Physics-informed Deep Learning for Wave Propagation and Full Waveform Inversions
Abstract
We propose a new deep learning approach based on Physics Informed Neural Networks (PINNs) to solve the wave equation and perform full waveform inversions (FWIs). PINNs by exploiting the knowledge of the governing physical laws reduce the need for labeled training data sets, and therefore greatly facilitate the learning process in applications where the training data is limited, such as FWIs in seismic inversion. Through various synthetic case studies, we test PINNs efficiency in solving the wave propagation (forward modeling) and FWI problems in 2D acoustic media with increasing degree of structural complexity. We show that PINNs can efficiently recover smooth as well as discontinuous material heterogeneities in the subsurface using both teleseismic plane waves and point seismic sources. Furthermore, we generalize PINNs formalism to include multiple sources of seismic energy in a single inversion, to increase ray coverage of the area under study. We discuss PINNs flexibility in handling free surface and absorbing boundary conditions, two main types of boundary conditions encountered in seismic wave propagation and show that PINNs can seamlessly respect each of these constraints. Additionally, we extend our framework to explore PINNs application to wave propagation and inverse problems in elastic media. Finally, we discuss the computational costs associated with the synthetic case studies performed here as well as potentials of PINNs for incorporating multiple geophysical data types (e.g., magnetotellurics, gravity) in a joint inversion framework.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.S35D0246R