Zone-specific prediction of specific charge in tunnel blasting with machine learning

Optimizing the specific charge in tunnel blasting is critical for excavation efficiency and stability. However, traditional approaches apply uniform charging across tunnel rounds, ignoring the distinct requirements of different blast zones. While previous studies have acknowledged these zone-specific needs, a comprehensive machine learning (ML) framework for predicting charge in each zone has been lacking. This study addresses that gap by developing a zone-specific framework to predict the optimal specific charge for the cut, stoping, lift, and contour zones. Using data from 208 tunnel blast rounds from 18 Korean sites representing diverse geological conditions, we evaluated various machine learning algorithms including linear models, support vector machines, neural networks, and tree-based ensembles. Our evaluation found that Random Forest was the optimal model for the cut zone (R² = 0.937), while XGBoost performed best for stoping (R² = 0.797), lift (R² = 0.925), and contour zones (R² = 0.927). Furthermore, SHAP analysis revealed each zone is governed by distinct parameters, with cut type dominating cut zone predictions, while spacing, round length, and rock type are most significant for stoping, lift, and contour zones, respectively. The proposed framework provides engineers with a practical, data-driven tool to improve fragmentation, reduce overbreak, and enhance both safety and cost-efficiency in tunnel excavation.