Deep learning of the precursory signatures in active source seismic data for improved prediction of laboratory earthquake

Abstract:

Small changes in seismic wave velocity and amplitude have been observed to carry precursory information about frictional failure in both laboratory experiments and nature. Laboratory studies report that wave velocities and amplitudes vary systematically along the seismic cycle, and show a distinct reduction prior to fault failure. While wave amplitude drops relatively early in the interseismic period, wave velocity decays later, close to the peak in shear stress just before failure. Despite the ubiquitous observations of these trends in the lab, the underlying physical mechanisms controlling these precursory signatures are still poorly understood. We hypothesize that reducing the full seismic waveform to a handful of features (such as the time-of-flight and amplitude of the first arrived wave packet) potentially disregards other important precursory information in the entirety of the signal. Exploiting such information could be crucial for improving failure prediction. In this study, we use deep learning (DL) to predict failure in a set of friction experiments where we simultaneously and continuously record the evolution of elastic wave characteristics throughout many laboratory seismic cycles. The resulting large dataset lends itself to the application of DL methods. We train, validate and test long short-term memory (LSTM), XGBoost and multilayer perceptron (MLP) models to predict the timing and size of laboratory earthquakes as well as the fault slip rate based on (1) handpicked velocity, amplitude and spectral features and (2) automatically extracted features from the full waveforms. We compare the performance of different models in predicting the laboratory earthquakes and report the significance rank of the used features. In addition, the transportability of models will be explored by using distinct datasets (from different experiments) for training and testing. The developed prediction models can be used for predicting fault failure in the context of seismic hazard assessment/warning in conjunction with continuous and long-term time-lapse monitoring of crustal faults, CO2 storage sites and unconventional reservoirs. Additionally, the automatically learned features from full seismic waveforms will potentially help understand physical processes leading to the nucleation of earthquakes.