Research on two-dimensional magnetotelluric inversion based on residual neural networks

The paper conducts a two-dimensional inversion study of magnetotelluric data based on residual neural networks. It designs and generates 50,000 models of electrical anomaly bodies. These models are subjected to forward calculations using a two-dimensional staggered grid finite difference batch parallel forward modeling program to compute the forward responses. The TE polarization mode, considering resistivity as the depth learning inversion model, is utilized as training samples for the deep learning model, employing the ResNet-50 network structure in the design. Finally, the reliability of the deep learning inversion model for magnetotelluric data is validated through inversion testing. The results of the inversion indicate that the neural network inversion model can provide real-time inversion outcomes that are reasonably accurate. It effectively delineates the boundaries of various anomaly bodies and accurately recovers the electrical parameters of these anomalous bodies.