Deep Learning Enhanced Joint Geophysical Inversion for Crosswell Monitoring

Abstract

A deep learning enhanced framework is proposed to jointly invert the crosswell DC resistivity and seismic travel time data. With the strong capability to extract the implicit patterns of the input data, our deep neural network is trained to fuse and extract the connections between separately inverted resistivity and velocity models by the conventional methods, while the structural similarity is imposed by letting the outputs of network approach the true resistivity and velocity models with the same structures. In the joint inversion framework, the well-trained network is adopted in an iterative way to generate the enhanced resistivity and velocity models to perform as the inputs for next round of inversion. Moreover, under our framework, multiple geophysical data can be used simultaneously to jointly invert the corresponding multiple properties. Numerical simulation demonstrates an improved accuracy of our method.