Deep Learning Enhanced Multi-physics Joint Inversion

Joint inversion has drawn considerable attention due to the availability of multiple geophysical datasets, ever-increasing computational resources, development of advanced algorithms, and its ability to reduce inversion uncertainty. A key issue of joint inversion is to develop effective strategies to link different geophysical data in a unified mathematical framework, where the information obtained from different models can complement each other. Traditionally, structural similarity constraints are pursued by joint inversion algorithms using manually crafted formulations (e.g. cross gradient). In this dissertation, we introduce a novel approach: a Deep Learning Enhanced (DLE) joint inversion framework. Within this framework, structural similarity is enforced using a well-trained deep neural network (DNN), enhancing the quality of joint inversion results through iterative improvements. The DNN is seamlessly integrated into existing independent inversion workflows, with the flexibility to extend its application to various multi-physics scenarios without requiring structural modifications. Within the DLE joint inversion framework, several key contributions are made. First, we design a double-channel U-Net for the simultaneous inversion of 2D DC resistivity data and seismic travel time. Extensive numerical experiments validate the efficacy of this method. Importantly, this learning-based approach exhibits impressive generalization capabilities when applied to datasets featuring diverse geological structures, sensing configurations, and nonconforming discretization. Second, we harness the power of deep perceptual loss as a regularization technique to further enhance structural similarity. Successive networks are trained with deep perceptual constraints, derived from a pre-trained network specializing in edge detection. The robustness of this approach is verified through experiments involving layered subsurface models, demonstrating its ability to jointly invert three types of geophysical data, including induced polarization data. Third, we simplify the DLE framework and apply it to tackle the challenging 3D joint inversion of magnetic and gravity gradient data. The proposed method is rigorously evaluated through synthetic and field cases, affirming its effectiveness and computational efficiency. In summary, this dissertation contributes to the advancement of multi-physics joint inversion by introducing the Deep Learning Enhanced framework. This innovative approach enhances both the accuracy and efficiency of geophysical inversion, promising broader applications and improved outcomes in the field.