Prediction of compressive strength and mix proportion design of coal gangue mixtures using ensemble learning and SHapley Additive exPlanations

The compressive strength of coal gangue (CG) mixtures is notoriously difficult to predict due to complex, nonlinear interactions among multiple composition factors, hindering its efficient utilization in construction materials. While machine learning (ML) has been applied to concrete property prediction, its adoption for CG systems—particularly regarding automated model development, interpretability, and inverse mix design—remains under-explored. To address this gap, this study proposes a comprehensive ensemble ML framework. Based on 142 experimental datasets, 14 ML models were developed and compared. The four top-performing ensemble models (CatBoost, GBDT, XGBoost, and ET) were further optimized via a random search methodology. SHapley Additive exPlanations (SHAP) analysis was employed to interpret the models, revealing the underlying relationships between mixture proportions and compressive strength. Subsequently, an inverse analysis approach was established for the intelligent design of optimal mix proportions. Among the optimized models, CatBoost demonstrated the best performance, achieving R² > 0.97 and MAE < 0.5 MPa on the test set. The practical applicability of the inverse model was confirmed through field tests on a highway section; the deviations between predicted and measured 7-day compressive strengths were within 10%. This study provides a reliable, interpretable, and data-driven solution for the performance prediction and mix design of CG mixtures, contributing to the sustainable reuse of industrial solid waste. • The CatBoost and GBDT ensemble models achieved highly accurate prediction of coal gangue mixture strength (R² > 0.97, MAE < 0.3 MPa). • SHAP analysis identified curing age and coal gangue content as the two most decisive factors affecting compressive strength. • An inverse design method enabled intelligent mix proportioning, validated by field tests with prediction errors within 10%. • The proposed data-driven framework provides a reliable and efficient tool for the sustainable utilization of industrial solid waste in construction.