Predicting Imminence of Analog Megathrust Earthquakes With Machine Learning: Implications for Monitoring Subduction Zones

Abstract

Subduction zones are monitored using space geodesy with increasing resolution, with the aim of better capturing the deformation accompanying the seismic cycle. Here, we investigate data characteristics that maximize the performance of a machine learning binary classifier predicting slip‐event imminence. We overcome the scarcity of recorded instances from real subduction zones using data from a seismotectonic analog model monitored with a spatially dense, continuously recording onshore geodetic network. We show that a 70–85 km‐wide coastal swath recording interseismic deformation gives the most important information on slip imminence. Prediction performances are mainly influenced by the alarm duration (amount of time that we consider an event as imminent), with density of stations and record length playing a secondary role. The techniques developed in this study are most likely applicable in regions of slow earthquakes, where stick‐slip‐like failures occur at time intervals of months to years.