Machine Learning Prediction of 28-Day Compressive Strength in Recycled Aggregate Concretes with Supplementary Cementitious Materials: Experimental Validation Using Metakaolin and Spent Fluid Catalytic Cracking Catalyst Residue

This study develops an interpretable machine learning framework for predicting the 28-day compressive strength of recycled aggregate concrete containing supplementary cementitious materials and validates it experimentally using metakaolin and spent fluid catalytic cracking catalyst residue. A broad multi-source database was used to train K-nearest neighbors, support vector regression, random forest, extremely randomized trees, and extreme gradient boosting models, together with a stacked ensemble. The stacked model achieved the highest internal cross-validation performance, with an R2 of 0.93. SHAP and partial dependence analyses identified cement content and water-to-binder ratio as the dominant predictors, whereas recycled coarse aggregate exerted a smaller effect conditioned by matrix quality. Independent experimental validation was conducted using two matched central composite designs produced with MK and FC3R under equivalent mixture-design conditions. MK consistently developed higher 28-day compressive strength than FC3R, with an average advantage of 6.06 MPa, while the predictive model reproduced the main material-specific trends for both systems. These findings confirm that 28-day strength is controlled by coupled binder–water–aggregate conditions and that interpretable ensemble learning can support preliminary mixture screening across distinct aluminosilicate systems within the investigated design domain.