SCEC Award Number 22053 View PDF
Proposal Category Individual Proposal (Integration and Theory)
Proposal Title Machine Learning-based Super-resolution Tomography of the Ridgecrest Region
Investigator(s)
Name Organization
Peter Gerstoft University of California, San Diego Kim Olsen San Diego State University
Other Participants Project Scientist Michael Bianco,
Graduate Student Zheng Zhou
SCEC Priorities 3d, 3e, 3a SCEC Groups Seismology, CXM, CS
Report Due Date 03/15/2023 Date Report Submitted 03/14/2023
Project Abstract
 Fusing various tomography models with different resolutions is desired when updating community models, we propose a novel approach to fuse multi-resolution seismic tomography maps with probability graphical models (PGMs). The PGMs are employed to provide segmentation results on the seismic tomography images with various resolutions, and the segmentation results here correspond to various velocity intervals. Furthermore, by taking the physical information (such
as ray-path density) into consideration, we introduce the physics-informed probability graphical
models (PIPGMs). We present the relation between subdomains with multiple resolutions, in terms of high-resolution (HR) and low-resolution (LR) components. By transferring the distribution information from the HR parts, the details in the LR areas can be enhanced by solving a maximum likelihood problem with prior knowledge from HR models informed.
To evaluate the efficacy of the proposed PGM fusion method, we employ the model on both synthetic checkerboard models and a fault zone structure imaged from the 2019 Ridgecrest, CA, earthquake sequence. The Ridgecrest fault zone image consists of shallow-scale (top 1 km) high-resolution surface wave models obtained from ambient noise tomography, which is embedded into the SCEC Community Velocity Model (CVM) version S4.26-M01. Our model is evaluated by the misfit between observed and calculated travel times. The evaluation results show that our proposed method can achieve better performance than the conventional Gaussian smoothing and the baseline machine learning methods.
Intellectual Merit Fusing various tomography models with different resolutions is desired when updating community models, we propose a novel approach to fuse multi-resolution seismic tomography maps with probability graphical models (PGMs). The PGMs are employed to provide segmentation results on the seismic tomography images with various resolutions, and the segmentation results here correspond to various velocity intervals. Furthermore, by taking the physical information (such
as ray-path density) into consideration, we introduce the physics-informed probability graphical
models (PIPGMs). We present the relation between subdomains with multiple resolutions, in terms of high-resolution (HR) and low-resolution (LR) components. By transferring the distribution information from the HR parts, the details in the LR areas can be enhanced by solving a maximum likelihood problem with prior knowledge from HR models informed.
To evaluate the efficacy of the proposed PGM fusion method, we employ the model on both synthetic checkerboard models and a fault zone structure imaged from the 2019 Ridgecrest, CA, earthquake sequence. The Ridgecrest fault zone image consists of shallow-scale (top 1 km) high-resolution surface wave models obtained from ambient noise tomography, which is embedded into the SCEC Community Velocity Model (CVM) version S4.26-M01. Our model is evaluated by the misfit between observed and calculated travel times. The evaluation results show that our proposed method can achieve better performance than the conventional Gaussian smoothing and the baseline machine learning methods.
Broader Impacts The work has supported one graduate student.
Exemplary Figure Figure 1: (a) The LR CVM-S4.26 model near Ridgecrest. (b) Result of directly inserting the HR 1 Hz Rayleigh wave tomography into the LR model. Combined LR and HR models, smoothed by (c) 7×7 and (d) 3×3 average filters. (e) Synthetic stations (red ’X’s) are deployed on the boundaries between HR and LR models for evaluation.