Probing Seismogenic Faults with Machine Learning

We analyze continuous seismic data with a variety of classical machine learning (ML) and deep learning (DL) models with the goal of identifying hidden signals connected to the earthquake cycle. In the laboratory, we find that continuous seismic waves originating in the fault zone are imprinted with fundamental information regarding the physics of the fault. Statistics of these low-amplitude, noise-like signals identified with supervised ML approaches can be used to estimate fault friction, fault displacement, and forecast upcoming failure with great accuracy. These results hold true for both stick-slip and slow-slip frictional regimes. Similarly, when we scale the approach to study slow-slip events in the Cascadia subduction zone and the San Andreas Fault, we find that continuous seismic waves contain information about the instantaneous fault displacement at all times. Direct application of these approaches to seismogenic faults in Earth is highly challenging to date. As a result, we are developing more generalized DL approaches where the model is trained on fault simulations and applied to laboratory fault data.