A Deep Learning-based Workflow for the Automated Focal Mechanism Determination in Italy

The determination of earthquake focal mechanisms is crucial for understanding tectonic processes and assessing seismic hazard. While traditional methods based on manual polarity picking have proven effective, they are time-consuming and potentially subject to analyst’s bias, particularly for small to moderate earthquakes. This study presents a machine learning-based approach for automating focal mechanism determination through P-wave polarity prediction. We develop and validate a methodology that combines a Convolutional Neural Network (CNN), based on the work by Ross et al. (2018) with the SKHASH algorithm (Skoumal et al. 2024), focusing on Italian earthquakes with magnitudes between 0 and 4.0, spanning the country's major active tectonic settings. The CNN, trained on a subset of the INSTANCE catalog by Michelini et al. (2021) containing over 450,000 Italian seismic waveforms, achieves 96% accuracy in polarity classification. To ensure the reliability of prediction, we implement stringent probability thresholds and validate our approach using a test dataset of 240,335 waveforms from 15,420 events. We then apply the methodology to 300 earthquakes with existing Time Domain Moment Tensor (TDMT) solutions finding a median Kagan angle (Kagan, 1991) between our FM and the TDMT solutions of 28.7°. This automated approach proves to be effective and may, in the future, offer a reliable means of determining focal mechanisms for lower magnitude events, where moment tensor inversions might not be feasible. This would contribute to a better understanding of regional seismotectonics and enhance seismic monitoring capabilities.