A Machine Learning Framework for Enhanced Site-Specific Ground Motion Modeling

This study presents a machine learning (ML) model aimed at capturing local site effects on seismic ground motion. Synthetic seismic spectrums are first generated using moment tensor solutions and a Green's Function Database from Pyrocko. Residuals between observed and synthetic data are computed in octave frequency bands, reflecting deviations introduced by site-specific conditions. These discrepancies are then modeled using a feedforward neural network trained on both normalized synthetic spectrums and site-specific parameters (e.g., bedrock depth, average shear-wave velocity, fundamental frequency). We demonstrate the effectiveness of this approach by applying it to Japan’s complex seismic environment, using strong-motion records from the K-NET and KiK-net networks. Once trained, the model accurately predicts and corrects these discrepancies, reconstructing spectrums that closely match real observations. This approach not only significantly enhances the interpretation of seismic data but also boosts earthquake hazard prediction in regions with complex site-effects. Overall, this framework provides a powerful tool for reducing the gap between simulated and actual ground motion, ultimately improving the reliability of seismic risk assessments.