Exciting news! We're transitioning to the Statewide California Earthquake Center. Our new website is under construction, but we'll continue using this website for SCEC business in the meantime. We're also archiving the Southern Center site to preserve its rich history. A new and improved platform is coming soon!

Earthquake Phase Association with Graph Neural Networks: Application to Seismicity in Northern California

Ian W. McBrearty, & Gregory C. Beroza

Submitted September 11, 2022, SCEC Contribution #12332, 2022 SCEC Annual Meeting Poster #048

We describe our Graph Neural Network based associator that is trained to both detect (determine the origin time, and location of) seismic sources, and associate picks to these sources from streaming discrete pick datasets recorded on arbitrary geometry seismic networks. The number of stations and their positions can change for any input, and each station can have any number of picks in time. The trained model is robust to high rates of false picks, missing and noisy picks, closely overlapping sources in space and time, and large aperture station networks. We train on synthetically simulated dense pick datasets, using an assumed velocity model. The predicted association likelihoods respect geometric considerations relevant to the seismic network used in the input. For instance, small earthquakes have associations only on the nearest stations, and large earthquakes have associations at greater distances. As a result, far fewer spuriously associated phases occur than in traditional association algorithms. Internally, the GNN solves the association problem by using one K-nearest-neighbor (k-NN) graph to represent the spatial region, and another k-NN graph to represent the station set. Iterative graph convolutions on this combined representation are used to parse the input pick data and map it to the source and source-arrival association predictions.

We demonstrate our method’s performance using picks produced by PhaseNet on EHZ components of the Northern California (NC) seismic network, using the USGS 3D velocity model of the greater San Francisco Bay region over a ~300 km x 250 km area that includes ~350 stations. We demonstrate the trained model recovers ~96% of all real events M>1 reported by the USGS on a random test of 500 days between the interval 2000 – 2020. The residuals of matched events have mean near zero, ~5 km std. spatial offsets, and ~1 s std. origin-time residuals. We further process a continuous 100 day interval in 2017 surrounding the time of a known M4.6 aftershock producing sequence. We find an increased recovery rate of 4-5x compared with the USGS catalog, and the newly detected sources closely follow the expected faults throughout the region. Nearly all of the new events have estimated magnitudes below the magnitude of completeness of the USGS catalog. In addition, we observe a significant number of aftershocks following the M4.6 event that are well localized near the epicenter, and decay in time following an Omori-style curve.

McBrearty, I. W., & Beroza, G. C. (2022, 09). Earthquake Phase Association with Graph Neural Networks: Application to Seismicity in Northern California. Poster Presentation at 2022 SCEC Annual Meeting.

Related Projects & Working Groups