Leveraging Full-Waveform Simulations to Learn Featured-Based Models for Bayesian Seismic Monitoring

Session: Innovative Seismo-Acoustic Applications to Forensics and Novel Monitoring Problems [Poster]
Type: Poster
Date: 4/29/2020
Time: 08:00 AM
Room: Ballroom
Description: 

Detecting, locating and characterizing weak seismic events pushes the limits of existing methods when the signal-to-noise ratio is small. To improve our monitoring capabilities, we must develop methods that simultaneously integrate waveform data from across the seismic network so that the combined signature of all weak signals can be detected. This requires modelling the seismic waveform for different candidate events while taking into account modeling uncertainty, sensor noise and signal background. Therefore, we take a Bayesian approach. Further, instead of building a model that predicts the seismic waveform itself, as in SIGVISA (Moore et al. 2017), we build a statistical model that predicts waveform features. By taking a feature-based approach, we do not require computationally expensive full-waveform simulations for inference, and we are less sensitive to assumptions about earth structure and seismic event characteristics, which can significantly influence of the overall shape of the waveform.

The key challenge to this approach is developing the feature-based model. To do this we utilize full-waveform simulations and historic waveform data. By simulating thousands of events with different event locations, sensor locations, focal mechanisms and earth model heterogeneity assumptions, we build a set of waveforms that incorporating our modeling uncertainty. From these waveforms, we build a statistical model of features that we expect to see for a candidate event. One example of useful features are seismic phase arrival times and the integrate signal power within the arrival window. Historic data then provides useful information for building a statistical model of sensor noise and the background environment that we incorporate into the feature model. We then incorporate these models into a Bayesian inference problem to identify seismic events and demonstrate this approach by analyzing events from the 2011 Circleville earthquake sequence in Utah.

SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525. SAND2020-0369 A

Presenting Author: Thomas A. Catanach

Authors