A 3-D TEM Inversion Method Based on Deep Learning With Transformer and U-Net

3-D inversion is the most challenging task for time-domain electromagnetic (TEM) studies due to its huge number of computations, especially when it comes to the real-data cases. The recent progress of deep learning (DL) method provides an alternative way for solving computational-extensive problems. This study focuses on developing efficient 3-D inversion scheme for inverting the TEM datasets based on DL methods. The U-Net, which has proved its effectiveness in TEM inversion, is revisited and redesigned here. The U-Net can extract local features from datasets though, and it struggles to capture the global ones that are important to ensure the consistency and reliability of the model parameters from different time channels. To address this limitation, we propose a novel approach (UT-Net) by integrating the self-attention mechanism of transformer into U-Net for TEM data inversion, which is rarely applied to geophysical inversion, especially for 3-D TEM problems. The UT-Net dynamically allocates different weights for each time channels, so that both the local and global features can be recovered. Combined with data augmentation strategies, an efficient and stable scheme for 3-D inversion of TEM data is established. The proposed method is validated by the synthetic and field examples.