Special issue on “Machine learning and AI in geotechnics”

The potential for machine learning and artificial intelligence to shape geotechnical engineering practice (and possibly theory) is immense. However, the agenda for machine learning in geotechnics should not be focused on applying or developing algorithms alone. The geotechnical context that gives rise to the data is important. The context can be related to statistics, physics, or experience. Statistics refer to the attributes of geotechnical data that depart significantly from the assumptions in classical statistics (large sample size, spatial/temporal/parametric independence, homogeneity, normality, etc.). Phoon, Ching, and Shuku (2022a) argued that geotechnical site data are “ugly”, because they are spatially varying, sparse, site-specific (or unique to some extent), and incomplete in the sense that a multivariate database is full of empty entries denoting lack of some measurements at certain locations/depths. The incompleteness attribute arises from an intent to maximize information on cross correlations between different soil parameters and geotechnical/geologic spatial correlations across a given site while minimizing the site investigation budget. At this point, this value of information optimization is an art rather than a science. The scientific challenge to draw useful insights from MUSIC-3X (Multivariate, Uncertain and Unique, Sparse, Incomplete, and potentially Corrupted with “3X” denoting 3D spatial variability) data was thought to be intractable until recently (Phoon, Ching, and Shuku 2022a). These “ugly” data attributes are the norm in a site investigation report. In rock engineering, data can be categorical rather than numerical. Phoon (2023) emphasized that “decision making in every discipline is supported by its own data with unique attributes and a tradition of successful practice (investigation, design, construction, testing, monitoring, and risk management methodology) that evolved to make the best use of these data and prevailing technologies”.

Physics refers to a body of rational knowledge that associates a “number” to “meaning”. Decisions supported by physics-informed results are “explainable” and “interpretable”. The finite element method is the most prevalent embodiment of physics in geotechnical engineering. Using finite element analysis, an engineer understands cause and effect (interpretability) and knows which input parameters affect the outputs (explainability). An engineer distinguishes between material and state parameters, between effective and total stress parameters, and between input and output parameters from a physical or numerical model. These distinctions exist when one approaches data from the lens of physics.

Experience refers to a body of empirical knowledge accrued from deliberate practice. It is restricted by the range of projects encountered by an engineer over his/her working life and it cannot be shared with other engineers efficiently. In contrast to statistics and physics, it is mainly subjective and qualitative. Nonetheless, many engineers regard experience as critical. For example, Simpson (2011) explained why Eurocode 7 (EC7) is worded to ensure an engineer always take full ownership in decision making: “EC7 attempts to do this by making the designer responsible for the selection of the characteristic values of materials, avoiding mathematical prescription of their derivation. Inevitably, such a process leads to values affected by the subjective experience, knowledge and judgement of the designer. The author would contest that it is better to accept such subjectivity than to discard the valuable information it provides”. In machine learning, experience is regarded as one type of “thick data” to distinguish it from the more well-known quantitative “big data”.

Decision making in current practice is based on physics and experience. There is no formal basis underpinning decision making other than qualitative guidelines such as Burland’s Triangle (Burland 1987; Phoon et al. 2022b) or Wroth rules (Wroth 1984; Phoon 2023). As such, geotechnical practice is regarded more of an “art” than a “science”. Phoon (2023) argued that decision making will be increasingly data-informed given the increasing power, ubiquity, and convergence of digital technologies and presented a data-informed decision support index (DIDI) to track this evolution. Geotechnical reliability is regarded as one stage with DIDI = 3 to 4. Higher DIDI stages will involve broader and deeper exploitation of data for decision making: (1) at real-time and system wide scale, (2) in hybrid intelligence (human, machine) mode, (3) for a specific project (cf. “precision construction” in Phoon 2018), and (4) to address more complex design goals beyond safety and economy (e.g., sustainability and resilience).

