Explainable hybrid stacking ensemble method for hard rock pillar stability prediction and engineering applications

Abstract Accurate prediction of hard rock pillar stability is vital for ensuring safety in underground mining operations. This study presents a novel hybrid stacking ensemble framework that integrates six machine learning models optimized by the Sparrow Search Algorithm (SSA) at both base and meta levels. A comprehensive dataset of 331 pillar cases was compiled from published case histories reported in 10 underground mines, characterized by five key geotechnical features. Seventy percent of the data were used for model training and the remaining 30% for testing. Performance was evaluated using five classification metrics, including accuracy, precision, recall, F 1‐score , and kappa. Statistical comparisons using the Friedman test and Nemenyi post‐hoc test confirmed the LightGBM‐meta stacking model as the top‐performing approach, achieving 93.0% accuracy, 0.937 F 1‐score , and a kappa of 0.858. SHapley Additive exPlanations (SHAP) identified pillar stress and uniaxial compressive strength ( UCS ) as the most influential features, enhancing model interpretability and engineering insight. A user‐friendly graphical user interface (GUI) was developed to enable practical deployment by mining engineers. The model was further validated on 11 field pillar cases from the Sanshandao Gold Mine, with parameters obtained via 3D laser scanning, in‐situ stress measurement, and laboratory UCS testing. All 11 field cases were correctly classified. Overall, the proposed SSA‐optimized stacking framework bridges data‐driven prediction and field deployment, offering practical support for pillar design and stability risk management.