Helmholtz Neural Operator for Full Waveform Inversion Tomography of California
Description:
Neural Operators (NOs) are a class of machine learning algorithms to learn the solutions of differential equations. NO models are trained with a large set of forward simulations run on high-performance computing (HPC). A recent study by Zou et al. (2024) described how Helmholtz NOs (HNOs) can simulate seismic waveforms in three-dimensions by considering solutions for a subset of the frequency bandwidth. Importantly, HNOs can be evaluated for new solutions with arbitrary 3D structure and source properties much faster than forward simulations. HNOs can be used to compute sensitivity kernels for full waveform inversion (FWI) by automatic differentiation and provide an alternative to adjoint-state simulations in waveform tomography to infer Earth’s seismic wavespeed structure. This study describes work-in-progress exploring the efficacy of HNOs for regional-scale FWI tomography of parts of California. We assembled a data set of broadband waveforms from moderate (MW 4.0-6.0) earthquakes with reported source properties spanning the broad plate boundary (Pacific Ocean to western Nevada; Gulf of California to the Mendocino Triple Junction). Conventional multiscale FWI based on the adjoint-state was performed with Salvus with a simple starting model building on our past experience with tomography of the region. The conventional FWI tomography provides a benchmark for which to compare tomography based on HNOs. Results will be presented with attention on comparison of the efficacy and accuracy of solutions using these different approaches.
Session: Scientific Machine Learning for Forward and Inverse Wave Equation Problems - I
Type: Oral
Date: 4/15/2025
Presentation Time: 04:30 PM (local time)
Presenting Author: Arthur
Student Presenter: No
Invited Presentation:
Poster Number:
Authors
Arthur Rodgers Presenting Author Corresponding Author rodgers7@llnl.gov Lawrence Livermore National Laboratory |
Claire Doody doody1@llnl.gov Lawrence Livermore National Laboratory |
Qingkai Kong kong11@llnl.gov Lawrence Livermore National Laboratory |
Caifeng Zou czou@caltech.edu California Institute of Technology |
Kamyar Azizzadenesheli kamyara@nvidia.com Nvidia |
Zachary Ross zross@gps.caltech.edu California Institute of Technology |
Robert Clayton clay@gps.caltech.edu California Institute of Technology |
|
|
Helmholtz Neural Operator for Full Waveform Inversion Tomography of California
Session
Scientific Machine Learning for Forward and Inverse Wave Equation Problems