Monitoring the 2020 Magna, Utah, Earthquake Sequence with Nodal Seismometers and Machine Learning

Abstract

Abstract Immediately following the 18 March 2020 Mww 5.7 Magna, Utah, earthquake, work began on installing a network of three-component, 5 Hz geophones throughout the Salt Lake Valley. After six days, 180 geophones had been sited within 35 km of the epicenter. Each geophone recorded 250 samples per second data onsite for ∼40 days. Here, we integrate the geophone data with data from the permanent regional seismic network operated by the University of Utah Seismograph Stations (UUSS). We use machine learning (ML) methods to create a new catalog of arrival time picks, earthquake locations, and P-wave polarities for 18 March 2020–30 April 2020. We train two deep-learning U-Net models to detect P waves and S waves, assigning arrival times to maximal posterior probabilities, followed by a two-step association process that combines deep learning with a grid-based interferometric approach. Our automated workflow results in 142,000 P picks, 188,000 S picks, and over 5000 earthquake locations. We recovered 95% of the events in the UUSS authoritative catalog and more than doubled the total number of events (5000 vs. 2300). The P and S arrival times generated by our ML models have near-zero biases and standard deviations of 0.05 s and 0.09 s, respectively, relative to corresponding analyst times picked at backbone stations. We also use a deep-learning architecture to automatically determine 70,000 P-wave first motions, which agree with 93% of 5876 hand-picked up or down first motions from both the backbone and nodal stations. Overall, the use of ML led to large increases in the number of arrival times, especially S times, that will be useful for future tomographic studies, as well as the discovery of thousands more earthquakes than exist in the UUSS catalog.