Abstract
Building a reliable geological model is essential for optimizing construction costs and mitigating risks from unforeseen ground conditions. Existing methods fail to couple soil types (geological structure) with their properties and lack the integration of multi-source data. This paper presents a novel deep-learning method using the K-Dimensional Tree-Graph Convolutional Neural Process (KDTree-GCNP) for structure–property coupled geological modeling. The KDTree is firstly proposed to efficiently generate graph nodes and edges in the Graph Convolutional Network (GCN) using the neighboring nodes aggregation procedure in the three-dimensional (3D) domain. Subsequently, the proposed GCNP aggregates the soil types and the properties for each graph node based on its adjacent nodes, followed by the message updating within the Neural Process (NP) so as to admit uncertainty quantification in geotechnical property predictions. Multi-source data including borehole logs, laboratory tests, in-situ tests, and geological profiles, are fused to the geological model. The proposed KDTree-GCNP method is verified using a benchmark study and applied to a tunnel project in Nanjing City. The results demonstrate that the proposed method is powerful in 3D coupled geological modeling, achieving high accuracy with coefficient of determination (R2) values of 0.82–0.97 for geotechnical property predictions and 97% accuracy for soil type classification. Finally, the current challenges and future opportunities are discussed in depth, including methodological insights on graph convolution in the spectral domain, physics-informed constraints, and uncertainty quantification challenges.