Double Difference Earthquake Location with Graph Neural Networks

• DOI: 10.48550/arxiv.2410.19323
• Publication Date: 2024-10-25
• Authors: Ian W. McBrearty; Gregory C. Beroza
• Keywords: Physics - Geophysics; FOS: Computer and information sciences; Computer Science - Machine Learning; FOS: Physical sciences; Geophysics (physics.geo-ph); Machine Learning (cs.LG)

Double difference earthquake relocation is an essential component of many catalog development workflows. This technique produces high-resolution relative relocations between events by minimizing differential measurements the arrival times waves from nearby sources, which highlights resolution faults and improves interpretation seismic activity. The inverse problem typically solved iteratively using conjugate-gradient minimization, however cost scales significantly with total number sources stations considered. Here we propose a Graph Neural Network (GNN) based double-difference framework, Difference (GraphDD), that trained to minimize residuals locate earthquakes. Through batching sampling method can scale arbitrarily large catalogs. Our architecture uses one graph represent stations, second creates Cartesian product two graphs capture relationships (e.g., travel time partial derivatives). key feature allows natural be used residuals. We implement our model on several distinct test cases including seismicity northern California, Turkiye, Chile, have highly variable data quality, station source distributions. obtain high in these tests, shows adaptability types loss functions location objectives, learning corrections mapping into reference frame different catalog. results suggest GNN approach promising direction for scaling very catalogs gaining new insights problem.