196 2.4. Validation Strategy To ensure the results are operationally relevant, we avoided simple random k-fold validation. Instead, we employed a True Spacing Analysis. The dataset was spatially decimated to simulate different drilling grids (e.g., 100m, 50m, 25m), and the model was tasked with predicting the removed "infill" holes. This rigorously tests the model's ability to predict unknown volumes rather than just interpolating between closely spaced points. 3. RESULTS 3.1. Global Estimation Accuracy The model was evaluated using a blind cross-validation split (80% training, 20% testing). The results demonstrate a substantial improvement in global accuracy metrics. Method RMSE (g/t) Improvement Ordinary Kriging (Baseline) 44.45 - GeaAI (Hybrid Model) 23.56 47.00% Table 1: Global Performance Metrics The hybrid model reduced the Root Mean Squared Error (RMSE) by nearly half. This indicates that approximately 47% of the error variance in Kriging is not random noise (nugget), but structured non-linear variability that the geometric features successfully captured. 3.2. Feature Importance and Explainability A critical requirement for mineral reporting is explainability. We analyzed the Gain metric from the XGBoost model to understand which features drove the decisions. Rank Feature Relative Importance (%) Geological Interpretation 1 Local Curvature 29.69% Detection of structural traps/shoots. 2 3D Wavelet Energy 23.19% Identification of textural complexity. 3 Local Variance 10.12% Measure of local heterogeneity. 4 Total Wavelet Energy 9.10% Signal intensity. 5 Local CV 3.70% Variability relative to mean. Table 2: Top 5 Feature Importance Notably, absolute coordinates (X, Y, Z) did not appear in the top 5. This confirms that the model is learning geological rules (e.g., "high curvature + high energy = high grade") rather than memorizing spatial locations. This "White Box" behavior is essential for building trust with resource geologists.
RkJQdWJsaXNoZXIy MTM0Mzk2