Application of deep learning toward predicting the size and timing of simulated megathrust earthquakes

Abstract:

Dealing with the suddenness and unknown magnitude of earthquakes is an ongoing challenge for earthquake response and preparedness. Recent progress has been made toward using machine learning algorithms to predict the timing of laboratory and analog earthquakes. Whether or not these methods scale to larger, more complex systems is still uncertain. Furthermore, the specific parameters that might be predictive of upcoming slip and size are poorly constrained. In this study, we employ two different deep learning approaches to make predictions of the timing and size of impending fault ruptures on a simulated subduction megathrust fault. We use the discrete element method (DEM) to construct a numerical analog of a megathrust fault, overlain by a preconditioned wedge. The fault interface irregularly experiences complex dynamic ruptures in response to steady displacement of the wedge backstop. Spatiotemporal fault displacement and stress data are then used to train a convolution neural network (CNN) to make predictions of the relative size of the next event. We also train a long short term memory network (LSTM) to predict whether or not the system is within 20 timesteps of earthquake rupture. Combined, these two independent algorithms that were trained on the same spatiotemporal data provide an estimate of timing and size of upcoming earthquakes in our DEM model. The results of this study provide support to the hypothesis that spatiotemporal patterns in evolving fault properties might be predictive of upcoming size and timing of earthquakes. The results underscore the value of high resolution time series data along modern subduction zones.