Bridging Heterogeneous Experimental Data and Soil Mechanics: An Interpretable Machine Learning Framework for Displacement-Dependent Earth Pressure

Classical earth pressure theories often struggle to account for the complex coupling effects of wall displacement and spatial non-uniformity under non-limit states. This study presents an interpretable machine learning framework designed to extract universal mechanical laws from heterogeneous experimental datasets. Using a multi-source database of rigid retaining walls with sandy backfill, a three-stage feature refinement strategy is proposed that incorporates Recursive Feature Elimination, Collinearity Analysis, and Interpretability Comparison to identify a parsimonious set of five fundamental physical parameters. A SHapley Additive exPlanations-Categorical Boosting (CatBoost-SHAP) framework is established to predict the active earth pressure coefficient (K) and interpret the underlying mechanisms across various movement modes (RB, RT, and T). Results demonstrate that the model effectively captures the progressive evolution of shear bands and the soil arching effect. Specifically, a critical displacement threshold of Δ/H ≈ 0.006 is identified, marking the transition from mode-dominated stress non-uniformity to magnitude-driven limit states. Leave-One-Dataset-Out Cross-Validation (LODOCV) confirms the model’s ability to maintain physical consistency over purely statistical fitting despite significant inter-literature heterogeneity. Finally, a Graphical User Interface (GUI) is developed to facilitate rapid, displacement-based design in engineering practice. This research bridges the gap between empirical laboratory observations and generalized mechanical logic, providing a data-driven foundation for refined geotechnical design.