Skip to content

Using Machine Learning to Enhance the Yellowstone Earthquake Catalog

Alysha Armstrong , Ben Baker, Keith Koper

Abstract

The University of Utah Seismograph Stations (UUSS) is responsible for earthquake monitoring in the Yellowstone volcanic region and has cataloged ~34K earthquakes with magnitudes between -1.29 and 4.83 from 2002 to 2022. While the UUSS catalog is generally complete down to M1.5, many small earthquakes and supplementary phase arrivals are not identified. Still, this information could help improve body-wave tomography models and provide insight into fault architectures, swarm evolution, dynamic triggering, and more. The nature of seismicity in Yellowstone, which is ~50% earthquake swarms, makes automatically cataloging these earthquakes challenging, however. Machine learning techniques can accurately process earthquakes and may be more reliable for periods of elevated seismicity than conventional methods. Here, we present the preliminary results for an enhanced Yellowstone earthquake catalog produced using machine learning models explicitly trained for Yellowstone. We apply a phase detection method that works well during periods of elevated seismicity to continuous Yellowstone seismic data and use an arrival time refinement method that produces Bayesian-derived uncertainties. When carefully evaluating these methods on a 10-day period containing an Mw 4.8 in Yellowstone, we identified 855 potentially new events, of which >99% were genuine, and recovered 83% of the UUSS catalog. We also use a feature-based magnitude estimation method that works well for small earthquakes and maintains consistency with the authoritative regional earthquake catalog. Ultimately, applying machine-learning methods will result in a comprehensive and detailed view of Yellowstone seismicity.