日本地球惑星科学連合2023年大会

講演情報

[E] 口頭発表

セッション記号 S (固体地球科学) » S-SS 地震学

[S-SS03] New trends in data acquisition, analysis and interpretation of seismicity

2023年5月23日(火) 10:45 〜 12:00 201A (幕張メッセ国際会議場)

コンビーナ:Enescu Bogdan(京都大学 大学院 理学研究科 地球惑星科学専攻 地球物理学教室)、Francesco Grigoli(University of Pisa)、青木 陽介(東京大学地震研究所)、Chairperson:Francesco Grigoli(University of Pisa)、Thomas P. Ferrand(Free University Berlin)、Bertrand Rouet-Leduc(Kyoto University)


11:45 〜 12:00

[SSS03-10] Earthquake Phase Association with Graph Neural Networks: Application to Northern California Seismicity

★Invited Papers

*Ian W McBrearty1Gregory C Beroza1 (1.Stanford University)


キーワード:Phase Association, Deep Learning, Seismic networks

Earthquake phase association is an essential component in the development of earthquake catalogs, yet automating this procedure for general monitoring conditions such as large aperture seismic networks, high pick rates, and foreshock and aftershock sequences, has remained an open challenge. The advent of deep learning phase pickers, which provide orders of magnitude more picks of small earthquakes than traditional methods, further highlights the need to improve association algorithms. Here, we describe our Graph Neural Network based associator (GENIE), that is trained to simultaneously predict both source space-time localization, and discrete source-arrival association likelihoods. Internally, the GNN uses one graph to represent the spatial source region, and another to represent the station set; the combination of which enables robustly parsing dense rates of input pick data, determining the location and timing of sources, and assigning picks to their respective sources and phase types. We apply our method on real data from the Northern California (NC) seismic network and re-detect 96% of all events M>1 reported by the USGS over 500 random days selected between 2000-2022. We additionally process a 100-day continuous interval in 2017-2018 and detect ~4x the number of total events reported by the USGS. Our new events have low magnitude estimates below the magnitude of completeness of the USGS catalog, and closely follow the expected faults in the region. Our results demonstrate that GENIE can effectively solve the association problem and enhance automatic processing workflows under complex seismic monitoring conditions. Current work includes evaluating the generalization and transferability of the model to new regions.