Seismic inversion by hybrid machine learning

Abstract

We present a new seismic inversion method that uses deep learning (DL) features for the subsurface velocity model estimation. The DL feature is a low-dimensional representation of the high-dimensional seismic data, which is automatically generated by a convolutional autoencoder (CAE) and preserved in the latent space. The low-dimensional DL feature contains the key information of the input seismic data. Therefore, instead of directly comparing the waveform differences between the observed and predicted data, such as full-waveform inversion (FWI). We measure their DL feature differences in the latent space of a CAE. The advantage of this low-dimensional comparison is that it is less prone to the cycle-skipping problem compared to FWI. The reason is that the DL features mainly contain the kinematic information, such as traveltime, of the input seismic data when the latent space dimension is small. However, more dynamic information, such as the waveform variations, can be preserved in the DL feature when the latent space dimension becomes larger. Therefore we propose a multiscale inversion approach that starts with inverting the low-dimensional DL features for the low-wavenumber information of the subsurface model. Then recover its high-wavenumber details through inverting the high-dimensional DL features. However, there is no governing equation that contains both the velocity and DL feature terms in the same equation. Therefore we use the automatic differentiation (AD) to numerically connect the perturbation of DL features to the velocity perturbation. In another word, we connect a deep learning network with the wave-equation inversion by using the AD. We denote this hybrid connection as hybrid machine learning (HML) inversion. Here, the AD replaces the complex math derivations of the gradient with a black box so anyone can do HML without having a deep geophysical background.