SCEC Award Number 23199 View PDF
Proposal Category Individual Proposal (Integration and Theory)
Proposal Title Improved Urban Seismic Monitoring Under Metropolitan Los Angeles
Investigator(s)
Name Organization
Gregory Beroza Stanford University Artemii Novoselov Stanford University
Other Participants
SCEC Priorities 3b, 3g SCEC Groups Seismology, CS, CXM
Report Due Date 03/15/2024 Date Report Submitted 11/15/2024
Project Abstract
 Work under this grant was motivated by by the results of Inbal et al. (2016) and follow-up work by Yang et al. (2021) to explore the evidence for deep seismicity under Los Angeles. For this purpose, we combined a probabilistic deep-learning-regression arrival time measurement model (PhaseHunter) illustrated in Figure 1, with a (soft-clustering) Gaussian-Mixture-Model-based phase association algorithm (GaMMA) to search for additional phases and events that would better illuminate deep seismicity in greater Los Angeles. The work was intended to be collaborative with Asaf Inbal of Tel Aviv University, but due to geopolitical events, the work was terminated before the collaboration could take effect.
Intellectual Merit The transform plate boundary in southern California extends from the Borderlands Offshore to the Eastern California Shear Zone in the Mojave Desert and the Walker Lane beyond. While much of the deformation occurs as strike-slip faulting, extension in the Gulf of California and Salton Trough as well as compression across the Transverse Ranges in the Big Bend of the San Andreas are important exceptions. The temperature and stress conditions in the lithosphere are thought to determine the yield strength envelope, which is bounded by friction in the upper crust, and by viscous flow or creep in the lower crust and upper mantle. Observations of this are related to the depth of deepest seismicity.
Broader Impacts The Probabilistic Deep Regression Model tested in this project is unique in that it estimates epistemic uncertainty in phase picks. That is useful for seismology, and potentially useful for other machine learning models.
Exemplary Figure Figure 1. PhaseHunter, which is available at: https://github.com/crimeacs/PhaseHunter, uses a Maxsemble ML architecture for probabilistic deep-learning measurement of phase arrival times. Figure shows workflow of Maxsembles for uncertainty estimation in seismic phase detection. (a) 3-component waveform input, (b) N copies of input are batched for inference, (c) neural network with subnetworks (Maxsembles), (d) Predicted P and S picks with mean and standard deviation, (e) Distribution of predictions for P and S picks.
Linked Publications

Add missing publication or edit citation shown. Enter the SCEC project ID to link publication.