A Unified Probabilistic Explainable AI Framework for Reliability Assessment of Caisson Foundations under Spatially Variable Soils

This study proposes a comprehensive probabilistic intelligent framework for evaluating the reliability and safety of Caisson foundations by integrating Monte Carlo Simulation (MCS), Subset Simulation (SS), and hybrid machine learning techniques. The probabilistic framework is based on lognormal random fields with Cholesky decomposition to realistically account for the subsurface uncertainty, the spatial variability of key soil parameters, namely angle of internal friction (ϕ) and unit weight (γ). The reliability analysis is performed by MCS and SS methods to evaluate the failure probability and the reliability index. To improve the predictive capability, LSSVM models optimised by PSO, HHO and GA are built to predict the factor of safety. Among the proposed models, LSSVM-HHO exhibits the best prediction accuracy, robustness and generalisation performance. SHAP-based interpretability offers insight into the effect of governing parameters, thus improving model transparency and engineering trust. The results show that an increase in the variability of soil parameters increases the probability of failure and decreases the reliability of the system, which indicates the need for probabilistic assessment in foundation design. The proposed framework offers a unified approach that combines reliability analysis, advanced optimization, and explainable artificial intelligence, providing a robust decision-support tool for safe and efficient geotechnical design under uncertainty.