus, Southport, Queensland 4215, Australia

A R T I C L E I N F O A B S T R A C T

Keywords: There has been an increasing demand for underground construction due to urbanization and limited land in
Artificial intelligence metropolitan cities in the recent years. However, the behavior of underground structures in soils and rocks is
Machine learning often not completely understood. The emergence of Artificial Intelligence (AI) techniques is envisaged to have a
Underground space
huge potential in addressing geotechnical problems that involve complex soil-structure interaction. This paper
Soil-structure interaction
thus aims at reviewing the applications of AI techniques in studying underground soil-structure interaction,
which focuses on aspects such as characterization of soils and rocks, pile foundations, deep excavations and
tunneling. An overview of different AI techniques is provided and a list of key AI applications in underground
works that have been published in the last ten years is also compiled to study the recent trend of machine
learning techniques in underground construction. The capabilities and limitations of these techniques are dis­
cussed throughout the paper, to help readers understand various techniques that are suitable for different un­
derground geotechnical applications. Lastly, some of the challenges that may be faced when applying the
techniques are identified, and recent development of AI in geotechnical engineering is discussed in which
possible countermeasures to overcome these limitations are highlighted.

1. Introduction process.
Recently, urbanization and limited land for infrastructure develop­
The complex behavior of soil and rock materials often limits our ment have led to an increasing demand for the construction of under­
understanding on their applications in engineering projects. Due to the ground infrastructure such as tunnels for mass rapid transit services,
complexity of their behavior, engineering design is usually carried out deep excavations and installation of pile foundations for high-rise
based on fundamental principles that are simplified with certain as­ buildings. The behavior of underground structures in soils and rocks is
sumptions. Sometimes, these simplifications may lead to underestima­ complex, highly non-linear, uncertain and not yet completely under­
tion or overestimation of ground behavior in the design works. The stood (W. Zhang et al., 2020c). With the emergence and development of
increasing use of Information Technology (IT) in data collection and AI techniques in geotechnical engineering, it is believed that these
data analysis has become a recent trend in construction industry. This techniques can be used to further understand the complex behavior of
results in the emergence of Artificial Intelligence (AI) techniques to underground construction. Therefore, this paper aims at reviewing the
solve complex engineering problems. This is achieved by learning pat­ applications of AI techniques related to soil-structure interaction in
terns of data inputs and outputs presented to the models to produce underground structures, with specific focus on the characterization of
meaningful interpretation. The models developed from such techniques soil and rock properties, pile foundations, underground space con­
are data-driven, as opposed to traditional approaches that adopt engi­ struction and tunneling works. An overview of various techniques that
neering principles to define functional relationships, which are usually have been successfully used in geotechnical applications is provided.
simplified with several assumptions (Shahin, 2016). The data obtained The applications of these techniques in underground soil-structure
are trained using selected AI technique to develop a robust design model interaction are then further discussed. The scope of this paper is sum­
that can analyze complex relationships between various input parame­ marized in Fig. 1 to provide readers with an overview for easy under­
ters and produce meaningful output to improve the engineering design standing. A list of AI applications in underground geotechnical works for

* Corresponding author at: Griffith University, School of Engineering and Built Environment, 170 Kessels Road, Nathan, Queensland 4111, Australia.
E-mail addresses: siawchian.jong@griffithuni.edu.au (S.C. Jong), d.ong@griffith.edu.au (D.E.L. Ong), y.oh@griffith.edu.au (E. Oh).

https://doi.org/10.1016/j.tust.2021.103946
Received 28 September 2020; Received in revised form 18 February 2021; Accepted 27 March 2021
Available online 16 April 2021
0886-7798/© 2021 Elsevier Ltd. All rights reserved.
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

the past ten years is also compiled to provide information on the recent values from the data to inform the accuracy of the developed model in
trend of AI applications in underground construction. Throughout the inferring the parameters of interest. The model output is compared with
paper, the capabilities and limitations of these techniques are discussed, the expected output by computing the error and adjusting it to obtain a
to help readers in selecting suitable methods for specific applications in mapping with the smallest error. On the other hand, in unsupervised
underground works. Lastly, some of the challenges that may be faced in learning, the data are usually divided into different clusters in which
the applications of AI techniques are reviewed and the recent develop­ “interesting patterns” are learned from the input data provided. The
ment of AI in geotechnical engineering is discussed to highlight the model predictions cannot be checked against the observed values
possible countermeasures for tackling some of the shortcomings of because they are not provided to the model. Since the aim of learning is
existing techniques. to discover meaningful patterns in the input data without corresponding
output, comparison between model output and expected output cannot
2. Overview of artificial intelligence techniques in geotechnical be carried out, hence, the problem is defined as “unsupervised”. In the
engineering field of geotechnical engineering, supervised learning is usually adopted
to solve the problems in hand.
In general, AI is a broad concept with the goal of developing human- In this section, the commonly used AI techniques in geotechnical
like intelligence in machines that can be achieved by using learning engineering are briefly discussed to provide a general understanding on
algorithms to mimic how a human brain learns. Machine learning is a their characteristics. These techniques include Artificial Neural Net­
branch of AI which allows machines to develop human-like intelligence works (ANNs), Genetic Programming (GP), Evolutionary Polynomial
without the need of explicit programming (Das et al., 2015). The terms Regression (EPR), Multivariate Adaptive Regression Splines (MARS),
“AI” and “machine learning” are sometimes used interchangeably. More Particle Swarm Optimization (PSO), Adaptive Neuro-Fuzzy Inference
specifically, it can be understood that AI techniques are machine System (ANFIS), Bayesian method, Support Vector Machines (SVM),
learning methods that can detect patterns in data automatically and Decision Tree (DT), Random Forest (RF), and Logistic Regression (LR).
make use of the discovered patterns to improve the prediction of future
trends or to facilitate better decision-making process. 2.1. Artificial Neural Networks (ANNs)
Machine learning can generally be divided into: (i) supervised
learning; and (ii) unsupervised learning. Supervised learning is also The concept of ANNs was first introduced by McCulloch and Pitts
known as predictive learning, with the goal of producing a mapping (1943). Later, Rumelhart et al. (1987) introduced the error propagation
from inputs to outputs using a set of training data whereas unsupervised training scheme. ANNs can compute a mapping from a multivariate
learning aims at exploring meaningful patterns in the data, which can be space of information to another, given a set of data (Garrett, 1994).
known as knowledge discovery (Murphy, 2012). The main difference Therefore, ANN models can process imperfect or incomplete data and
between supervised and unsupervised learning is that in supervised capture non-linear relationships among the variables in the system
learning, the model predictions can be compared with the observed (Javadi and Rezania, 2009). An ANN model can be considered as a

Fig. 1. Scope of the review of AI applications in underground soil-structure interaction.

2
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

“black box”, in which data is provided to the model to be processed and f(.) = the function defined by the user;
then produces an output that maps the input. Hence, the model is only aj = the constant;
able to describe the end-to-end relationship but does not provide any a0 = the intercept.
physical deduction process while analyzing the data (Jin et al., 2018). The detailed steps of developing an EPR model are provided in
Typically, the structure of ANNs consists of a series of nodes, that are Giustolisi and Savic (2006).
arranged in three main layers: an input layer, one or more hidden layers,
and an output layer, as shown in Fig. 2. The details of development of an 2.4. Multivariate Adaptive Regression Splines (MARS)
ANN model can be found in Shahin et al. (2008).
MARS was introduced by Friedman (1991) for flexible regression
2.2. Genetic Programming (GP) modeling of high dimensional data. It is a data-driven technique that
does not require any assumption about the underlying relationship be­
GP is an evolutionary computing method that generates a structured tween the input variables and output (Zhang and Goh, 2013). MARS
representation of data provided. GP was first introduced by Cramer models comprise of an expansion in product spline basis functions,
(1985) and further developed by Koza (1992). It is an extension of Ge­ where the number of basis functions and parameters associated with the
netic Algorithm (GA), which are search techniques based on natural functions are determined by the data (Friedman, 1991). The connection
evolutionary mechanisms such as selection, reproduction, cross-over points between the segments are known as knots, which mark the end of
and mutation, to solve function identification problems. The develop­ one region of the data and the beginning of another (Zhang and Goh,
ment of a typical GP model involves the creation of an initial population 2013). The splines are connected smoothly together to produce a flexible
of computer models that consist of a set of functions and a set of ter­ model that can capture both linear and non-linear behavior (W. Zhang
minals as defined by the user, which are arranged in a tree-like structure et al., 2020c). A MARS model can be generally expressed as Eq. (2)
containing a root node, branches of functional nodes, and terminals (Samui and Kim, 2013):
(Fatehnia and Amirinia, 2018). Further details on how to develop a GP

M
model can be found in Koza (1992). Y = θ0 + θm ψ m (X) + ε (2)
m=1

2.3. Evolutionary Polynomial Regression (EPR)


where Y = the model output;
EPR is a hybrid regression method that was developed by Giustolisi θ0 = the constant of the model;
and Savic (2006) to combine the best features of numerical regression θm = the unknown coefficient for the mth basis function m = (1,2,⋯,
with GP symbolic regression technique. There are two stages for EPR M);
model development: (i) structure identification, and (ii) parameter ψ m (X) = the basis function for m = (1, 2, ⋯, M);
estimation. EPR searches for symbolic structures by conducting an ε = the error distribution.
evolutionary search using GA in the first stage and then estimates the The details of modeling process of a MARS model are described in
coefficients of selected polynomial terms by conducting numerical Friedman (1991).
regression using least squares method in the second stage. The general
expression of an EPR model is defined as Eq. (1) (Giustolisi and Savic, 2.5. Particle Swarm Optimization (PSO)
2006):
PSO is a global optimization technique that was first proposed by

m
( ) Kennedy and Eberhart (1995). The algorithm is inspired by the social
y= F X, f (X), aj + a0 (1)
j=1 behavior displayed by organisms that move in swarm. The population is
known as the swarm whereas the individuals within the population are
where y = the estimated vector of output of the process; known as the particles (Himanshu and Burman, 2019). A PSO model is
m = the number of terms of the target expression; developed by placing the particles in the search space. Each particle
F(.) = the function constructed by the process; represents a potential solution and predicts the objective function at its
X = the matrix of input variables; current position. The particles move to the next position by combining

Fig. 2. Typical structure of ANNs.

3
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

their current and best position at each iteration. Eventually, the swarm The detailed description of a Bayesian model is provided in Houlsby
will move to the best location with the optimum fitness function after and Houlsby (2013).
several iterations (Gordan et al., 2016). The steps of building a PSO
model can be found in Himanshu and Burman (2019). 2.8. Support Vector Machines (SVM)

2.6. Adaptive Neuro-Fuzzy Inference System (ANFIS) SVM was first proposed by Vapnik (1995) for pattern classification
and later developed for regression by Smola and Schölkopf (1998). SVM
ANFIS is a hybrid AI technique that was developed by Jang (1993) is a non-parametric technique based on statistical learning theory that
which integrates the linguistic representation of a fuzzy inference sys­ utilizes kernels to solve non-linear high-dimensional problems by con­
tem (FIS) with the learning ability of ANNs (Cabalar et al., 2012). FIS is a trolling model complexity (Micheletti et al., 2014). The aim of SVM is to
system that can consider expert knowledge in designing a process with find an optimal separation hyperplane that maximizes the isolation
input and output fuzzy sets that are controlled by if-then rules (Ghor­ margin between two types of samples with minimal error (Ren et al.,
banzadeh et al., 2018). By combining fuzzy logic and ANNs in function 2019). The linear classifier that maximizes the distance between the
approximation, ANFIS can produce a mapping between input and output nearest data point of each class is known as the optimal separating hy­
parameters through hybrid learning (Armaghani et al., 2015). The perplane whereas the nearest data points that define the margin are
development process of an ANFIS model can be found in Jang (1993). known as support vectors (see Fig. 3). Further details on the modelling
steps of a SVM model can be found in Smola and Schölkopf (1998).

2.7. Bayesian method 2.9. Decision Tree (DT)

Bayesian method is a machine learning technique that is developed DT is a machine learning technique that can be used to solve both
based on Bayes’ theorem proposed by Thomas Bayes (Bayes and Price, classification and regression problems by segmenting the predictor
1763). Bayes’ theorem forms the foundation of Bayesian inference, space into a number of regions, resulting in a tree-like structure that is
which allows uncertainty of parameters to be quantified using proba­ developed based on a set of splitting rules (James et al., 2013). The tree
bility (Murphy, 2012). In Bayesian perspective, probability is inter­ consists of a root node (containing all data), a set of internal nodes
preted as the degree of belief. Hence, it can be considered as the (splits) and a set of terminal nodes (leaves). Each path from root node to
evolution of the degree of belief in the parameters, before observing the the leave node implies a decision rule that represents the relationship
data, and after considering the data in the analysis (Jin et al., 2018). The between the input and output variables clearly, allowing easy inter­
probability distribution of parameters of interest, θ given the observed pretation to the solution provided (Salimi et al., 2018a). In DT, features
data, D can be described as p(θ|D), which is known as posterior distri­ of data are made up of predictor variables whereas the class to be
bution of the parameters. It can be calculated using Bayes’ theorem mapped is known as target variable (Xu et al., 2005). For continuous or
expressed in Eq. (3): discrete variable, the resulting tree is known as regression tree whereas
p(θ)p(D|θ) for categorical variable, the resulting tree is known as classification tree.
p(θ|D) = (3) The detailed description of a DT model is provided in Xu et al. (2005).
p(D)

where p(θ|D) = the posterior distribution of parameters; 2.10. Random Forest (RF)
p(θ) = the prior distribution of parameters;
p(D|θ) = the likelihood function; RF is an ensemble machine learning technique developed by Breiman
p(D) = the model evidence. (2001). Instead of building individual DTs, RF employs an ensemble of

Fig. 3. Support vectors with optimal separating hyperplane and maximum margin.

4
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

randomly trained DTs that can improve the prediction accuracy by Significant findings from some of the literature that are reviewed are
combining a large set of DTs (Behnia and Blais-Stevens, 2018). RF can be further highlighted and discussed to assess the performance/accuracy of
applied to both regression and classification problems. Fig. 4 shows the the techniques in tackling different underground geotechnical problems.
typical development process of a RF model. Bagging, also known as This will provide readers with an indirect measure of the confidence
bootstrap aggregating, is integrated with DT when building a RF model level or reliability of the applied methods.
(P. Zhang et al., 2019). First, n bootstrap sets are created by random From the literature compiled by the authors for AI applications in
sampling with replacement N training samples from the training dataset. underground works for the past ten years (see Table 1), it is found that
Then, a DT is built using a bootstrap sample. Finally, the result is ob­ most of the techniques discussed in Section 2 have been applied to
tained by aggregating the outputs from all the trees (Zhou et al., 2017). different parts of underground works (see Fig. 5). All the techniques
The details of building a RF model can be found in Breiman (2001). have been applied to study soil and rock properties as well as pile
foundations, except for LR in which no application has been found to
date. In deep excavations and tunneling, a wide range of techniques
2.11. Logistic Regression (LR)
have also been used, except for EPR. In addition, Fig. 6 shows the total
citations accumulated for different techniques based on the compiled
LR is a multivariate statistical learning technique that is developed
literature (until time of writing). It is evident that ANNs and Bayesian
for dealing with categorical response variable that represents a binary
method are the most widely used techniques in underground construc­
event rather than a continuous parameter (Zhang and Goh, 2013). It is
tion, followed by other techniques such as MARS, GP and SVM,
used to model the relationship between a dichotomous dependent var­
respectively. When comparing the total citations obtained between year
iable and a set of independent variables that can contain both contin­
2011 to 2015 and year 2016 to date (see Figs. 7 and 8, respectively), the
uous and categorical variables (Xu et al., 2013). The result of a LR model
results show that applications of ANNs and Bayesian approach remain
is interpreted as the probability of an event, with the model developed
highly cited, although ‘decline’ is observed. GP and MARS are still
based on logistic function expressed as Eq. (4) (Bui et al., 2011):
moderately popular among the techniques available while decline in
1 citations based on EPR, PSO and SVM is observed in the last five years.
P(event) = (4)
1 + e− z One possible reason for the decline in citations is the time of publication.
For literature that is published within the last five years, it is considered
where P(event) = the probability of an event occurring; z = the linear ‘very recent’ as compared to the work published before year 2016, and
sum of a constant and the product of the independent variables and their therefore accumulates relatively lesser citations. On the other hand,
respective coefficients. citations based on DT, RF and LR have risen within the last five years,
In LR, the model coefficients are estimated using maximum likeli­ showing that there are increasing applications based on these methods
hood method, in which the coefficients that make the observed results in underground construction.
“most likely” are selected for the model (Shirzadi et al., 2012). The In summary, all the AI techniques are generally robust and can be
details of developing a LR model are further described in James et al. readily applied to a wide range of geotechnical problems efficiently
(2013). assuming that the researchers have prior programming knowledge and
skills. The recent trend of AI applications in underground construction is
3. Applications in underground soil-structure interaction revealed from the comparison analysis. Particularly, ANNs and Bayesian
method are popular in underground geotechnical studies. Methods such
In this section, the successful AI applications in studying under­ as DT, RF and LR have also become increasingly popular in this field
ground soil-structure interaction are reviewed and discussed. The ap­ within the last five years.
plications cover various aspects, including characterization of soils and
rocks, pile foundations, deep excavations and tunneling works.

Fig. 4. Typical development process of a RF model.

