Abstract Predicting large earthquakes remains a significant challenge due to the complexity of fault systems and the variability of preparatory processes. We introduce an unsupervised machine learning framework to categorize seismicity patterns and identify, when present, seismicity transients preceding large earthquakes. We focus on five large earthquakes and extract seismo-mechanical features per families of events, defined as clustered events in space, time and magnitude. Here we show that for those cases displaying a preparatory phase, specific long-lasting families belonging to a critical category signalling an upcoming earthquake occur during the preparatory phase. Compared to other periods, critical categories reflect a higher spatial-temporal localization, earthquake interaction and strain release. The method will not detect such a transient for earthquakes with no detectable seismic preparatory phase. Finally, we demonstrate that the method is capable of identifying preparatory phases (when present), showing potential for operational earthquake forecasting.
Karimpouli et al. (Mon,) studied this question.