Machine Learning-Based Prediction of Shaft Sinking Unit Costs from Geotechnical Parameters: A Case Study from the Birimian Greenstone Belt, Ghana

Shaft sinking is among the most capital-intensive phases of underground mine development, yet reliable cost estimation at the pre-feasibility stage remains difficult, empirical rules struggle to propagate site-specific geotechnical uncertainty into cost predictions, and no machine learning (ML) study has previously addressed this gap. This paper develops and compares three ML models like Random Forest (RF), XGBoost, and Artificial Neural Network (ANN) for predicting shaft sinking unit costs (USD/m) from geotechnical parameters derived from three vertical shafts being sunk to 1,500 m depth in the Birimian greenstone belt, Upper East Region, Ghana. A bottom-up component cost model, validated against West African industry benchmarks, generated training labels for 450 synthetic depth-zone records augmented from 18 base records by Monte Carlo sampling within ±15% of zone means. Five-fold cross-validation was applied throughout. XGBoost achieved the best performance (R² = 0.94, RMSE = USD 1,850/m, MAPE = 6.2%), followed by RF (R² = 0.91) and ANN (R² = 0.88). SHapley Additive exPlanation (SHAP) analysis identified depth midpoint, UCS, and hydrogeological risk as the three most influential features, collectively accounting for 72% of model variance, thus, finding that is physically interpretable through the depth-dependent cost escalation mechanism of drill-and-blast shaft sinking. The XGBoost model predicts unit costs within ±8% of the component-based estimate across all 18 validation points, meeting AACE Class 3 pre-feasibility accuracy of ±15%, and provides a rapid data-driven screening tool for shaft sinking cost estimation in geologically variable terrains.