Toward Safe and Sustainable Mine Backfilling: Interpretable UCS Prediction of Cemented Backfill Using IGSO-Optimized Ensemble Learning

Accurate prediction of the 28-day unconfined compressive strength (UCS) of cemented backfill is essential for proportion design and quality control in mine backfill engineering. In this study, 119 groups of experimental data were used to develop a prediction framework, in which the ratio of cement to tailings and backfill concentration were adopted as the only two input features. Six ensemble learning models, including random forest (RF), gradient boosting regressor (GBR), extreme gradient boosting (XGBoost), categorical boosting (CatBoost), light gradient boosting machine (LightGBM), and adaptive boosting (AdaBoost), were established to predict 28-day UCS. To improve model performance, an improved glider snake optimizer (IGSO) with logistic chaotic initialization was proposed for hyperparameter optimization. Among the optimized models, IGSO-CatBoost achieved the best overall prediction accuracy, with coefficient of determination (R2) = 0.8838, root mean square error (RMSE) = 0.1956 MPa, mean absolute error (MAE) = 0.1614 MPa, and variance accounted for (VAF) = 88.8232%. Error-distribution analysis further indicated that AdaBoost had the highest proportion of small-error predictions, whereas CatBoost showed the best overall error control. Shapley additive explanatory (SHAP) and PAWN analyses consistently identified backfill concentration as the dominant input variable. Partial dependence plot (PDP) analysis showed that the effect of backfill concentration was nonlinear, with a marked variation in the middle concentration range rather than a simple monotonic trend. In general, the proposed IGSO-CatBoost framework provides an accurate and interpretable approach for rapid UCS prediction of cemented backfill.