Foresight review

The Special Issue on “Machine Learning and AI in Geotechnics” is led by a foresight review paper that attempts to contextualize the future of machine learning in geotechnics against the distinctive features of geotechnical engineering practice that exist for good reasons since they have stood the test of time (Phoon and Zhang 2023). The authors opined that machine learning in geotechnics (MLIG) should be approached with an appropriate balance of three elements: (1) data centricity, (2) fit for (and transform) practice, and (3) geotechnical context. The first element dominates the literature in MLIG. The second and third elements need more emphasis. The authors argued that this agenda underpins a new interdisciplinary field (termed “data-centric geotechnics” in Phoon, Ching, and Cao 2022c) involving genuine integration of information, data, techniques, tools, perspectives, concepts, theories, and/or experience from both geotechnical engineering and machine learning in computer science. Numerous foundational challenges beyond methodological ones can be identified when research is pursued under the broader “data first practice central” agenda of data-centric geotechnics. Ultimately, success is measured by how practice is improved or even transformed in the longer term. A simple “value” matrix is proposed to guide future research: (1) Type 1 (incremental value) involving available data and existing conventional applications, (2) Type 2a (potentially high value) involving available data and new applications, (3) Type 2b (high value) involving new data and existing applications, and (4) Type 3 (disruptive value) involving new data and completely novel applications.

The remaining 12 papers present the application of machine learning to site characterization (2 papers), soil behavior (1 paper), soil dynamics (1 paper), landslides (2 papers), tunnelling (4 papers), and mining (2 papers). The value of ML in these papers is likely Type 2a or Type 2b. Four papers exploit physics to improve the interpretability and/or generalizability of ML methods (Li et al. 2023; Shioi et al. 2023; Tan et al. 2023; Yan et al. 2023). Engineers ultimately bear the responsibility of a design decision and they are arguably more comfortable with interpretable ML results. In the geotechnical context, generalizability is in part related to the applicability of ML methods trained at one spatial location (or site) to another location. This is part of the site recognition challenge (Phoon, Ching, and Shuku 2022a).

Site characterization

Given that a knowledge of the site is central to any geotechnical or rock engineering project, data-driven site characterization (DDSC) must constitute one key application domain in data-centric geotechnics, although other infrastructure lifecycle phases such as project conceptualization, design, construction, operation, and decommission/reuse would benefit from data-informed decision support as well. One part of DDSC that addresses numerical soil data in a site investigation report and soil property databases is pursued under Project DeepGeo (Phoon and Ching 2021). In principle, the source of data can also go beyond site investigation and the type of data can go beyond numerical such as categorical, text, audios, images, videos, and expert opinion. Note that the data source is related but not identical to the data attribute. The purpose of Project DeepGeo is to produce a 3D stratigraphic map of the subsurface volume below a full-scale project site and to estimate relevant engineering properties at each spatial point based on actual site investigation data and other relevant Big Indirect Data (BID). An intrinsic DeepGeo challenge is the estimation of a 3D trend function that can be varying smoothly or exhibits jumps at layer boundaries. It is likely to form the basis for an Underground Information Modelling (underground part of Building Information Modelling) and digital twins (Phoon and Zhang 2023).

Shi and Wang (2023) extended research in Project DeepGeo to training image databases. A systematic framework is proposed to develop these databases based on three factors, namely, geological origin, site location, and application scenario. As a pilot study, 54 geological cross-sections, mainly interpreted by experienced engineers, for weathered granite and tuff slopes in Hong Kong are collected and compiled as two training image databases. These collected training images serve as prior geological knowledge for typical weathering profiles in Hong Kong. Following the proposed criteria, multiple compatible training images can be identified and combined with an image-based Bayesian learning method for ensemble learning of subsurface stratigraphy and quantification of stratigraphic uncertainty for a specific site with limited investigation data.

Shuku and Phoon (2023a) proposed a DDSC method for 2D/3D stratigraphic mapping considering geological uncertainty using a Potts model which is a form of Markov random field (MRF) models. The Potts model requires only a few site-specific profiles of soil types. Other geological prior knowledge such as geological cross-sections interpreted by geologists, orientation of strata, and stratigraphic dips are not necessary. The proposed method is demonstrated through benchmark examples (Phoon et al. 2022d; Shuku and Phoon 2023b) and a real case history. The Potts model provides reasonable stratigraphic inference based solely on limited borehole data for both 2D and 3D ground scenarios. The uncertainty of the inference (geological uncertainty) can be also visualized in the proposed method. This visualization is useful, because it directs the engineer to focus on mechanically important but low accuracy zones where the critical failure mechanism passes through. Project DeepGeo focused primarily on the estimation of spatially variable soil properties with smooth trends, rather than the estimation of stratigraphic interfaces demonstrated in this study.