5
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

Table 1
Applications of AI techniques in underground soil-structure interaction for the past ten years.
Author(s) AI Techniques Description of Applications

Characterization of Soil/Rock Properties


Jalalifar et al. (2011) ANFIS Prediction of rock mass rating
Armaghani et al. (2015) ANFIS, MLR, ANNs Prediction of unconfined compressive strength (UCS) and Young’s modulus (E) of granite
Shahri (2016) ANNs Prediction of clay sensitivity parameter
Debnath and Dey (2017) ANNs Prediction of peak shear stress along the cohesive soil-geosynthetic interface
Armaghani et al. (2016) ANNs, ANFIS, NLMR Prediction of UCS of granitic rock
Bejarbaneh et al. (2018) ANNs, FIS Prediction of E of sandstone
Khandelwal et al. (2018) ANNs, GA-ANNs, MLR Prediction of cohesion of rock
Khuntia et al. (2015) ANNs, LSSVM, MARS Prediction of compaction parameters of sandy soils
Houlsby and Houlsby (2013) Bayesian Fitting of design profile to undrained strength of soils
Wang and Cao (2013) Bayesian Characterization of undrained E of clay
Ching and Phoon (2014) Bayesian Characterization of clay parameters
Liu et al. (2016) Bayesian Analysis of multivariate correlation between resilient modulus and CPTu indices of clayey soils
Wang et al. (2016) Bayesian Characterization of site-specific soil parameters
Ching et al. (2017) Bayesian Characterization of parameters of coarse-grained soils
Wang and Zhao (2017) Bayesian Characterization of soil properties and evaluation of site-specific performance of design charts
Gong et al. (2017) Bayesian Optimization of site investigation
Yang et al. (2019) Bayesian Prediction of E of granite
Hu et al. (2020) Bayesian Interpolation of site-specific geotechnical data from sparse measurements
L. Wang et al. (2018) Bayesian Determination of site-specific soil–water characteristic curve
X. Wang et al. (2018) Bayesian Prediction of sub-surface soil/rock stratification and the pertinent soil properties
Ching and Phoon (2019) Bayesian Construction of site-specific multivariate probability distribution models
Feng (2015) Bayesian Prediction of E of rock, deformation modulus of rock mass, and tunnel squeezing
Wang et al. (2015) Bayesian Modeling of site-specific variability and prediction of geotechnical parameters
Wang and Aladejare (2016) Bayesian Characterization of UCS and E of rock
Wang and Akeju (2016) Bayesian Characterization of effective cohesion and effective friction angle of soils
Bozorgzadeh et al. (2019) Bayesian Prediction of rock strength parameters
Jin et al. (2019) Bayesian Development of selection method for advanced soil models accounting for soil uncertainty
Hu and Wang (2020) Bayesian Consideration of uncertainty in CPT-based subsurface soil classification and stratification
Shi and Wang (2021a) Bayesian Interpolation of subsurface soil stratigraphy from sparse measurements
Wang et al. (2020) BCS Soil classification and zonation in a 2D vertical cross-section using CPT
Y. Wang et al. (2018) BCS-KL Development of random field samples from sparse measurements
Hu et al. (2019) BCS-KL Development of 2D isotropic or anisotropic random field samples from sparse measurements
Zhao et al. (2019) BNGR Evaluation of physical and chemical properties of natural fibres in soil reinforcement
Cui and Jing (2019) BPNN Prediction of geotechnical parameters
Kurnaz and Kaya (2018) BRNN, SVM, ELM Prediction of compression index of clay
Shi and Wang (2021b) Ensemble RBFNN Incorporation of geotechnical anisotropy and quantification of uncertainty in spatial
interpolation
Faramarzi et al. (2012) EPR Prediction of volume change behavior of soils
Javadi et al. (2012) EPR Modeling of behavior of soils under cyclic loading
Ahangar-Asr et al. (2015) EPR Modeling of stress–strain and volume change behavior of unsaturated soils
Ghaderi et al. (2019) GFNN Prediction of soil types
Mola-Abasi et al. (2013) GMDH-Type NN-GA Prediction of shear wave velocity of soils
Asadi et al. (2011) GP Prediction of intact rock strength
Shen and Jimenez (2018) GP Prediction of Mohr-Coulomb shear strength parameters of intact sandstone
Z. L. Cheng et al. (2020a) GP Prediction of variation in soil suction
Z. L. Cheng et al. (2020b) GP Prediction of variation in soil suction
Khandelwal et al. (2017) GP, NLMR Prediction of brittleness of intact rock
Ren et al. (2019) KNN, NB, RF, ANNs, SVM Prediction of rock compressive strength
Li and Tan (2017) LSSVM Prediction of UCS and shear strength of rock
Çelik (2019) LSSVM, MLR Prediction of UCS of carbonate rock
Ru et al. (2019) LSSVM-MCS Prediction of drilling rate index of rock
Samui (2013) MARS Prediction of E of jointed rock mass
Samui and Kurup (2012) MARS, LSSVM Prediction of over-consolidation ratio of clay
Soleimani et al. (2018) MGGP Prediction of collapse settlement of sandy gravel soils
Mohammadzadeh et al. (2016) MGGP Prediction of compression index of fine-grained soils
Ghasemi et al. (2018) Model Trees Prediction of UCS and E of carbonate rock
Hussain et al. (2019) PSO, ICA Prediction of rock brittleness index
Mohamad et al. (2015) PSO-ANNs Prediction of UCS of rock
Liu et al. (2015) RBFNN, BPNN, GRNN, SVM, ELM Prediction of UCS of rock
Yang and Li (2019) SA-DE Back-analysis of rock creep parameters
Garg et al. (2014) SVR, ANNs, MGGP Prediction of stress-dependent soil water retention curves
Pal and Deswal (2014) SVR, ELM Prediction of resilient modulus of cohesive soils
Fattahi (2017) SVR-ABC, ANFIS-SCM Prediction of UCS of rock from Schmidt hammer rebound values

Pile Foundations
Ghorbani et al. (2018) ANFIS Prediction of ultimate axial load bearing capacity of piles
Harandizadeh et al. (2019) ANFIS-GMDH-GSA, FP-GMDH Prediction of ultimate bearing capacity of piles
Alkroosh and Nikraz (2011a) ANNs Simulation of load-settlement behavior of pile foundations
Tarawneh (2013) ANNs Prediction of pipe pile setup
Ghorbani and Niavol (2017) ANNs, EPR Prediction of settlement of coupled piled raft foundations
Saeedi Azizkandi et al. (2014) ANNs, GP Prediction of uplift pile displacement
Tarawneh and Imam (2014) ANNs, MLR Prediction of pile setup for pipe piles, concrete piles and H-piles
Pham et al. (2020) ANNs, RF Prediction of ultimate axial bearing capacity of driven piles
J. Zhang et al. (2020) BN Calibration of resistance factor for the design of pile foundations
(continued on next page)

6
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

Table 1 (continued )
Author(s) AI Techniques Description of Applications

Ismail and Jeng (2011) BPNN, RBFNN, GRNN, HON Simulation of pile load-settlement curve
Ebrahimian and Movahed (2017) EPR Prediction of ultimate axial bearing capacity of pile foundations
Ebrahimian et al. (2015) EPR Prediction of ε50 and lateral behavior of pile foundations
Alkroosh and Nikraz (2011b) GEP Prediction of the correlation of CPT data with axial pile capacity
Alkroosh and Nikraz (2012) GEP Prediction of axial capacity of pile foundations driven into cohesive soils
Alkroosh and Nikraz (2014) GEP Prediction of pile dynamic capacity
Armaghani et al. (2018) GEP, MLR Prediction of settlement of rock-socketed piles
Muduli et al. (2015) GP, MARS, BRNN, DENN Prediction of lateral load capacity of piles in clay
Samui and Kim (2013) LSSVM, MARS Prediction of lateral load capacity of piles
Singh et al. (2019) M5 Model Tree, RF, ANNs Prediction of oblique load carrying capacity of batter pile groups
Samui (2012) MARS Prediction of ultimate capacity of driven piles in cohesionless soils
Zhang and Goh (2016) MARS, BPNN Prediction of pile drivability
Das and Suman (2015) MARS, FN Prediction of lateral load capacity of piles in clay
Mohanty et al. (2018) MOGP, MARS, FN Prediction of capacity of driven piles in cohesionless soils
Kardani et al. (2020) PSO-(DT, KNN, MLPNN, RF, SVR, XGBoost) Prediction of bearing capacity of piles in cohesionless soils
Armaghani et al. (2020) PSO-ANNs Estimation of pile settlement
Ismail et al. (2013) PUNN-PSO-BP, BP, PSO Prediction of load-deformation behavior of axially loaded piles
Shahin (2014a) RNN Simulation of load-settlement behavior of drilled shafts
Shahin (2014b) RNN Simulation of load-settlement behavior of steel driven piles
Moayedi and Hayati (2019) SVM, ReGEPSVM, TWSVM, ANFIS, GP Prediction of friction capacity of driven piles installed in clay
Singh et al. (2018) SVR, Gaussian process Prediction of oblique load capacity of batter pile groups

Deep Excavations / Tunneling


Mottahedi et al. (2018) ANFIS, ANFIS-PSO Prediction of overbreak in underground excavations due to blasting
Elbaz et al. (2019) ANFIS-GA, ANFIS Prediction of earth pressure balance (EBP) shield performance during tunneling
Gordan et al. (2019) ANNs, ABC Prediction of factor of safety of retaining wall under both static and dynamic conditions
Wang et al. (2012) Bayesian Prediction of excavation-induced maximum ground settlement
Juang et al. (2013) Bayesian Back-analysis of soil parameters in braced excavations
Qi and Zhou (2017) Bayesian Updating of soil parameters for braced excavations
Jin et al. (2018) Bayesian Back-analysis of most probable soil parameters in braced excavations
Li et al. (2020) Bayesian Back-analysis of uncertain geomechanical parameters of shield tunnels probabilistically
Miro et al. (2015) Bayesian Evaluation of subsoil parameter uncertainties on tunnel-induced ground movements
Zhou et al. (2013) Bayesian Prediction of maximum pullout shear stress of grouted soil nails
Ghasemi and Gholizadeh (2019a) BLR, LDA Prediction of tunnel squeezing
Chung et al. (2019) BN Identification of potential risk events that may occur during tunnel construction
Gao et al. (2013) BPNN Prediction of deformation of soil nailing in deep excavations
Salimi et al. (2018a) CART, GP, PCA, NLMR Analysis of tunnel boring machine (TBM) performance
Salimi et al. (2018b) CART Prediction of specific rock mass boreability index
Faradonbeh and Taheri (2019) GA-ENN, GEP, C4.5 DT Prediction of occurrence of rock burst
Shi et al. (2012) GA-SVM, BPNN Prediction of deformation in deep excavations
Naghadehi et al. (2018) GEP Correlation of hard rock TBM performance with routine ground properties
Koopialipoor et al. (2019) GMDH Prediction of TBM penetration rate
Ghasemi and Gholizadeh (2019b) KNN, C5.0 DT Prediction of tunnel squeezing
Ozdemir (2016) LR Prediction of sinkhole susceptibility
Li and Jimenez (2018) LR Prediction of rock burst hazard
Zhang and Goh (2017) MARS Assessment of ultimate limit state of twin caverns
Goh et al. (2018) MARS Prediction of maximum surface settlement due to tunneling
Zhang et al. (2018) MARS Prediction of wall deflection profile caused by deep braced excavations in clay
Zheng et al. (2019) MARS Prediction of uplift displacement of underground structures
W. Zhang et al. (2019) MARS Prediction of wall deflection profile caused by deep braced excavations in clay
Adoko et al. (2013) MARS, ANNs Prediction of diameter convergence of high-speed railway tunnel in weak rock
W. Zhang et al. (2020b) MARS, DT Prediction of maximum bending moment within tunnel’s linings
Goh et al. (2017) MARS-LR Prediction of stability of underground entry-type excavations
Xue and Xiao (2017) PSO-LSSVM Evaluation of rock deformation of underground caverns
P. Zhang et al. (2019) PSO-RF Prediction of tunneling-induced settlement and EPB shield operational parameters
Zhou et al. (2017) RF Prediction of ground settlement above tunnels
Xie and Peng (2019) RF Prediction of excavation damaged zones of underground mines
Zhou et al. (2016) RF Prediction of ground movement above tunnels
Zhou et al. (2019) RF, SVM Prediction of safety risks of deep foundation pits in subway stations
Mahdevari et al. (2013) SVM Prediction of tunnel convergence during excavation
Shi et al. (2019) SVM Prediction of deformation of surrounding rock in shallow-buried tunnels
Cong et al. (2016) SVM Prediction of allowable deformation for surrounding rock of underground caverns
Wu et al. (2014) SVM, ANNs Prediction of tunnel surrounding rock displacement
Mahdevari et al. (2014) SVR Prediction of TBM penetration rate in hard rock conditions
Fattahi and Babanouri (2017) SVR-DE, SVR-ABC, SVR-GSA Prediction of TBM penetration rate
Fattahi (2016) SVR-HS, SVR-DE, SVR-PSO Prediction of deformation modulus of rock mass
W. Zhang et al. (2020a) XGBoost, ANNs, SVM, MARS Prediction of surface settlement caused by EPB tunneling
W. Zhang et al. (2020d) XGBoost, RF Prediction of factor of safety against basal heave for deep braced excavations
R. Zhang et al. (2020) XGBoost, RF, DT, MARS Prediction of maximum lateral wall deformation for braced excavations

Abbreviations
ABC Artificial Bee Colony
ANFIS Adaptive Neuro-Fuzzy Inference System
ANNs Artificial Neural Networks
BCS Bayesian Compressive Sampling
BLR Binary Logistic Regression
(continued on next page)

7
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

Table 1 (continued )
BN Bayesian Network
BNGR Bayesian Nonparametric General Regression
BP Back Propagation
BPNN Back-Propagation Neural Network
BRNN Bayesian Regularization Neural Network
CART Classification and Regression Tree
DE Differential Evolution
DENN Differential Evolution Neural Network
DT Decision Tree
ELM Extreme Learning Machine
ENN Emotional Neural Network
EPR Evolutionary Polynomial Regression
FIS Fuzzy Inference System
FN Functional Network
FP Fuzzy Polynomial
GA Genetic Algorithm
GEP Gene Expression Programming
GFNN Generalized Feed-Forward Neural Network
GMDH Group Method of Data Handling
GP Genetic Programming
GRNN Generalized Regression Neural Network
GSA Gravitational Search Algorithm
HON High-Order Neural Network
HS Harmony Search
ICA Imperialism Competitive Algorithm
KL Karhunen-Loève Expansion
KNN K-Nearest Neighbor
LDA Linear Discriminant Analysis
LR Logistic Regression
LSSVM Least Squares Support Vector Machines
MARS Multivariate Adaptive Regression Splines
MCS Monte Carlo Simulation
MGGP Multi-Gene Genetic Programming
MLPNN Multi-Layer Perceptron Neural Network
MLR Multiple Linear Regression
MOGP Multi-Objective Genetic Programming
NB Naïve Bayes
NLMR Non-Linear Multiple Regression
NN Neural Network
PCA Principal Component Analysis
PSO Particle Swarm Optimization
PUNN Product-Unit Neural Network
RBFNN Radial Basis Function Neural Network
ReGEPSVM Regularized Generalized Proximal Support Vector Machines
RF Random Forest
RNN Recurrent Neural Network
SA Simulated Annealing
SCM Subtractive Clustering Method
SVM Support Vector Machines
SVR Support Vector Regression
TWSVM Twin Support Vector Machines
XGBoost eXtreme Gradient Boosting

3.1. Characterization of soil/rock properties

Characterization of Soils/Rocks Pile Foundations Characterization of soils and rocks forms the basis in any geotech­
(All techniques except LR) (All techniques except LR) nical design works but not without challenges faced in laboratory tests
(Mehdizadeh et al., 2017; 2016; Wong et al., 2017) and inferior soil
conditions requiring conditioning and improvement (Leong et al.,
AI Applications in 2018a, 2018b, 2016a, 2016b, 2015; Omoregie et al., 2020, 2019a,
Underground Soil- 2019b, 2019c, 2017, 2016). AI techniques thus become helpful in
Structure Interaction making useful inferences based on collected data so that the character­
istic design values for various parameters can be predicted to improve
the understanding on site-specific ground behavior. There have been
Deep Excavations Tunneling
many successful applications of AI techniques in studying soil and rock
(All techniques except EPR) (All techniques except EPR) properties.

3.1.1. Bayesian method


Fig. 5. AI applications in underground soil-structure interaction. Bayesian method is commonly used in characterizing soil and rock
parameters in underground geotechnical design. Ching and Phoon
(2019) employed Bayesian machine learning approach to develop a site-
specific multivariate model for clay. The model could handle incomplete
data apart from quantifying parameter uncertainty. More recently, Y.

8
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

Total Citations vs. AI Techniques (From Year 2011-2021)


1200
Web of Science
1000 1082
Scopus
800 849
832
Citations
769
600
550
400 450
380 400
331 344
200 287 284
213200
46 63 77 94 96 40 60 82
0

AI Techniques

Fig. 6. Total citations for compiled literature on AI applications in underground soil-structure interaction from year 2011 to 2021 (until time of writing).

Total Citations vs. AI Techniques (Web of Science)


500
450 2011-2015
400 450
403 2016-2021
350 382 366
Citations

300
250
200 246
209
150 183
160
100 122 134 120 127
50 93 101 96
40 6 0 77 0 0 60
0

AI Techniques

Fig. 7. Comparison of total citations (Web of Science) for compiled literature on AI applications in underground soil-structure interaction for year 2011–2015 versus
year 2016–2021 (until time of writing).

Total Citations vs. AI Techniques (Scopus)


700
600 2011-2015
600 2016-2021
500
482
Citations

400 446
403
300 336
200 263
237
213 214 211
100 133 133 137
48 15 67 0 94 0 40 0 82
0

AI Techniques

Fig. 8. Comparison of total citations (Scopus) for compiled literature on AI applications in underground soil-structure interaction for year 2011–2015 versus year
2016–2021 (until time of writing).

