A Hybrid Deep Learning Framework for 1D Transient Electromagnetic Inversion with Multi-Scale Convolution and Transformer-Based Global Modeling

Abstract Transient electromagnetic (TEM) inversion aims to estimate subsurface resistivity structures from measured transient responses. Although one-dimensional (1D) inversion methods are widely used, accurately resolving deep resistivity variations remains challenging due to the complex temporal dependence between early- and late-time responses. To address this issue, we propose a deep learning framework for efficient 1D TEM resistivity inversion. The model combines a multiscale convolution module to capture response features at different temporal scales, a Transformer encoder to model long-range dependencies, and a U-Net-based encoder-decoder architecture with skip connections to preserve structural information during feature reconstruction. Experiments on synthetic datasets show that our proposed method outperforms conventional CNN, Transformer, and MLP models in both accuracy and robustness. Quantitative results indicate RMSE values of 0.0633 and 0.0442 and RE values of 0.0358 and 0.0404 on the 7-layer and 30-layer datasets, corresponding to RMSE reductions of 55.7% and 28.5% compared with the CNN baseline. In particular, the model shows enhanced capability in recovering deeper resistivity structures. Application to field TEM data further confirms its practical effectiveness. These results suggest that our proposed framework provides an effective data-driven strategy for improving the efficiency and reliability of 1D TEM resistivity inversion.