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

Summary In this abstract, we introduced a 3D TEM inversion network, UT-Net, that integrates the self-attention mechanism of Transformer into U-Net. UT-Net addresses the limitation of U-Net in capturing global features by dynamically allocating weights to different time channels. Additionally, data augmentation methods are employed to reduce the time required for generating new training datasets. Experimental results on synthetic data demonstrate that UT-Net with data augmentation achieves highly accurate inversion results in terms of the positions and boundaries of subsurface anomalies, significantly outperforming traditional U-Net. Moreover, UT-Net is also successfully applied to field data, accurately identifying shallow aquifer collapse zones. The results demonstrate that UT-Net demonstrates strong effectiveness and offers a novel method for 3D electromagnetic inversion in complex geological environments.