Structurally constrained graph convolutional networks for 3D geochemical anomaly identification

The formation of surface geochemical anomalies in hydrothermal gold deposits is controlled by deep structure-alteration-mineralization and hydrothermal activities. Structural constraints on fluid migration cause lateral fluid deflection, leading to spatial decoupling between surface geochemical signals and deep orebodies. Conventional geospatial AI anomaly identification methods merely adopt pure data-based spatial correlation, failing to describe anisotropic fluid attenuation. This work develops a structurally constrained graph convolutional network for 3D geochemical anomaly extraction. The heterogeneous graph architecture integrates anisotropic spatial decay formulations and structural gated feature fusion. Taking the Xiadian gold deposit as a case, we build a cross-dimensional graph linking 2D surface geochemical samples and 3D undersurface structural voxels, with edges weighted by self-learned anisotropic decay tensors and spatial deflection vectors. Quantitative tests confirm the model accurately reproduces deep geochemical chimney effects: the structural permeability anisotropy ratio reaches 4.37, and lateral fluid deflection extends approximately 446 m, enabling objective projection of surface anomalies to deep structures. Supported by deep anomaly signals, the prospectivity model achieves an AUC of 0.9292 and accuracy of 0.8825. This quantitative interpretable method solves the anomaly-orebody decoupling issue and improves deep blind ore prediction accuracy, providing an interpretable intelligent tool for deep mineral exploration.