Intelligent cemented paste backfill-strength design framework: Interpretable machine learning and mix-proportion optimization

The rapid advancement of artificial intelligence has introduced new vitality into cemented paste backfill (CPB) technology. However, current machine learning models for CPB-strength prediction are generally forward-only and single-output, and lack clarity on multi-feature interactions and an integrated full-process design framework. To this end, this study proposes a light gradient boosting machine (LightGBM) model, optimized by Optuna, for predicting the CPB strength at multiple curing ages (3, 7, and 28 d). The model dataset comprised the unconfined compressive strength (UCS) results of 738 CPB specimens prepared with various types of tailings and mix proportions. SHapley Additive exPlanations (SHAP) were employed to elucidate the influence patterns and relative importance of input features while inherently considering their complex, multi-feature interaction effects on the output of the model. Additionally, a simulated annealing (SA) algorithm was integrated with the predictive model to enable the inverse process from the target UCS value to the optimal material-mix proportions. The results demonstrated the effectiveness of Optuna in hyperparameter tuning, leading to an optimized LightGBM model that accurately predicted the multi-age CPB UCS ( R 2 > 0.98). SHAP analysis identified key features, notably the high correlation of the water–cement ratio and CaO content in the tailings with the CPB strength. The SA algorithm effectively provided optimal CPB mix proportions that met the target 28-d UCS value, balancing multiple conflicting objectives such as the solid content and cement dosage. Finally, a user-friendly graphical user interface was developed to integrate these models and provide an accessible, visual, machine-learning-based CPB-strength design tool for mining engineers.