Ensemble learning for CO₂ footprint prediction of waste glass powder-based UHPC

Accurate forecasting of the carbon dioxide footprint (CO₂-FP) associated with ultra-high-performance concrete (UHPC) that incorporates waste glass powder (WGP) is crucial for the advancement of sustainable construction materials. This investigation proposes an ensemble learning framework that amalgamates gradient boosting, random forest, and extreme learning machine models to accurately estimate the CO₂-FP of WGP-enriched UHPC mixtures, thereby enhancing both precision and stability. The ensemble model employs data-driven optimization techniques to elucidate complex nonlinear interrelationships between the constituents of the mix and various environmental impact indicators. The interpretability of the model is augmented through the application of SHAP and partial dependence analyses, which elucidate the roles of cement, WGP, and superplasticizer contents as the predominant factors influencing CO₂ emissions. The framework exhibits superior generalization capabilities in comparison to standalone models, thereby underscoring its robustness for both predictive and diagnostic purposes. In addition to its predictive prowess, the research introduces a systematic methodology that correlates material composition, binder optimization, and sustainability outcomes. The results yield actionable insights into mitigating embodied carbon in UHPC through the valorization of industrial glass waste, thereby contributing to practices aligned with the circular economy and the advancement of environmentally sustainable concrete design strategies.