Full Seismic Waveform Modeling and Inversion with Neural Operators
Abstract
Full waveform modeling is powerful in fine-tuning the earth structure models and unraveling earthquake source processes, but it is usually computationally expensive. Here, we propose to accelerate full waveform modeling using a recently proposed machine learning paradigm called neural operator. Once trained, the neural operator can simulate the wavefield at negligible cost. We train a U-shaped neural operator (U-NO) model on an ensemble of numerical simulations performed with random velocity models and source locations to learn the solution to the 2D elastic wave equation. We demonstrate that the trained model enables rapid and accurate simulation for arbitrary source locations, velocity structures and mesh discretization far off from the training data. The trained U-NO also allows reverse-mode automatic differentiation for efficient full-waveform inversion.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.S16B..07C