An Interpretable XGBoost-SHAP framework for rapid regional assessment of liquefaction-induced lateral spreading in mildly sloping ground

Liquefaction-induced lateral spreading poses significant risks to infrastructure on mildly sloping ground during earthquakes. This study develops an interpretable machine learning (ML) framework for rapid regional-scale prediction of liquefaction-induced displacement using an Extreme Gradient Boosting (XGBoost) approach. A comprehensive database is constructed from nonlinear fully coupled finite element simulations calibrated against centrifuge test data and expanded using Latin Hypercube sampling to generate 50,000 cases. Based on this database, an XGBoost model is trained using key geometric and soil parameters together with cumulative absolute velocity (CAV), with hyperparameters optimized via grouped 5-fold cross-validation. Furthermore, the XGBoost model is validated using five centrifuge tests not used for ML training, providing additional experimental validation beyond the simulation database. To interpret the relationships between input features and model predictions, SHapley Additive exPlanations (SHAP) analysis is conducted, revealing CAV, relative density, and liquefiable layer thickness as the dominant factors, with higher CAV and layer thickness and lower relative density generally associated with increased predicted displacement. Ultimately, the validated model is applied to a hypothetical regional domain to demonstrate its capability for rapid regional-scale prediction of liquefaction-induced displacement and seismic fragility assessment under different damage states. Overall, the developed XGBoost-SHAP framework provides a computationally efficient and interpretable modeling approach for regional-scale liquefaction hazard assessment, facilitating rapid seismic vulnerability evaluation and supporting performance-based geotechnical design.