Scaling Up Large Fourier Neural Operator Training in 3D Seismic Waveform Modeling
Recent developments in using Fourier Neural Operators (FNO) to solve 2D and 3D elastic wave equations provide the basis for using machine learning tools for seismic wave modeling, thus enabling fast full waveform inversion in various applications. However, training these models usually requires heavy computation both in time and hardware, such as HPCs, with limitations in scaling the model up. In this presentation, we build on the prior efforts (Yang et al., 2023; Zou et al., 2023) to experiment with ways to improve the larger model training to facilitate the use of this approach in large-scale real-world applications. We test transfer learning, model parallelization, physics-informed neural operators to make training larger models easier. This evaluation of different approaches provides guidance for future training of large neural operator models for full waveform inversion applications.
Session: Machine Learning for Full Waveform Inversion: From Hybrid to End-to-End Approaches - I
Type: Oral
Room: Kahtnu 2
Date: 5/3/2024
Presentation Time: 05:15 PM (local time)
Presenting Author: Qingkai Kong
Student Presenter: No
Additional Authors
Qingkai Kong Presenting Author Corresponding Author kong11@llnl.gov Lawrence Livermore National Laboratory |
Eric Matzel matzel1@llnl.gov Lawrence Livermore National Laboratory |
Caifeng Zou czou@caltech.edu California Institute of Technology |
Youngsoo Choi choi15@llnl.gov Lawrence Livermore National Laboratory |
Zachary Ross zross@gps.caltech.eud California Institute of Technology |
Kamyar Azizzadenesheli kamyara@nvidia.com Nvidia |
Arthur Rodgers rodgers7@llnl.gov Lawrence Livermore National Laboratory |
Robert Clayton clay@gps.caltech.edu California Institute of Technology |
|
Scaling Up Large Fourier Neural Operator Training in 3D Seismic Waveform Modeling
Category
Machine Learning for Full Waveform Inversion: From Hybrid to End-to-End Approaches
Description