Explainable stacked machine learning-enabled approach for mechanical properties prediction of eco-friendly RCC

Roller-compacted concrete (RCC) has gained increasing attention in sustainable pavement and infrastructure applications due to its high load-bearing capacity, rapid construction and cost efficiency. However, the complex and nonlinear interactions among material constituents make the reliable prediction of its mechanical properties challenging. This paper proposes an integrated, interpretable and application-oriented artificial intelligence (AI) framework for predicting the compressive strength (CS) and flexural strength (FS) of RCC. An extensive literature-based database was compiled, initially comprising 938 CS and 464 FS data points, which was subsequently refined through Isolation Forest–based outlier detection to yield a final dataset of 927 CS and 445 FS data points. Five advanced machine learning (ML) models—random forest, extra trees regressor, gradient boosting machine, extreme gradient boosting and categorical boosting—were developed as base learners, followed by the construction of 20 stacked ML models. Hyperparameter optimization was performed using grid search combined with five-fold cross-validation. The results demonstrate that stacked ML models significantly outperform individual base learners. The optimal stacked ML configurations achieved test performances of R² = 0.90 and RMSE = 4.2 MPa for CS, and R² = 0.86 and RMSE = 0.63 MPa for FS, indicating strong predictive accuracy and generalization capability. To enhance transparency and engineering relevance, SHapley Additive exPlanations (SHAP) and individual conditional expectation (ICE) analyses were employed to identify dominant parameters and nonlinear trends. The analyses revealed that CS is primarily governed by the water-to-cement ratio, curing age and aggregate composition, whereas FS is more sensitive to coarse aggregate content, curing age and steel fiber dosage. Based on ICE results, practical optimum design ranges were proposed to support a performance-oriented and sustainable RCC mixture design. Lastly, a GUI-guided predictive software integrating the best-performing base and stacked models was developed to facilitate practical implementation, enabling rapid decision-making and reduced experimental effort.