Soil behavior

He et al. (2023) used a multi-fidelity neural network (MRNN) to model the rate-dependent behaviour of soft clays. The MRNN was proposed for reducing the dependency on the number of high-fidelity data (Zhang et al. 2022). First, the low-fidelity (LF) synthetic data generated by an existing elasto-viscoplastic Modified Cam-clay model were used to develop a LF model for capturing the underpinning stress-strain-strain rate relationship. The prediction error of the LF model on the high-fidelity (HF) experimental data of interest is calibrated by an HF model with final refinement by linear regression. The feasibility and generalization of MRNN are verified on Hong Kong marine deposits and Merville clay. It is shown to be more accurate compared to conventional constitutive models.

Soil dynamics

Shioi et al. (2023) conducted a fundamental study on applying dynamic mode decomposition (DMD) for efficient wide-area ground seismic response analysis. DMD is a method that reduces the dimensionality of spatiotemporal information and discovers dynamical systems. However, DMD assumes a linear time-invariant system, which poses a limitation when dealing with non-stationarity and nonlinearity in seismic responses. To overcome these challenges, this study introduced two solutions: DMDc, which adds a control term to DMD to address non-stationarity, and a dynamic deformation model based on domain knowledge from geotechnical engineering to address nonlinearity. Although the study targeted a one-dimensional seismic response analysis, in which the geotechnical and input seismic models are straightforward, it is still reassuring to see that DMDc could properly extract the dynamical profiles. The possibility that DMD can contribute to the realization of data-driven forecasting with high mechanistic interpretability was presented.

Landslides

Xi et al. (2023) found that several deep learning approaches can predict landslide displacement that is an early warning index of an impending failure. Current prediction methods mainly focus on the temporal correlation of the displacement data and ignore the spatial correlation. To fully consider the spatiotemporal features of displacement data, this paper first established a fully connected graph to obtain the spatial correlation of multiple monitoring points and then developed three deep learning models, temporal graph convolutional network-long short term memory (TGCN-LSTM), temporal graph convolutional network-gate recurrent unit (TGCN-GRU), and Attention-temporal graph convolutional network (Attention-TGCN), based on graph convolutional networks to learn the complex nonlinear features of the time-series displacement data of the Huanglianshu landslide. The trained models accurately predict the landslide displacements, with a maximum coefficient of determination (R²) of 0.85 at the individual monitoring points.

Zhang et al. (2023) compiled a global landslide database with displacements of 55 slopes measured up to the time of failure. The time of slope failure is often predicted based on displacement measurements using the inverse velocity methods. Based on such a comprehensive database, the performances of the linear and nonlinear inverse velocity methods are assessed. It is found that the non-linear inverse velocity method may be subjected to numerical pitfalls like a ill-conditioned Hessian matrix and a saddle point. However, the linear inverse velocity method is free of these two pitfalls. Overall, the linear inverse velocity method is not only more stable, but also more accurate. The results provide insights on how to predict slope failure time based on displacement measurements.

Tunnelling

Li et al. (2023) proposed a new approach for sharing the knowledge of tunnel boring machine (TBM) data from various projects. By adopting some mechanical conversion relations, TBM data from various projects can be normalized and then used to build a machine learning model. A massive dataset consisting of over 20 billion records in the Yin-Song Diversion Project in China is used to train a machine learning model of TBM performance prediction following this procedure. Further, the machine learning model is applied to an on-going TBM construction project with limited data and provides reliable predictions of key rock breaking parameters such as cutterhead torque and cutterhead thrust. This study highlights the importance of developing new approaches to normalize TBM data and improve the accuracy of machine learning models. This approach has the potential to be applied to other TBM projects, helping to reduce costs and improve efficiency in the construction of tunnels.

