Optimized Source and Station Graphs for Spatially Adaptive Earthquake Phase Association with Graph Neural Networks

Ian W. McBrearty, & Gregory C. Beroza

Submitted August 30, 2026, SCEC Contribution #15412, 2026 SCEC Annual Meeting Poster #151

Earthquake phase association is an essential step for developing high-quality earthquake catalogs. Many algorithms have recently been proposed, however most still struggle on dense seismicity sequences, high density seismic networks, and large geographic regions. Further, most are highly sensitive to hyper-parameter choices and require significant tuning for each new application/setting. Here we show progress on a scale-adaptive variant of the GENIE earthquake phase associator, which is a graph neural network-based model to directly predict earthquake hypocenters (location and origin time) and phase associations from pick data. By using a combination of optimized low-discrepancy source graph sampling, graph re-wiring, and spectral graph sampling methods, we design a variant of GENIE that is applicable from local to regional scales without parameter tuning or re-training. We demonstrate this scale-adaptive variant of the model on local seismicity detection during the dense Cahuilla swarm sequence in southern California in 2018, as well as detections across of all northern California spanning the interval of the 2022 Ferndale earthquake. We also demonstrate robust detection and association performance on an `out-of-distribution’ test during the 2014 Iquique earthquake in northern Chile, where we accurately recover seismicity along the subducting Nazca plate down to ~250 km depth. These results highlight the applicability and adaptability of our new scale-adaptive association model, which enables simple and efficient association of seismic data across highly variable station geometries and spatial domains without retraining or hyper-parameter tuning.

Citation
McBrearty, I. W., & Beroza, G. C. (2026, 08). Optimized Source and Station Graphs for Spatially Adaptive Earthquake Phase Association with Graph Neural Networks. Poster Presentation at 2026 SCEC Annual Meeting.

Related Projects & Working Groups
Seismology