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

Product carbon footprint estimation can support early screening in the food and beverage sector, but life cycle assessment requires detailed inventory data that are often unavailable. Machine learning can estimate reference-food impacts from food descriptors; however, whether grouping training outcomes improves prediction remains uncertain when unseen foods must be routed using predictors alone. To examine this gap, we present and evaluate a Gaussian mixture model (GMM)-based framework combining outcome-specialized regressors with neighbor-based routing. Multilingual LEAF names are linked to AGRIBALYSE 3.1.1 climate-change values and enriched with non-environmental descriptors and CIQUAL nutrients, yielding 471,970 observations across 1562 reference codes. Within each training fold, a univariate GMM defines hard or overlapping alpha-cut subsets for K-nearest-neighbor, random-forest, and extreme gradient boosting (XGBoost) regression. Five predictor-space neighbors determine majority or weighted routing. Evaluation uses ten repetitions of five-fold cross-validation grouped by reference code, with training-only preprocessing and fitting. Weighted routing lowers mean root mean squared error (RMSE) relative to majority routing across matched configurations. Weighted fuzzy random forest at alpha = 0.1 reduces mean RMSE from 3.465 to 3.136 kg CO2e/kg, while global XGBoost achieves the lowest RMSE of 2.773. No primary comparison is significant under corrected repeated cross-validation tests with Holm adjustment. Target-value-grouped sensitivity analysis retains the principal learner-dependent pattern. These findings identify limits to outcome-based specialization rather than general superiority, supporting further investigation of reference-impact screening without replacing product-specific life cycle assessment.