Explainable tree-based machine learning for sustainable mix optimization and compressive strength prediction of recycled aggregate fiber reinforced self-compacting concrete

The reliable prediction and sustainable optimization of the compressive strength (CS) of recycled aggregate fiber-reinforced self-compacting concrete (RA-FRSCC) represent significant challenges in modern concrete engineering, primarily because of the intricate and highly nonlinear relationships among mix-design parameters, curing conditions, fiber characteristics, and environmental performance indicators, including CO₂ emission constraints. An explainable tree-based machine learning (ML) framework was developed to predict CS and optimize mixture proportions while simultaneously considering mechanical performance and environmental impact. The framework integrates a diverse set of tree-based ML algorithms, including decision tree (DT), random forest (RF), extra trees (ET), gradient boosting (GB), extreme gradient boosting (XGB), and light gradient boosting machine (LGBM), evaluated under both default and systematically optimized hyperparameter configurations. A comprehensive experimental dataset consisting of cement (C), fiber length (FL), fiber content (fiber), natural coarse aggregate (NCA), recycled coarse aggregate (RCA), fine aggregate (FA), water (W), and curing age (days) was employed to predict CS. Model predictive performance was comprehensively evaluated using an 80:20 training–testing data split in conjunction with 10-fold cross-validation. The coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE) were employed as statistical performance metrics to comprehensively evaluate model robustness, generalization capability, and predictive reliability. Through systematic hyperparameter optimization, the XGB model exhibited superior predictive performance, achieving near-perfect accuracy on the training dataset (R² = 0.999, RMSE = 0.554) and excellent generalization on the testing dataset (R² = 0.978, RMSE = 1.664), thereby demonstrating a robust balance between model fidelity, predictive reliability, and generalization capability. To enhance model interpretability, a comprehensive explainable artificial intelligence (XAI) framework incorporating shapley additive explanations (SHAP), permutation-based feature importance, partial dependence plots (PDP), accumulated local effects (ALE), and individual conditional expectation (ICE) analyses was employed, consistently identifying RCA content and curing age as the dominant variables influencing the CS of RA-FRSCC. Furthermore, multi-objective optimization algorithms were integrated into the proposed framework to identify environmentally sustainable mix designs that maximize CS while minimizing associated CO₂ emissions. Unlike previous studies that primarily focus on predictive accuracy or limited interpretability, this study presents a novel integrated predictive–interpretable–optimization framework that combines advanced ensemble learning, multi-level XAI, and sustainability-driven mix design within a unified methodology. The proposed framework provides high predictive accuracy, enhanced model transparency, and practical decision-support capability for the development of sustainable and high-performance RA-FRSCC systems.