GNGTS 2023 - Atti del 41° Convegno Nazionale

Session 1.1 GNGTS 2023 Forecasting megathrust earthquakes: what can we learn from the laboratory? Mastella G. 1 , Corbi F. 2 , Bedford J. 3 , Rosenau M. 4 , Funiciello F. 1 1 Università “Roma TRE”, Rome, Italy, Dip Scienze, Laboratory of Experimental Tectonics 2 Istituto di Geologia Ambientale e Geoingegneria - CNR c/o Dipartimento di Scienze della Terra, Sapienza Università di Roma, Rome, Italy 3 Institut für Geologie, Mineralogie und Geophysik, Ruhr-Universität Bochum, Bochum, Germany 4 GFZ Helmholtz Centre Potsdam, German Research Centre for Geosciences, Potsdam, Germany Abstract Earthquakes are considered unpredictable. Whether earthquake prediction will someday be feasible remains an unanswered question, which stems from the debate about their stochastic or deterministic nature. Despite the increasing number of geophysical observations at natural faults, one of the apparently insurmountable drawbacks limiting scientific progress in earthquake prediction is the lack of sufficiently long seismic catalogs needed to derive statistics and develop forecasts. Producing stick-slip events (labquakes) provides an analog of natural seismic cycles in well-controlled settings. The peculiar characteristic of producing long catalogs of small labquakes through laboratory geodesy and seismology makes lab experiments an ideal playground for testing Machine Learning (ML) potentialities. Over the last few years, ML has been used to predict earthquake-like failures in various experiments [Ren et al. 2020]. In double direct shear experiments, acoustic emissions can be used to predict the time to the next labquake, fault zone shear stress, and labquake stress drop of rock samples [e.g. Rouet Leduc et al., 2017; Rouet Leduc et al., 2018; Laurenti et al., 2022]. Similarly, for seismotectonic analog models of the subduction megathrust, analog earthquakes have been successfully predicted through time to failure [Corbi et al., 2019] as well as through a classification approach [Corbi et al., 2020]. Although these studies highlight the ability of ML to predict various features of an impending labquake [Hulbert et al., 2019], they are limited by the modelers’ arbitrary choice in selecting features deemed informative of the underlying process [Van Klaveren et al., 2020]. Trying to overcome some of these limitations, we here propose a novel spatiotemporal regression framework that forecasts future surface velocity fields from past ones [Mastella et al., 2022b]. We use geodetic-like surface deformation data from two different analog setups of megathrust seismic cycles [Corbi et al., 2013; Mastella et al., 2022a], exploring the predictability of analog earthquakes at different prediction horizons in the future. To perform the forecasting, we