One‐dimensional deep neural network inversion of airborne transient electromagnetic data

Abstract The airborne transient electromagnetic method (ATEM), as a large‐scale and efficient geophysical exploration method, applies to exploration in areas with complex terrain and harsh environments. However, the ATEM method generates a large amount of data, which demonstrates a non‐linear relationship with the resistivity model. Additionally, the non‐uniqueness of the inversion problem further complicates the computational difficulty of the inversion problem. These problems raise the calculation cost of traditional inversion. Thus, our research group introduced an ATEM one‐dimensional inversion method based on a convolutional neural network (CNN) and gate recurrent unit (GRU). Specifically, the nonlinear relationship between the resistivity model and the electromagnetic response was obtained through the excellent nonlinear processing ability of the deep neural network. The test results suggest that this method had higher accuracy than the traditional shallow neural network inversion method. Compared with the traditional regularization inversion method, the deep learning method exhibited less dependence on the initial model and did not involve complex regularization parameter selection. It commonly needs to only train a large number of diverse data samples and can obtain an inversion result with high accuracy. After the deep neural network training was completed, ATEM field real‐time data were processed under the same hardware parameters. Additionally, the reliability of the CNN–GRU with field data was verified to confirm that the CNN–GRU deep neural network possessed good practicability. The test unveils that the inversion result of the deep neural network was close to the actual model, and the inversion accuracy was high. The one‐dimensional inversion results were combined to obtain the pseudo‐two‐dimensional inversion results, contributing to the improved visualization of inversion results.