Machine Learning Framework for Predicting Critical Mix-Design and Strength Properties of Eco-Efficient Recycled Aggregate Concrete

Abstract Efforts to reduce concrete’s embodied carbon and the environmental impacts of construction-and-demolition (C&D) waste have increasingly shifted toward using recycled-aggregate concrete (RAC) to replace natural aggregate, either partially or entirely. Over the past two decades, dozens of studies have been conducted on the properties of RAC mixes. These studies provide a solid foundation for prediction models that eliminate the reliance on actual tests, which can be time-consuming. Machine learning (ML) models predict concrete properties from interacting variables, such as water-to-cement ratio, cement content, RAC replacement level, aggregate density, etc. These data-driven predictions are more reliable than traditional empirical formulas. ML model can capture nonlinear relationships among diverse input features, providing a framework for optimizing mix design and evaluating mechanical performance. In this study, a database of 358 mix designs from the literature was used to predict three key mechanical properties (compressive strength, split-tensile strength, and modulus of elasticity) of RAC-based concrete. The decision tree was used as the principal predictive model, and its robustness was verified through K-fold cross-validation. Model performance was high in training (R² = 0.93, 0.92, and 0.94, respectively) and remained acceptable in testing (R² = 0.75, 0.78, and 0.76). SHAP (SHapley Additive exPlanations) Analysis shows cement content and aggregate density are the most influential features across all three properties, followed by water-to-cement ratio and RAC replacement level. This interpretable ML framework streamlines mix-design optimization, reduces laboratory work, and guides production of low-carbon RAC with dependable mechanical performance.