An Accelerated Magnetotelluric 2D Forward Modeling Network Model: Transformer+Unet

Abstract The The accuracy and efficiency of two-dimensional electromagnetic forward simulation calculations are crucial for the success of the inversion process. However, traditional numerical simulation methods used for forward modeling are computationally intensive and slow, especially on personal computer systems. To solve this problem and improve computational efficiency, we propose a scheme based on Transformer-Unet (T-Unet) to accelerate the establishment of a two-dimensional magnetotelluric model. The purpose of constructing a T-Unet neural network is to establish a mapping between the geoelectric model and the apparent resistivity while also creating the corresponding dataset. Through network training and iteration, a neural network weight model is obtained, enabling the prediction of apparent resistivity and phase values for forward modeling results. Experimental results demonstrate that compared to traditional finite element forward modeling, T-Unet not only significantly reduces computational time but also achieves high forward calculation accuracy. Furthermore, T-Unet is applied to NLCG inversion. Similar to the finite element forward modeling results for NLCG, T-Unet inversion results accurately determine the location and structure of anomalous bodies. We firmly believe that deep learning neural networks have the potential to accelerate the improvement of forward computation time efficiency for inversion solutions. In summary, the proposed T-Unet scheme provides an effective solution for enhancing the efficiency of two-dimensional electromagnetic forward modeling. By utilizing deep learning technology, it is expected to advance the inversion process and accelerate forward calculations in the field of geophysics.