Supervised Deep Learning Models to Improve the Yellowstone Seismic Catalog With Application to the 2013/2014 Norris Geyser Basin Earthquake Swarm

Supervised deep learning is becoming a common method of detecting and processing earthquakes as it has accuracy comparable to seismic analysts while being relatively quick and requiring little human intervention. Such a method is ideal for areas of abundant seismicity that often have events occurring close together in time, complex sources, and noisy phase arrivalssuch as Yellowstone National Park. In Yellowstone, ~50% of the seismicity occurs in swarms and is influenced by the tectonics of the Basin and Range, Intermountain Seismic Belt, and the movement of hydrothermal fluids relating to the volcanic system. These facets of Yellowstone seismicity result in an incomplete earthquake catalog, especially for smaller magnitude events. Increasing the size of the earthquake catalog available for Yellowstone allows for detailed analysis of swarm migration, improved tomographic imaging, mapping of fault architectures, and more. Using a supervised deep learning approach, we produce an improved catalog of phase arrival times and P-wave first motions for the Yellowstone region. To the best of our knowledge, similar studies for other regions train their models using only three-component data. Here, we develop a method to detect phase arrivals using only the vertical component because this is often the primary data available in Yellowstone. We also work to emphasize the success of our models in periods of dense seismic activity through data augmentation. Additionally, we intend to develop a method of approximating the quality of the phase arrival picks that reflects the judgment of seismic analysts and not just the uncertainty of the deep learning models. We validate our methodology by applying it to a seismic swarm occurring in the Norris Geyser Basin of Yellowstone National Park from September 2013 into June 2014.