GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 2.2 GNGTS 2024 Development of a hybrid method for ground shaking map reconstrucHon in near-real Hme S. F. Fornasari , V. Pazzi , G. Costa Dipar9mento di Matema9ca, Informa9ca e Geoscienze (MIGE - Università degli Studi di Trieste, Italia) IntroducHon Real-Cme seismic monitoring is of primary importance for rapid and targeted emergency operaCons aper potenCally destrucCve earthquakes. A key aspect in determining the impact of an earthquake is the reconstrucCon of the ground-shaking field, usually expressed as the ground moCon parameter. TradiConal algorithms (e.g. ShakeMap®) compute the ground- shaking fields from the punctual data at the staCons relying on ground-moCon predicCon equaCons (GMPEs) computed on esCmates of the earthquake locaCon and magnitude when the instrumental data are missing. The results of such algorithms are then subordinate to the evaluaCon of locaCon and magnitude, which can take several minutes. Since machine learning techniques have already been proven capable of esCmaCng the ground moCon parameters (Fornasari et al., 2023), a hybrid method has been developed to integrate neural networks in the ShakeMap® workflow to speed up the current ground- shaking map evaluaCon process. The core idea is to adopt the ShakeMap® mulCvariate normal distribuCon (MVN) method for the intensity measure (IM) interpolaCon and use a neural network, in place of the ground moCon predicCon equaCons (GMPEs), to esCmate the IM condiConal expected value and uncertainty at the target sites based only on data available in real-Cme and thus do not wait for the magnitude and locaCon esCmates. Furthermore, by reusing the ShakeMap® framework, the complexity of the model is reduced with improvements in the interpretability of the results. Method The proposed hybrid method consists of two steps: first, the expected IM values (and their uncertainCes) are computed at the staCons and target locaCons; then the recorded and expected IMs are passed to the MVN to compute the ground-shaking map (and its uncertainty). The approach adopted to replace the GMPE is called ConvoluConal CondiConal Neural Process (ConvCNP, Gordon et al., 2019): starCng from sparse randomly sampled observaCons, a funcConal representaCon of them is computed, discreCzed to a regular grid and fed to a backbone neural network whose outputs are converted from the funcCon space to the