9
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

Wang et al. (2018) and Hu et al. (2019) proposed a novel method to 3.1.5. Discussion
simulate random field parameters from sparse measurements using From the review, it is found that Bayesian approach is a popular
Bayesian compressive sampling. The proposed approach considered method that is used to study soil and rock properties. This may be due to
spatial variability of project sites in site characterization. In addition, the reason that Bayesian method considers soil parameters as random
Wang et al. (2020) applied Bayesian method to classify subsurface soil variables in the model, which allows the distribution of the parameters
stratigraphy based on limited cone penetration test (CPT) data. The to be captured probabilistically (Ching and Phoon, 2019; Jin et al.,
proposed method was further extended to quantification of uncertainty 2018). Thus, the uncertainty inherent in soil properties can be consid­
in the classification of subsurface soil stratigraphy in Shi and Wang ered and subsequently quantified in the model (Ching and Phoon, 2019;
(2021a) and Hu and Wang (2020). There are also other successful ap­ Ozdemir, 2011; Qi and Zhou, 2017; Wang et al., 2012). Besides random
plications of Bayesian method in soil and rock characterization, such as variable modelling, Bayesian method has recently been successfully
characterization of soil and rock parameters (Wang and Cao, 2013; Yang extended to random field modelling by considering spatial variability in
et al., 2019), study of resilient modulus using CPT parameters of cohe­ site characterization (Hu et al., 2019; Y. Wang et al., 2018). Subsurface
sive soils (Liu et al., 2016), determination of soil–water characteristic soil stratigraphy can be classified probabilistically using Bayesian
curve (L. Wang et al., 2018), site characterization based on CPT data (X. method, allowing the associated uncertainty to be reasonably quantified
Wang et al., 2018), and interpolation of site-specific geotechnical pa­ (Hu and Wang, 2020; Shi and Wang, 2021a). This can be helpful for
rameters (Hu et al., 2020). geotechnical problems with sparse and limited measurements collected
from site characterization.
3.1.2. ANNs Apart from Bayesian method, techniques such as ANNs and SVM are
ANNs have also been widely used to predict various soil and rock also commonly used in the study of soil and rock properties. This is
parameters. Shahri (2016) developed an ANN model to predict the clay because ANN and SVM models are robust in solving high-dimensional
sensitivity parameter in a high landslide prone area, by utilizing data problems that are non-linear efficiently (Micheletti et al., 2014; Nour­
collected from piezocone penetration test (CPTu). Khandelwal et al. ani et al., 2014). However, since ANNs cannot present the functional
(2018) developed a hybrid GA-ANN model to estimate the cohesion of relationships between the input parameters and output for interpreta­
limestone based on three input parameters, i.e. P-wave velocity, Un­ tion, method such as GP has been employed due to the capability of GP
confined Compressive Strength (UCS) and Brazilian tensile strength of models to present the output in the form of tree structures or mathe­
limestone samples. Bejarbaneh et al. (2018) employed ANNs and fuzzy matical equations, making the results easily understood by users
logic in predicting Young’s modulus of sandstone based on a set of index (Muduli et al., 2015; Sadrossadat et al., 2018).
rock properties obtained from laboratory tests conducted on 96 sand­ Some significant findings from the collected literature are summa­
stone samples. Other ANN applications in predicting soil and rock rized in Table 2. A widely used performance metric in assessing model
properties include the estimation of UCS of granite (Armaghani et al., fit, known as coefficient of determination, R2, is used to compare model
2016), prediction of peak shear stress along the cohesive soil- performance. Table 2 shows that all the techniques can provide direct or
geosynthetic interface (Debnath and Dey, 2017), and prediction of indirect estimation of soil/rock parameters with acceptable accuracy,
compressive index of clay (Kurnaz and Kaya, 2018). ranging from R2 of 0.73 reported by SVM models to accuracy as high as
90% as reported by Bayesian method. Bayesian method can also
3.1.3. SVM consider spatial variability in site characterization and quantify
SVM is another popular technique for estimating soil and rock pa­ parameter uncertainty, which cannot be achieved by using other tech­
rameters. Fattahi (2016) applied SVM to predict the deformation niques. It is noted that Bayesian method can effectively handle incom­
modulus of rock mass. Several hybrid models were developed by plete/missing and limited data and the proposed Bayesian framework in
training the SVM models with Harmony Search (HS) algorithm, Differ­ some of the applications are applicable to projects with different
ential Evolution (DE) algorithm and PSO. Li and Tan (2017) developed a geological conditions (Ching and Phoon, 2019; Wang and Cao, 2013)
Least Squares SVM (LSSVM) model to estimate the UCS and shear while other techniques such as ANNs, SVM and GP require new data to
strength of rock. Çelik (2019) also built a predictive model to estimate better generalize the models to different geology (Z. L. Cheng et al.,
the UCS of carbonate rock by using LSSVM technique. 90 rock samples 2020a, 2020b; Khandelwal et al., 2017; Shahri, 2016).
were prepared in which P-wave velocity, Schmidt hardness rebound
number and cubic sample size were used as the input parameters in 3.2. Pile foundations
predicting the strength of carbonate rock. Other soil and rock properties
that have been studied using SVM technique include the prediction of The complex behavior of pile foundations is another aspect of un­
over-consolidation ratio of clay based on CPTu data (Samui and Kurup, derground soil-structure interaction that has been widely studied (Leung
2012), prediction of UCS of rock using Schmidt hammer rebound values et al., 2006; Ong et al., 2015, 2009, 2006, 2003; Ong and Choo, 2011).
(Fattahi, 2017), and evaluation of drilling rate index of rock and its Many AI techniques have been applied successfully in modeling the
uncertainty (Ru et al., 2019). behavior of pile foundations, including: (i) prediction of ultimate
bearing capacity, (ii) estimation of pile settlement, and (iii) study of
3.1.4. GP load-settlement response.
GP is also commonly used in characterizing soil and rock properties.
Khandelwal et al. (2017) developed a GP model to predict the brittleness 3.2.1. Pile bearing capacity
of intact rock using rock samples collected from 48 tunnel projects. Shen The prediction of bearing capacity of piles is the most popular area of
and Jimenez (2018) applied GP to predict the shear strength parameters AI applications in pile foundations since pile capacity usually governs
of intact sandstone using UCS and uniaxial tensile strength (UTS) of the design of pile foundations (Pham et al., 2020; Shahin, 2016).
sandstone under different stress conditions. More recently, Z. L. Z.L. Ghorbani et al. (2018) employed ANFIS to predict the ultimate axial
Cheng et al. (2020a, 2020b)) employed GP successfully to estimate load bearing capacity of piles by using 108 datasets that include pile
spatiotemporal variations of soil suction based on field monitoring data geometry, material, installation, full-scale static pile load test and CPT
with reasonable error. Other successful applications of GP in soil and results. Harandizadeh et al. (2019) also applied ANFIS approach to
rock characterization include the prediction of strength of jointed rock predict the bearing capacity of piles based on CPT results. Pham et al.
mass (Asadi et al., 2011), prediction of compression index of fine- (2020) adopted ANNs to predict the ultimate bearing capacity of driven
grained soils (Mohammadzadeh et al., 2016), and prediction of piles based on 2314 driven pile static load test results. Alkroosh and
collapse settlement of sandy gravel soils (Soleimani et al., 2018). Nikraz (2011b) developed Gene Expression Programming (GEP) models

10
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

Table 2
Summary of findings of AI applications in soil/rock characterization.
Techniques Findings

Direct/Indirect Estimation of Consideration of Spatial Quantification of Model Highlights of Outcome


Soil Parameters Variability Uncertainty Performance

Bayesian Yes Yes Yes Accuracy > 90% Can handle incomplete and limited data; applicable to
other project sites
ANNs Yes No No R2 > 0.81 Applicable to similar conditions; can be generalized to
SVM Yes No No R2 > 0.73 different geology with new data
GP Yes No No R2 > 0.84

Note:
R2 denotes coefficient of determination of the model.

to correlate CPT results with axial capacity of piles for both bored piles with non-linear characteristics (Choi and Lee, 2010; Jalalifar et al.,
and driven piles. A total of 50 bored pile load tests and 58 driven pile 2011; Nourani et al., 2014) whereas techniques such as GP, MARS, and
load tests as well as CPT results were collected for training the GEP EPR can provide well-structured mathematical expressions for easy
models. Mohanty et al. (2018) adopted a variant of GP, known as Multi- interpretation of results (Arthur et al., 2020; Javadi et al., 2010; Muduli
Objective Genetic Programming (MOGP) to predict the bearing capacity et al., 2015). Availability of massive case studies in the form of experi­
of driven piles in cohesionless soils. Other successful AI applications in mental and field tests that have been published allows a comprehensive
predicting pile capacity include EPR (Ebrahimian et al., 2015; Ebrahi­ database to be compiled for the analysis of pile foundations, as shown by
mian and Movahed, 2017), MARS (Das and Suman, 2015; Samui, 2012; the studies discussed in Sections 3.2.1–3.2.3. Therefore, AI techniques
Samui and Kim, 2013), and SVM (Moayedi and Hayati, 2019; Samui and such as ANNs that are adaptive and can be generalized may be more
Kim, 2013; Singh et al., 2018). suitable. With the improvement in AI techniques over the years, the
trend then shifts towards other methods such as GP, MARS and EPR that
3.2.2. Pile settlement can both handle data complexity and provide simple mathematical ex­
Pile settlement is another important element in the design of pile pressions that are easily understood by users.
foundations (Alkroosh and Nikraz, 2011a; Armaghani et al., 2018; Xia In pile foundations, R2 is commonly used to assess model perfor­
and Zou, 2017). It is governed by factors such as pile load, shear mance. Therefore, some of the results in the literature are summarized in
modulus of soils, pile length, pile diameter and radial distance at which Fig. 9 to compare the accuracy of different techniques. Fig. 9 presents
shear stress becomes negligible (Armaghani et al., 2018). Saeedi Aziz­ the R2 reported by some of the literature that are collected by the au­
kandi et al. (2014) developed both ANN and GP models to predict the thors. It can be observed from the figure that AI techniques generally
uplift displacement of piles based on CPT data. A database made up of perform well in their predictions, with R2 greater than 0.80 reported in
157 pile tensile load test results was used to develop the models. all the literature. Based on the comparison, it is noted that ANFIS and
Armaghani et al. (2018) developed a GEP model to predict the settle­ MARS seem to outperform other methods, with R2 that is greater than
ment of rock-socketed piles by using a database of 96 piles socketed in 0.90 in the reported studies. It is understood that the techniques have
different rock types. Later, Armaghani et al. (2020) developed a hybrid different capabilities and limitations in their applications, hence, hybrid
model of PSO-ANNs to estimate pile settlement. The ANN model was models that combine two or more techniques or modified methods that
optimized by utilizing PSO algorithm in which the most important PSO enhance the efficiency of the original algorithms have been adopted in
parameters for predicting pile settlement were identified through the recent studies. The hybrid/modified models can outperform tech­
sensitivity analyses so that an optimum hybrid model could be obtained. niques such as ANNs, GP, SVM and RF and provide performance com­
parable to EPR, ANFIS and MARS, with R2 greater than 0.85 in their
3.2.3. Pile load-settlement response predictions.
The strength and serviceability criteria in the design of pile foun­
dations are generally satisfied by designing piles with adequate capacity
and within acceptable settlement limits. Thus, establishing a good un­ 3.3. Underground space construction (deep excavations and tunneling)
derstanding on the load-settlement relationship of piles becomes
important (Alkroosh and Nikraz, 2011a; Ong et al., 2018; Shahin, Deep excavations for underground infrastructure due to rapid pop­
2014a, 2014b). Alkroosh and Nikraz (2011a) developed ANN models to ulation growth as well as underground tunnel construction for rail
simulate the load-settlement behavior of pile foundations in sand and transit to relieve urban traffic have become common practice in
mixed soils, subjected to axial loads based on in-situ pile load tests and metropolitan cities in the recent years (Jin et al., 2018; Li et al., 2020).
CPT results. Ismail and Jeng (2011) adopted High-Order Neural However, ground movements due to excavations and tunneling remain a
Network (HON) to simulate the load-settlement curve of axially loaded major concern that need to be addressed because such movements may
piles using pile properties as inputs. Using a variant of HON, known as result in significant displacements in adjacent buildings (Chong and
Product-Unit Neural Network (PUNN), Ismail et al. (2013) proposed an Ong, 2020; Hashash et al., 2011, 2006; Kung et al., 2007). The devel­
ANN model that was developed using hybrid PSO-Back Propagation opment of machine learning techniques enables the complex soil-
algorithm to predict the load-deformation behavior of axially loaded structure interaction in underground construction to be studied more
piles. Shahin (2014a, 2014b) also employed Recurrent Neural Network effectively. The AI applications in underground space construction are
(RNN) to model the load-settlement response of axially loaded drilled discussed in three aspects: (i) updating of soil parameters, (ii)
shafts and driven steel piles using CPT data, respectively. excavation/tunneling-induced deformations, and (iii) stability of un­
derground space.
3.2.4. Discussion
It is found that AI techniques that are capable in finding complex 3.3.1. Updating of soil parameters
data mapping as well as generating simple mathematical representations Back-analysis based on field observations has been widely used to
are popular in the study of pile foundations. ANNs and ANFIS are “black- predict soil parameters in braced excavations or tunneling projects
box” models that are efficient in handling high-dimensional datasets (Finno and Calvello, 2005; Gens et al., 1996; Gioda and Locatelli, 1999;
Rechea et al., 2008). Bayesian method has been used to update soil

11
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

Fig. 9. Comparison of model performance for various AI applications in pile foundations.

parameters based on monitoring data collected from soil-structure re­ settlements. Wang et al. (2012) adopted Bayesian approach to predict
sponses. Juang et al. (2013) proposed a Bayesian framework to back the maximum ground settlement caused by excavations using centrifuge
analyze and update soil parameters in a multistage braced excavation by test data. W. Zhang et al. (2019) and Zhang et al. (2018) adopted MARS
using maximum wall deflection or maximum ground settlement. Then, to predict the lateral wall deflection profiles caused by deep excavations
the framework was extended to update soil parameters using both types in clay. The model could predict the value and location of maximum
of response observations simultaneously. Qi and Zhou (2017) adopted wall deflection for the braced excavation system. Goh et al. (2018) also
Bayesian approach to update soil parameters by using wall deflection applied MARS to predict the maximum surface settlement caused by
data collected for 49 wall sections from 11 case histories. The uncer­ Earth Pressure Balance (EPB) tunneling based on monitoring data
tainty in soil parameters was reduced whereas the accuracy of predicted collected for ground deformations and shield operation. Zhou et al.
maximum deflection increased through the updating process. Jin et al. (2017, 2016) applied RF to predict ground settlements induced by
(2018) applied sequential Bayesian inference to infer the most probable shield-driven tunnels. The model considered input variables such as
estimates of soil parameters and predict the wall deformations for sub­ tunnel geometry, geological properties, and operation parameters to
sequent excavation stages accurately. Miro et al. (2015) presented a predict the maximum surface settlement above tunnels. P. Zhang et al.
probabilistic back-analysis to evaluate the impact of uncertainty of (2019) presented RF approach to predict tunneling-induced ground
subsoil parameters on tunneling-induced ground movements. The un­ settlement and operational parameters such as face pressure and grout
certainty of soil parameters was quantified and updated using Bayesian filling for EPB shield tunneling. SVM has also been applied in the pre­
method to increase the reliability of model output. Li et al. (2020) also diction of tunnel surrounding rock displacement (Shi et al., 2019; Wu
proposed a Bayesian approach to back analyze the uncertain geo­ et al., 2014), and prediction of surrounding rock deformation of un­
mechanical parameters of shield tunnels by using time series monitoring derground caverns (Cong et al., 2016; Xue and Xiao, 2017).
data. The proposed Bayesian framework could reduce the uncertainty of
parameters during shield tunnel excavation and help to mitigate the 3.3.3. Stability of underground space
potential risk of ground collapse. Assessment of stability of underground excavations and tunnels is
another important aspect of underground construction. Many AI tech­
3.3.2. Excavation/tunneling-induced deformations niques have been used to evaluate tunnel squeezing, tunnel conver­
Prediction of excavation/tunneling-induced deformations forms gence, stability of underground openings, and rock burst hazard.
another essential part of the design of underground construction. Tunnel squeezing is the time-dependent deformation around tunnels
Empirical equations, design charts and numerical methods have been that occurs when limiting shear stress is exceeded and creep starts. The
widely used in the prediction of excavation/tunneling-induced ground deformation may stop during construction or continue over a long time
movements (Shi et al., 2019; Zhang et al., 2018; Zhou et al., 2017). period (Barla, 1995). Ghasemi and Gholizadeh (2019a) adopted LR to
However, empirical methods are limited to the applications in similar predict squeezing conditions by using 220 tunneling case histories
ground conditions whereas numerical methods are time-consuming due collected and overburden depth and rock mass quality as input vari­
to the complexity in model development (W. C. Cheng et al., 2020; Choo ables. Besides, 115 tunneling case histories were compiled by Ghasemi
and Ong, 2020, 2017, 2015; Ong and Choo, 2018, 2016; Peerun et al., and Gholizadeh (2019b) to develop a DT model to predict the potential
2020, 2019; Shi et al., 2019; Zhou et al., 2017). AI techniques, which can of tunnel squeezing.
process complex datasets thus gain attention in the prediction of de­ The monitoring of tunnel convergence is essential to prevent tunnel
formations induced by excavations and tunneling. instability, excessive surface settlement and unforeseen incidents
Bayesian method can be used to predict ground movements in un­ (Mahdevari et al., 2013). Adoko et al. (2013) proposed to predict
derground works, as reported by Miro et al. (2015), Jin et al. (2018) and diameter convergence of a high-speed railway tunnel in weak rock
Li et al. (2020), in which soil parameters were back-analyzed to predict conditions based on MARS and ANNs. Both the models showed excellent
wall deflections in braced excavations and/or tunneling-induced ground prediction accuracy which could assist in decision-making process

