Deciphering influential features in the seismic catalog for large earthquake occurrence from a machine learning perspective

The spatiotemporal distribution and magnitude of seismicity collected over decades are crucial for understanding the stress interactions underlying large earthquakes. In this study, machine learning (ML) explainers identify and rank the features that distinguish Large Earthquake Occurrence (LEO) from non-LEO spatiotemporal windows. Seventy-eight statistics related to time, latitude, longitude, depth, and magnitude were extracted from the earthquake catalog (Global Centroid Moment Tensor) to produce 202,706 spatiotemporally discretized windows. ML explainers trained on these windows revealed the maximum magnitude ( M max ) as the most influential feature. Classification performance improved when the maximum inter-event time, the average inter-event time, and the minimum ratio of focal depth to magnitude were jointly trained with M max . The top five features showed weak-to-moderate correlations, providing complementary information to the ML explainers. Our explainable ML framework can be extended to different earthquake catalogs, including those with focal mechanisms and small-magnitude events. • Our explainable machine learning (ML) pipeline, which includes feature extraction, XGBoost classifiers, and SHAP values, is applied to primary seismic catalog data. • The most influential factor identified for the large earthquake recognition is the maximum magnitude of spatiotemporally clustered earthquake events. • The top five important factors significantly improve the ML classification performance. • Our methodology has the potential to be applied to focal mechanisms, regional earthquake catalogs, and numerical simulations.