Speaker: Ian McBrearty, Stanford University
Title: Enhanced Earthquake Detection with Graph Neural Networks
Time: Friday, Oct 24 at 12:00pm PST
Location: EMS B210
Abstract: Developing accurate earthquake catalogs is an essential goal in seismology, where improved catalogs can enhance our view of active seismogenic processes, fault network distributions, velocity models of the Earth, and earthquake statistics. However, developing these catalogs comes with a number of challenges, such as high event rates and aftershock sequences, high false pick rates and noise levels, time-varying station networks, and highly heterogenous source and station distributions. Automating the processing of these datasets and determining when, where, and the magnitude of each event accurately requires involved algorithms, and improvements have recently been obtained with deep learning methods. Here I review two graph neural networks used for phase association and double-difference relocation. I highlight how the particular GNN architecture choice closely conforms with the structure of the problems being solved, which assists with learning reliable models. The utility of both techniques is demonstrated in the construction of a long-term, deep learning enhanced catalog of northern California (~2 million events) that shows high spatial resolution and consistency with the existing catalog, while increasing the detection rate by ~3.5x. The improved catalog illuminates many major and minor fault systems throughout the region and records the active seismic processes of notable events, including the >M6 Parkfield, San Simeon, Ridgecrest, Ferndale, and Napa Valley earthquakes.
