|
Hello Network Seismology Community and Beyond!
Mailing
List Sign Up
Join us for our next seminar!
Thursday August 20th, 2 pm MT: Yijian Zhou - Caltech - Toward an AI-Enhanced Near-Real-Time Earthquake Cataloging System for the Southern California Seismic Network
Abstract:
Rapid and complete microseismicity monitoring is increasingly important as modern seismic networks evolve, supporting scientific investigations, hazard assessment, and timely situational awareness following significant earthquakes. Although recent AI/ML algorithms
have substantially improved phase picking, their integration into routine regional network operations has remained limited. Here, we present the development of an AI-enhanced near-real-time cataloging framework for the Southern California Seismic Network (SCSN).
This framework separates offline model training from online inference. Continuous waveform data are processed using modular AI phase pickers, followed by subnetwork-based phase association and optional refinement of arrival times using pickers pre-trained on
a California-wide positive dataset. The resulting phase files are subsequently processed by the existing SCSN location and magnitude modules, ensuring compatibility with current operational products while allowing AI components to evolve independently. The
modular design also enables straightforward integration and comparison of multiple phase picking and association algorithms.
The offline workflow constructs an enhanced reference catalog for 2020–2025 by processing continuous waveform data archived in the SCEDC AWS Dataset. We first apply a rule-based workflow Phase, Association, and Location (PAL; Zhou et al. 2021) to generate positive
and negative training samples for individual station-day waveform segments, which are used to train four AI pickers: Self-Attention RNN (SAR), PhaseNet, Res-UNet, and Frame-Transformer. These SCSN-optimized pickers are deployed in the online workflow while
simultaneously producing an enhanced retrospective catalog for 2020–2025.
The online workflow processes continuously incoming waveform data using overlapping 10-minute windows with 5-min steps. Within each window, every picker is applied using a short 2.5-s sliding stride, so that the same seismic arrival is evaluated multiple times
at different 25-s window positions. Candidate arrivals are retained when they are detected repeatedly by the same model across overlapping windows and/or by multiple pickers. Phase association is then performed independently within overlapped subnetworks,
after which detections are merged to remove duplicates. This modular strategy provides a pathway for incorporating future advances in earthquake detection and phase association into operational cataloging.
Do you have an idea for a seminar topic? Get in touch with the seminar committee (Will Yeck, Gabrielle Tepp, Mairi Litherland, Kyren Bogolub, and Bill Barnhart)
|
