Joint 2D inversion of AMT and seismic traveltime data with deep learning constraint

Abstract

We apply deep learning techniques to assist the joint inversion of audio-magnetotelluric (AMT) and seismic travel time data. Based on the assumption that the resistivity-velocity relationship is known a priori, deep neural networks are used to learn the nonlinear maps between these two parameters. In this manner, the correlation between resistivity and velocity in both structure and value can be implicitly established through the neural network. During the inversion, we alternatively update resistivity and velocity using the Gauss-Newton method. Moreover, both resistivity and velocity are updated based on the reference model mapped from the other parameter by deep residual convolutional neural networks (DRCNNs). Numerical example shows that this joint inversion scheme can achieve more accurate inversion and converges faster than separate inversion.