A Robust GDF-ML Framework for Dynamic Grade Modeling: Adaptive Resource Estimation in Complex Porphyry Systems

Porphyry copper deposits are strategic resources characterized by complex mineralization and high spatial non-stationarity. Traditional linear estimation methods like Ordinary Kriging often fail due to "smoothing effects" and a reliance on manual partitioning. This study proposes the GDF-ML framework, which integrates Geological Distance Fields (GDF) with machine learning to achieve high-fidelity grade modeling. By utilizing cKDTree for efficient spatial indexing, absolute coordinates are translated into Signed Distance Fields (SDF), creating "Spatial Fingerprints" that encode topological relationships between samples and geological entities. Results demonstrate that GDF-ML significantly outperforms Ordinary Kriging, increasing the R2 score from 0.3080 to 0.7696 on an independent test set while accurately reproducing the "barren core" and irregular high-grade zones. Furthermore, SHAP analysis validates the model's decision-making logic, aligning "intra-domain gain" effects with established metallogenic theories. This framework provides a scalable, automated methodology for dynamic resource evaluation, eliminating the inefficiencies of traditional workflows and supporting the digital evolution of modern smart mines.