Data-driven multi-objective optimization of recycled aggregate concrete mixes via LightGBM strength modelling and SHAP interpretability

This paper presents a data-driven framework for predicting the compressive strength of recycled aggregate concrete (RAC) under design constraints, an environmentally sustainable construction material. The proposed method uses a LightGBM regression model integrated with multi-objective optimisation via the NSGA-II algorithm, improving prediction accuracy, robustness, and interpretability. While LightGBM is compared against XGBoost, CatBoost, and a stacking-based ensemble for benchmarking purposes, all reported results, including optimisation, uncertainty analysis, and SHAP-based feature importance, correspond exclusively to the final LightGBM-based model, ensuring clarity and consistency. A tolerance-based accuracy criterion of $$\pm 10\%$$ is adopted to reflect practical engineering requirements. The model development and validation are carried out using 535 concrete mix samples obtained from experiments, representing a wide range of material proportions and curing ages. The proposed LightGBM–NSGA-II method shows very consistent and trustworthy results. It achieved an average cross-validation $$R^2$$ of 0.9859, an adjusted $$R^2$$ of 0.9604, a tolerance-based prediction accuracy of 98.08%, and a cross-validation accuracy of 94.29%. Prediction errors remain low, with an RMSE of approximately 12.2 MPa and an MAE of approximately 9.3 MPa. Reliability is confirmed by a Prediction Interval Coverage Probability of 97.87% and a Pareto dominance count of zero. SHAP and gain analysis identify the water-binder ratio, curing age, and superplasticiser content as the most influential factors. Statistical validation using paired t-tests and Wilcoxon signed-rank tests supports the method’s stability. Therefore, the method is suitable for the sustainable concrete mix design.