Machine Learning‐Enhanced Ground Motion Prediction Incorporating 2D Pile–Soil–Structure Interaction Effects for Seismic Site Amplification Assessment

ABSTRACT The Pile–Soil–Structure Interaction (PSSI) effect plays a critical role in seismic wave propagation from bedrock to the ground surface, significantly influencing site amplification and introducing uncertainty in structural input motions. However, its complex nonlinear behavior has traditionally been difficult to capture in empirical ground motion prediction models. To address this, the present study develops a machine learning‐enhanced framework for seismic site amplification prediction that explicitly incorporates PSSI effects. A verified 2D finite element model of a pile–soil–structure system was used to generate a dataset comprising 88,023 simulations under varying ground motion, structure and site conditions. Out of 11 input features, seven key parameters were selected, including structural period ( T ), pile‐to‐structure stiffness ratio ( K ratio ), bedrock PGA, and soil shear wave velocity parameters. Five machine learning algorithms—Backpropagation Neural Network (BP), Random Forest (RF), Support Vector Regression (SVR‐ε and SVR‐ν), and Genetic Algorithm‐optimized BP (GA‐BP)—were trained and evaluated using multiple metrics (R 2 , RMSE, MAE, MAPE, and Dynamic Time Warping). The GA‐BP model outperformed others in predictive accuracy and was further interpreted using Shapley Additive exPlanations (SHAP) to gain physical insight into parameter influence. Results indicate that bedrock PGA controls short‐period amplification but wanes at longer periods; bedrock depth ( Z 1.0 ) dominates long‐period amplification, with peak effects at intermediate depths; the site fundamental period ( SFP ) drives resonance when matching the predominant input period; V s30 governs short‐period response; impedance contrast ( V ratio ) enhances amplification when exceeding 1.36; K ratio shows a threshold at 2.4; and T exhibits a period‐matching effect with contributions shifting from negative to positive. This study presents a physically interpretable, data‐driven framework for modeling seismic site amplification under PSSI conditions. The results provide valuable insights for ground motion modeling, site classification, and performance‐based seismic design of pile‐supported structures, bridging the gap between numerical modeling and practical engineering applications.