Joint Inversion of Audio-Magnetotelluric and Seismic Travel Time Data Using Attribute Fusion Based on Deep Learning

We propose a deep learning scheme to assist joint inversion of audio-magnetotelluric(AMT) and seismic travel time data. A deep convolutional neural network is designed to fuse the separately inverted multi-physics models. An implicit relationship between inverted and true models can be established. During the inversion, the reconstruction resistivity and velocity models are updated with the Gauss-Newton method, based on the references generated by the trained neural network. Compared with separate inversion, this scheme can automatically incorporate interpretation knowledge into inversion with the help of deep learning. Numerical experiments show that it can achieve better inversion accuracy compared with separate inversion.