A novel normalization method of transient electromagnetic data for efficient neural network training

Groundwater and mineral resources play a crucial role in human society and have distinct geo-electrical properties in the subsurface. Transient Electromagnetic (TEM) method is effective in determining these properties, but its unique features, such as its dynamic range spanning orders of magnitude and the possibility of negative values, can challenge the optimization of neural networks and result in longer training times and lower accuracy. To address this challenge, we propose a novel normalization method to improve the information of TEM data for the training of neural networks. We apply our proposed method to an airborne TEM data forward modeling problem and show that it achieves high accuracy as compared to the commonly used normalization techniques, and improved computational efficiency as compared to numerical methods. The trained network can accelerate both deterministic and stochastic inversion schemes, enabling efficient modeling of large datasets.