Editorial for Advances and applications of deep learning and soft computing in geotechnical underground engineering

We are privileged to be invited by the Honorary Editor-in-Chief,Professor Qihu Qian, Editor-in-Chief, Professor Xia-Ting Feng, and the editorial staff of the Journal of Rock Mechanics and Geotechnical Engineering(JRMGE),to serve as Guest Editors for this Special Issue (SI).

The purpose of this SI is to review the latest development of machine learning (ML) techniques including the soft computing (SC)and deep learning (DL) methods as well as their key applications in geotechnical underground engineering problems. The publication of this SI is timely given the significant interest and progress in data-centric geotechnics (Phoon and Ching, 2021). The International Society for Soil Mechanics and Geotechnical Engineering(ISSMGE) TC309/TC304/TC222 3rd Machine Learning in Geotechnics Dialogue(3MLIGD)was convened on 3 December 2021 to foster greater connectivity between researchers and practitioners to accelerate progress in this nascent field of data-centric geotechnics.The SI contains 22 invited papers covering different ML models,their performance, and the challenges faced in real world applications.

Zhu et al. (2021) proposed a robust probabilistic ML model based on a natural gradient boosting (NGBoost) algorithm combined with XGBoost to process local spatial geographic information to predict the rockhead distribution.

Isleyen et al. (2021) presented a convolutional neural network(CNN) with a transfer learning approach for detection of the roof fall hazards caused by high horizontal stress.

Duan et al.(2021)compared four ML methods,i.e.random forest(RF), support vector machine (SVM), CNN and residual neural network (ResNN) in classifying the signals obtained from the recorded seismic dataset.

Tang and Na (2021) examined the hybrid use of ML methods including SVM, RF, BPNN, as well as DNN and optimization techniques such as particle swarm optimization (PSO) and grid search(GS) for estimation of the maximum surface settlement induced by tunnelling.

Parsajoo et al.(2021) improved the adaptive neuro-fuzzy inference system (ANFIS) model by adopting the artificial bee colony(ABC) as its optimization algorithm for performance assessment of tunnel boring machine (TBM).

Abbaszadeh Shahri et al. (2021) explored the use of artificial neural network (ANN) with different training algorithms and activation functions to develop a three-dimensional (3D) depth to bedrock(DTB) model for a case study in Stockholm, Sweden.

Wu et al. (2021) applied a TBM-rock mutual feedback perception method based on 10,807 tunneling cycles measured from the Songhua River water conveyance tunnel.

Xiao et al. (2021) systematically analyzed pre-grouting data from a large underground project and provided a visual reference to support future surrounding projects.

Liu et al.(2021)built recurrent neural networks(RNNs)and convolutional neural networks (CNNs) using the TBM vibration data obtained from Metro Line R2 in Jinan,China,to identify the ground condition at the working face.

Zhang et al. (2021) proposed a DL-based method for analyzing the influence of inherent spatial variability of soil on the deformation of embedded tunnels.

Moon et al. (2021) presented several regression models to improve the analytical accuracy of fault gouge to better understand the reactivation of fault,behavior of earthquake,and mechanism of slope failure based on physico-mechanical properties measured from 224 specimens collected and tested in 62 fault zones across South Korea.

Li et al. (2021) combined support vector regression (SVR) and five different optimization methods including grey wolf optimization (GWO), PSO, genetic algorithm (GA), salp swarm algorithm(SSA)and GS to improve the prediction performance and optimize the hyper-parameters for accurate prediction of mean fragmentation size.

Bardhan et al. (2021) proposed a hybrid ensemble soft computing (HENSM) approach for prediction of TBM penetration rate.

Murlidhar et al.(2021)developed a variety of empirical models and five ML models including multi-layer perceptron (MLP), RF,SVM, Harris Hawks optimization (HHO)-MLP and whale optimization algorithm(WOA)-MLP,based on collected data of 152 blasting events in three open pit granite mines in Johor, Malaysia.

Fathipour-Azar (2021) presented Gaussian process (GP), K-star,RF, and XGBoost models as well as several alternative data-driven methods for evaluation of joint roughness coefficient (JRC) values,based on 112 rock joint profile datasets.

Jamei et al. (2021) developed kernel extreme learning machine(KELM)to predict flyrock distance based on blasting data obtained from three quarry sites in Malaysia.

Xie et al.(2021)established different ANFIS-based models optimized by multi-verse optimizer(MVO),equilibrium optimizer(EO),simulated annealing (SA), and Henry gas solubility optimization(HGSO)to investigate the stability of the roadways in underground coal mines exploited by longwall mining.

