A Gaussian Mixture Model-Based Machine Learning Framework for Product Carbon Footprint Estimation in the Food and Beverage Sector

Source abstract

The global food supply chain generates approximately 13.7 billion metric tons of CO₂-equivalents annually, yet accurate Product Carbon Footprint (PCF) estimation remains inaccessible to most manufacturers. Life Cycle Assessment (LCA), the recognized gold standard, imposes prohibitive data demands and costs, while increasingly stringent European Union (EU) Product Environmental Footprint requirements intensify the urgency for scalable alternatives. Machine learning (ML) offers a promising substitute by utilizing publicly accessible product attributes, but the structural diversity of the food and beverage sector generates heterogeneous emission profiles that undermine single-model approaches trained on limited per-category data. This study proposes a novel ML framework combining multi-source data integration with Gaussian Mixture Model (GMM)-based clustering. Seven publicly available datasets were harmonized into a unified corpus of 3,933 samples, and GMMs were employed to cluster products by underlying emission patterns rather than predetermined taxonomies. Cluster-specific Random Forest Regression, K-Nearest Neighbors Regression, and eXtreme Gradient Boosting models were trained on each subset, with unseen samples assigned at inference via K-Nearest Neighbors voting. Results from ten repetitions of stratified five-fold cross-validation confirm that the proposed framework significantly outperforms standard single-model counterparts across all major product categories. The findings demonstrate that the proposed framework offers a scalable, data-driven alternative to LCA, enhancing PCF transparency for manufacturers and supporting informed sustainability decisions across the food and beverage sector.