Traditional gradient-based inversion methods usually suffer from the problems of falling into local minima and relying heavily on initial guesses. Deep-learning methods have received increasing attention due to their excellent nonlinear fitting ability. However, given the recent application of deep-learning methods in the field of magnetotelluric (MT) inversion, there are currently challenges associated with achieving high inversion resolution and extracting sufficient features. We develop a neural network model (called MT2DInv-Unet) based on the deformable convolution for 2D MT inversion to approximate the nonlinear mapping from the MT response data to the resistivity model. The deformable convolution is achieved by adding an offset to each sample point of the conventional convolution operation, which extracts hidden relationships and allows the flexible adjustment of the size and shape of the feature region. Meanwhile, we design the network structure with multiscale residual blocks, which effectively extract the multiscale features of the MT response data. This design not only enhances the network performance but also alleviates issues such as vanishing gradients and network degradation. The results of synthetic models indicate that our network inversion method has stable convergence, good robustness, and generalization performance, and it performs better than the fully convolutional neural network and U-Net network. Finally, the inversion results of field data show that MT2DInv-Unet can effectively obtain a reliable underground resistivity structure and has a good application prospect in MT inversion.
The conventional magnetotelluric inversion method is subject to the influence of the initial model, which leads to an unstable inversion process and a tendency to get trapped at local optimal solutions. In contrast, deep learning technology relies on its powerful non-linear fitting capability and can construct complex non-linear mappings directly from observation data (input) to model (output). In recent years, it has received extensive attention from researchers. Due to the difficulties in creating a sufficiently large dataset and performing extensive neural network training, most current magnetotelluric inversion methods for geophysical exploration remain limited to one-dimensional (1D) or two-dimensional (2D) scenarios. To the best of our knowledge, for deep learning-based three-dimensional (3D) magnetotelluric inversion, currently there is no reported work in the literature. In this work, we propose a 3D magnetotelluric inversion method based on deep learning technology. By designing a neural network architecture for 3D structures (MT3D-Net), we achieve an end-to-end mapping from the network input to output. To alleviate the excessive dependence of the network on the training set, we introduce a joint weighted loss function based on data-driven and physics-driven method, allowing the network to follow the physical constraints of magnetotelluric data during the training process and thus more reasonably guide the update of network parameters. Numerical experiments show that this method combines the advantages of traditional and data-driven inversions, significantly improving the stability and accuracy of magnetotelluric inversion. The proposed method has been successfully applied to synthetic models and measured field data, and has good application prospects.
1 January 2021 in Journal of Computational Physics