Magnetotelluric Inversion Using Dual-Channel Deep Learning With Physical Information Constraints: TM Case

In traditional magnetotelluric (MT) inversion, there is the problem of insufficient inversion resolution, which cannot fully meet the needs of geological exploration for the analysis of high-resolution electrical structures. Although deep learning (DL)-based direct inversion methods have achieved remarkable results in synthetic data inversion, they encounter numerous difficulties when dealing with actual MT measurement data. Especially in the absence of physical information, these methods cannot effectively adapt to the complex characteristics of actual MT measurement data. This study focuses on developing a high-resolution and highly robust MT neural network inversion method. Considering the gradual change of the underground resistivity structure, we use cubic spline interpolation to generate random synthetic models with continuously varying resistivity values, constructing training datasets that closely resemble actual underground models to enhance the practicality of the inversion. Additionally, synthetic relative noise following a Gaussian distribution is added to the training data. A U-shaped DL model (TUNet) is constructed using a transformer, and then a MT inversion neural network with physical information constraints dual-channel DL (PDI-TUNet) is realized. Experimental results show that PDI-TUNet performs excellently in synthetic data tests, regardless of whether the data are noisy or noise-free. In the processing of actual field data, PDI-TUNet can reconstruct the subsurface resistivity structure in better accordance with the observed data, fully demonstrating its great potential in the processing of MT field data.