Predicting the Carbon Footprint of Building Material Using Machine Learning: Production Phase, Secondary Material Use, and Recycling Potential

Environmental product declarations (EPDs) provide carbon footprint data for construction products. This study evaluated machine learning predictions of global warming potential (GWP–total) for the production stage (A1–A3) and module D from EPD resource-use and waste indicators, focusing on secondary material use (SM). After interquartile range removal of target outliers, the analysis used 2214 A1–A3 records and 1765 EPDs with module D data from ÖKOBAUDAT. Extreme Gradient Boosting (XGBoost), Random Forest, and Support Vector Regression (SVR) were tuned by 5-fold cross-validation; predictor importance was assessed with Shapley Additive Explanations values and ablation. XGBoost achieved the highest test coefficient of determination (R2) 0.963 for A1–A3, 0.791 for module D, with relative mean absolute errors of 13% and 27%, respectively. Its advantage over Random Forest was mostly within bootstrap uncertainty. Total use of non-renewable primary energy resources (PENRT) dominated A1–A3 predictions, whereas radioactive waste disposed (RWD), total use of renewable primary energy resources (PERT), and SM gained importance for module D. Removing SM changed module D R2 by −0.03 (95% confidence interval: −0.07 to 0.00). Grouped validation reduced module D R2 to approximately 0.4 for unseen groups. The models may support plausibility checks and estimation of missing module D values.