Stacking Ensemble Learning for Copper Grade Prediction at the Sungun Porphyry Copper Deposit: Statistical Performance Versus Geological Coherence

Mineral resource estimation has traditionally relied on geostatistical techniques such as Ordinary Kriging (OK), which explicitly account for spatial continuity but may introduce smoothing effects and require extensive manual parameterization. In parallel, Machine Learning (ML) methods have emerged as flexible alternatives capable of modeling complex nonlinear relationships; however, their ability to generate geologically reliable block models remains insufficiently evaluated. This study assesses stacking ensemble learning for copper grade prediction at the Sungun porphyry deposit (Iran), comparing its performance against individual ML models, including Artificial Neural Networks (ANNs), Random Forest (RF), and Support Vector Regression (SVR), as well as domain-constrained and global unconstrained OK. Two ML scenarios were evaluated, using spatial coordinates either alone or in combination with mineralogical domains. Model performance was assessed using random and spatially buffered cross-validation together with spatial and geological validation criteria. Under random cross-validation, RF and Stacking 1 (ANN–RF–SVR) achieved the highest statistical performance among the ML models; however, all ML approaches showed substantial performance deterioration under spatially buffered cross-validation, indicating that the stronger performance observed under random cross-validation was strongly influenced by spatial autocorrelation. The incorporation of mineralogical domains had a limited and model-dependent effect under random cross-validation, but improved the spatially buffered cross-validation R2 values of all ML models, although these remained negative, while also enhancing geological compartmentalization and boundary preservation. Stacking 2 (ANN–SVR) produced the most geologically coherent ensemble model, although it overestimated tonnage relative to the domain-constrained OK benchmark. Overall, within the evaluated framework, the results indicate that predictive accuracy alone is insufficient for evaluating ML-based grade-prediction models and that spatial and geological validation criteria provide essential complementary information for assessing the reliability of the resulting block models.