Generalizable deep learning models for predicting laboratory earthquakes

Abstract

Machine learning models can predict laboratory earthquakes using Acoustic emission, the lab equivalent of microseismicity, and changes in fault zone elastic properties during the lab seismic cycle. Applying them to natural earthquakes requires testing their generalizability across lab settings and stress conditions. Here, we show a fine-tuned convolutional neural network (CNN) model effectively transfer across different conditions. Our model employs techniques from natural language processing, including decoder techniques, to capture the relationship between AE and fault stress. We fine-tune the regression head of a deep CNN while fixing the decoder’s weights and successfully predict lab seismic events for a range of conditions. With fine-tuning, CNN models trained on one lab fault configuration predict time to failure and shear stress for another configuration at varying fault slip rates. These results demonstrate the potential of extending lab-based methods to different conditions that could eventually include tectonic earthquakes and seismic forecasting.