Numerical analysis of a deep learning formulation of elastic full waveform inversion with high order total variation regularization in different parameterization

Abstract

We have formulated elastic seismic full waveform inversion (FWI) within a deep learning environment. Training this network with a single seismic data set is equivalent to carrying out elastic FWI. There are three main motivations for this effort. The first is an interest in developing an inversion algorithm which is more or less equivalent to standard elastic FWI but which is ready-made for a range of cloud computational architectures. The second is an interest in algorithms which can later (i.e., not in this paper) be coupled with a more involved training component, involving multiple datasets. The third is a general interest in developing the idea of theory-guiding within machine learning solutions for large geophysical problems, wherein the number of degrees of freedom within a network, and the 1