Explainable ensemble machine learning framework with particle swarm optimization for predicting compressive strength of CO₂-cured concrete

The increasing demand for low-carbon construction materials has led to growing interest in carbonation-cured concrete, which enables permanent sequestration of carbon dioxide while influencing material performance. However, predicting the compressive strength of CO₂-cured concrete remains challenging due to complex nonlinear interactions between mixture composition and curing conditions. Unlike previous studies that employed single algorithms with grid search optimization, this study integrates seven advanced ensemble models with metaheuristic optimization, multi-level interpretability analysis, and probabilistic uncertainty estimation within a unified framework . Seven ensemble learning models, including Random Forest, CatBoost, LightGBM, NGBoost, AdaBoost, Extra Trees, and DeepGBM, were optimized using Particle Swarm Optimization to improve predictive performance. The Random Forest model achieved the highest predictive accuracy on the testing dataset with R 2 = 0.955 and RMSE = 3.9251 MPa. All seven models demonstrated strong predictive performance, with R 2 values exceeding 0.92, confirming the effectiveness of PSO-based hyperparameter optimization across all evaluated algorithms . SHapley Additive exPlanations, accumulated local effects, permutation feature importance, and counterfactual analysis were applied to evaluate feature contributions and interactions. Cement content, coarse aggregate content, and water content were identified as the most influential variables, while carbonation curing parameters showed secondary effects. A graphical user interface was developed to enable real-time compressive strength prediction based on user-defined input parameters. The proposed framework provides accurate prediction and interpretable insights into the influence of mixture composition and curing conditions within the scope of the studied dataset