Fast forward modeling of 3D magnetotelluric via neural networks

Magnetotelluric (MT) forward modeling is the basis of inversion interpretation, and the process of inversion calculation needs to call forward modeling constantly, and MT forward modeling is also the main means to analyze the underground electrical structure. Traditional 3D MT forward modeling is mainly realized by numerical simulation methods such as finite element method (FEM) and finite difference method (FDM). The mesh size and the accuracy of the dissection determine the time required for calculation, and the realization of fast forward modeling helps to improve the efficiency of electromagnetic exploration. The rise of deep learning technology provides a new strategy to realize this goal, and the mapping process of forward modeling can be constructed by neural networks. In this paper, we realize 3-D MT fast forward modeling by constructing a deep neural network to accurately predict the apparent resistivity and impedance phase in different polarization directions of the theoretical geoelectric model. Compared with the traditional FDM, the rate of 3-D forward modeling of MT based on deep learning is significantly improved and provides a new solution for the traditional MT inversion.