AI-PAL: Self-Supervised AI Phase Picking via Rule-Based Algorithm for Generalized Earthquake Detection

Delineating fault structures through microseismicity is crucial for earthquake hazard assessment, yet constructing high-resolution catalogs over extended periods remains challenging. This study introduces AI-PAL, a novel deep learning-driven workflow that employs a Self-Attention RNN (SAR) model trained with detections from PAL, an established rule-based algorithm (Zhou, Yue, et al., 2021, https://doi.org/10.1785/0220210111), for generalized earthquake detection. PAL utilizes short-term-average over long-term-average algorithm for event detection, ensuring consistent performance across different datasets. AI-PAL leverages these rule-based picks as training labels, enabling self-supervised learning of the SAR model across arbitrary regions, thereby enhancing PAL's detection capabilities. We applied SAR-PAL to two distinct regions that are featured by recent large earthquakes: (a) the preseismic period of the Ridgecrest-Coso region (2008–2019), and (b) the pre-to-postseismic period of the East Anatolian Fault Zone (EAFZ, 2020–2023/04). Our results demonstrate that SAR-PAL offers slightly higher detection completeness than the quake template matching matched filter catalog, while boosts over 100 times faster processing and a superior temporal stability, avoiding detection gaps during background periods. Compared to PhaseNet and GaMMA, two widely recognized phase picker and associator, SAR-PAL proved more scalable, achieving ∼2.5 times more event detections in the EAFZ case, along with a ∼7 times higher phase association rate. We further experimented training PhaseNet and SAR with PAL detections and routine catalogs, and found that no other combinations matched the detection performance of SAR-PAL. The enhanced catalogs built by SAR-PAL reveals geometrical complexities of the Ridgecrest faults and the Erkenek-Pütürge segment of EAFZ, offering insights into their contrasting roles during the large earthquake.