Machine Learning for Understanding Lab Earthquake Prediction and Precursors

I summarize a broad suite of laboratory data sets showing that stick-slip failure events –lab earthquakes– are commonly preceded by both measurable changes in fault zone properties and acoustic emission (AE) events that foretell catastrophic failure.  These works show that both types of data can be used to predict labquakes with machine learning (ML) methods and deep learning (DL) approaches.  The first works used continuous measurements of AE to predict the timing of labquakes and the fault zone shear stress. Subsequent studies showed that catalogs of AE events could also predict labquakes and that ML approaches could also predict stress drop, peak fault slip velocity and the duration of failure. Recently, DL has been used to predict and autoregressively forecast labquakes and fault zone shear stress. Consistent with previous works, we see that seismic b-value begins to decrease as faults unlock and start to creep.  This provides a sensible connection between the ML-based predictions, fault zone elastic properties, and the physics of failure.  In the lab, AE events represent a form of foreshock and, not surprisingly, the rate of foreshock activity correlates with fault slip rate and its acceleration toward failure.  Our work shows precursory changes in wave speed prior to labquakes, consistent with many well known past studies, but the early studies did not provide a method to predict impending failure.  ML and DL predicts with fidelity the time of impending failure and other aspects of it. This suggests the possibility of physics-based models for prediction. We are working to connect ML prediction of labquakes with the evolution of fault zone elastic properties, frictional contact mechanics and constitutive laws.  A central goal is to learn from lab earthquake prediction to improve forecasts of earthquake precursors and tectonic faulting.