Application of machine learning to identify key physicochemical predictors for mimicking UHT cow milk

Plant-based milk alternatives (PBMAs) differ from cow milk in their physicochemical properties, creating a significant challenge in their ability to replicate dairy-like functionality. In this study, machine learning was applied to identify the key physicochemical predictors associated with dairy-like properties in PBMAs. Two models, Random Forest Regression (RFR) and Gradient Boosting Regression (GBR), were trained using physicochemical, chemical composition and ingredients data from cow (10), soy (10), oat (10), almond (10), coconut (10), rice (3), and plant-hybrid (1) milk types. Model performance was evaluated using the coefficient of determination (R²), root mean square error (RMSE), learning curves, and compared using the Wilcoxon signed-rank test. Both models demonstrated strong predictive performance on testing data, RFR (R² = 0.80 ± 0.05; RMSE = 1.29 ± 0.19), and GBR (R² = 0.76 ± 0.05; RMSE = 1.46 ± 0.27), with no statistical difference in predictive accuracy. Permutation feature importance identified fat, carbohydrate, and xanthan gum content as the key predictors of physicochemical similarity to cow milk, reflecting their roles in determining volatile compounds, viscosity, particle size, whiteness, and physical stability. Volatile compounds, including δ-octalactone, also contributed to predictive accuracy. The performance of the models was limited by the small dataset size, the absence of sensory data, and commercial formulations. This study demonstrated the value of machine learning in predicting key physicochemical predictors in PBMAs mimicry of dairy. Future work should focus on formulation prediction, increasing dataset size, and the integration of sensory data to improve model robustness.