A Novel Self‐Supervised Deep Learning Inversion Method Incorporating a Fast Forward Network for Transient Electromagnetic Data

Abstract Deep learning (DL) is an increasingly popular method for inverting the transient electromagnetic (TEM) data. Currently, DL‐based inversion for TEM data primarily employs supervised learning methods that rely on labeled data sets to train the network. The generalization of networks depends on the quantity of qualified and realistic data sets which require thousands of huge amount of forward simulation. We propose a new self‐supervised learning method by incorporating a fast forward network into the DL inversion network to reduce the reliance on labeled data sets. After pre‐training the forward and inversion network with only hundreds of labeled data sets, the inversion network starts considering the observed response and optimizes its parameters based on the loss between the observed response and the response from the forward network. To improve the accuracy of the forward network, we perform a secondary training on it during the inversion process. Additionally, we develop a fully convolutional inversion network based on autoencoder structure to address discontinuities between resistivity layers. The synthetic data test demonstrates that the proposed method can recover anomalous bodies more quickly and achieve smaller fitting residuals with 400 labeled data sets, outperforming the supervised learning method and traditional IRLS inversion method. Finally, in analyzing two field data sets, the method obtains more accurate underground structures, showing good application prospects in TEM inversion.