Sustainable Mix Design of Recycled Aggregate Concrete: Machine Learning-Assisted Multi-Objective Optimization of Strength, Life-Cycle Cost, and Net Carbon Emissions

Abstract

Recycled aggregate concrete (RAC) mix design requires simultaneous consideration of mechanical performance, environmental impacts, and economic costs, yet these objectives are often evaluated separately. This study developed an integrated framework combining machine-learning-based strength prediction, life-cycle assessment, life-cycle cost analysis, constrained three-objective optimization, and preference-sensitive decision analysis. Using 407 RAC mixtures, Optuna-tuned Random Forest, XGBoost, and LightGBM models were compared, and SHAP was applied for interpretation. LightGBM achieved the best test performance, with an R2 of 0.8822, an RMSE of 4.0276 MPa, and an MAE of 2.8865 MPa. The water-to-cement ratio, sand ratio, and superplasticizer dosage were the three leading predictors, together accounting for 68.5% of the normalized SHAP importance. A 100-generation NSGA-II optimization produced 150 feasible Pareto solutions spanning 33.87–75.25 MPa in compressive strength, 456.80–616.38 CNY/m3 in life-cycle cost, and 248.84–395.00 kg CO2e/m3 in net carbon emissions. Higher-strength solutions generally required more cement and lower water-to-cement and recycled aggregate replacement ratios. Equal-weight TOPSIS selected P006, whereas the SMAA–TOPSIS simulation identified P007 as the alternative with the highest first-rank acceptability of 35.92%. By treating compressive strength as an explicit objective rather than a predefined constraint, the framework maps the continuous strength–cost–carbon trade-off within a volumetrically feasible mix-design space and identifies preference-dependent RAC design strategies.