Machine learning-based ground motion model with finite-fault distance metrics

Abstract This study introduces a ground motion model based on a machine-learning approach, the Extreme Gradient Boosting (XGBoost) algorithm, to enhance near-fault seismic predictions. The model is trained on over 7700 strong-motion records from the ITalian ACcelerometric Archive (ITACAext 2.0) database. The machine-learning model integrates predictors representing source characteristics (Mw, style of faulting), path effects including finite-fault distance metrics and site effects. The model is developed through a systematic evaluation of multiple predictor combinations and optimized using Bayesian hyperparameter tuning within a Leave-One-Event-Out cross-validation framework. Model performance is assessed on four moderate-to-strong Italian earthquakes (Mw 6.0–6.6) and is compared against the Italian ground-motion regression model (ITA18). The machine learning-based approach yields lower root mean square error values across all tested ground motion intensity measures, particularly for peak ground acceleration, peak ground velocity, and spectral accelerations at short and intermediate periods, and provides a better agreement with observations than ITA18, with about a 56.54% overall RMSE reduction and roughly 41.70% for stations above the threshold. Feature importance, assessed using Shapley values, varies with period. Geometric and distance-related parameters are dominant at short periods, whereas magnitude and site effects, including deeper structure, become more relevant at longer periods. Spatial patterns of Shapley values align with physical expectations, highlighting rupture directivity, hanging-wall effects, and near-fault saturation. The combination of physically grounded predictors with interpretable machine learning offers a robust and flexible framework for ground-motion prediction, with potential applications in near real-time seismic emergency response and risk-based decision-making.