Explainable machine learning surrogate for rapid prediction of abnormal settlement in high-filled canal embankments

• An RSM-FEM-XAI framework is proposed for rapid prediction and user-friendly interpretation. • The framework is verified on a real-world abnormal settlement of high-filled canal embankment. • Feature selection is conducted using ANOVA, followed by XGBoost as a surrogate for FEM. • SHAP analysis interpretably evaluates impact of input features on canal embankment settlement. • A user-friendly web-based platform is developed to interactively implement prediction results. High-filled canal embankments are prone to abnormal settlement during service, risking urban lifeline infrastructure. Finite element (FE) analysis captures hydro mechanical processes and settlement, but it is time consuming and opaque to non-specialists/decision-makers. This study introduces a surrogate model that retains FE‑level fidelity while enabling settlement prediction with engineering-grade accuracy. A calibrated FE model of a 12-m‑high canal embankment was coupled with central composite design to generate 2,002 synthetic scenarios with eight systematically varied features: groundwater depth, groundwater drawdown, mound height, elastic modulus, Poisson’s ratio, cohesion, internal friction angle, and the position at high-filled canal embankment. Feature selection was conducted using analysis of variance (ANOVA), followed by training an extreme gradient boosting (XGBoost) model to serve as a surrogate for the FE model. The surrogate achieved R² = 0.98, RMSE = 0.02 m, MAE=0.01m, MAPE=5.81%, and RMSRE=8.93%, cutting per‑scenario runtime from nearly 30 min (FEA) to less than 1 s. Shapley additive explanation (SHAP) analysis exposed groundwater drawdown as the dominant driver and reveal actionable thresholds (e.g., drawdown > 5.7 m markedly increases predicted settlement), with notable interactions between mound height, elastic parameters, and local position; cohesion and internal friction angle exerted secondary but non‑negligible effects consistent with soil mechanics. The trained model is publicly available at https://gtzhdqhhko3d3bktnuugfw.streamlit.app/ through an web-based graphical user interface platform, which accepts user-defined input variables and instantly provides settlement predictions along with SHAP-based interpretability diagnostics. The proposed framework provides a practical pathway for continuous settlement monitoring of high-filled embankments and enhancing regional water security and transport resilience.