GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 1.1 GNGTS 2024 Supervised and unsupervised machine learning approaches for identfying the preparatory process of moderate earthquakes at The Geysers, California A.G. Iaccarino 1 , M. Picozzi 1,2 1 Università degli Studi di Napoli “Federico II”, Dipartmento di Fisica “Etore Pancini”, Napoli, Italy 2 Natonal Insttute of Oceanography and Applied Geophysics, OGS, Sgonico, Italy Earthquakes predicton is considered the holy grail of seismology. Afer almost a century of eforts without convincing results, the recent raise of machine learning (ML) methods in conjuncton with the deployment of dense seismic networks has boosted new hope in this feld. Even if large earthquakes stll occur unantcipated, recent laboratory, feld and theoretcal studies support the existence of a preparatory phase preceding earthquakes, where small and stable ruptures progressively develop into an unstable and confned zone around the future hypocenter (Dresen et al., 2020; Kato & Ben-Zion, 2020; Mignan, 2012). Here, we present two works focused on the induced seismicity at The Geysers geothermal feld in California. Due to its complex geological structure, the industrial operatons for energy producton and the existence of a dense seismic network, the Geysers area represents a natural laboratory for seismicity studies (Bentz et al., 2019; Holtzman et al., 2018; Martnez-Garzón et al., 2020; Trugman et al., 2016). We address the preparatory phase of Mw≥3.5 earthquakes identfcaton problem by developing ML approaches to analyze tme-series of physics-related features extracted from catalog informaton and estmated for events that occurred before the mainshocks. Specifcally, we study the temporal evoluton of the b-value from the Gutenberg-Richter (b), the magnitude of completeness (Mc), the fractal dimension (Dc), the inter-event tme (dt), and the moment rate (Mr). In a frst work (Picozzi & Iaccarino, 2021), we use a supervised technique (Recurrent Neural Network, RNN) to reveal the preparaton of 8 Mw≥3.9 earthquakes. In the second one (Iaccarino & Picozzi, 2023), we apply an unsupervised K-means clustering technique on 19 Mw≥3.5 events. The results of the frst work show that the preparatory phase for the three testng Mw≥3.9 earthquakes lasted from few hours to few days, in agreement with the short-tme preparaton process (~1 day) observed for a similar magnitude natural earthquake (De Barros et al., 2020).