158 mineralized cells creates strong class imbalance, which may inflate AUC and recall while increasing false positives, potentially raising exploration validation costs. The heuristic Empirical Weighted Mean (Wmd) exhibited negligible explanatory power (R² = 0.0045) and extremely low precision (0.0017), highlighting the limitations of linear, expert-driven combinations in modeling complex geological interactions. Similarly, while KNN and LGB achieved moderate AUC values—suggesting a latent ability to rank prospective zones—they failed to reach meaningful binary classification performance (F1-score = 0.0000). This discrepancy suggests that while nonlinear machine learning models are better equipped to capture the intricate interactions among hydrothermal signatures and structural controls, they require careful threshold optimization to overcome the inherent class imbalance in mineral prospectivity mapping. Consequently, the most reliable prospectivity maps for operational deployment were generated by SVR (conservative targeting) and XGB (exhaustive screening). These maps not only successfully delineated known mineralized zones but also identified new, previously unrecognized high-potential areas, validating the utility of this machine learning-based approach for mineral exploration. Table 1 – Metrics and Hyperparameters associated with the predictive methods Parameter Wmd RF GAM XGB LGB KNN SVR R² 0.0045 0.1001 0.0224 0.3750 0.0823 0.0763 0.7764 MAE 0.0373 0.0007 0.0014 0.0006 0.0010 0.0006 0.1498 RMSE 0.0569 0.0138 0.0144 0.0118 0.0139 0.0139 0.1499 AUC 0.9338 0.9238 0.9823 0.9989 0.8560 0.8309 0.9730 Precision 0.0017 0 0 0.0036 0 0 0.9800 Recall 0.8750 0 0 1 0 0 0.5000 F1-score 0.0034 0 0 0.0700 0 0 0.6700 Threshold 0.0930 0 0.0031 0.0040 0.0068 0.0101 0.7200 Wmd: Weighted Mean, RF: Random Forest, GAM: Generalized Additive Model, XGB: Extreme Gradient Boosting, LGB: Light Gradient Boosting Machine, KNN: K-Nearest Neighbors, SVR: Support Vector Regression. 4. DISCUSSION The results demonstrate the superiority of non-linear machine learning approaches over heuristic weighted combinations for mineral prospectivity mapping. While the empirical weighted mean failed to capture meaningful relationships (R² = 0.0045), SVR and XGB successfully modeled complex interactions among geological, geochemical, structural, and remote sensing variables. Similar improvements of ML over heuristic or linear methods have been reported in porphyry and gold prospectivity studies (Abedini et al., 2023; Mitra et al., 2025). The distinct performance profiles of SVR and XGB reveal complementary exploration strategies. SVR achieved the highest precision (0.98) and F1-score (0.6667), making it suitable for advanced exploration stages where minimizing false positives is critical. Its higher optimal threshold (0.72) reflects a conservative classification strategy that prioritizes confidence in predicted targets. In limited training scenarios, SVMbased approaches have shown sensitivity to overfitting but can perform robustly when carefully tuned (Lachaud et al., 2023). This behavior is particularly advantageous in capital-intensive drilling campaigns
RkJQdWJsaXNoZXIy MTM0Mzk2