Accurate prediction of compressive strength is vital for designing sustainable concretes incorporating construction and demolition waste (CDW). However, conventional approaches often struggle due to the inherent variability of CDW. This study introduces an innovative stacked ensemble machine learning framework that integrates multiple heterogeneous regressors—gradient boosting, XGBoost, random forest, and decision tree—through a meta-learning approach to achieve ultra-precise strength prediction of CDW-based concretes. Using a dataset of 149 mix designs, the stacked ensemble achieved on the held-out test set \(R^2 = 0.9878\) , RMSE \(= 1.082\) MPa, MAE \(= 0.822\) MPa; five-fold CV showed similar dispersion, reducing error by nearly 50% compared with the best single model (Wilcoxon \(p < 10^{-7}\) ). Uncertainty was quantified via 95% confidence intervals (paired bootstrapping on the test set and fold-wise dispersion in 5-fold CV) to support statistical comparisons. Model interpretability via SHAP identified curing time and cement content as dominant factors, while higher water–cement ratios and CDW proportions were associated with strength reduction. A user-friendly Tkinter-based graphical interface was developed to enable practitioners to perform mix design predictions and “what-if” analyses. While the model demonstrates excellent accuracy, its performance is limited by the dataset size and exclusion of durability-related variables. Future research should focus on expanding the database, integrating durability indices, and exploring uncertainty quantification to enhance generalization. These findings highlight a pathway toward intelligent, data-driven mix design for sustainable construction materials.
