• ground motion
• machine learning
• site effects prediction

This thesis investigates how local site effects influence seismic ground motion, with the objective of improving spectral ground motion predictions. Focusing on Japan, where high-quality seismic data from KiK-net and K-NET are available, the study employs a machine learning (ML) framework to isolate and model site-specific contributions to observed spectral variability. The methodology begins by generating synthetic spectra (source and path-related) using the Pyrocko Waveform Simulation (PWS) framework, which relies on Green's function databases and moment tensor solutions. These simulations are computed at real station locations and represent the theoretical ground motion spectra that include source and path effects, but not site amplification. Observed spectra, on the other hand, are extracted from recorded seismic events at the same locations. The difference between the observed and synthetic spectra is defined as the site-related spectral value, which captures the amplification effects due to local geological and geotechnical conditions. To model these site-related spectral values, a ML approach is developed using a fully connected neural network implemented in TensorFlow. The input features include normalized synthetic spectra and site-specific parameters such as bedrock depth, average shear-wave velocity, fundamental frequency, maximum spectral ratio, epicentral distance, and geological/geographical information primarily sourced from J-SHIS. The model is trained to predict site-related spectral values across nine frequency bands, effectively learning the nonlinear relationships between local site proxies and amplification patterns. Once trained, the network predicts the site-related spectral values, which are then added back to the synthetic spectra (source and path-related) to reconstruct the full predicted spectra. This allows the model to bridge the gap between theory and observation by correcting synthetic spectra for local site effects. To validate the model’s generalizability, it was also applied to another dataset from Europe provided by the European Strong-Motion (ESM) database. Despite the smaller dataset (compared to the Japanese one) and differing geological conditions, the model showed good performance. It was able to predict site-related spectral values that, once added to the synthetic spectra, yielded reconstructed spectra closely resembling the observed ones. This demonstrates that the model retains its predictive capability even under reduced data availability and across different regional contexts. The results demonstrate that the ML framework accurately reproduces predicted spectra that successfully capture the influence of complex site conditions. Since the framework has been successfully validated using other data from a different geographical context, it could theoretically be applied to other regions worldwide, offering a scalable approach for site effect modeling in areas with limited data availability This work provides a robust basis for future applications in ground motion modeling and seismic hazard assessment in regions affected by strong site amplification.