Three-dimensional voxel geological modelling for subsurface stratigraphy: A graph convolutional network approach

Abstract

Three-dimensional (3D) geological modelling enhances the understanding and visualisation of complex subsurface stratigraphy, which underpins geotechnical digital twin and resilience design. Existing methods for 3D geological modelling suffer from either high computational burden or low modelling accuracy in large-scale region modelling with complex subsurface stratigraphy. This paper presents a novel deep learning method that applies the graph convolutional network (GCN) to 3D voxel geological modelling using limited boreholes. A topological graph is firstly constructed, with spatial points encoded as graph nodes. The strata types and spatial coordinates are incorporated into the feature vector of each node. Spatial correlations are quantified through weighted edges by connecting pairs of nodes within a cuboid neighbouring system. Besides, the occurrence probability of strata in all boreholes is embedded into the feature vector of each graph node to further improve the model robustness. A series of comparisons shows that the proposed method outperforms traditional TPS and MPS methods in terms of modelling accuracy. The proposed method is finally applied to a real tunnel engineering in Changsha City, which demonstrates the effectiveness of the proposed method in complex 3D geological settings.