Machine learning models and SHAP-based interpretability for predicting the strength of cemented rock fill in an underground mine in Peru

Abstract

The determination of uniaxial compressive strength (UCS) of cemented rock fill (CRF) in underground mining commonly depends on laboratory testing and curing periods, which delay mix-design decisions and quality control under operational constraints. This study evaluated and compared machine learning models combined with SHapley Additive exPlanations (SHAP) to predict CRF UCS at an underground mine in Peru. The experimental dataset included mix-design variables such as cement (C), water (W), waste rock (WR), screened aggregate (SCR) and derived parameters including cement content (C%), water-to-cement ratio (w/c), waste-rock-to-cement ratio (WR/c) and curing ages of 7, 14 and 28 days. Data preprocessing involved interquartile range capping and Yeo–Johnson transformation, followed by a 70/30 train-test split (n = 105/45). Five machine learning models NGBoost, Explainable Boosting Machine (EBM), CatBoost, LightGBM and Extra Trees were trained and optimised using 5-fold cross-validation. Among them, EBM achieved the best predictive performance on the test set (R² = 0.98, RMSE = 0.07, MAE = 0.045, MAPE = 2.59%, sMAPE = 2.61%, VAF = 98.31%). SHAP analysis identified curing age as the most influential predictor. Overall, the proposed approach provides an accurate and interpretable tool for optimising CRF mix design and quality control.