12
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

regarding tunnel stability during construction. Mahdevari et al. (2013) 2017; Kohestani et al., 2015; Micheletti et al., 2014; Zhou et al., 2019).
adopted SVM to predict tunnel convergence in which soil parameters Therefore, when analyzing ground deformations that involve compli­
and monitored tunnel displacements were used as the input parameters. cated interaction between the surrounding ground and supported
Goh et al. (2017) developed a hybrid MARS-LR model to evaluate the excavation systems and/or tunnels, they can capture the complex
conditions of underground entry-type excavations by classifying the behavior of the entire soil-structure system with high accuracy.
conditions of underground excavations into stable, potentially stable Techniques such as SVM, LR, DT and RF have been widely applied to
and unstable. Mottahedi et al. (2018) proposed a hybrid ANFIS-PSO the evaluation of stability of underground space. This may be due to the
model to predict overbreak in underground excavations in which 270 reason that these techniques can solve both regression and classification
datasets were collected for model training. Xie and Peng (2019) adopted problems efficiently (Behnia and Blais-Stevens, 2018; James et al., 2013;
RF to predict the tunnel excavation damaged zones in deep underground Sitharam et al., 2008). Particularly, LR is efficient in predicting proba­
mines based on monitoring data. bility of event occurrence and making predictions in binary conditions
Rock burst hazard is defined as the instantaneous release of strain or multi-level categories (Ghasemi and Gholizadeh, 2019a, 2019b; Goh
energy stored in surrounding rock mass that could lead to injuries, et al., 2017; Li and Jimenez, 2018).
damage to underground structures and potential collapse of under­ Some of the findings from the collected literature are summarized in
ground excavations (Faradonbeh and Taheri, 2019). Li and Jimenez Table 3. Apart from R2, other numerical measures such as relative error
(2018) adopted LR to predict the probability of rock burst occurrence and root mean squared error (RMSE) are also used to assess model ac­
based on tunnel depth, maximum tangential stress, elastic energy index, curacy. The models generally provide acceptable accuracy, ranging from
UCS and UTS of rock. Faradonbeh and Taheri (2019) predicted rock R2 of 0.72 to 0.94 in various studies. Techniques such as MARS, RF, SVM
burst hazard in a binary condition by using Emotional Neural Network and DT are efficient in studying ground deformations as well as assessing
(ENN), GEP and DT. A total of 134 rock burst events were compiled from the conditions of underground space. It is noted that only Bayesian
case histories for training. method has been employed in multistage updating of soil parameters in
deep excavations due to its superiority in back-analysis based on field
3.3.4. Discussion observations, and LR as an efficient classification technique, is mainly
The review shows that Bayesian method is commonly used to back used in assessing the stability of underground space. Bayesian method
analyze and update soil parameters in underground space construction. can also reduce parameter uncertainty through updating process,
This is because most of the methods such as ANNs, ANFIS and GP are increasing the model accuracy in making forward predictions. This en­
deterministic, where soil parameters are treated as constants and thus ables observational method to be employed more effectively in multi­
produce results with fixed values that may not necessarily reflect the stage excavation projects. One shortcoming that has been reported in
true values of the parameters (Qi and Zhou, 2017). Therefore, Bayesian adopting RF and DT is that more data are required for training process to
method which treats soil parameters as random variables has gained produce more comprehensive models.
increased attention as a robust approach in back analyzing and updating
soil parameters (Ching et al., 2010; Juang et al., 2013; Wang et al., 3.4. Tunneling
2010).
It is found that Bayesian method can also be used to predict ground The AI applications in tunneling are further discussed in terms of
settlements and wall deflections caused by excavations and/or tunnels. assessment of tunnel boring machine (TBM) performance and project
This is probably due to the ability of Bayesian approach to incorporate risk management. The evaluation of TBM performance is essential in
prior information such as case studies, field observations and engi­ tunneling projects to ensure that the projects can meet the planned
neering judgement into the model (Wang and Cao, 2013). Hence, in schedule and cost (Elbaz et al., 2019; Koopialipoor et al., 2019; Salimi
underground construction where monitoring data are collected pro­ et al., 2018b). Risk management is always important in construction to
gressively, the project conditions can be updated easily by integrating ensure that potential risks of a project are properly assessed so that
previous data as prior into the model, to obtain updated model param­ unforeseen circumstances can be mitigated effectively.
eters, and then forward predict the deformations caused by underground
works. 3.4.1. TBM performance
Techniques such as MARS, RF and SVM are also widely used to AI techniques such as GP, DT and SVM have been used to predict the
predict excavation/tunneling-induced deformations. These techniques performance of TBM. Naghadehi et al. (2018) adopted GEP to correlate
can analyze high-dimensional problems that are non-linear (Fattahi, hard rock TBM performance with routine ground properties to predict

Table 3
Summary of findings of AI applications in underground space construction.
Techniques Findings

Multistage Updating of Evaluation of Excavation/ Assessment of Stability of Model Highlights of Outcome


Soil Parameters Tunneling-Induced Deformations Underground Space Performance

Bayesian Yes Yes No Relative error < Can reduce parameter uncertainty through
10% updating process
MARS No Yes Yes R2 > 0.72 Can provide relative importance of
variables; easy to interpret; adaptive
RF No Yes Yes RMSE < 3.93 More data are required to produce a more
comprehensive model
SVM No Yes Yes R2 > 0.94 Model complexity is independent of input
dimensionality
DT No Yes Yes Accuracy > 94% More data are required to produce a more
comprehensive model
LR No No Yes Accuracy > 87% Can solve classification problems
efficiently

Notes:
R2 denotes coefficient of determination of the model.
RMSE denotes root mean squared error of the model.

13
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

the penetration rate and field penetration index of TBM. Elbaz et al. Some findings from the collected literature are summarized in
(2019) proposed a hybrid ANFIS-GA model to predict the EPB shield Table 4. The models developed based on different techniques can pro­
performance. The shield operational parameters were used as input duce output with reasonable accuracy, from the lowest R2 reported as
variables to compute the TBM advance rate. Salimi et al. (2018a) 0.72 to prediction accuracy as high as 98% in the studies. The models
applied a variant of DT, known as Classification and Regression Tree can be easily applied to problems with similar geological conditions, but
(CART) and GP to evaluate the relationship between geological condi­ more data are needed to better generalize the models to other problems.
tions and TBM operation. Salimi et al. (2018b) also applied CART to As an ensemble learning technique, RF can minimize overfitting and
predict the advance rate of TBM in underground excavations. A database improve the shortcomings of DT. For SVM models, the model complexity
was compiled to predict the specific rock mass boreability index. Mah­ does not depend on input dimensionality, but the number of support
devari et al. (2014) adopted a variant of SVM, Support Vector Regression vectors defined in the model, which requires the optimal selection of a
(SVR) to predict the penetration rate of TBM in hard rock conditions. few control parameters only (Mahdevari et al., 2014; Xue and Xiao,
Queens Water Tunnel in New York City was chosen as the case study to 2017).
validate the proposed model. Fattahi and Babanouri (2017) also applied
SVR to predict TBM performance by using different algorithms such as 3.5. Summary
DE algorithm, Artificial Bee Colony (ABC) algorithm, and Gravitational
Search Algorithm (GSA). The review of AI applications in the study of underground soil-
structure interaction is summarized in Fig. 10 for better visualization
3.4.2. Risk management of the techniques that are commonly used in different aspects of un­
To properly manage the risks of shield TBM tunneling projects, derground construction. The capabilities of the techniques that are
Chung et al. (2019) proposed a risk management framework based on discussed in Sections 3.1 to 3.4, which lead to the selection of certain
Bayesian Network (BN). The approach considered geological factors and techniques in different applications are also highlighted in Fig. 10 to
types of TBM systematically to identify the potential risk events that provide readers with guidance towards further understanding and
could occur during tunnel construction. The degree of risk for the choosing an AI technique that may be more suitable for their problems.
identified potential risks was subsequently quantified by predicting the
cost of mitigation measures against event occurrence. The proposed 4. Challenges and recent development of artificial intelligence
framework could minimize drawbacks of TBM tunneling method, such
as difficulty in substituting the machine type once selected, and exces­ It is evident from the review that AI techniques are superior in
sive project delay due to unexpected incidents. The safety risks in un­ different aspects, resulting in the selection of certain techniques for
derground structures can also be analyzed using RF. Zhou et al. (2019) studying specific geotechnical problems. As the discussion in Section 3
developed a RF model to predict the risks of deep foundation pits in mainly focuses on the capabilities/advantages of various methods, this
subway stations using monitoring data. The model could estimate and section aims to review some limitations of the techniques that have been
identify the safety risks to allow the implementation of appropriate found in past studies so that the challenges that may be faced when
mitigation measures. adopting these methods can be understood. Then, recent development of
the techniques is provided to highlight some of the improvements that
3.4.3. Discussion have been made to enhance the efficiency of existing AI techniques and/
GP, DT and SVM have been successfully applied to predict TBM or to overcome the limitations that have been found in their
performance. This is possibly because methods such as GP and DT can applications.
produce models that are simple and easy to interpret, in the forms of
mathematical expressions or tree-based structures (Ghasemi et al., 2018; 4.1. Limitations/challenges of AI techniques
Muduli et al., 2015; Nefeslioglu et al., 2010) whereas SVM is robust in
solving non-linear high-dimensional problems (Fattahi, 2017; Micheletti Some of the limitations of the AI techniques, together with their main
et al., 2014). To manage the potential risks that could arise from features and capabilities that are discussed in previous sections are
tunneling, BN can be used to identify and quantify the potential risks. By summarized in Table 5 for comparison. In general, most AI techniques
developing a BN model, knowledge from different sources can be produce deterministic models. Bayesian method, however, presents a
incorporated into the model to identify the cause of an event occurrence probabilistic approach in analyzing the problems. In deterministic
and estimate its probability (Feng, 2015). Thus, it is useful as a decision- models, uncertainty is neglected by the models and there is lack of
making tool in managing different types of risks that may arise from a ability to assess model accuracy by using a measure of confidence (Jin
tunneling project, so that proper contingency plans can be implemented et al., 2018; Phoon et al., 2003). Meanwhile, Bayesian method can
where necessary to prevent potential injuries, schedule delay and cost provide a probability distribution over the predictions and allow the
overrun. assessment of confidence associated with model output (Shahin, 2016).
Based on the review of literature and Table 5, some of the shortcomings
of the techniques are discussed as follows:
Table 4
a) Applicability/generalization capability
Summary of findings of AI applications in tunneling works.
One challenge in the applications of AI techniques is the applicability
Techniques Findings or generalization capability of the techniques. Since AI techniques are
Model Highlights of Outcome data-driven, they significantly depend on the data provided to make
Performance predictions. Developed AI models are applicable to problems with
GP R2 > 0.72 Applicable to similar geological conditions similar soil/rock types, but caution should be exercised to apply them to
SVM R2 > 0.95 Model complexity is independent of input different geological conditions because extrapolation is required for new
dimensionality data that are outside the range of the training data. Hence, more data
DT R2 > 0.79 More data are required to produce a more
comprehensive model
that include other geological conditions are required to train the models
BN – Can incorporate information from different to adapt to the new data and improve their applicability. This has been
sources for quantification of risks reported in literature such as GP (Z. L. Cheng et al., 2020a, 2020b;
RF Accuracy > Can minimize overfitting problems and improve Khandelwal et al., 2017; Naghadehi et al., 2018), MARS (Goh et al.,
98% shortcomings of DT
2018; W. Zhang et al., 2019; Zhang et al., 2018), DT (Faradonbeh and
Note: R2 denotes coefficient of determination of the model Taheri, 2019; Ghasemi and Gholizadeh, 2019b; Salimi et al., 2018a),

14
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

Fig. 10. Summary of the review of AI applications in underground soil-structure interaction.

and LR (Ghasemi and Gholizadeh, 2019a; Goh et al., 2017). training process may be required to identify the optimal model structure
However, some methods can be general. The Bayesian framework for ANNs because it is unknown a priori (Adoko et al., 2013; Bhatta­
proposed by Wang and Cao (2013) and Ching and Phoon (2019) are charya et al., 2018; Choi and Lee, 2010). GP is less powerful in finding
applicable to different project-specific data. The proposed approach is constants while searching for optimal solutions and tends to produce
independent of the nature of the data, implying that it can be applied to functions that grow in length with time (Davidson et al., 1999; Javadi
different soil types. Using Bayesian method, previously trained model and Rezania, 2009). PSO also has a slow search process and less effective
can be used to supplement new data through the updating process by in finding local optima for the problems (Gordan et al., 2016).
incorporating previous output as prior into the model. Further, uncer­ d) Overfitting
tainty associated with model parameters can be reduced through Some AI techniques may experience overfitting during training
Bayesian updating, improving the model reliability to provide more process, resulting in a model that may seem to fit the data very well but
accurate predictions (Jin et al., 2018; Juang et al., 2013; Li et al., 2020; does not generalize well when applied to other problems. The problem
Miro et al., 2015; Qi and Zhou, 2017). of overfitting has been found in techniques such as EPR (Giustolisi and
b) Processing of sparse data Savic, 2006), MARS (W. Zhang et al., 2020c), and GP (Armaghani et al.,
Generally, the larger the datasets, the better the performance of AI 2018).
techniques. As the number of datasets increases, the “noise” in the data e) Computational complexity
may also increase with the increasing number of parameters or infor­ When dealing with complex datasets, some techniques may require
mation collected, resulting in sparse data for model training. Sometimes, higher computational capability to deal with the training process. Jin
sparse data may be collected from project sites with highly variable et al. (2018) highlighted that sufficient computational capability is
geological conditions. Some techniques are found to be more susceptible required to train Bayesian models efficiently in a more complex setting.
to sparse data, such as MARS (W. Zhang et al., 2020c) and DT (James Kohestani et al. (2015) also reported that SVM models may face issue of
et al., 2013; Tso and Yau, 2007). RF may also be sensitive to the selection computational complexity when dealing with more complex datasets.
of datasets used for model training (Zhou et al., 2017).
c) Slow convergence 4.2. Recent development of AI techniques
Techniques such as ANNs and ANFIS are slow in convergence and
tend to be trapped in local minima while searching for optimal solutions To overcome some of the challenges that may be faced when
(Armaghani et al., 2020; Mottahedi et al., 2018). In addition, long applying AI techniques, the existing techniques have been improved or

15
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

Table 5
Comparison of commonly used AI techniques in geotechnical engineering.
Techniques Main Features Capabilities Limitations References

ANNs A network model that is processed by Adaptive; can be generalized; and can Considered as a “black-box” model which cannot Armaghani et al.
several interconnected elements known as handle high-dimensional datasets present an explicit relationship between the (2015) 1
nodes or neurons, positioned in the network with non-linear characteristics input and output; the optimal model structure Choi and Lee (2010)
2, 3
layers must be identified prior to the analysis
Nourani et al. (2014)
2

Gordan et al (2016) 3
Bhattacharya et al.
(2018) 3
GP Establishes functional relationships among Provides explicit and structured Not powerful in finding constants, hence, optimal Shen and Jimenez
the variables by using a tree structure representation of the problem in the values and good structures of function can be (2018) 1
consisting of terminals and functions form of tree structure or equations missed in the search process; tends to produce Muduli et al. (2015) 2
functions that grow in length with time Sadrossadat et al.
(2018) 2
Davidson et al. (1999)
3

Javadi and Rezania


(2009) 3
EPR Captures non-linear and complex Generates transparent and well- Overfitting problems may occur due to the Javadi et al. (2010) 1
interactions between the variables without structured mathematical expressions flexibility of the model Javadi and Rezania
the need to assume the form of relationship (2009) 2
between the variables Shahin (2015) 2
Giustolisi and Savic
(2006) 3
MARS Non-linear and non-parametric approach Can find complex data mapping in Susceptible to overfitting, limited to handling Mohanty et al. (2018)
1
that defines relationships between variables high-dimensional data and produce large data and less accurate for sparse data
by using a series of piecewise splines simple models for interpretation Das and Suman
(2015) 2
Arthur et al. (2020) 2
W. Zhang et al.
(2020c) 3
PSO Inspired by the social behavior of swarms Easy to apply and computationally The search process may be slow and ineffective in Zhang et al. (2009) 1
which uses the potential of each individual effective; strong in finding global searching for local optima Himanshu and
particle to achieve the optimum goal in a optimum Burman (2019) 1
coordinated way Gordan et al. (2016)
2, 3

ANFIS Combines the properties of neural networks High prediction capacity in solving Considered as a “black-box” model which does Jalalifar et al. (2011)
1, 2
and fuzzy systems to learn the features from complex and non-linear problems; can not provide practical predictive equations
the datasets develop the model without data pre- Cabalar et al. (2012)
1, 2
processing
Ghorbani et al.
(2018) 2, 3
Sadrossadat et al.
(2018) 2, 3
Bayesian Considers parameters as random variables Can incorporate existing knowledge Requires sufficient computational capability and Jin et al. (2018) 1, 2, 3
and captures the distribution of input into the model and quantify number of observations for more complex Ozdemir (2011) 2
parameters and the output probabilistically uncertainty rigorously; superior in problems; reasonable estimate of prior Wang et al. (2012) 2
back-analysis distribution may be required Qi and Zhou (2017) 2
Juang et al. (2013) 3
SVM Finds the optimal separation hyperplane Robust in solving non-linear high- High computational complexity; sensitive to Ren et al. (2019) 1
through training to maximize the isolation dimensional problems by using kernel missing and unbalanced data Micheletti et al.
margin between two kinds of samples functions (2014) 2
Fattahi (2017) 2
Kohestani et al.
(2015) 3
DT Splits the data recursively into subsets by Easy to interpret and provides Does not perform as well as other techniques for Nefeslioglu et al.
sequentially partitioning in random trees so understandable rules or structures; non-linear data and susceptible to noisy data (2010) 1, 2 Lombardo
that each subset contains homogenous provides visual interpretations et al. (2015) 1
states of the target variable Ghasemi et al. (2018)
2

Tso and Yau (2007) 3


James et al. (2013) 3
RF Uses an ensemble of randomly trained Easy to understand and robust against Sensitive to the selection of the datasets with de Oliveira et al.
decision trees to separate the data into sub- overfitting with the ability to work respect to the resulting tree structure (2019) 1
samples and build a decision tree for each with high-dimensional data Micheletti et al.
(2014) 2
Kohestani et al.
(2015) 2
Zhou et al. (2017) 3
LR Finds the best model that describes the Can handle any type of variables Constrained to minimize the gradients of the Lombardo et al.
relationship between the outcome and a set without requiring any specific response curves of each single independent (2015) 1, 2, 3
of independent variables properties for their distributions variable Pham et al. (2017) 1
Bui et al. (2011) 2
1
Reference for Column 2 “Main Features”.
2
Reference for Column 3 “Capabilities”.

