Rapid Imaging of Subsurface Media With Magnetotellurics Based on Pix2Pix GAN

Rapidly imaging underground electrical structures is very challenging, especially for complex structures. Conventional inversion algorithms require repeated solving of large-scale forward problems, where the main computing cost goes. We have implemented an underground resistivity imaging method based on the Pix2Pix (P2P) generative adversarial network (GAN) structure. Our approach assembles phase information with regular resistivity observation and has superior imaging accuracy. We utilize Gaussian random fields (GRFs) to construct synthetic resistivity models. This practice provided meaningful geological representation and introduced some preferable variations to our GAN model. By comparing the accuracy of the prediction results, we determine the form of the loss function. Detailed qualitative and quantitative evaluations indicate that our multiple-parameter method is the most effective for inversion compared to using only a single parameter (apparent resistivity or impedance phase). We further introduce the objective function of conventional inversion into the loss function of GAN to process noisy data. The loss function with conventional geophysical constraints greatly enhances the resistance to noise. Finally, compared with the nonlinear conjugate gradient (NLCG) inversion method, our method demonstrates faster prediction capabilities and better performance in revealing high-resistivity bodies beneath low-resistivity layers on synthetic data, and the validation of real-world data highlights its practical applicability.