Seismo Lab Seminar

Earthquake phase association is the long-standing and challenging seismological problem of determining how to connect sets of seismic arrival time measurements to the a priori unknown population of events that generated them. The challenge of this problem grows as the spatial scale and number of stations considered increases, both due to the computational effort required, and due to complications that can arise with multiple simultaneous earthquakes, missing and false picks, noisy picks, heterogenous station coverage and uncertain velocity structure. All of these factors are also increasingly challenging at low magnitudes near the detection threshold where phase measurements become less reliable.

Here we present applications of a graph neural network associator (GENIE) proposed in McBrearty and Beroza (2023) at the local, regional and global scales. GENIE uses a station graph to represent station geometry, a source graph to represent potential seismic sources, and learns to associate phase measurements as expressed through the graph formed by the Cartesian product of those two graphs. An advantage of graph neural networks compared to other deep learning frameworks is that they can naturally handle non-regularly sampled data, such as occurs with arbitrary station networks, and in addition, can be applied to time-varying graphs (or station networks) without requiring retraining. These capabilities make our model effective for the real-world challenges that are pervasive in the association problem. When comparing our model against non-DL associators, such as Earthworm, and GaMMA, we find favorable performance, both at the local and regional scales. Our findings highlight the importance of addressing the association problem accurately and the potential of graph neural networks for improving seismological processing pipelines.