Deep learning for instantaneous inversion of SkyTEM data

Abstract

The widespread application of Airborne Time-domain electromagnetic (ATEM) data, including the SkyTEM system (Sorensen and Auken 2004), to imaging shallow geothermal fluid pathways, is limited by the computational expense of inverting these large datasets. We present an approach using a physics-informed neural network to invert SkyTEM data in 1D and show early results of the application of the method to survey datasets. Neural network model predictions are several orders of magnitude faster than traditional inversion and produce very similar results. Future work will expand the neural networks to higher dimensionality, providing opportunities for large-scale 2D inversion, which is currently not computationally feasible yet is required for improved management of complex geothermal and hydrothermal systems.