Surf3DNet1.0: A deep learning model for regional-scale 3D subsurface structure mapping

Abstract. This study introduces an efficient deep learning approach based on convolutional neural networks with joint autoencoder and adversarial structures for 3D subsurface mapping from surface observations. The method was applied to delineate palaeovalleys in an Australian desert landscape. The neural network was trained on a 6,400 km2 domain by using a land surface tomography as 2D input and an airborne electromagnetic (AEM)-derived probability map of palaeovalley presence as 3D output. The trained neural network has a maximum square error < 0.10 and produces a square error < 0.10 across 93 % of the validation areas, demonstrating that it is reliable in reconstructing 3D palaeovalley patterns beyond the training area. Due to its generic structure, the neural network structure designed in this study and the training algorithm have broad application potential to construct 3D geological features (ore bodies, aquifer) from 2D land surface observations.