Miah (2021) proposed data-driven connectionist models with least square support vector machine(LSSVM)using a global optimization, coupled simulated annealing (CSA) for prediction of shear wave velocity for clastic sedimentary rocks.

Zhang et al. (2021) presented various ML-based methods including ensemble learning (EL)method, decision tree regression(DTR)and conventional regression for prediction of the maximum ground surface settlement induced by braced excavation in anisotropic clays.

Yang et al. (2021) formulated surrogate models based on gated recurrent unit (GRU) neural network and Nesterov-accelerated Adam (Nadam) algorithm for prediction of braced excavation deformation.

Atangana Njock et al. (2021) developed ANN model optimized by differential evolution (DE) for robust and reliable prediction of jet grouted column diameters based on 209 filed records of single,double and triple jet grouting systems.

Chwala and Kawa(2021)presented a ML-based surrogate model based on a random failure mechanism method (RFMM) for more efficient bearing capacity assessment.

Though this SI has assembled 22 of the leading research papers in the field,and provided an opportunity to assess progress to date,the authors would like to propose the future developments and challenges of deep learning and soft computing in geotechnical underground engineering herein:

(1) ML models are developed on data while DL methods rely greatly on big data, and by design they do not incorporate any physical law (such as mass and energy balance) and do not extrapolate well beyond the range of the training data.Historically, physical modeling and ML have generally been treated as two different fields or avenues with very different scientific paradigms(theory-driven versus data-driven).Yet,in fact these two approaches are complementary, with physical approaches in principle being directly interpretable and offering the potential of extrapolation beyond observed conditions, whereas data-driven approaches are highly flexible in adapting to data and are amenable to finding unexpected patterns. The foremost challenges facing geoengineers and geoscientists are how to extract as much useful information as possible and how to gain new insights from both ML and numerical/physical simulations as well as the interplay between them.

(2) Development of benchmark dataset. Benchmark dataset,especially dataset of big data,used for ML model calibration and verification is essential. In this regard, several efforts have been made in recent years to catalog and publish benchmark datasets to support new model development and running data science challenges, such as “Kaggle Competitions” (https://www.kaggle.com/), “Open Images Challenge”(https://www.ai.google/),“Tianchi data science competition”(https://tianchi.aliyun.com/), “Lotus contests for TBM predictive performances” organized in China, “Data science competitions to build a better world” organized by DRIVENDATA company (https://www.drivendata.org), etc. While development and sharing of benchmark training datasets are essential,research and investment in techniques that require less training data (such as semi-supervised learning and active learning) are required.

(3) For geoengineering of various complexities, different MLs and optimizations should be adopted for cross validation.For geoengineering of multivariate big data with noises, algorithms of error correction and considerable robustness are desirable. While for problems of limited features with incomplete inputs, the choice should be quite different.Based on Phoon et al.(2019),the MUSIC-X characteristics of geotechnical data (MUSIC-X: multivariate, uncertain and unique,sparse,incomplete,and potentially corrupted,with X denoting the time and/or spatial dimension)decided that the adoption of the methods should not be blind.In other words,optimal results can never be achieved by blindly applying ML models.

(4) Another trend in application of geoengineering and geosciences is the visual computing,which handles with images and 3D models, i.e. computer graphics, image processing,visualization, computer vision, virtual and augmented reality, video processing, but also includes aspects of pattern recognition, human computer interaction, ML and digital libraries. The core challenges are the acquisition, processing,analysis and rendering of visual information(mainly images and video).Wang et al.(2021)developed CNN EXPLAINER to help users more easily understand the inner workings of CNNs,which can run locally in users’web browsers without the need for installation or specialized hardware,broadening the public’s education access to modern deep learning techniques.Recently Stefan Sietzen successfully realized the CNN visualization via Unity 3D and plenty of codes(https://vimeo.com/stefsietz).

(5) It is essential to regulate the standards for data collection,validation,testing and analysis in establishing the systematic criterion of original data acquisition. Recently, Liu et al.(2022) developed a software system for big data management, which fulfills the tasks of collecting, transmitting,classifying, screening, managing and analyzing, based on more than 80,000 sets of standard geo-material physicomechanical data, laying a good example for data standards.In particular, primary data in engineering practice are required to be in conformity with the uniform management standards, which needs to be mandatory by the official management departments and authorities (data privacy).Otherwise, the research progress will be stagnated due to absence of sufficient data samples for ML regardless of great efforts in algorithm development.Consequently,engineering application of ML,DL and optimization is not only a scientific issue, but also a management and system topic requiring cooperation of multiple parties involved.

Again, we are grateful to the authors for their generous contributions and patience during the review process. Our heartfelt thanks also go to the dedicated reviewers for their useful comments.