Testing the ability of machine learning models to predict the timing of seismogenic nucleation

Abstract:

Advances in machine learning (ML)-based data processing have proven successful in predicting the timing of fault-slip behavior using seismic data recorded in controlled laboratory experiments. Similar ML analysis of regional seismic and geodetic data from the Cascadia subduction zone on Vancouver Island show predictable ground displacements from the slowly slipping portion of the fault. The full potential of this technology to provide novel insight into seismogenic faults is still unknown. To date, no successful attempt has been made to apply a similar conceptual framework to shallow crustal faults systems. In this work we explore the possibility of applying ML models to predict the onset of slip in the seismogenic zone of the central San Andreas Fault. We use seismic waveform data from Parkfield, California to train a suite of ML regression models to estimate the onset time of seismogenic events. Training the ML models requires evaluating the data features input to the model and establishing the correct model target, known as metadata tuning, and selecting the optimal model parameters, known as hyperparameter tuning. We describe the results for the model optimization procedure that show the prediction limits of the data set explored in this study and the best modeling results. Insights learned from the prediction modeling effort are discussed in the context of furthering our understanding of fault mechanics and earthquake nucleation.