[S-CG40] Science of slow-to-fast earthquakes

[SCG40-05] Machine learning predicts earthquakes in the continuum model of a rate-and-state fault and in a meter-scale laboratory experiment

*Reiju Norisugi¹, Yoshihiro Kaneko¹, Bertrand Rouet-Leduc¹ (1.Kyoto University)

Keywords:We developed the network representation of a synthetic earthquake catalog., The trained machine-learning model can predict time remaining before simulated earthquakes., The network representation of seismic moment and event interval provides the predictability., We found that our approach is also applicable to a meter-scale laboratory experiment.

Machine learning (ML) has been used to study the predictability of centimeter-scale laboratory earthquakes. However, the question remains whether or not this approach can be applied to earthquakes in nature where one may have to rely on sparse earthquake catalogs, and where important timescales vary by orders of magnitude. We first apply ML to a synthetic seismicity catalog, generated by continuum models of a rate-and-state fault with frictional heterogeneities, containing foreshocks, mainshocks, and aftershocks that nucleate similarly. 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 towards upcoming earthquakes. The output from the trained ML model indicates that the increase in seismic moment and the decrease in recurrence interval averaged over certain networks have predictive power of the time-to-mainshock. The developed approach enables us also to predict the shear stress averaged over the local fault patches on the fault. We then apply our method to the laboratory earthquake data of Yamashita et al (2021). Remarkably, we find that the trained ML model via network representation of the laboratory earthquake catalog can also predict the time-to-failure of meter-scale mainshocks with great accuracy, from the scale of tens of seconds to milliseconds towards upcoming mainshocks. 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.