SCEC Award Number 20140 View PDF
Proposal Category Individual Proposal (Integration and Theory)
Proposal Title Machine Learning-Based Tomography of Ridegecrest region
Investigator(s)
Name Organization
Peter Gerstoft University of California, San Diego Kim Olsen San Diego State University
Other Participants Michael Bianco (postdoc)
Dylan Snover (Graduate student)
SCEC Priorities 3d, 3e, 3a SCEC Groups Seismology, CS
Report Due Date 03/15/2021 Date Report Submitted 05/08/2021
Project Abstract
 We perform ambient noise tomography using data recorded from Nodal arrays within a ~50 km by 50 km area including the July 2019 M7.1 and M6.4 Ridgecrest, CA, earthquakes. The imaging uses a locally sparse tomography (LST) approach with unsupervised dictionary learning and least-squares regularization that directly learns from local patches without requiring a large volume of training data. We resolve a 1-5 km wide region of shallow low S-wave velocities, reduced by about 40% relative to the surrounding host rock, and localizing from the surface to at least 1 km depth. This low-velocity zone (LVZ), surrounding the fault traces for the M7.1 and M6.4 events, is surprisingly well correlated with the extent of the distributed faulting as mapped by differences between daily passes of the PlanetLabs Satellite imagery, suggesting a possible causative relation between the imaged LVZ and the rupture sequence. Our imaged LVZ likely consists of a combination of shallow sediments/weathered rock and a damage zone associated with the causative faults for the M6.4 and 7.1 ruptures. Our imaging results are consistent with a heterogeneous flower structure of (predominantly shallow, <~ 1 km) rock damage, as found in nonlinear models caused by plastic reduction in velocities. Our results from LST tomography are consistent with those from conventional least squares while the LST method achieves a smaller misfit error.
Intellectual Merit We perform ambient noise tomography using data recorded from Nodal arrays within a ~50 km by 50 km area including the July 2019 M7.1 and M6.4 Ridgecrest, CA, earthquakes. The imaging uses a locally sparse tomography (LST) approach with unsupervised dictionary learning and least-squares regularization that directly learns from local patches without requiring a large volume of training data. We resolve a 1-5 km wide region of shallow low S-wave velocities, reduced by about 40% relative to the surrounding host rock, and localizing from the surface to at least 1 km depth. This low-velocity zone (LVZ), surrounding the fault traces for the M7.1 and M6.4 events, is surprisingly well correlated with the extent of the distributed faulting as mapped by differences between daily passes of the PlanetLabs Satellite imagery, suggesting a possible causative relation between the imaged LVZ and the rupture sequence. Our imaged LVZ likely consists of a combination of shallow sediments/weathered rock and a damage zone associated with the causative faults for the M6.4 and 7.1 ruptures. Our imaging results are consistent with a heterogeneous flower structure of (predominantly shallow, <~ 1 km) rock damage, as found in nonlinear models caused by plastic reduction in velocities. Our results from LST tomography are consistent with those from conventional least squares while the LST method achieves a smaller misfit error.
Broader Impacts The work has supported one graduate student and one Postdoctoral researcher.
Exemplary Figure Figure 1 from report
A composite 3D image of Rayleigh wave velocities in the Ridgecrest area, obtained from inversion of Rayleigh waves dispersion. Low velocity zones can be observed around the flower-shaped structures. A composite 3D image of Rayleigh wave velocities obtained by the sparse arrays is shown in Figure below. Generally, the surface traces from the M6.4 and M7.1 Ridgecrest events intersect the cross sections at the center of the LVZs, and the shallow cross sections reveal distinct LVZ flower structures, as observed by Zigone [2019]. However, the complexity of the fault zone, as characterized from the variation of particularly the width of the LVZ, is remarkable. For example, the fault zone at arrays A1 and B1 appear to delineate two or more separate low-velocity parts of the fault zone, which may represent a concentration of damage along different locations of the rupture for past events.