Challenges of incorporating Machine Learning and DAS into the Southern California Seismic Network (SCSN) and Southern California Earthquake Data Center (SCEDC)

Allen L. Husker, Gabrielle Tepp, Ellen Yu, Ettore Biondi, Rayomand Bhadha, Ryan Tam, Aparna Bhaskaran, Benjamin Shimota, Nytica Artiaga, & Zhongwen Zhan

Published September 8, 2024, SCEC Contribution #14001, 2024 SCEC Annual Meeting Poster #222 (PDF)

Research and Development (R&D) is a core component of the SCSN and SCEDC. The most important goal of the SCSN is to accurately record earthquakes, especially large ones, and catalog them in real-time. The SCEDC then stores the data in perpetuity. All R&D efforts must not hinder these goals and should help to advance them. The scientific community has made great strides in machine learning (ML) techniques for earthquake detection and with distributed acoustic sensing (DAS). Here we present some of our efforts to incorporate them into our systems.

It is not straightforward to put ML code into a real-time earthquake monitoring (RT) system. ML algorithms are designed for research purposes and often overlook factors important for real-time monitoring, such as processing time, computational requirements, and data delays and incompleteness. Moreover, our RT system, Earthworm/AQMS, is now very old and was not written in a way to easily test or change individual parts. We therefore are using a 3-pronged approach to incorporate ML: 1) use ML in postprocessing to get the benefits of additional and more accurate picks for already-detected earthquakes, 2) provide RT data files to replay so that researchers can train ML codes on it, and 3) incorporate ML into the RT system by revising code to meet requirements of operation. Tests of the ML picker show that it outperforms the RT system only for earthquakes M<5.

DAS can be incorporated in the RT system in several ways that are currently being tested or already in use. The Ridgecrest DAS array is sending 5000 channels at 100 sps every second to the SCSN. We plan to incorporate 18 of these channels into our earthquake detection process. We have been acquiring DAS in real-time for many months, and the system keeps up with a feed of 5000 channels. The next step is to send 1 channel into the RT system as the equivalent of a seismic station to test real-time DAS in our system. However, the best option for earthquake monitoring is to use the benefits of the entire DAS array. We recently succeeded in using the ML picker PhaseNet-DAS for real-time phase picking. Eventually, we plan to submit picks directly into the pick-ring that goes to the associator and earthquake locator. The final steps for the SDEDC are to archive DAS data including station metadata. DAS data is large and will be saved as triggered data per event. Then it will be made accessible through similar methods used with seismic data through the cloud.

Citation
Husker, A. L., Tepp, G., Yu, E., Biondi, E., Bhadha, R., Tam, R., Bhaskaran, A., Shimota, B., Artiaga, N., & Zhan, Z. (2024, 09). Challenges of incorporating Machine Learning and DAS into the Southern California Seismic Network (SCSN) and Southern California Earthquake Data Center (SCEDC) . Poster Presentation at 2024 SCEC Annual Meeting.

Related Projects & Working Groups
Research Computing (RC)