Joint Inversion of Seismic and Resistivity Data Powered by Deep Learning

Incorporating multiple perspectives makes joint inversion of multiple geophysical data sets an effective way for improving the accuracy of imaging complex geological structures. In this article, drawing inspiration from the inherent nonlinear mapping abilities of deep learning (DL), we introduce a groundbreaking joint inversion framework and network named JointInvNet. Unlike end-to-end networks that directly map geophysical data to models, we propose a hybrid inversion framework that combines insights from the physical laws with data-driven learning, which iteratively updates the independently inverted results simultaneously via JointInvNet. In particular, it is assumed that different geophysical parameters change on both sides of the geological boundary, and the Laplace convolution operator is used to extract boundary information and provide structural constraints for the loss function. To demonstrate the advantages over traditional separate inversion and cross-gradient inversion, numerical experiments are performed on seismic and resistivity data. As illustrated by visual and quantitative comparisons, JointInvNet could lead to satisfactory inversion results, with excellent agreement with ground-truth models and good generalization ability to more complex models. Moreover, weight settings between seismic and resistivity model parameters and applicability when structural similarity assumptions do not hold are discussed to illustrate the potential of the proposed method.