Deep learning inversion for Tipper response of Z-axis Tipper electromagnetic

• Published in: Journal of applied geophysics Vol. 241; p. 105794
• Main Authors: Yao, Yu, Zhang, Zhihou
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.10.2025
• Subjects: 2D Tipper response inversion; Dilated convolution; Fully convolution neural network; Multi-scale feature fusion; ZTEM; Fully convolution neural network; ZTEM; Dilated convolution; Multi-scale feature fusion; 2D Tipper response inversion
• ISSN: 0926-9851
• DOI: 10.1016/j.jappgeo.2025.105794

The massive amount of tipper data generated by Z-axis Tipper Electromagnetic (ZTEM) system requires a more sophisticated and fast inversion algorithm to precisely link the 2D tipper data with the geological structures. Traditional inversion methods are prone to getting stuck in local extremes and rely on the initial model. Exploiting the remarkable nonlinear mapping ability of deep learning, we propose the TipInv-net framework to establish the nonlinear relationship between tipper data and underground electrical parameters. In terms of data set construction, we have built single resistivity anomaly body model, combined resistivity anomaly body model and fault model data set for TipInv-net training. In the aspect of network construction, based on the classic U-net network framework, we use the average pooling at the input end to achieve multi-scale tipper response feature extraction to construct multi-scale input in the encoder path. At the same time, dense jump connection is used to further fuse and extract multi-scale features. In order to more accurately extract and integrate the tipper response characteristics extracted at each stage, TipInv-net adopts the Atrous Spatial Pyramid Pooling (ASPP) to enhance the detailed description of the abnormal body and fault model, thus improving the accuracy of inversion. The superior performance of TipInv-net is demonstrated through both the theoretically simulated data set and the geologically acquired data set in West Area of China. The experimental results reveal that the TipInv-net could not only get accurate inversion results with high efficiency, but also has outperforming robustness. •A novel DL-based framework called TipInv-net was designed for 2D ZTEM inversion, and a diverse simple dataset was constructed for TipInv-net training.•The research explored the effectiveness of DL intelligent modules in the inversion process, and provided some ideas for designing neural network models.•The method proposed in the research was applied to the measured tipper data to serve the selection and construction of railway tunnels, and the inversion results have good correspondence with geological data.