16
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

3
Reference for Column 4 “Limitations”.

modified to enhance their efficiency. The improved or modified AI developed are further discussed as follows:
techniques are sometimes known as “variants” of the techniques. Hybrid a) ANNs
AI models that combine two or more techniques in one model have also As one of the earliest developed AI techniques, many variants have
become increasingly popular to improve model performance while been developed to improve the original ANN algorithm. For example,
tackling the limitations of existing methods. In this section, the variants Shahri (2016) stated that Levenberg-Marquardt algorithm is an
of AI techniques and hybrid models are discussed to provide readers advanced non-linear optimization algorithm that can speed up the
with possible countermeasures that can be considered to solve some of training process of ANNs whereas Ismail and Jeng (2011) highlighted
the issues imposed by selected techniques. that High-Order Neural Network (HON) can achieve high prediction
accuracy with a smaller network size by increasing the information ca­
4.2.1. Variants of techniques pacity and reducing the number of processing elements required in the
There are many modifications that have been done to improve model to form higher-order combinations of the inputs.
existing AI techniques (see Fig. 11). Some of the variants that have been b) GP

• Back-Propagation Neural Network (BPNN)


• Bayesian Regularization Neural Network (BRNN)
• Differential Evolution Neural Network (DENN)
• Feed-Forward Neural Network (FFNN)
• Generalized Feed-Forward Neural Network (GFNN)
ANNs • Generalized Regression Neural Network (GRNN)
• High-Order Neural Network (HON)
• Levenberg-Marquardt Neural Network (LMNN)
• Multi-Layer Perceptron Neural Network (MLPNN)
• Radial Basis Function Neural Network (RBFNN)
• Recurrent Neural Network (RNN)

• Gene Expression Programming (GEP)


• Linear Genetic Programming (LGP)
• Multi-Expression Programming (MEP)
GP
• Multi-Gene Genetic Programming (MGGP)
• Multi-Objective Genetic Programming (MOGP)
• Weighted Genetic Programming (WGP)

• Bayesian Network (BN)


Bayesian • Bayesian Nonparametric General Regression (BNGR)
• Bayesian Regression

• Least Squares Support Vector Machines (LSSVM)


• Regularized Generalized Proximal Support Vector Machines (ReGEPSVM)
SVM • Support Vector Classification (SVC)
• Support Vector Regression (SVR)
• Twin Support Vector Machines (TWSVM)

• Classification and Regression Tree (CART)


DT • Classification Tree (CT)
• Regression Tree (RT)

• Random Forest Classification (RFC)


RF
• Random Forest Regression (RFR)

• Binary Logistic Regression (BLR)


LR • Kernel Logistic Regression (KLR)
• Logistic Multiple Regression (LMR)

Fig. 11. Summary of some variants developed for improving the performance of AI techniques.

17
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

Modifications have also been implemented to GP algorithm to (ABC) and Gravitational Search Algorithm (GSA) to form three hybrid
improve its performance. Armaghani et al. (2018) employed a variant of models (i.e., SVR-DE, SVR-ABC, and SVR-GSA) to predict TBM pene­
GP, Gene Expression Programming (GEP), to improve the efficiency of tration rate. The algorithms improved the generalization capability of
GP by tackling problems related to applying genetic operators in the SVR models, while significantly increased the speed and accuracy of
model and overfitting. Naghadehi et al. (2018) also adopted GEP to searching for optimal model parameters.
produce a more robust model that can be better generalized to similar In summary, the use of ensemble learning methods such as RF and
rock types and geological conditions. XGBoost as well as combinations of AI techniques with optimization
c) Bayesian method algorithms to form hybrid models have shown to be effective in dealing
Bayesian Nonparametric General Regression (BNGR), as one of the with issues such as processing of sparse data, slow convergence, over­
variants of Bayesian method, has recently been applied successfully by fitting, and computational complexity of existing AI techniques.
Zhao et al. (2019) in studying soil reinforcement. The model harnessed
the advantages of Bayesian inference in proper selection of input vari­ 5. Conclusions
ables for the model and powerful prediction capacity of general
regression neural networks to evaluate the significance of various factors The review of AI applications in underground soil-structure inter­
in soil reinforcement. action shows that:
d) SVM
Moayedi and Hayati (2019) adopted two variants of SVM, Regular­ (a) A wide range of AI techniques have been applied to different
ized Generalized Proximal SVM (ReGEPSVM) and Twin SVM (TWSVM), aspects of underground works based on the literature collected by
to predict the friction capacity of driven piles in clay. ReGEPSVM was the authors for the last ten years. The increase of AI applications
able to reduce the runtime of the model by half while showing better in underground construction indicates that they are in great de­
generality on different datasets as compared to other regression mand due to their capability in handling high-dimensional
methods. datasets. AI techniques are particularly powerful in processing
In summary, the development of variants of different techniques has complex datasets and efficient in capturing non-linear relation­
proven to be effective in tackling some of the issues that may be faced ships amongst the variables in the model, with high prediction
when existing AI methods are used, such as slow convergence, over­ accuracy.
fitting, and computational complexity. (b) In terms of citation count, observation shows that ANNs and
Bayesian method are most widely applied to underground works.
4.2.2. Hybrid techniques However, the comparison analysis reveals that methods such as
Other than modifying existing techniques, different techniques can DT, RF and LR have become increasingly popular in this field
be combined in a single model to overcome some of the limitations that while ANNs and Bayesian method remain widely used in under­
have been identified. They can be combined using ensemble learning ground geotechnical studies.
methods, which integrate several algorithms in one model, or combined (c) Some AI techniques have certain drawbacks, such as the lack of
as hybrid models by using optimization algorithms to improve model transparency and inability in knowledge extraction in developed
performance. models, resulting in “black-box” models such as ANNs and ANFIS.
a) Ensemble models Other techniques like EPR, MARS, Bayesian method and LR can
An ensemble model that combines several machine learning algo­ provide well-structured mathematical expressions of the output;
rithms can be effective in overcoming the challenges of existing tech­ or visual representations presenting the solutions in tree-based
niques while excels in its performance. RF is one of the ensemble structures such as GP, DT and RF.
learning algorithms developed to improve model performance by (d) Some of the challenges that have been identified in AI applica­
combining a large set of randomly trained DTs (Behnia and Blais- tions include applicability/generalization capability, suscepti­
Stevens, 2018). Zhou et al. (2019) explained that by integrating Clas­ bility to sparse data, slow convergence, overfitting, and
sification and Regression Tree (CART) and bagging, a learning method computational complexity when dealing with complex datasets.
based on bootstrapping, RF model can overcome overfitting and limi­ (e) Developed AI models are applicable to problems with similar
tations of single DTs with strong data mining capability and high pre­ conditions. For problems with geological conditions that vary
diction accuracy. substantially, more data are required for training to better
Extreme Gradient Boosting (XGBoost) is another ensemble learning generalize the developed models. By supplying the models with
method proposed by Chen and Guestrin (2016) to improve model ac­ new data, they can be trained to adapt to the new data and
curacy through integration of DT and allow the model to process sparse improve their applicability.
data. W. Zhang et al. (2020a) applied XGBoost to estimate surface set­ (f) To address some of the limitations of existing techniques, recent
tlement caused by tunneling. The model was more computationally development of AI shows that variants of different techniques
efficient than methods such as MARS, SVM and ANNs and could be used have been developed to improve the original algorithms. Apart
reliably in solving multivariate non-linear geotechnical problems. W. from modifying existing techniques, ensemble machine learning
Zhang et al. (2020d) also applied RF and XGBoost to assess basal heave algorithms such as RF and XGBoost are shown to be effective in
stability of braced excavations. The models showed excellent perfor­ tackling overfitting and susceptibility to sparse data with excel­
mance, with R2 of 0.988 and 0.994 reported by RF and XGBoost models, lent model performance. Hybrid models that combine existing
respectively. methods with optimization algorithms such as PSO and GA can
b) Hybrid models overcome issues of slow convergence and computational
Optimization algorithms such as PSO and GA can be effective in complexity of AI techniques.
optimizing AI techniques. Ismail et al. (2013) combined PSO with ANNs
to achieve an efficient global search process and prevent the model from Declaration of Competing Interest
being trapped at local minima or missing the best solution. Elbaz et al.
(2019) proposed a hybrid ANFIS-GA model to predict TBM performance, The authors declare that they have no known competing financial
which improved the original ANFIS model as GA enhanced the runtime interests or personal relationships that could have appeared to influence
for parameter tuning in ANFIS. Fattahi and Babanouri (2017) proposed the work reported in this paper.
to optimize Support Vector Regression (SVR) by using evolutionary al­
gorithms such as Differential Evolution (DE), Artificial Bee Colony

18
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

References Ching, J., Phoon, K.K., 2019. Constructing site-specific multivariate probability
distribution model using Bayesian machine learning. J. Eng. Mech. 145, 1–15.
https://doi.org/10.1061/(ASCE)EM.1943-7889.0001537.
Adoko, A.C., Jiao, Y.Y., Wu, L., Wang, H., Wang, Z.H., 2013. Predicting tunnel
Ching, J., Phoon, K.K., 2014. Correlations among some clay parameters - the multivariate
convergence using multivariate adaptive regression spline and artificial neural
distribution. Can. Geotech. J. 51, 686–704. https://doi.org/10.1139/cgj-2013-0353.
network. Tunn. Undergr. Sp. Technol. 38, 368–376. https://doi.org/10.1016/j.
Ching, J., Phoon, K.K., Chen, Y.C., 2010. Reducing shear strength uncertainties in clays
tust.2013.07.023.
by multivariate correlations. Can. Geotech. J. 47, 16–33. https://doi.org/10.1139/
Ahangar-Asr, A., Javadi, A.A., Khalili, N., 2015. An evolutionary approach to modelling
T09-074.
the thermomechanical behaviour of unsaturated soils. Int. J. Numer. Anal. Methods
Choi, M., Lee, G., 2010. Decision tree for selecting retaining wall systems based on
Geomech. 39, 539–557. https://doi.org/10.1002/nag.2323.
logistic regression analysis. Autom. Constr. 19, 917–928. https://doi.org/10.1016/j.
Alkroosh, I., Nikraz, H., 2014. Predicting pile dynamic capacity via application of an
autcon.2010.06.005.
evolutionary algorithm. Soils Found. 54, 233–242. https://doi.org/10.1016/j.
Chong, E.E.M., Ong, D.E.L., 2020. Data-driven field observational method of a
sandf.2014.02.013.
contiguous bored pile wall system affected by accidental groundwater drawdown.
Alkroosh, I., Nikraz, H., 2012. Predicting axial capacity of driven piles in cohesive soils
Geosci. 10, 268. https://doi.org/10.3390/geosciences10070268.
using intelligent computing. Eng. Appl. Artif. Intell. 25, 618–627. https://doi.org/
Choo, C.S., Ong, D.E.L., 2020. Assessment of non-linear rock strength parameters for the
10.1016/j.engappai.2011.08.009.
estimation of pipe-jacking forces. Part 2. Numerical modeling. Eng. Geol. 265,
Alkroosh, I., Nikraz, H., 2011a. Simulating pile load-settlement behavior from CPT data
105405 https://doi.org/10.1016/j.enggeo.2019.105405.
using intelligent computing. Cent. Eur. J. Eng. 1, 295–305. https://doi.org/10.2478/
Choo, C.S., Ong, D.E.L., 2017. Impact of highly weathered geology on pipe-jacking
s13531-011-0029-2.
forces. Geotech. Res. 4, 94–106. https://doi.org/10.1680/jgere.16.00022.
Alkroosh, I., Nikraz, H., 2011b. Correlation of pile axial capacity and CPT data using
Choo, C.S., Ong, D.E.L., 2015. Evaluation of pipe-jacking forces based on direct shear
gene expression programming. Geotech. Geol. Eng. 29, 725–748. https://doi.org/
testing of reconstituted tunneling rock spoils. J. Geotech. Geoenvironmental Eng.
10.1007/s10706-011-9413-1.
141, 04015044. https://doi.org/10.1061/(asce)gt.1943-5606.0001348.
Armaghani, D.J., Asteris, P.G., Fatemi, S.A., Hasanipanah, M., Tarinejad, R., Rashid, A.S.
Chung, H., Lee, I.M., Jung, J.H., Park, J., 2019. Bayesian networks-based shield TBM risk
A., Huynh, V.V., 2020. On the use of neuro-swarm system to forecast the pile
management system: methodology development and application. KSCE J. Civ. Eng.
settlement. Appl. Sci. 10, 1904. https://doi.org/10.3390/app10061904.
23, 452–465. https://doi.org/10.1007/s12205-018-0912-y.
Armaghani, D.J., Faradonbeh, R.S., Rezaei, H., Rashid, A.S.A., Amnieh, H.B., 2018.
Cong, M., Yuehu, T., Erbing, L., Meng, Y., Jianli, D., 2016. Allowable deformation
Settlement prediction of the rock-socketed piles through a new technique based on
prediction for surrounding rock of underground caverns based on support vector
gene expression programming. Neural Comput. Appl. 29, 1115–1125. https://doi.
machine. Period. Polytech. Civ. Eng. 60, 361–369. https://doi.org/10.3311/
org/10.1007/s00521-016-2618-8.
PPci.7732.
Armaghani, D.J., Tonnizam Mohamad, E., Hajihassani, M., Yagiz, S., Motaghedi, H.,
Cramer, N.L., 1985. A representation for the adaptive generation of simple sequential
2016. Application of several non-linear prediction tools for estimating uniaxial
programs. International Conference on Genetic Algorithms and the Applications.
compressive strength of granitic rocks and comparison of their performances. Eng.
183–187.
Comput. 32, 189–206. https://doi.org/10.1007/s00366-015-0410-5.
Cui, K., Jing, X., 2019. Research on prediction model of geotechnical parameters based
Armaghani, D.J., Tonnizam Mohamad, E., Momeni, E., Narayanasamy, M.S., Mohd
on BP neural network. Neural Comput. Appl. 31, 8205–8215. https://doi.org/
Amin, M.F., 2015. An adaptive neuro-fuzzy inference system for predicting
10.1007/s00521-018-3902-6.
unconfined compressive strength and Young’s modulus: a study on Main Range
Das, S., Dey, A., Pal, A., Roy, N., 2015. Applications of artificial intelligence in machine
granite. Bull. Eng. Geol. Environ. 74, 1301–1319. https://doi.org/10.1007/s10064-
learning: review and prospect. Int. J. Comput. Appl. 115, 31–41. https://doi.org/
014-0687-4.
10.5120/20182-2402.
Arthur, C.K., Temeng, V.A., Ziggah, Y.Y., 2020. Multivariate adaptive regression splines
Das, S.K., Suman, S., 2015. Prediction of lateral load capacity of pile in clay using
(MARS) approach to blast-induced ground vibration prediction. Int. J. Mining.
multivariate adaptive regression spline and functional network. Arab. J. Sci. Eng. 40,
Reclam. Environ. 34, 198–222. https://doi.org/10.1080/17480930.2019.1577940.
1565–1578. https://doi.org/10.1007/s13369-015-1624-y.
Asadi, M., Eftekhari, M., Bagheripour, M.H., 2011. Evaluating the strength of intact rocks
Davidson, J.W., Savic, D., Walters, G.A., 1999. Method for the identification of explicit
through genetic programming. Appl. Soft Comput. J. 11, 1932–1937. https://doi.
polynomial formulae for the friction in turbulent pipe flow. J. Hydroinformatics 1,
org/10.1016/j.asoc.2010.06.009.
115–126. https://doi.org/10.2166/hydro.1999.0010.
Barla, G., 1995. Squeezing rocks in tunnels. ISRM News J. 2, 44–49.
de Oliveira, G.G., Ruiz, L.F.C., Guasselli, L.A., Haetinger, C., 2019. Random forest and
Bayes, T., Price, R., 1763.. An essay towards solving a problem in the doctrine of chances
artificial neural networks in landslide susceptibility modeling: a case study of the Fão
by the Late Rev. Mr. Bayes. Philos. Trans. 53, 370–418.
River Basin, Southern Brazil. Nat. Hazards 99, 1049–1073. https://doi.org/10.1007/
Behnia, P., Blais-Stevens, A., 2018. Landslide susceptibility modelling using the
s11069-019-03795-x.
quantitative random forest method along the northern portion of the Yukon Alaska
Debnath, P., Dey, A.K., 2017. Prediction of laboratory peak shear stress along the
Highway Corridor. Canada. Nat. Hazards 90, 1407–1426. https://doi.org/10.1007/
cohesive soil–geosynthetic interface using artificial neural network. Geotech. Geol.
s11069-017-3104-z.
Eng. 35, 445–461. https://doi.org/10.1007/s10706-016-0119-2.
Bejarbaneh, B.Y., Bejarbaneh, E.Y., Amin, M.F.M., Fahimifar, A., Armaghani, D.J.,
Ebrahimian, B., Movahed, V., 2017. Application of an evolutionary-based approach in
Majid, M.Z.A., 2018. Intelligent modelling of sandstone deformation behaviour using
evaluating pile bearing capacity using CPT results. Ships Offshore Struct. 12,
fuzzy logic and neural network systems. Bull. Eng. Geol. Environ. 77, 345–361.
937–953. https://doi.org/10.1080/17445302.2015.1116243.
https://doi.org/10.1007/s10064-016-0983-2.
Ebrahimian, B., Nazari, A., Pasha, A.Y., 2015. Evaluating ε50 for lateral load-
Bhattacharya, S., Murakonda, P., Kumar Das, S., 2018. Prediction of uplift capacity of
displacement behavior of piles in clay. Ocean Eng. 96, 149–160. https://doi.org/
suction caisson in clay using functional network and multivariate adaptive
10.1016/j.oceaneng.2014.12.027.
regression spline. Sci. Iran. 25, 517–531. https://doi.org/10.24200/sci.2017.4192.
Elbaz, K., Shen, S.L., Zhou, A., Yuan, D.J., Xu, Y.S., 2019. Optimization of EPB shield
Bozorgzadeh, N., Harrison, J.P., Escobar, M.D., 2019. Hierarchical Bayesian modelling of
performance with adaptive neuro-fuzzy inference system and genetic algorithm.
geotechnical data: application to rock strength. Geotechnique 69, 1056–1070.
Appl. Sci. 9, 780. https://doi.org/10.3390/app9040780.
https://doi.org/10.1680/jgeot.17.P.282.
Faradonbeh, R.S., Taheri, A., 2019. Long-term prediction of rockburst hazard in deep
Breiman, L.E.O., 2001. Random forests. Mach. Learn. 45, 5–32. https://doi.org/10.1023/
underground openings using three robust data mining techniques. Eng. Comput. 35,
A:1010933404324.
659–675. https://doi.org/10.1007/s00366-018-0624-4.
Bui, D.T., Lofman, O., Revhaug, I., Dick, O., 2011. Landslide susceptibility analysis in the
Faramarzi, A., Javadi, A.A., Alani, A.M., 2012. EPR-based material modelling of soils
Hoa Binh province of Vietnam using statistical index and logistic regression. Nat.
considering volume changes. Comput. Geosci. 48, 73–85. https://doi.org/10.1016/j.
Hazards 59, 1413–1444. https://doi.org/10.1007/s11069-011-9844-2.
cageo.2012.05.015.
Cabalar, A.F., Cevik, A., Gokceoglu, C., 2012. Some applications of adaptive neuro-fuzzy
Fatehnia, M., Amirinia, G., 2018. A review of genetic programming and artificial neural
inference system (ANFIS) in geotechnical engineering. Comput. Geotech. 40, 14–33.
network applications in pile foundations. Int. J. Geo-Engineering 9. https://doi.org/
https://doi.org/10.1016/j.compgeo.2011.09.008.
10.1186/s40703-017-0067-6.
Çelik, S.B., 2019. Prediction of uniaxial compressive strength of carbonate rocks from
Fattahi, H., 2017. Applying soft computing methods to predict the uniaxial compressive
nondestructive tests using multivariate regression and LS-SVM methods. Arab. J.
strength of rocks from schmidt hammer rebound values. Comput. Geosci. 21,
Geosci. 12, 193. https://doi.org/10.1007/s12517-019-4307-2.
665–681. https://doi.org/10.1007/s10596-017-9642-3.
Chen, T., Guestrin, C., 2016. XGBoost: a scalable tree boosting system, in: Proceedings of
Fattahi, H., 2016. Application of improved support vector regression model for
the ACM SIGKDD International Conference on Knowledge Discovery and Data
prediction of deformation modulus of a rock mass. Eng. Comput. 32, 567–580.
Mining. pp. 785–794. https://doi.org/10.1145/2939672.2939785.
https://doi.org/10.1007/s00366-016-0433-6.
Cheng, W.C., Li, G., Ong, D.E.L., Chen, S.L., Ni, J.C., 2020. Modelling liner forces
Fattahi, H., Babanouri, N., 2017. Applying optimized support vector regression models
response to very close-proximity tunnelling in soft alluvial deposits. Tunn. Undergr.
for prediction of tunnel boring machine performance. Geotech. Geol. Eng. 35,
Sp. Technol. 103, 103455 https://doi.org/10.1016/j.tust.2020.103455.
2205–2217. https://doi.org/10.1007/s10706-017-0238-4.
Cheng, Z.L., Zhou, W.H., Ding, Z., Guo, Y.X., 2020a. Estimation of spatiotemporal
Feng, X., 2015. Application of Bayesian Approach in Geotechnical Engineering.
response of rooted soil using a machine learning approach. J. Zhejiang Univ. Sci. A
Polytechnic University of Madrid.
21, 462–477. https://doi.org/10.1631/jzus.A1900555.
Finno, R.J., Calvello, M., 2005. Supported excavations: observational method and
Cheng, Z.L., Zhou, W.H., Garg, A., 2020b. Genetic programming model for estimating
inverse modeling. J. Geotech. Geoenvironmental Eng. 131, 826–836. https://doi.
soil suction in shallow soil layers in the vicinity of a tree. Eng. Geol. 268, 105506
org/10.1061/(ASCE)1090-0241(2005)131:7(826).
https://doi.org/10.1016/j.enggeo.2020.105506.
Friedman, J.H., 1991. Multivariate adaptive regression splines. Ann. Stat. 19, 1–141.
Ching, J., Lin, G.H., Phoon, K.K., Chen, J., 2017. Correlations among some parameters of
https://doi.org/10.1214/aos/1176347963.
coarse-grained soils — the multivariate probability distribution model. Can.
Geotech. J. 54, 1203–1220. https://doi.org/10.1139/cgj-2016-0571.

