Instability risk assessment of submarine strata based on data-driven modelling and interpretable machine learning algorithms

Submarine stratum instability poses a potential threat to subsea tunnels and other marine infrastructure in geologically complex areas. This study develops an interpretable data-driven framework for assessing submarine stratum instability in the Miaodao Islands region of the Bohai Strait, China. A database of 120 borehole-derived and laboratory-tested submarine-soil samples was established using ten geological, physical, and mechanical indicators. GA-optimized K-means clustering was employed to extract instability-risk grades from the mechanical parameters and establish a four-level grading system. Based on the derived labels, BP, SVM, ELM, and PSO-ELM models were developed and compared. Their testing accuracies were 0.444, 0.500, 0.444, and 0.639, respectively, with PSO-ELM achieving the best overall performance and a macro-averaged F1-score of 0.641. SHAP analysis showed that submarine-stratum instability is jointly controlled by geological structure, burial conditions, physical state, and mechanical properties. In the PSO-ELM model, void ratio, depth, and elastic modulus contributed approximately 23.5%, 19.0%, and 18.7%, respectively, while the two fault-related variables together contributed approximately 26%. These results highlight the importance of fault-influenced, deeply buried, loose, and low-stiffness strata in subsea engineering. The proposed framework integrates data-driven risk-grade extraction, nonlinear prediction, and interpretable analysis, providing a quantitative basis for submarine-stratum assessment and risk-informed decision-making.