GNGTS 2023 - Atti del 41° Convegno Nazionale

Session 1.1 GNGTS 2023 employed multiple Deep Learning-based (DL) models, following a self-supervised regression, meaning that no feature extraction is required. In such a framing, the whole seismic cycle is forecast, including interseismic, coseismic, and postseismic stages. By comparing the performances obtained with forecasting baselines, such as the Persistence model and a Random Forest (RF) regressor, we quantified forecasting improvements that can be achieved with DL. From all architectures tested in this study, convolutional recurrent neural networks provide the best predictions although their performances depend on experiment characteristics and hyperparameters tuning. We demonstrate that analog seismic cycles can be forecasted in both the two different seismotectonic megathrust models, despite the quality of the predictions depending on experimental configurations. The onset, magnitude and space-time propagations of analog earthquakes can be successfully predicted up to a horizon of the order of their duration. This study introduces a novel framing of the laboratory forecasting that may open up new perspectives for application to real Earth. Given the similarities between the spatiotemporal rupture patterns generated by seismotectonic analog models and the observed behaviors along natural subduction zones [Corbi et al., 2022; Philibosian and Meltzner, 2020], learning in the laboratory how to predict fast and/or slow earthquakes from geodetic data is a challenging evolution of this study, that has the potentiality to revolutionize hazard assessment at subduction zones. One of the best candidate locations to test our approach is the Cascadia subduction zone (North America) , where geodetic records contain tens of cycles of slow earthquakes [Michel et al., 2019]. Although the physical mechanism of slow earthquakes might be different from that of large megathrust earthquakes, our intuition is that DL can learn the slip evolution of natural megathrust systems simply by deciphering the spatiotemporal kinematic evolution of surface deformation, as it has been able to do in the laboratory. References Corbi, F., Funiciello, F., Moroni, M., Van Dinther, Y., Mai, P. M., Dalguer, L. A., & Faccenna, C. (2013). The seismic cycle at subduction thrusts: 1. Insights from laboratory models. Journal of Geophysical Research: Solid Earth , 118 (4), 1483–1501. https://doi.org/10.1029/2012JB009481 Corbi, F., Sandri, L., Bedford, J., Funiciello, F., Brizzi, S., Rosenau, M., & Lallemand, S. (2019). Machine learning can predict the timing and size of analog earthquakes. Geophysical Research Letters , 46 (3), 1303–1311. https://doi.org/10.1029/2018GL081251 Corbi, F., Bedford, J., Sandri, L., Funiciello, F., Gualandi, A., & Rosenau, M. (2020). Predicting imminence of analog megathrust earthquakes with machine learning: Implications for monitoring subduction zones. Geophysical Research Letters , 47 (7), e2019GL086615. https://doi. org/10.1029/2019GL086615. Corbi, F., Bedford, J., Poli, P., Funiciello, F., & Deng, Z. (2022). Probing the seismic cycle timing with coseismic twisting of subduction margins. Nature Communications , 13 (1), 1911. https://doi.org/10.1038/s41467-022-29564-2.