19
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

Gao, L., Tang, Z.W., Liu, Q.Z., 2013. Displacement prediction for soil nailing based on Risk Uncertain. Eng. Syst. Part A Civ. Eng. 6, 1–12. https://doi.org/10.1061/
ANN. Appl. Mech. Mater. 353–354, 614–618. https://doi.org/10.4028/www. AJRUA6.0001059.
scientific.net/AMM.353-356.614. Hu, Y., Zhao, T., Wang, Y., Choi, C., Ng, C.W.W., 2019. Direct simulation of two-
Garg, A., Garg, A., Tai, K., 2014. A multi-gene genetic programming model for estimating dimensional isotropic or anisotropic random field from sparse measurement using
stress-dependent soil water retention curves. Comput. Geosci. 18, 45–56. https:// Bayesian compressive sampling. Stoch. Environ. Res. Risk Assess. 33, 1477–1496.
doi.org/10.1007/s10596-013-9381-z. https://doi.org/10.1007/s00477-019-01718-7.
Garrett, J.H., 1994. Where and why artificial neural networks are applicable in civil Hussain, A., Surendar, A., Clementking, A., Kanagarajan, S., Ilyashenko, L.K., 2019. Rock
engineering. J. Comput. Civ. Eng. 8, 129–130. https://doi.org/10.1061/(ASCE) brittleness prediction through two optimization algorithms namely particle swarm
0887-3801(1994)8:2(129). optimization and imperialism competitive algorithm. Eng. Comput. 35, 1027–1035.
Gens, A., Ledesma, A., Alonso, E.E., 1996. Estimation of parameters in geotechnical https://doi.org/10.1007/s00366-018-0648-9.
backanalysis - II. Application to a tunnel excavation problem. Comput. Geotech. 18, Ismail, A., Jeng, D.S., 2011. Modelling load-settlement behaviour of piles using high-
29–46. https://doi.org/10.1016/0266-352X(95)00022-3. order neural network (HON-PILE model). Eng. Appl. Artif. Intell. 24, 813–821.
Ghaderi, A., Abbaszadeh Shahri, A., Larsson, S., 2019. An artificial neural network based https://doi.org/10.1016/j.engappai.2011.02.008.
model to predict spatial soil type distribution using piezocone penetration test data Ismail, A., Jeng, D.S., Zhang, L.L., 2013. An optimised product-unit neural network with
(CPTu). Bull. Eng. Geol. Environ. 78, 4579–4588. https://doi.org/10.1007/s10064- a novel PSO-BP hybrid training algorithm: applications to load-deformation analysis
018-1400-9. of axially loaded piles. Eng. Appl. Artif. Intell. 26, 2305–2314. https://doi.org/
Ghasemi, E., Gholizadeh, H., 2019a. Development of two empirical correlations for 10.1016/j.engappai.2013.04.007.
tunnel squeezing prediction using binary logistic regression and linear discriminant Jalalifar, H., Mojedifar, S., Sahebi, A.A., Nezamabadi-pour, H., 2011. Application of the
analysis. Geotech. Geol. Eng. 37, 3435–3446. https://doi.org/10.1007/s10706-018- adaptive neuro-fuzzy inference system for prediction of a rock engineering
00758-0. classification system. Comput. Geotech. 38, 783–790. https://doi.org/10.1016/j.
Ghasemi, E., Gholizadeh, H., 2019b. Prediction of squeezing potential in tunneling compgeo.2011.04.005.
projects using data mining-based techniques. Geotech. Geol. Eng. 37, 1523–1532. James, G., Witten, D., Hastie, T., Tibshirani, R., 2013. An Introduction to Statistical
https://doi.org/10.1007/s10706-018-0705-6. Learning with Applications in R, Springer. Springer New York. https://doi.org/
Ghasemi, E., Kalhori, H., Bagherpour, R., Yagiz, S., 2018. Model tree approach for 10.1007/978-1-4614-7138-7.
predicting uniaxial compressive strength and Young’s modulus of carbonate rocks. Jang, J.R., 1993. ANFIS : adaptive-network-based fuzzy inference system. IEEE Trans.
Bull. Eng. Geol. Environ. 77, 331–343. https://doi.org/10.1007/s10064-016-0931- Syst. Man. Cybern. 23, 665–685. https://doi.org/10.1109/21.256541.
1. Javadi, A.A., Faramarzi, A., Ahangar-Asr, A., 2012. Analysis of behaviour of soils under
Ghorbani, A., Niavol, M.F., 2017. Evaluation of induced settlements of piled rafts in the cyclic loading using EPR-based finite element method. Finite Elem. Anal. Des. 58,
coupled static-dynamic loads using neural networks and evolutionary polynomial 53–65. https://doi.org/10.1016/j.finel.2012.04.005.
regression. Appl. Comput. Intell. Soft Comput. 2017 https://doi.org/10.1155/2017/ Javadi, A.A., Johari, A., Ahangar-Asr, A., Faramarzi, A., Toll, D.G., 2010. A new
7487438. approach to constitutive modelling of unsaturated soils using evolutionary
Ghorbani, B., Sadrossadat, E., Bolouri Bazaz, J., Rahimzadeh Oskooei, P., 2018. polynomial regression. Unsaturated Soils Theor. Numer. Adv. Unsaturated Soil
Numerical ANFIS-based formulation for prediction of the ultimate axial load bearing Mech. - Proc. 4th Asia Pacific Conf. Unsaturated Soils 1, 661–666.
capacity of piles through CPT data. Geotech. Geol. Eng. 36, 2057–2076. https://doi. Javadi, A.A., Rezania, M., 2009. Applications of artificial intelligence and data mining
org/10.1007/s10706-018-0445-7. techniques in soil modeling. Geomech. Eng. 1, 53–74. https://doi.org/10.12989/
Ghorbanzadeh, O., Rostamzadeh, H., Blaschke, T., Gholaminia, K., Aryal, J., 2018. A new gae.2009.1.1.053.
GIS-based data mining technique using an adaptive neuro-fuzzy inference system Jin, Y., Biscontin, G., Gardoni, P., 2018. A Bayesian definition of ‘most probable’
(ANFIS) and k-fold cross-validation approach for land subsidence susceptibility parameters. Geotech. Res. 5, 130–142. https://doi.org/10.1680/jgere.18.00027.
mapping. Nat. Hazards 94, 497–517. https://doi.org/10.1007/s11069-018-3449-y. Jin, Y.F., Yin, Z.Y., Zhou, W.H., Shao, J.F., 2019. Bayesian model selection for sand with
Gioda, G., Locatelli, L., 1999. Back analysis of the measurements performed during the generalization ability evaluation. Int. J. Numer. Anal. Methods Geomech. 43,
excavation of a shallow tunnel in sand. Int. J. Numer. Anal. Methods Geomech. 23, 2305–2327. https://doi.org/10.1002/nag.2979.
1407–1425. https://doi.org/10.1002/(SICI)1096-9853(199911)23:13<1407::AID- Juang, C.H., Luo, Z., Atamturktur, S., Huang, H., 2013. Bayesian updating of soil
NAG934>3.0.CO;2-Q. parameters for braced excavations using field observations. J. Geotech.
Giustolisi, O., Savic, D.A., 2006. A symbolic data-driven technique based on evolutionary Geoenvironmental Eng. 139, 395–406. https://doi.org/10.1061/(ASCE)GT.1943-
polynomial regression. J. Hydroinformatics 8, 207–222. https://doi.org/10.2166/ 5606.0000782.
hydro.2006.020b. Kardani, N., Zhou, A., Nazem, M., Shen, S.L., 2020. Estimation of bearing capacity of
Goh, A.T.C., Zhang, W., Zhang, Y., Xiao, Y., Xiang, Y., 2018. Determination of earth piles in cohesionless soil using optimised machine learning approaches. Geotech.
pressure balance tunnel-related maximum surface settlement: a multivariate Geol. Eng. 38, 2271–2291. https://doi.org/10.1007/s10706-019-01085-8.
adaptive regression splines approach. Bull. Eng. Geol. Environ. 77, 489–500. https:// Kennedy, J., Eberhart, R., 1995. Particle swarm optimization. IEEE International
doi.org/10.1007/s10064-016-0937-8. Conference on Neural Networks - Conference Proceedings. 1942–1948. https://doi.
Goh, A.T.C., Zhang, Y., Zhang, R., Zhang, W., Xiao, Y., 2017. Evaluating stability of org/10.1109/ICNN.1995.488968.
underground entry-type excavations using multivariate adaptive regression splines Khandelwal, M., Marto, A., Fatemi, S.A., Ghoroqi, M., Armaghani, D.J., Singh, T.N.,
and logistic regression. Tunn. Undergr. Sp. Technol. 70, 148–154. https://doi.org/ Tabrizi, O., 2018. Implementing an ANN model optimized by genetic algorithm for
10.1016/j.tust.2017.07.013. estimating cohesion of limestone samples. Eng. Comput. 34, 307–317. https://doi.
Gong, W., Tien, Y.M., Juang, C.H., Martin, J.R., Luo, Z., 2017. Optimization of site org/10.1007/s00366-017-0541-y.
investigation program for improved statistical characterization of geotechnical Khandelwal, M., Shirani Faradonbeh, R., Monjezi, M., Armaghani, D.J., Majid, M.Z.B.A.,
property based on random field theory. Bull. Eng. Geol. Environ. 76, 1021–1035. Yagiz, S., 2017. Function development for appraising brittleness of intact rocks using
https://doi.org/10.1007/s10064-016-0869-3. genetic programming and non-linear multiple regression models. Eng. Comput. 33,
Gordan, B., Armaghani, D.J., Hajihassani, M., Monjezi, M., 2016. Prediction of seismic 13–21. https://doi.org/10.1007/s00366-016-0452-3.
slope stability through combination of particle swarm optimization and neural Khuntia, S., Mujtaba, H., Patra, C., Farooq, K., Sivakugan, N., Das, B.M., 2015. Prediction
network. Eng. Comput. 32, 85–97. https://doi.org/10.1007/s00366-015-0400-7. of compaction parameters of coarse grained soil using multivariate adaptive
Gordan, B., Koopialipoor, M., Clementking, A., Tootoonchi, H., Tonnizam Mohamad, E., regression splines (MARS). Int. J. Geotech. Eng. 9, 79–88. https://doi.org/10.1179/
2019. Estimating and optimizing safety factors of retaining wall through neural 1939787914Y.0000000061.
network and bee colony techniques. Eng. Comput. 35, 945–954. https://doi.org/ Kohestani, V.R., Hassanlourad, M., Ardakani, A., 2015. Evaluation of liquefaction
10.1007/s00366-018-0642-2. potential based on CPT data using random forest. Nat. Hazards 79, 1079–1089.
Harandizadeh, H., Toufigh, M.M., Toufigh, V., 2019. Application of improved ANFIS https://doi.org/10.1007/s11069-015-1893-5.
approaches to estimate bearing capacity of piles. Soft Comput. 23, 9537–9549. Koopialipoor, M., Nikouei, S.S., Marto, A., Fahimifar, A., Armaghani, D.J., Mohamad, E.
https://doi.org/10.1007/s00500-018-3517-y. T., 2019. Predicting tunnel boring machine performance through a new model based
Hashash, Y.M.A., Marulanda, C., Ghaboussi, J., Jung, S., 2006. Novel approach to on the group method of data handling. Bull. Eng. Geol. Environ. 78, 3799–3813.
integration of numerical modeling and field observations for deep excavations. https://doi.org/10.1007/s10064-018-1349-8.
J. Geotech. Geoenvironmental Eng. 132, 1019–1031. https://doi.org/10.1061/ Koza, J., 1992. Genetic programming: on the programming of computers by means of
(ASCE)1090-0241(2006)132:8(1019). natural selection. MIT Press.
Hashash, Y.M.A., Song, H., Osouli, A., 2011. Three-dimensional inverse analyses of a Kung, G.C., Juang, C.H., Hsiao, E.L., Hashash, Y.A., 2007. Simplified model for wall
deep excavation in Chicago clays. Int. J. Numer. Anal. Methods Geomech. 35, deflection and ground-surface settlement caused by braced excavation in clays.
1059–1075. https://doi.org/10.1002/nag.949. J. Geotech. Geoenvironmental Eng. 133, 731–747. https://doi.org/10.1061/(ASCE)
Himanshu, N., Burman, A., 2019. Determination of critical failure surface of slopes using 1090-0241(2007)133:6(731).
particle swarm optimization technique considering seepage and seismic loading. Kurnaz, T.F., Kaya, Y., 2018. The comparison of the performance of ELM, BRNN, and
Geotech. Geol. Eng. 37, 1261–1281. https://doi.org/10.1007/s10706-018-0683-8. SVM methods for the prediction of compression index of clays. Arab. J. Geosci. 11,
Houlsby, N.M.T., Houlsby, G.T., 2013. Statistical fitting of undrained strength data. 770. https://doi.org/10.1007/s12517-018-4143-9.
Geotechnique 63, 1253–1263. https://doi.org/10.1680/geot.13.P.007. Leong, H.Y., Ong, D.E.L., Sanjayan, J.G., Nazari, A., 2018a. Strength development of
Hu, Y., Wang, Y., 2020. Probabilistic soil classification and stratification in a vertical soil–fly ash geopolymer: assessment of soil, fly ash, alkali activators, and water.
cross-section from limited cone penetration tests using random field and Monte Carlo J. Mater. Civ. Eng. 30, 04018171. https://doi.org/10.1061/(asce)mt.1943-
simulation. Comput. Geotech. 124, 103634 https://doi.org/10.1016/j. 5533.0002363.
compgeo.2020.103634. Leong, H.Y., Ong, D.E.L., Sanjayan, J.G., Nazari, A., 2016a. Suitability of Sarawak and
Hu, Y., Wang, Y., Zhao, T., Phoon, K.K., 2020. Bayesian supervised learning of site- Gladstone fly ash to produce geopolymers: a physical, chemical, mechanical,
specific geotechnical spatial variability from sparse measurements. ASCE-ASME J.

