Machine learning framework for predicting mechanical properties of fiber-reinforced recycled aggregate concrete with sustainability optimization

The construction industry's significant environmental footprint necessitates the development of sustainable concrete. Incorporating recycled concrete aggregates (RCA) and natural fibers presents a promising solution, but complex non-linear interactions between mix design parameters make predicting mechanical properties challenging. This study develops an integrated machine learning framework with four principal innovations: (1) a comprehensive database of 519 mix designs compiled from published literature; (2) After data cleaning using a four-level stratified interpolation and interquartile range outlier detection , feature engineering created eight composite features. (3) systematic comparison of five machine learning models including random forest (RF), gradient boosting decision tree (GBDT), deep neural network (DNN), multi-output neural network (MONN) optimised via Bayesian optimisation, and a stacking ensemble; (4) model interpretability using SHAP analysis and integration of sustainability indicators into a decision-support tool. The GBDT model achieved the highest accuracy: R² = 0.9687 (compressive) and 0.9562 (tensile), with RMSE of 3.317 MPa and 0.251 MPa respectively. Ten-fold cross-validation confirmed robustness (average R² = 0.9607 for compressive, 0.9477 for tensile). SHAP analysis confirmed water-to-binder ratio as the most critical factor and revealed that fiber type significantly influences tensile properties. Fibers were shown to mitigate RCA-induced strength loss. The decision support system identified a 100% RCA formulation reducing embodied carbon by 42.7% while maintaining 28.5 MPa compressive strength. This work provides an accurate, interpretable machine learning framework for sustainable fiber-reinforced RCA concrete design, bridging research and practice.