Forecasting Surface Velocity Fields Associated With Laboratory Seismic Cycles Using Deep Learning

• Published in: Geophysical research letters 2022-08, Vol.49 (15), p.n/a
• Main Authors: Mastella, G., Corbi, F., Bedford, J., Funiciello, F., Rosenau, M.
• Format: Article
• Language: English
• Subjects: Algorithms; analog modeling; Analogs; Artificial neural networks; Cycles; Deep learning; Earthquake forecasting; Earthquake prediction; Earthquakes; Experiments; Feature extraction; Fields; Forecasting; Laboratories; laboratory earthquake forecasting; Learning algorithms; Machine learning; megathrust earthquakes; Neural networks; Performance prediction; Propagation; Recurrent neural networks; Seismic activity; Structure-function relationships; Subduction; Subduction (geology); Subduction zones; Surface velocity; Transfer learning; Velocity; Velocity distribution
• ISSN: 0094-8276; 1944-8007
• DOI: 10.1029/2022GL099632

It has been recently demonstrated that Machine Learning (ML) can predict laboratory earthquakes. Here we propose a prediction framework that allows forecasting future surface velocity fields from past ones for analog experiments of megathrust seismic cycles. Using data from two types of experiments, we explore the prediction performances of multiple Deep Learning (DL) and ML algorithms. In such a self‐supervised regression, no feature extraction is required and the entire seismic cycle is forecasted. The onset, magnitude, and propagation of analog earthquakes can thus be predicted at different prediction horizons. From all architectures tested in this study, convolutional recurrent neural networks (CNN‐LSTM and CONVLSTM) provide the best predictions although their performances depend on experiment characteristics and hyperparameters tuning. Analog earthquakes can be successfully anticipated up to a horizon of the order of their duration. This laboratory‐based study may open new avenues for transfer learning applications with data from natural subduction zones. Plain Language Summary In the last few years scientists have shown their ability to predict the occurrence of earthquakes simulated in the laboratory using Machine Learning, a group of algorithms useful to learn hidden structure in data, complete tasks, and make predictions. By applying methods inspired by the structure and function of the brain (so‐called “Neural Networks”), we here introduce an approach aiming to forecast the temporal evolution of the surface velocity field of analog experiments of the subduction megathrust seismic cycle. We show that our approach allows forecasting not only the onset and the size of laboratory earthquakes but also their preparatory phase and their propagation. Our success in laboratory earthquake forecasting provides optimism that one day similar results may be achieved with natural earthquakes. Key Points A spatiotemporal regression framework is used to forecast surface velocity associated with seismic cycles reproduced in the laboratory The onset, magnitude, and propagation of analog earthquakes can be predicted up to a temporal horizon of the order of their duration Deep Learning outperforms standard Machine Learning, although their performances depend on data characteristics and model configurations