Shaking Things Up with Deep Learning: From Training to Model Benchmarking and Seismic Catalog Building

Deep learning is playing an increasingly important role in seismology, particularly for the automatic picking of P and S waves, enabling the creation of more complete and higher-quality seismic catalogs. This work explores both the potential and the challenges of applying these models in different seismic contexts. We present two case studies focused on catalog building: the 2009 L’Aquila earthquake sequence and the 2023 Mugello swarm in Tuscany. In the L’Aquila case, we constructed a new catalog using deep learning techniques and compared it with the reference catalog obtained through traditional methods (Valoroso et al., 2013), highlighting improvements in the number and quality of phase picks. For the Mugello swarm, we generated a one-week catalog to investigate the aftershock migration pattern. We also examine how training datasets influence model performance. For the Fucino Basin, we compare catalogs produced using models trained on a California dataset versus the Italian INSTANCE dataset, showing the impact of regional training data on phase picking accuracy. Finally, we address the complexity of training neural networks in seismology, particularly when dealing with imbalanced datasets like AQ2009, where P-phase picks significantly outnumber S-phase picks. We present key training tests and introduce new functions and strategies that improve model robustness in these conditions.