20
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

mineralogical and microstructural analysis. Ceram. Int. 42, 9613–9620. https://doi. Murphy, K.P., 2012. Machine Learning: A Probabilistic Perspective. MIT Press.
org/10.1016/j.ceramint.2016.03.046. Naghadehi, M.Z., Samaei, M., Ranjbarnia, M., Nourani, V., 2018. State-of-the-art
Leong, H.Y., Ong, D.E.L., Sanjayan, J.G., Nazari, A., 2016b. The effect of different Na2O predictive modeling of TBM performance in changing geological conditions through
and K2O ratios of alkali activator on compressive strength of fly ash based- gene expression programming. Meas. J. Int. Meas. Confed. 126, 46–57. https://doi.
geopolymer. Constr. Build. Mater. 106, 500–511. https://doi.org/10.1016/j. org/10.1016/j.measurement.2018.05.049.
conbuildmat.2015.12.141. Nefeslioglu, H.A., Sezer, E., Gokceoglu, C., Bozkir, A.S., Duman, T.Y., 2010. Assessment
Leong, H.Y., Ong, D.E.L., Sanjayan, J.G., Nazari, A., 2015. A genetic programming of landslide susceptibility by decision trees in the metropolitan area of Istanbul.
predictive model for parametric study of factors affecting strength of geopolymers. Turkey. Math. Probl. Eng. 2010 https://doi.org/10.1155/2010/901095.
RSC Adv. 5, 85630–85639. https://doi.org/10.1039/c5ra16286f. Nourani, V., Pradhan, B., Ghaffari, H., Sharifi, S.S., 2014. Landslide susceptibility
Leong, H.Y., Ong, D.E.L., Sanjayan, J.G., Nazari, A., Kueh, S.M., 2018b. Effects of mapping at Zonouz Plain, Iran using genetic programming and comparison with
significant variables on compressive strength of soil-fly ash geopolymer: variable frequency ratio, logistic regression, and artificial neural network models. Nat.
analytical approach based on neural networks and genetic programming. J. Mater. Hazards 71, 523–547. https://doi.org/10.1007/s11069-013-0932-3.
Civ. Eng. 30, 04018129. https://doi.org/10.1061/(asce)mt.1943-5533.0002246. Omoregie, A.I., Khoshdelnezamiha, G., Senian, N., Ong, D.E.L., Nissom, P.M., 2017.
Leung, C.F., Ong, D.E., Chow, Y.K., 2006. Pile behavior due to excavation-induced soil Experimental optimisation of various cultural conditions on urease activity for
movement in clay. II: Collapsed wall. J. Geotech. Geoenvironmental Eng. 132, isolated Sporosarcina pasteurii strains and evaluation of their biocement potentials.
45–53. https://doi.org/10.1061/(asce)1090-0241(2006)132:1(45). Ecol. Eng. 109, 65–75. https://doi.org/10.1016/j.ecoleng.2017.09.012.
Li, C., Jiang, S.H., Li, J., Huang, J., 2020. Bayesian approach for sequential probabilistic Omoregie, A.I., Ngu, L.H., Ong, D.E.L., Nissom, P.M., 2019a. Low-cost cultivation of
back analysis of uncertain geomechanical parameters and reliability updating of Sporosarcina pasteurii strain in food-grade yeast extract medium for microbially
tunneling-induced ground settlements. Adv. Civ. Eng. 2020 https://doi.org/ induced carbonate precipitation (MICP) application. Biocatal. Agric. Biotechnol. 17,
10.1155/2020/8528304. 247–255. https://doi.org/10.1016/j.bcab.2018.11.030.
Li, N., Jimenez, R., 2018. A logistic regression classifier for long-term probabilistic Omoregie, A.I., Ong, D.E.L., Nissom, P.M., 2019b. Assessing ureolytic bacteria with
prediction of rock burst hazard. Nat. Hazards 90, 197–215. https://doi.org/ calcifying abilities isolated from limestone caves for biocalcification. Lett. Appl.
10.1007/s11069-017-3044-7. Microbiol. 68, 173–181. https://doi.org/10.1111/lam.13103.
Li, W., Tan, Z., 2017. Research on rock strength prediction based on least squares support Omoregie, A.I., Palombo, E.A., Ong, D.E.L., Nissom, P.M., 2020. A feasible scale-up
vector machine. Geotech. Geol. Eng. 35, 385–393. https://doi.org/10.1007/s10706- production of Sporosarcina pasteurii using custom-built stirred tank reactor for in-
016-0114-7. situ soil biocementation. Biocatal. Agric. Biotechnol. 24, 101544 https://doi.org/
Liu, S., Zou, H., Cai, G., Bheemasetti, T.V., Puppala, A.J., Lin, J., 2016. Multivariate 10.1016/j.bcab.2020.101544.
correlation among resilient modulus and cone penetration test parameters of Omoregie, A.I., Palombo, E.A., Ong, D.E.L., Nissom, P.M., 2019c. Biocementation of sand
cohesive subgrade soils. Eng. Geol. 209, 128–142. https://doi.org/10.1016/j. by Sporosarcina pasteurii strain and technical-grade cementation reagents through
enggeo.2016.05.018. surface percolation treatment method. Constr. Build. Mater. 228, 116828 https://
Liu, Z., Shao, J., Xu, W., Wu, Q., 2015. Indirect estimation of unconfined compressive doi.org/10.1016/j.conbuildmat.2019.116828.
strength of carbonate rocks using extreme learning machine. Acta Geotech. 10, Omoregie, A.I., Senian, N., Phua, Y.L., Ngu, L.H., Ong, D.E.L., Ginjom, I.R.H., Nissom, P.
651–663. https://doi.org/10.1007/s11440-014-0316-1. M., 2016. Ureolytic bacteria isolated from Sarawak limestone caves show high
Lombardo, L., Cama, M., Conoscenti, C., Märker, M., Rotigliano, E., 2015. Binary logistic urease enzyme activity comparable to that of Sporosarcina pasteurii (DSM 33).
regression versus stochastic gradient boosted decision trees in assessing landslide Malays. J. Microbiol. 12, 463–470.
susceptibility for multiple-occurring landslide events: application to the 2009 storm Ong, D.E.L., Choo, C.S., 2018. Assessment of non-linear rock strength parameters for the
event in Messina (Sicily, southern Italy). Nat. Hazards 79, 1621–1648. https://doi. estimation of pipe-jacking forces. Part 1. Direct shear testing and backanalysis. Eng.
org/10.1007/s11069-015-1915-3. Geol. 244, 159–172. https://doi.org/10.1016/j.enggeo.2018.07.013.
Mahdevari, S., Shahriar, K., Yagiz, S., Akbarpour Shirazi, M., 2014. A support vector Ong, D.E.L., Choo, C.S., 2016. Back-analysis and finite element modeling of jacking
regression model for predicting tunnel boring machine penetration rates. Int. J. Rock forces in weathered rocks. Tunn. Undergr. Sp. Technol. 51, 1–10. https://doi.org/
Mech. Min. Sci. 72, 214–229. https://doi.org/10.1016/j.ijrmms.2014.09.012. 10.1016/j.tust.2015.10.014.
Mahdevari, S., Shirzad Haghighat, H., Torabi, S.R., 2013. A dynamically approach based Ong, D.E.L., Choo, C.S., 2011. Sustainable construction of a bored pile foundation system
on SVM algorithm for prediction of tunnel convergence during excavation. Tunn. in erratic phyllite. In: ASEAN Australian Engineering Congress. Kuching, Sarawak,
Undergr. Sp. Technol. 38, 59–68. https://doi.org/10.1016/j.tust.2013.05.002. Malaysia, pp. 30–45.
McCulloch, W.S., Pitts, W., 1943. A logical calculus of the ideas immanent in nervous Ong, D.E.L., Leung, C.E., Chow, Y.K., 2006. Pile behavior due to excavation-induced soil
activity. Bull. Math. Biophys. 5, 115–133. https://doi.org/10.1007/BF02478259. movement in clay. I: Stable wall. J. Geotech. Geoenvironmental Eng. 132, 36–44.
Mehdizadeh, A., Disfani, M.M., Evans, R., Arulrajah, A., Ong, D.E.L., 2017. Mechanical https://doi.org/10.1061/(asce)1090-0241(2006)132:1(36).
consequences of suffusion on undrained behaviour of a gap-graded cohesionless soil Ong, D.E.L., Leung, C.F., Chow, Y.K., 2009. Behavior of pile groups subject to excavation-
- an experimental approach. Geotech. Test. J. 40, 1026–1042. https://doi.org/ induced soil movement in very soft clay. J. Geotech. Geoenvironmental Eng. 135,
10.1520/GTJ20160145. 1462–1474. https://doi.org/10.1061/(asce)gt.1943-5606.0000095.
Mehdizadeh, A., Disfani, M.M., Evans, R., Arulrajah, A., Ong, D.E.L., 2016. Discussion of Ong, D.E.L., Leung, C.F., Chow, Y.K., 2003. Piles subject to excavation-induced soil
“development of an internal camera-based volume determination system for triaxial movement in clay. In: 13th European Conference on Soil Mechanics and
testing” by S. E. Salazar, A. Barnes, and R. A. Coffman. The technical note was Geoetechnical Engineering. Czech Republic, Prague, pp. 777–782.
published in Geotechnical Testing Journal, Vol. 38, No. 4, 2015. [DOI: 10.1520/. Ong, D.E.L., Leung, C.F., Chow, Y.K., Ng, T.G., 2015. Severe damage of a pile group due
Geotech. Test. J. 39, 165–168. https://doi.org/10.1520/GTJ20150153. to slope failure. J. Geotech. Geoenvironmental Eng. 141, 04015014. https://doi.org/
Micheletti, N., Foresti, L., Robert, S., Leuenberger, M., Pedrazzini, A., Jaboyedoff, M., 10.1061/(asce)gt.1943-5606.0001294.
Kanevski, M., 2014. Machine learning feature selection methods for landslide Ong, D.E.L., Sim, Y.S., Leung, C.F., 2018. Performance of field and numerical back-
susceptibility mapping. Math. Geosci. 46, 33–57. https://doi.org/10.1007/s11004- analysis of floating stone columns in soft clay considering the influence of dilatancy.
013-9511-0. Int. J. Geomech. 18, 04018135. https://doi.org/10.1061/(asce)gm.1943-
Miro, S., König, M., Hartmann, D., Schanz, T., 2015. A probabilistic analysis of subsoil 5622.0001261.
parameters uncertainty impacts on tunnel-induced ground movements with a back- Ozdemir, A., 2016. Sinkhole susceptibility mapping using logistic regression in Karapınar
analysis study. Comput. Geotech. 68, 38–53. https://doi.org/10.1016/j. (Konya, Turkey). Bull. Eng. Geol. Environ. 75, 681–707. https://doi.org/10.1007/
compgeo.2015.03.012. s10064-015-0778-x.
Moayedi, H., Hayati, S., 2019. Artificial intelligence design charts for predicting friction Ozdemir, A., 2011. Landslide susceptibility mapping using Bayesian approach in the
capacity of driven pile in clay. Neural Comput. Appl. 31, 7429–7445. https://doi. Sultan Mountains (Akşehir, Turkey). Nat. Hazards 59, 1573–1607. https://doi.org/
org/10.1007/s00521-018-3555-5. 10.1007/s11069-011-9853-1.
Mohamad, E.T., Jahed Armaghani, D., Momeni, E., Alavi Nezhad Khalil Abad, S.V., 2015. Pal, M., Deswal, S., 2014. Extreme learning machine based modeling of resilient modulus
Prediction of the unconfined compressive strength of soft rocks: a PSO-based ANN of subgrade soils. Geotech. Geol. Eng. 32, 287–296. https://doi.org/10.1007/
approach. Bull. Eng. Geol. Environ. 74, 745–757. https://doi.org/10.1007/s10064- s10706-013-9710-y.
014-0638-0. Peerun, M.I., Ong, D.E.L., Choo, C.S., 2019. Interpretation of geomaterial behavior
Mohammadzadeh, S.D., Bolouri Bazaz, J., Vafaee Jani Yazd, S.H., Alavi, A.H., 2016. during shearing aided by PIV technology. J. Mater. Civ. Eng. 31, 04019195. https://
Deriving an intelligent model for soil compression index utilizing multi-gene genetic doi.org/10.1061/(asce)mt.1943-5533.0002834.
programming. Environ. Earth Sci. 75, 1–11. https://doi.org/10.1007/s12665-015- Peerun, M.I., Ong, D.E.L., Choo, C.S., Cheng, W.C., 2020. Effect of interparticle behavior
4889-2. on the development of soil arching in soil-structure interaction. Tunn. Undergr. Sp.
Mohanty, R., Suman, S., Das, S.K., 2018. Prediction of vertical pile capacity of driven pile Technol. 106, 103610 https://doi.org/10.1016/j.tust.2020.103610.
in cohesionless soil using artificial intelligence techniques. Int. J. Geotech. Eng. 12, Pham, B.T., Tien Bui, D., Prakash, I., 2017. Landslide susceptibility assessment using
209–216. https://doi.org/10.1080/19386362.2016.1269043. bagging ensemble based alternating decision trees, logistic regression and J48
Mola-Abasi, H., Eslami, A., Shourijeh, P.T., 2013. Shear wave velocity by polynomial decision trees methods: a comparative study. Geotech. Geol. Eng. 35, 2597–2611.
neural networks and genetic algorithms based on geotechnical soil properties. Arab. https://doi.org/10.1007/s10706-017-0264-2.
J. Sci. Eng. 38, 829–838. https://doi.org/10.1007/s13369-012-0525-6. Pham, T.A., Ly, H.B., Tran, V.Q., Giap, L. Van, Vu, H.L.T., Duong, H.A.T., 2020.
Mottahedi, A., Sereshki, F., Ataei, M., 2018. Overbreak prediction in underground Prediction of pile axial bearing capacity using artificial neural network and random
excavations using hybrid ANFIS-PSO model. Tunn. Undergr. Sp. Technol. 80, 1–9. forest. Appl. Sci. 10, 1871. https://doi.org/10.3390/app10051871.
https://doi.org/10.1016/j.tust.2018.05.023. Phoon, K.K., Kulhawy, F.H., Grigoriu, M.D., 2003. Development of a reliability-based
Muduli, P.K., Das, M.R., Das, S.K., Senapati, S., 2015. Lateral load capacity of piles in design framework for transmission line structure foundations. J. Geotech.
clay using genetic programming and multivariate adaptive regression spline. Indian Geoenvironmental Eng. 129, 798–806. https://doi.org/10.1061/(ASCE)1090-0241
Geotech. J. 45, 349–359. https://doi.org/10.1007/s40098-014-0142-2. (2003)129:9(798).

