Deep learning framework for determining earthquake focal mechanisms with application to the Campi Flegrei caldera.

We present DeepFoc, a novel deep learning framework for estimating earthquake focal mechanisms from P-wave polarities and amplitudes, specifically designed for low-to-moderate magnitude events in complex tectonic settings. Trained entirely on synthetic data generated by using an on-the-fly simulation strategy, DeepFoc learns the nonlinear relationship between seismic observables and focal parameters (strike, dip, rake). We apply it to the Campi Flegrei caldera (Italy), a densely populated volcanic area experiencing ongoing unrest. Compared to classical inversion methods, DeepFoc provides more stable and accurate solutions, especially under degraded data conditions. Validation against synthetic and real seismic events shows improved agreement with observed data and better robustness to noise and incomplete inputs. By capturing focal mechanism variability at small scale, DeepFoc demonstrates high sensitivity and generalisation ability. Its computational efficiency and performance make it suitable for real-time integration in seismic monitoring workflows, offering timely source characterisation during seismic swarms and volcanic crises.