Effective Inversion of Yellowstone Airborne TEM Data Using Deep Learning

Abstract Deep learning methodologies can significantly accelerate the interpretation of airborne transient electromagnetic (ATEM) data. Nevertheless, it remains challenging for deep learning methods to deal with data vectors with missing values. This study introduces innovative processing techniques for transient electromagnetic data, enabling the trained neural network to effectively manage data vectors with missing values. Furthermore, it presents a comprehensive analysis within the Yellowstone National Park study area, comparing the performance of networks trained on real field data sets and synthetic data sets in ATEM data inversion. The results strongly support the superiority of networks trained on field data sets over those trained on synthetic ones. In addition, the research highlights two key factors differentiating these data sets—noise levels and the distribution of resistivity models. It examines the variations in the distribution of resistivity models across data set types and their consequential effects on inversion results. This study underscores the critical importance of utilizing real field data on network training, demonstrating its remarkable effectiveness in deciphering intricate geological structures and achieving detailed imaging of the subsurface conductivity.