21
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

Qi, X.H., Zhou, W.H., 2017. An efficient probabilistic back-analysis method for braced Sitharam, T.G., Samui, P., Anbazhagan, P., 2008. Spatial variability of rock depth in
excavations using wall deflection data at multiple points. Comput. Geotech. 85, Bangalore using geostatistical, neural network and support vector machine models.
186–198. https://doi.org/10.1016/j.compgeo.2016.12.032. Geotech. Geol. Eng. 26, 503–517. https://doi.org/10.1007/s10706-008-9185-4.
Rechea, C., Levasseur, S., Finno, R., 2008. Inverse analysis techniques for parameter Smola, A.J., Schölkopf, B., 1998. A tutorial on support vector regression, in: NeuroCOLT
identification in simulation of excavation support systems. Comput. Geotech. 35, Technical Report NC-TR-98-030. University of London, UK.
331–345. https://doi.org/10.1016/j.compgeo.2007.08.008. Soleimani, S., Jiao, P., Rajaei, S., Forsati, R., 2018. A new approach for prediction of
Ren, Q., Wang, G., Li, M., Han, S., 2019. Prediction of rock compressive strength using collapse settlement of sandy gravel soils. Eng. Comput. 34, 15–24. https://doi.org/
machine learning algorithms based on spectrum analysis of geological hammer. 10.1007/s00366-017-0517-y.
Geotech. Geol. Eng. 37, 475–489. https://doi.org/10.1007/s10706-018-0624-6. Tarawneh, B., 2013. Pipe pile setup: database and prediction model using artificial
Ru, Z., Zhao, H., Zhu, C., 2019. Probabilistic evaluation of drilling rate index based on a neural network. Soils Found. 53, 607–615. https://doi.org/10.1016/j.
least square support vector machine and Monte Carlo simulation. Bull. Eng. Geol. sandf.2013.06.011.
Environ. 78, 3111–3118. https://doi.org/10.1007/s10064-018-1327-1. Tarawneh, B., Imam, R., 2014. Regression versus artificial neural networks: predicting
Rumelhart, D., Hinton, G., Williams, R., 1987. Learning Internal Representations by pile setup from empirical data. KSCE J. Civ. Eng. 18, 1018–1027. https://doi.org/
Error Propagation. In: Parallel Distributed Processing: Explorations in the 10.1007/s12205-014-0072-7.
Microstructure of Cognition: Foundations. MIT Press, pp. 318–362. Tso, G.K.F., Yau, K.K.W., 2007. Predicting electricity energy consumption: a comparison
Sadrossadat, E., Ghorbani, B., Oskooei, R., Kaboutari, M., 2018. Use of adaptive neuro- of regression analysis, decision tree and neural networks. Energy 32, 1761–1768.
fuzzy inference system and gene expression programming methods for estimation of https://doi.org/10.1016/j.energy.2006.11.010.
the bearing capacity of rock foundations. Eng. Comput. (Swansea Wales) 35, Vapnik, V.N., 1995. Introduction: Four Periods in the Research of the Learning Problem.
2078–2106. https://doi.org/10.1108/EC-07-2017-0258. In: The Nature of Statistical Learning Theory. Springer, New York. https://doi.org/
Saeedi Azizkandi, A., Kashkooli, A., Baziar, M.H., 2014. Prediction of uplift pile 10.1007/978-1-4757-2440-0_1.
displacement based on cone penetration tests (CPT). Geotech. Geol. Eng. 32, Wang, L., Cao, Z.J., Li, D.Q., Phoon, K.K., Au, S.K., 2018. Determination of site-specific
1043–1052. https://doi.org/10.1007/s10706-014-9779-y. soil-water characteristic curve from a limited number of test data – a Bayesian
Salimi, A., Faradonbeh, R.S., Monjezi, M., Moormann, C., 2018a. TBM performance perspective. Geosci. Front. 9, 1665–1677. https://doi.org/10.1016/j.
estimation using a classification and regression tree (CART) technique. Bull. Eng. gsf.2017.10.014.
Geol. Environ. 77, 429–440. https://doi.org/10.1007/s10064-016-0969-0. Wang, L., Ravichandran, N., Juang, C.H., 2012. Bayesian updating of KJHH model for
Salimi, A., Rostami, J., Moormann, C., Hassanpour, J., 2018b. Examining feasibility of prediction of maximum ground settlement in braced excavations using centrifuge
developing a rock mass classification for hard rock TBM application using non-linear data. Comput. Geotech. 44, 1–8. https://doi.org/10.1016/j.compgeo.2012.03.003.
regression, regression tree and generic programming. Geotech. Geol. Eng. 36, Wang, X., Wang, H., Liang, R.Y., Zhu, H., Di, H., 2018. A hidden Markov random field
1145–1159. https://doi.org/10.1007/s10706-017-0380-z. model based approach for probabilistic site characterization using multiple cone
Samui, P., 2013. Multivariate adaptive regression spline (Mars) for prediction of elastic penetration test data. Struct. Saf. 70, 128–138. https://doi.org/10.1016/j.
modulus of jointed rock mass. Geotech. Geol. Eng. 31, 249–253. https://doi.org/ strusafe.2017.10.011.
10.1007/s10706-012-9584-4. Wang, Y., Akeju, O.V., 2016. Quantifying the cross-correlation between effective
Samui, P., 2012. Determination of ultimate capacity of driven piles in cohesionless soil: a cohesion and friction angle of soil from limited site-specific data. Soils Found. 56,
multivariate adaptive regression spline approach. Int. J. Numer. Anal. Methods 1055–1070. https://doi.org/10.1016/j.sandf.2016.11.009.
Geomech. 36, 1434–1439. https://doi.org/10.1002/nag.1076. Wang, Y., Aladejare, A.E., 2016. Bayesian characterization of correlation between
Samui, P., Kim, D., 2013. Least square support vector machine and multivariate adaptive uniaxial compressive strength and Young’s modulus of rock. Int. J. Rock Mech. Min.
regression spline for modeling lateral load capacity of piles. Neural Comput. Appl. Sci. 85, 10–19. https://doi.org/10.1016/j.ijrmms.2016.02.010.
23, 1123–1127. https://doi.org/10.1007/s00521-012-1043-x. Wang, Y., Au, S.K., Cao, Z., 2010. Bayesian approach for probabilistic characterization of
Samui, P., Kurup, P., 2012. Multivariate adaptive regression spline (MARS) and least sand friction angles. Eng. Geol. 114, 354–363. https://doi.org/10.1016/j.
squares support vector machine (LSSVM) for OCR prediction. Soft Comput. 16, enggeo.2010.05.013.
1347–1351. https://doi.org/10.1007/s00500-012-0815-7. Wang, Y., Cao, Z., 2013. Probabilistic characterization of Young’s modulus of soil using
Shahin, M.A., 2016. State-of-the-art review of some artificial intelligence applications in equivalent samples. Eng. Geol. 159, 106–118. https://doi.org/10.1016/j.
pile foundations. Geosci. Front. 7, 33–44. https://doi.org/10.1016/j. enggeo.2013.03.017.
gsf.2014.10.002. Wang, Y., Cao, Z., Li, D., 2016. Bayesian perspective on geotechnical variability and site
Shahin, M.A., 2015. Use of evolutionary computing for modelling some complex characterization. Eng. Geol. 203, 117–125. https://doi.org/10.1016/j.
problems in geotechnical engineering. Geomech. Geoengin. 10, 109–125. https:// enggeo.2015.08.017.
doi.org/10.1080/17486025.2014.921333. Wang, Y., Hu, Y., Zhao, T., 2020. Cone penetration test (CPT)-based subsurface soil
Shahin, M.A., 2014a. Load-settlement modeling of axially loaded drilled shafts using classification and zonation in two-dimensional vertical cross section using Bayesian
CPT-based recurrent neural networks. Int. J. Geomech. 14, 1–7. https://doi.org/ compressive sampling. Can. Geotech. J. 57, 947–958. https://doi.org/10.1139/cgj-
10.1061/(ASCE)GM.1943-5622.0000370. 2019-0131.
Shahin, M.A., 2014b. Load-settlement modeling of axially loaded steel driven piles using Wang, Y., Zhao, T., 2017. Bayesian assessment of site-specific performance of
CPT-based recurrent neural networks. Soils Found. 54, 515–522. https://doi.org/ geotechnical design charts with unknown model uncertainty. Int. J. Numer. Anal.
10.1016/j.sandf.2014.04.015. Methods Geomech. 41, 781–800. https://doi.org/10.1002/nag.2658.
Shahin, M.A., Jaksa, M.B., Maier, H.R., 2008. State of the art of artificial neural networks Wang, Y., Zhao, T., Cao, Z., 2015. Site-specific probability distribution of geotechnical
in geotechnical engineering. Electron. J. Geotech. Eng. Bouquet 08, 1–26. properties. Comput. Geotech. 70, 159–168. https://doi.org/10.1016/j.
Shahri, A.A., 2016. An optimized artificial neural network structure to predict clay compgeo.2015.08.002.
sensitivity in a high landslide prone area using piezocone penetration test (CPTu) Wang, Y., Zhao, T., Phoon, K.K., 2018. Direct simulation of random field samples from
data: a case study in Southwest of Sweden. Geotech. Geol. Eng. 34, 745–758. https:// sparsely measured geotechnical data with consideration of uncertainty in
doi.org/10.1007/s10706-016-9976-y. interpretation. Can. Geotech. J. 55, 862–880. https://doi.org/10.1139/cgj-2017-
Shen, J., Jimenez, R., 2018. Predicting the shear strength parameters of sandstone using 0254.
genetic programming. Bull. Eng. Geol. Environ. 77, 1647–1662. https://doi.org/ Wong, S.T.Y., Ong, D.E.L., Robinson, R.G., 2017. Behaviour of MH silts with varying
10.1007/s10064-017-1023-6. plasticity indices. Geotech. Res. 4, 118–135. https://doi.org/10.1680/
Shi, C., Wang, Y., 2021a. Nonparametric and data-driven interpolation of subsurface soil jgere.17.00002.
stratigraphy from limited data using multiple point statistics. Can. Geotech. J. 58, Wu, Q., Yan, B., Zhang, C., Wang, L., Ning, G., Yu, B., 2014. Displacement prediction of
261–280. https://doi.org/10.1139/cgj-2019-0843. tunnel surrounding rock: a comparison of support vector machine and artificial
Shi, C., Wang, Y., 2021b. Non-parametric machine learning methods for interpolation of neural network. Math. Probl. Eng. 2014 https://doi.org/10.1155/2014/351496.
spatially varying non-stationary and non-Gaussian geotechnical properties. Geosci. Xia, Z.Q., Zou, J.F., 2017. Simplified approach for settlement analysis of vertically loaded
Front. 12, 339–350. https://doi.org/10.1016/j.gsf.2020.01.011. pile. J. Eng. Mech. 143, 1–11. https://doi.org/10.1061/(ASCE)EM.1943-
Shi, S., Zhao, R., Li, S., Xie, X., Li, L., Zhou, Z., Liu, H., 2019. Intelligent prediction of 7889.0001334.
surrounding rock deformation of shallow buried highway tunnel and its engineering Xie, Q., Peng, K., 2019. Space-time distribution laws of tunnel excavation damaged zones
application. Tunn. Undergr. Sp. Technol. 90, 1–11. https://doi.org/10.1016/j. (EDZs) in deep mines and EDZ prediction modeling by random forest regression.
tust.2019.04.013. Adv. Civ. Eng. 2019 https://doi.org/10.1155/2019/6505984.
Shi, X., Chen, W., Lv, X.J., 2012. Deformation prediction of deep excavation using Xu, C., Xu, X., Dai, F., Wu, Z., He, H., Shi, F., Wu, X., Xu, S., 2013. Application of an
support vector machine. Appl. Mech. Mater. 157–158, 66–69. https://doi.org/ incomplete landslide inventory, logistic regression model and its validation for
10.4028/www.scientific.net/AMM.157-158.66. landslide susceptibility mapping related to the May 12, 2008 Wenchuan earthquake
Shirzadi, A., Saro, L., Hyun Joo, O., Chapi, K., 2012. A GIS-based logistic regression of China. Nat. Hazards 68, 883–900. https://doi.org/10.1007/s11069-013-0661-7.
model in rock-fall susceptibility mapping along a mountainous road: Salavat Abad Xu, M., Watanachaturaporn, P., Varshney, P.K., Arora, M.K., 2005. Decision tree
case study, Kurdistan. Iran. Nat. Hazards 64, 1639–1656. https://doi.org/10.1007/ regression for soft classification of remote sensing data. Remote Sens. Environ. 97,
s11069-012-0321-3. 322–336. https://doi.org/10.1016/j.rse.2005.05.008.
Singh, T., Pal, M., Arora, V.K., 2019. Modeling oblique load carrying capacity of batter Xue, X., Xiao, M., 2017. Deformation evaluation on surrounding rocks of underground
pile groups using neural network, random forest regression and M5 model tree. caverns based on PSO-LSSVM. Tunn. Undergr. Sp. Technol. 69, 171–181. https://
Front. Struct. Civ. Eng. 13, 674–685. https://doi.org/10.1007/s11709-018-0505-3. doi.org/10.1016/j.tust.2017.06.019.
Singh, T., Pal, M., Arora, V.K., 2018. Modeling of oblique load test on batter pile group Yang, L., Feng, X., Sun, Y., 2019. Predicting the Young’s modulus of granites using the
based on support vector machines and Gaussian regression. Geotech. Geol. Eng. 36, Bayesian model selection approach. Bull. Eng. Geol. Environ. 78, 3413–3423.
1597–1607. https://doi.org/10.1007/s10706-017-0413-7. https://doi.org/10.1007/s10064-018-1326-2.

22
S.C. Jong et al. Tunnelling and Underground Space Technology incorporating Trenchless Technology Research 113 (2021) 103946

Yang, L., da Li, Z., 2019. Inverse analysis of rock creep model parameters based on Zhang, W., Zhang, R., Wu, C., Goh, A.T.C., Wang, L., 2020d. Assessment of basal heave
improved simulated annealing differential evolution algorithm. Geotech. Geol. Eng. stability for braced excavations in anisotropic clay using extreme gradient boosting
37, 639–649. https://doi.org/10.1007/s10706-018-0634-4. and random forest regression. Undergr. Sp. 1–9 https://doi.org/10.1016/j.
Zhang, J., Hu, J., Li, X., Li, J., 2020. Bayesian network based machine learning for design undsp.2020.03.001.
of pile foundations. Autom. Constr. 118, 103295 https://doi.org/10.1016/j. Zhang, W.G., Goh, A.T.C., 2013. Multivariate adaptive regression splines for analysis of
autcon.2020.103295. geotechnical engineering systems. Comput. Geotech. 48, 82–95. https://doi.org/
Zhang, P., Chen, R.P., Wu, H.N., 2019. Real-time analysis and regulation of EPB shield 10.1016/j.compgeo.2012.09.016.
steering using random forest. Autom. Constr. 106, 102860 https://doi.org/10.1016/ Zhang, Y., Gallipoli, D., Augarde, C.E., 2009. Simulation-based calibration of
j.autcon.2019.102860. geotechnical parameters using parallel hybrid moving boundary particle swarm
Zhang, R., Wu, C., Goh, A.T.C., Böhlke, T., Zhang, W., 2020. Estimation of diaphragm optimization. Comput. Geotech. 36, 604–615. https://doi.org/10.1016/j.
wall deflections for deep braced excavation in anisotropic clays using ensemble compgeo.2008.09.005.
learning. Geosci. Front. https://doi.org/10.1016/j.gsf.2020.03.003. Zhao, L.-S., Zhou, W.-H., Su, L.-J., Garg, A., Yuen, K.-V., 2019. Selection of physical and
Zhang, W., Goh, A.T.C., 2017. Reliability assessment of ultimate limit state of twin chemical properties of natural fibers for predicting soil reinforcement. J. Mater. Civ.
caverns. Geomech. Geoengin. 12, 48–59. https://doi.org/10.1080/ Eng. 31, 04019212. https://doi.org/10.1061/(asce)mt.1943-5533.0002850.
17486025.2016.1162331. Zheng, G., Yang, P., Zhou, H., Zeng, C., Yang, X., He, X., Yu, X., 2019. Evaluation of the
Zhang, W., Goh, A.T.C., 2016. Multivariate adaptive regression splines and neural earthquake induced uplift displacement of tunnels using multivariate adaptive
network models for prediction of pile drivability. Geosci. Front. 7, 45–52. https:// regression splines. Comput. Geotech. 113, 103099 https://doi.org/10.1016/j.
doi.org/10.1016/j.gsf.2014.10.003. compgeo.2019.103099.
Zhang, W., Li, H.R., Wu, C.Z., Li, Y.Q., Liu, Z.Q., Liu, H.L., 2020a. Soft computing Zhou, J., Sh, X., Du, K., Qiu, X., Li, X., Mitri, H.S., 2017. Feasibility of random-forest
approach for prediction of surface settlement induced by earth pressure balance approach for prediction of ground settlements induced by the construction of a
shield tunneling. Undergr. Sp. https://doi.org/10.1016/j.undsp.2019.12.003. shield-driven tunnel. Int. J. Geomech. 17, 1–12. https://doi.org/10.1061/(ASCE)
Zhang, W., Li, Y., Wu, C., Li, H., Goh, A.T.C., Lin, H., 2020b. Prediction of lining response GM.1943-5622.0000817.
for twin tunnels constructed in anisotropic clay using machine learning techniques. Zhou, J., Shi, X., Du, K., Qiu, X., Li, X., Mitri, H.S., 2016. Development of ground
Undergr. Sp. https://doi.org/10.1016/j.undsp.2020.02.007. movements due to a shield tunnelling prediction model using random forests.
Zhang, W., Zhang, R., Goh, A.T.C., 2018. Multivariate adaptive regression splines Geotechnical Special Publication. 108–115. https://doi.org/10.1061/
approach to estimate lateral wall deflection profiles caused by braced excavations in 9780784480106.014.
clays. Geotech. Geol. Eng. 36, 1349–1363. https://doi.org/10.1007/s10706-017- Zhou, W.-H., Yuen, K.-V., Tan, F., 2013. Estimation of maximum pullout shear stress of
0397-3. grouted soil nails using Bayesian probabilistic approach. Int. J. Geomech. 13,
Zhang, W., Zhang, R., Wang, W., Zhang, F., Goh, A.T.C., 2019. A multivariate adaptive 659–664. https://doi.org/10.1061/(asce)gm.1943-5622.0000259.
regression splines model for determining horizontal wall deflection envelope for Zhou, Y., Li, S., Zhou, C., Luo, H., 2019. Intelligent approach based on random forest for
braced excavations in clays. Tunn. Undergr. Sp. Technol. 84, 461–471. https://doi. safety risk prediction of deep foundation pit in subway stations. J. Comput. Civ. Eng.
org/10.1016/j.tust.2018.11.046. 33, 1–14. https://doi.org/10.1061/(ASCE)CP.1943-5487.0000796.
Zhang, W., Zhang, R., Wu, C., Goh, A.T.C., Lacasse, S., Liu, Z., Liu, H., 2020c. State-of-
the-art review of soft computing applications in underground excavations. Geosci.
Front. 11, 1095–1106. https://doi.org/10.1016/j.gsf.2019.12.003.

23

You might also like