Sensory and Physicochemical Evaluation of Dairy-Like Plant-Based Milk Formulations Predicted by Machine Learning

Plant-based milk alternatives (PBMAs) are now increasingly used by some consumers in similar ways to cow milk. However, they differ in their physicochemical and sensory properties. Improving PBMAs to better mimic dairy is limited by the time and cost of development trials. Machine learning offers a promising approach to predict optimal formulations without the need for extensive trial and error. In this study, Random Forest regression (RFR) and gradient boosting regression (GBR) models were used to formulate PBMAs (Formulation 1 and 2, respectively) that mimic cow milk. Sensory analysis, viscosity, stability, color, particle size, and headspace volatiles were compared to cow milk to assess model accuracy. Both formulations were sensorially distinct from coconut, oat, soy, almond, pasteurized, and ultra-high temperature (UHT) cow milk, associated with vanilla, caramel, astringent, bitter, metallic, and nutty notes, compared to the sweet, dairy, and coconut notes of dairy. The formulations contained aldehydes, including hexanal and pentanal from soy, although compounds including vanillin, octanal, and 1-pentanol were shared with cow milk. UHT cow (0.68 µm) had the smallest average particle size (D₅₀), followed by Formulation 2 and pasteurized cow (1.04 µm), with Formulation 1 (1.62 µm) as the largest. The UHT and pasteurized cow milk had higher whiteness indices compared with Formulation 1 and Formulation 2. Formulations exhibited faster sedimentation compared to cow milk. The models correctly predicted viscosity and some volatiles. However, predictive performance was limited by the small dataset and restricted access to commercial formulations. A larger dataset and access to formulations would improve model accuracy. PRACTICAL APPLICATIONS: This research demonstrates the potential of predictive modelling for the development of plant-based alternatives to animal-derived ingredients. Such an approach would reduce reliance on trial and error in product development, saving time and development costs while meeting consumer needs.