GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 1.1 GNGTS 2024 ML vs semi-automatc seismic catalogues: the L’Aquila 2009 earthquake sequence example R. Fonzet 1, 2 , L. Valoroso 1 , A. Govoni 1 , P. De Gori 1 , C. Chiarabba 1 1 Isttuto Nazionale di Geofsica e Vulcanologia (INGV), Rome (RM) 2 Università degli Studi Roma Tre , Rome (RM) Nowadays the use of neural networks and artfcial intelligence in seismology delivers high- resoluton seismic catalogues including very small magnitude events that remained undetected by human analysts and standard monitoring room procedures. Here, we test the performance of such ML methods for their reliability in the high-seismic hazard area of the Central Apennines, Italy. In this work, we apply the QuakeFlow (QF) workfow (Zhu et al., 2023), based on the PhaseNet (Zhu et al., 2019) and GaMMA (Zhu et al., 2022) modules for event detecton and associaton respectvely, to the L’Aquila 2009 seismic sequence involving 90 seismic statons and we benchmark our new catalogue against the high-resoluton seismic catalogue of Valoroso et al., (2013) that used a semi-automatc procedure including detecton and an automatc picking procedure for P and S waves (i.e., Manneken Pix algorithm, Aldersons et al., 2009). We analysed the earthquakes that occurred during the entre 2009 year and obtained approximately 336,000 events vs the 182,000 from Valoroso et al., (2013) catalogue, obtaining about 85% more earthquakes. Thus, our new catalogue detected 154,452 more events (~ 85%) with respect to the Valoroso et al., (2013) catalogue. We selected all events having at least 6 P- and 4 S- arrival tmes (i.e. about 222,000 earthquakes) and we computed 1D-locatons using Hypoellipse (Lahr, 1999) with an ad-hoc velocity models for the Central Apennines (Fig. 1). The 1D-locatons clearly highlight the geometric characteristcs of the seismogenic faults actvated during the sequence: the Paganica fault with its set of synthetc and antthetc minor faults (sectons 13-19) and the Campotoso fault (sectons 1-9). Finally, in the last sectons (sectons 19-23) a low-angle fault was actvated during the sequence. The comparison with the Valoroso et al., (2013) catalogue was obtained by using only those seismic statons which were in common between the two datasets (purple triangles in Fig. 2) and selectng the common earthquakes (i.e., events having P- and S- waves arrival tmes at a common staton within 2 seconds, Fig. 2a-c). The ML- catalogue presents a larger number of phases with respect to the Valoroso et al., (2013) catalogue and lower GAP (°) values (Fig. 2d). The applicaton of this new methodology could speed up the tme to analyse seismic sequences even in real tme. Our fndings are expected to help scientsts to understand the earthquake generaton mechanisms of the 2009 L’Aquila earthquake sequence in terms of nucleaton processes, the underlying physical triggering processes leading to a richer afershock catalogue, and revealing any hidden faults in the vicinity of well-known and mapped structures that remain unseen for the last two decades.