PINNtomo: Seismic tomography using physics-informed neural networks
Abstract
Seismic tomography has been used over the years as a pre-eminent tool for subsurface model building at various scales ranging from global and regional scales in earthquake seismology to local scales in exploration seismology. Conventional tomography methods suffer from a number of limitations, including the use of some form of smoothing regularization to compensate for the ill-posedness of the problem. This ends up limiting the resolution of the inverted velocity model. Moreover, these methods typically need an initial model with some general background features of the Earth represented, like a constant depth gradient. The choice of the initial model may affect the final solution and is usually not obvious prior to inversion. Furthermore, for models with irregular topography, considerable grid and algorithmic adaptations are needed to account for the free-surface topography. We propose a novel algorithm for the seismic tomography problem based on developments in the field of scientific machine learning. We use the emerging paradigm of physics-informed neural networks (PINNs) that overcomes the limitation of deep learning associated with sparse data by incorporating the governing partial differential equation into the neural network's loss function. Specifically, we develop a PINN-based tomography algorithm to invert for the velocity model. Given traveltimes at seismic stations covering part of the computational domain, we use neural networks to approximate the traveltime factor and the velocity fields, subject to the physics-informed regularizer based on the factored eikonal equation. Doing so allows us to better compensate for the poorly determined aspects of the velocity model compared to conventional smoothing regularizers. Also, we find the performance of the method to be independent of the initial velocity model. Moreover, the method is easily adaptable to models with irregular topography. Additional advantages of the method include ease of deployment across a variety of platforms (CPUs, GPUs) and architectures (desktops, clusters) without any modification. We demonstrate the efficacy of the proposed algorithm in solving the tomography problem by obtaining a velocity model that produces traveltimes matching those observed at seismic stations while honoring the physics of wave propagation.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.S12B..03W