Physics-Informed and Explainable Machine Learning for Expansive Soil Swell–Shrink Dynamics: A Critical Review of Computational Architectures, Constraints, and Validation Paradigms in Geotechnical AI
Chukwuemeka Uchenna Anosike
*
Information Technology and Project Management, University of Denver, Colorado, United States.
Chidera Kingsley Eze
Civil Engineering Department, School of Computing, Engineering & Digital Technologies, Teesside University, Middlesbrough, United Kingdom.
Onyeama Odirachukwu Godswill
Department of Civil Engineering, Enugu State University of Science and Technology, Enugu State, Nigeria.
*Author to whom correspondence should be addressed.
Abstract
Expansive soil prediction involves more than estimating a swelling index from classification properties. Engineering decisions depend on path-dependent deformation, evolving water retention, restraint, fabric change and uncertain environmental forcing. Physics-informed and explainable machine learning offer complementary ways to address these difficulties, but their benefits depend on what is constrained, what is observed and how generalisation is tested. This critical narrative review integrates expansive-soil mechanics, direct machine-learning applications, scientific machine learning, explainability and predictive validation. Literature eligible through 29 July 2026 was selected through live scholarly searching, bibliographic verification and qualitative methodological appraisal. The synthesis distinguishes endpoint prediction, constitutive response learning, hydraulic inversion and field-scale forecasting rather than treating them as interchangeable tasks. Direct evidence includes recurrent learning constrained by a void-ratio-dependent retention relationship, interpretable swelling-index models and geographically broader validation of swelling–consolidation envelopes. These advances do not establish general reliability for multicycle, field-scale swell–shrink dynamics. Four recurrent weaknesses limit interpretation: heterogeneous test labels, dependence within compiled datasets, confusion between empirical regularisation and governing physical laws, and attribution analyses that are mistaken for causal explanations. Conservation-based networks and thermodynamic architectures provide stronger structural controls, yet remain vulnerable to constitutive misspecification, parameter non-identifiability and optimisation failure. A mechanism–constraint–validation alignment framework is developed to connect each intended engineering claim to suitable observables, architectural restrictions and independent tests. Priorities include specimen- and source-separated benchmarks, controlled hydromechanical reversal experiments, explicit treatment of model discrepancy, calibrated uncertainty under domain shift and prospective field evaluation. The most defensible near-term role is an auditable hybrid component within conventional geotechnical investigation and analysis, not an autonomous replacement for laboratory testing or engineering judgement. Progress should be assessed through transferable physical fidelity and decision-relevant reliability rather than isolated improvements in predictive fit.
Keywords: Unsaturated soil mechanics, hydraulic hysteresis, constitutive learning, model discrepancy, Shapley attribution, domain shift, soil-structure interaction