Machine Learning Predicts Earthquakes in the Continuum Model of a Rate‐And‐State Fault With Frictional Heterogeneities

Machine learning (ML) has been used to study the predictability of laboratory earthquakes. However, the question remains whether or not this approach can be applied in a tectonic setting where one may have to rely on sparse earthquake catalogs, and where important timescales vary by orders of magnitude. Here, we apply ML to a synthetic seismicity catalog, generated by continuum models of a rate‐and‐state fault with frictional heterogeneities, which contains foreshocks, mainshocks, and aftershocks that nucleate in a similar manner. We develop a network representation of the seismicity catalog to calculate input features and find that the trained ML model can predict the time‐to‐mainshock with great accuracy, from the scale of decades to minutes. Our results offer clues as to why ML can predict laboratory earthquakes and how the developed approach could be applied to more complex problems where multiple timescales are at play. Plain Language Summary Accurately forecasting the timing of an impending large earthquake is one of the main but elusive goals of seismology. Recent studies show that machine learning (ML), a powerful tool that can draw inferences from hidden patterns in data, can predict the timing of laboratory earthquakes that mimic faulting in nature. However, the question remains whether or not an ML approach can be applied to natural faults where one may have to rely on sparse earthquake catalogs, and where important timescales vary by orders of magnitude. In this study, we develop and apply an ML approach to a synthetic earthquake catalog generated by physics‐based computer simulations. We find that the trained model can predict the time remaining before the mainshock with great accuracy, from the scale of decades, long before the upcoming earthquake, down to the scale of hours to minutes, right before the mainshock. Our results provide clues as to why ML can predict laboratory earthquakes and how to apply an ML‐based approach to real earthquakes where multiple timescales are at play. Key Points We developed a network representation of a synthetic earthquake catalog to train a machine‐learning Machine learning (ML) model The trained ML model can predict the time remaining before simulated mainshocks with great accuracy The network representation of seismic moment and event interval enables us to accurately predict the timing of mainshocks