Interpretable machine learning for predicting cumulative plastic strain of marine soils under multi-factor coupling effects

Accurate prediction of cumulative plastic strain (CPS) is essential for assessing the deformation stability of marine strata subjected to artificial ground freezing (AGF). This study aims to clarify how variations in clay content and pore-water salinity regulate cyclic responses after freeze–thaw (FT) disturbance. Cyclic triaxial tests were conducted to investigate CPS under coupled effects of clay content, salinity, FT disturbance, and cyclic stress ratio (CSR). The resulting observations were used to develop Bayesian-optimized support vector regression (SVR), random forest (RF), extreme gradient boosting (XGBoost), and categorical boosting (CatBoost) models using grouped cross-validation, followed by SHapley Additive exPlanations (SHAP) interpretation of the optimal model. Clay content and salinity exerted coupled and nonmonotonic effects on CPS. Within the test conditions, the salinity associated with minimum CPS shifted from 2% to 1% with increasing clay content. FT reduced CPS at clay contents of 10% and 15% but amplified it at 20% and 30%, indicating a reversal between 15% and 20%. CatBoost achieved the best predictive performance, while SHAP identified loading cycles, CSR, and clay content as the dominant contributors. These findings support the long-term stability assessment of heterogeneous marine strata in subsea and coastal underground engineering.