← Back to Sessions

Machine Learning and the 2020 Magna, Utah Earthquake Sequence

Session: Network Seismology: Keeping the Network Running While Integrating New Technologies II

Type: Oral

Date: 4/22/2021

Presentation Time: 02:00 PM Pacific

Description: 

Responding to a notable seismic event is an important task for any regional seismic network (RSN). Complicating the RSN’s response is that a large event may be followed by increases in both the volume of data collected and the number of events recorded. Such challenges can slow the generation of preliminary catalogs to be used by future scientific studies. Naturally, automated catalog creation methods become highly desirable. To this end, recent successes in machine learning (ML) may represent a paradigm shift in operational algorithms for signal detection, first-motion picking, and association. In this study, we apply an ML pipeline to the 18 March 2020 M_ww 5.7 earthquake in Magna, Utah. The aftershock sequence was well-recorded by ~40 three-component stations operated by the University of Utah Seismograph Stations (UUSS) as well as 180 temporary nodal geophones. We then create an earthquake catalog from the aforementioned dataset using a U-Net deep-learning architecture for P and S wave detection and picking, a convolutional neural network for first-motion picking, and a recurrent neural network for the preliminary association of arrivals. To test the promised generalizability of these ML methods, no examples from the Magna sequence were included in the training datasets. Yet we still recovered 95% of the events in the UUSS’s authoritative catalog and observed the same distribution of seismicity. We also compared the ML data products against products generated by our well-tuned real-time system. While our P picking accuracy was only marginally better than our production STA/LTA picker’s accuracy our ML methodology generated a substantial number of S picks. Moreover, the ML first motion estimates led to focal mechanisms that were highly consistent with manually generated focal mechanisms computed in an auxiliary study. This suggests that pre-training ML models on existing network data and deploying these models during periods of enhanced seismicity can be a viable methodology in an RSN’s event response strategy.

Presenting Author: Ben Baker

Student Presenter: No

Authors