Seismic Inversion by Hybrid Machine Learning

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