Ma et al. (2023) proposed a real-time intelligent classification method for drill and blast tunneling in rocks using a machine learning algorithm. A new database is compiled from 286 case studies in China. The data include rock hardness, weathering degree, rock mass structure, structural plane integrity, rock mass integrity, in-situ stress condition, groundwater condition, and surrounding rock grades. The intelligent classification method for the surrounding rock is established using machine learning, which is shown to be robust in learning and generalizing for small and imbalanced samples. A tunnel information management system for drilling and blasting tunnels is built by combining the intelligent classification method and cloud technology, which is applied to the Lexi Expressway project. The tunnel information management system can provide a timely and accurate reference for the dynamic design and construction of tunnels and are of practical significance in promoting the development of tunnel informatization and intelligence.

Yan et al. (2023) noted that developing a generalized model for axial displacement prediction for immersed tunnel joints is challenging, because of the complex geo-environment and limited monitoring data. This is a typical small data problem in geotechnics since only a few features are available for training machine learning models. The one-year monitoring data in this study including axial displacement, structure temperature, and water level show seasonal trends and their correlations vary spatially along the tunnel. The study proposed a hybrid physical data (HPD) informed Deep Neural Network (DNN) to improve spatial generalization for axial displacement prediction. The HPD is created based on physical analysis and contributes to the DNN as a substituting feature rather than an additional feature. The proposed model is shown to outperform conventional regression methods in terms of spatial generalization with improved physical interpretation.

Zhou et al. (2023) proposed a tunnel lining crack detection algorithm named improved You Only Look Once version X (YOLOX), which can detect tunnel cracks in a complex environment efficiently and accurately. The images taken during the damage detection study phase conducted in several tunnels are compiled and expanded to obtain a tunnel crack damage image dataset that can represent the complex physical state of the tunnel. These crack images can be used for model training and testing, and after experiments, the improved YOLOX model can identify these images and can achieve high speed, high accuracy, and real-time dynamic detection of tunnel cracks.

Mining

Morgenroth et al. (2023) observed that stress model updating for mine-scale finite difference models is generally time consuming and tedious, requiring a large volume of engineering hours and computational power. Their study presents a novel machine learning approach, using a microseismic database, to update a stress model at Garson Mine in Ontario, Canada. A Long-Short Term Memory (LSTM) network is trained on the microseismic events and the geomechanical parameters from a previously manually calibrated FLAC3D model. Two LSTM networks are trained for comparison, one to predict the principal stresses and one to predict the six-component stress tensor. The predicted values are compared to the values computed by the FLAC3D model (“ground truth”).

Tan et al. (2023) mapped the spatial stress distribution over a coalmine roof by using a transfer convolutional neural network (TCNN) model. A convolutional neural network (CNN) model is first pre-trained based on the simulated data from the finite difference method. This CNN can be refined for a specific coalmine roof using limited structural health monitoring (SHM) data through transfer learning. The SHM data are recorded by a stress sensor installed in the roof of the Dongtan coalmine over a range of 120 m ahead of the working face. This study demonstrated that practical site-specific predictions can be made even in the presence of limited field measurements by adopting a physics-informed rather than pure data-driven approach.

ISSMGE TC309/TC304/TC222 Third Machine Learning in Geotechnics Dialogue

This special issue concludes with a report summarizing the discussion outcomes from the ISSMGE TC309/TC304/TC222 Third Machine Learning in Geotechnics Dialogue (3MLIGD), which was hosted online by the Norwegian Geotechnical Institute on 3 December 2021 (Phoon et al. 2023). There is a consensus that the potential of digital transformation in geotechnical site characterization is significant. Digital transformation is expected to change the “rules of the game” in the context of Industry 4.0 that is rapidly evolving in tandem with emerging technologies. The 3MLIGD covers the emerging agenda in data-centric geotechnics, challenges, ideas for collaboration, and a glimpse of current and future developments with a focus on data-driven site characterization (DDSC). It provides some recommendations for developing data-centric geotechnics revolving around some basic issues: data access, data standardization, data quality, data protection, and education. The 3MLIGD concluded that practitioners and researchers who are interested in data-centric geotechnics can achieve greater impact and hasten progress by fostering more collaborations and working more closely together on or shadowing real industry projects.

The guest editors thanked the authors and the reviewers for their valuable contributions to the emerging field of data-centric geotechnics.

Correction Statement

This article has been corrected with minor changes. These changes do not impact the academic content of the article.