Speaker
Description
Ground motion prediction for intermediate depth earthquakes remains challenging as conventional log linear empirical equations cannot capture complex source, path, and site interactions. This study presents the first machine learning ground motion model for intermediate depth earthquakes in Romania, using 1,627 recordings from 156 events (M 4.0-5.98, 2006-2025) across 19 stations operated by the National Institute for Earth Physics (INFP). The eXtreme Gradient Boosting (XGBoost) achieved R² = 0.885 for peak ground acceleration (PGA) and R² = 0.875 for peak ground velocity (PGV), outperforming the empirical baseline (R² = 0.495). Residual σ = 0.291 falls within the range of published models (0.28 to 0.35). Temporal validation on post 2020 events yielded R² = 0.724. A ResNet18 CNN on wavelet scalograms achieved R² = 0.684, lower than XGBoost. Three explainability methods, SHapley Additive exPlanations (SHAP), LIME, and Grad-CAM, were applied independently. SHAP identified magnitude and epicentral distance as dominant predictors, with azimuthal dependence consistent with Moesian Platform waveguide amplification. Grad-CAM revealed ResNet18 focused on the S wave arrival window. Convergent findings establish a triple XAI framework for seismic hazard assessment in data scarce intermediate depth regions.
Keywords: ground motion prediction; XGBoost; deep learning; explainability; SHAP; Vrancea; intermediate depth earthquakes