Seismic Inversion by Hybrid Machine Learning
Abstract
We develop a new seismic inversion based on hybrid machine learning (HML) method to invert for the subsurface velocity model using the skeletonized information that is automatically generated by the convolutional autoencoder (CAE). A colinear constraint guarantees that CAE can compress the key features of complex input seismic signals into the latent space, rather than mapping the inputs directly to the outputs. We measure the data misfit in the low dimensional latent space of CAE. The new misfit measure is less prone to cycle skipping when the latent space dimension is small. We use an automatic differentiation (AD) schema to compute the derivative of new misfit function with respect to the subsurface model parameters by backpropagating the gradient from latent space to the model parameters. Numerical results on both synthetic and field crosswell seismic data show that the HML method can recover both the low- and high-wavenumber information of the subsurface velocity model effectively. Our work provides a general framework using the wave equation to invert the skeletal features generated by deep learning networks, which combine the wave equation physics with the pattern learning ability of deep learning.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMS064.0016C
- Keywords:
-
- 3225 Numerical approximations and analysis;
- MATHEMATICAL GEOPHYSICS;
- 3260 Inverse theory;
- MATHEMATICAL GEOPHYSICS;
- 7260 Theory;
- SEISMOLOGY;
- 7290 Computational seismology;
- SEISMOLOGY