GNGTS 2023 - Atti del 41° Convegno Nazionale

Session 2.2 GNGTS 2023 Real-time ground-shaking maps reconstruction of late 2022 Italian earthquakes S. F. Fornasari, V. Pazzi, G. Costa Department of Mathematics and Geosciences, University of Trieste, Italy INTRODUCTION Real-time seismic monitoring provides crucial information for emergency operations after severe seismic events. Ground shaking maps, usually expressed in terms of one of the many ground motion parameters (GMPs) proposed in the literature, are used to evaluate the impact of an earthquake. Traditional approaches, such as ShakeMap (Wald, Quitoriano, Heaton, Kanamori, Scrivner, and Worden, 1999; Wald, Worden, Thompson, and Hearne, 2021), use ground motion prediction equations (GMPEs) constrained by the instrumental GMPs to produce a gridded representation of the ground motion and thus are dependant on the estimate of location and magnitude of the event. The automated ShakeMap procedure of the Istituto Nazionale di Geofisica e Vulcanologia (INGV) for the Department of Civil Protection (DPC) is triggered, for events of , when the revised ≥ 3. 0 location and magnitude are available, usually several minutes after the event (Margheriti et al., 2021). ShakeRec (Fornasari, Pazzi, and Costa, 2022) has been developed for the DPC to reconstruct ground-shaking maps in real-time leveraging only the data from the stations, specifically the ones of the Italian Strong Motion Network (RAN, Costa et al., 2022). As such, ShakeRec can fill the temporal gap between the arrival of the seismic data at the data centre and the evaluation of GMPE-based ground-shaking maps providing information during the early stages of emergency management. The reconstruction capabilities of ShakeRec have been tested for two recent Italian earthquakes: the 22 September 2022 Bargagli earthquake (Genova, North-West Italy, ) and the 9 4. 1 November 2022 Marche earthquake (offshore, Central Italy, ). 5. 7 METHODS Considering a single GMP, the reconstruction is performed using an ensemble of five convolutional neural networks, as shown in the central panel in Figure 1. The architecture of each ensemble member has been adapted from Fukami et al. (2021) who proposed a data-driven spatial field