The knowledge of the crustal stress field is crucial for understanding the seismic activity in an area that, in turn, requires an in-depth knowledge on the dynamics of the crust. To that end, the reconstruction of focal mechanisms of earthquakes as reliable as possible is a preliminary and basic requirement to infer proper source mechanisms. Currently, the fault plane solution method, using P-wave polarities, is still frequently used. Anyway, manually determining the polarities of P-waves is time-consuming and susceptible to human error. These issues can be solved by automated processes thorough the application of machine learning techniques.In our study, the Convolutional First Motion (CFM) network, a Deep Convolutional Neural Network, is presented. It is utilized to categorize seismic traces based on the polarity of the P-waves' first motions. We used waveforms from two datasets: the Italian seismic catalogue INSTANCE and waveforms from earthquakes that occurred in the Mount Pollino region of Italy between 2010 and 2014.We developed a method based on Principal Component Analysis and Self-Organising Maps, which enabled a clustering process to identify sets of appropriate traces. The network was trained using 130·000 time windows centered on P-wave arrival times relative to waveforms in the INSTANCE catalogueThe network achieved accuracies of 95.7% and 98.9% on two test sets that were generated using the datasets for Mt. Pollino and a portion of the INSTANCE catalogue, respectively.This work has been partially supported by PRIN-2017 MATISSE project, No 20177EPPN2, funded by Italian Ministry of Education and Research.

The 28th IUGG General Assembly (IUGG2023) (Berlin 2023)