Abstract
Abstract
Accurate prediction of compressive strength (CS), flexural strength (FS), and splitting tensile strength (STS) of recycled aggregate concrete (RAC) remains challenging due to the complex nonlinear relationships among mixture components and the high resource demands of traditional laboratory testing, which conflict with sustainability goals. To address these challenges, this study proposes an explainable multi-target machine learning (ML) framework for the simultaneous prediction of RAC mechanical properties. The proposed approach integrates ensemble learning models, including Random Forest, AdaBoost, CatBoost, LightGBM, Gradient Boosting, and XGBoost, with advanced hyperparameter optimization strategies, Random Search, Successive Halving, and Bayesian Optimization, to enhance predictive performance. Missing data are handled using imputation techniques to ensure model reliability. Among the evaluated imputation-model combinations, MissForest-Random Forest (
R
2
= 0.86,
RMSE
= 1.53 MPa) achieved the highest predictive performance and outperformed traditional linear regression. Additionally, to improve model interpretability, explainable machine learning techniques, including SHapley Additive exPlanations (SHAP) and Individual Conditional Expectation (ICE), were employed to quantify feature importance and reveal nonlinear relationships between input variables and target outputs. The results indicate that the Effective Water-Cement Ratio and Nominal Maximum Recycled Concrete Aggregate (RCA) Size were the primary factors governing RAC mechanical behavior. The proposed framework provides both high predictive accuracy and interpretability, offering a practical and reliable tool for the design of sustainable concrete materials. This study contributes to advancing data-driven approaches in construction engineering by enabling efficient, interpretable, and resource-conscious prediction of concrete performance.