Physics-informed deep learning for magnetotelluric 2D forward modeling

Abstract

Summary Magnetotelluric method is widely used in mineral and oil & gas exploration. The forward modeling and inversion for 1D MT have been quite mature. For 2D cases, however, numerical simulation is needed because few analytical solutions are available. In recent years, the rapid development of machine learning (ML) has facilitated the extraction of information from massive data. In many physics and engineering fields, however, the problem to be solved often satisfies certain partial differential equations (PDEs) and this kind of prior knowledge is not reflected in classic ML algorithms. Physics-informed neural networks were proposed for the solution of PDEs with physical laws serving as the regularization term of the loss function. Combined with the boundary conditions, the physical field at any point in the domain of interest can be predicted with the trained NNs. In this paper, we use PINNs for 2D MT forward modeling. After training the network, the real and imaginary parts of the magnetic field at any location in space can be obtained. Numerical examples prove that PINNs can fulfil effective MT 2D forward modeling.