Deep neural network-based interpretation of frequency-domain electromagnetic data: Dual output of 2D resistivity models and electromagnetic data

Abstract

The assessment of manmade structures such as dams and embankments is vital for protecting property and human life. Consequently, regular and accurate geophysical monitoring of these infrastructures is widely practiced. However, the iterative numerical procedures typically used for geophysical data inversion can become computationally intensive, especially in resource-constrained field surveys. Recent advances in machine learning, particularly deep neural networks (DNNs), offer an alternative solution by significantly reducing the time required to generate subsurface images once the network has been trained. In this study, we develop a DNN-based interpretation method for frequency-domain electromagnetic (FDEM) data. Beyond producing inverted images of the subsurface, our approach also estimates uncertainty by comparing the predicted electromagnetic (EM) response with the original input data. We highlight a key pitfall in traditional network architectures: if a data predictor is trained jointly and from scratch alongside the model predictor, it may replicate the input data - giving the misleading impression that every inversion is highly accurate. To address this, we develop a transfer learning-based training procedure where a pre-trained data predictor's weights are fixed. This ensures that the predicted EM response genuinely reflects the inverted subsurface model rather than reproducing the input data. We validate our method using FDEM data collected at the Saemangeum seawall in South Korea, comparing the DNN inversion results with those from conventional inversion. Our technique successfully detects potential hazards in the seawall while also providing a practical measure of reliability for the interpreted results.