Flexible and Accurate Prior Model Construction Based on Deep Learning for 2-D Magnetotelluric Data Inversion

The conventional magnetotelluric (MT) data inversion methods, such as the nonlinear conjugate gradient method, quasi-Newton method, and Gauss–Newton method and so on, can converge robustly, but their results are easily affected by the initial model and regularization term. Although supervised learning can break through the resolution limitation by directly learning the nonlinear relationship between the model and the data, it cannot guarantee the data fitting without considering physical constraints. Here, we propose a novel prior model generation method using deep learning for conventional inversion to jointly take advantages of the two techniques. We first combine Gaussian random rough surface scheme and random polygon generation algorithm to construct practical 2-D geoelectric models, in which the prior information on the geoelectric structure can be flexibly integrated. Then, a fast 2-D MT forward modeling method is applied to calculate the forward responses and establish the training set. Finally, we use the training set to complete the parameter optimization of U-shaped network (U-NET) and run the conventional inversion with prior model generated by the trained U-NET. Numerical experiments with synthetic data show that the proposed method can effectively integrate the advantages of conventional inversion and supervised learning, and remarkably improve the resolution in the inversion if proper training sets are used. The inversions of the USArray data also prove that our method can retain high-resolution structures predicted by the U-NET in the final inversion results with a data fitting as good as the traditional Gauss–Newton method.