Chemometrics and machine learning for efficient grading and sensory prediction of raw milk flavor

Dairy labeling has historically focused on health information. However, consumer surveys have identified flavor as a key driver of purchase decisions, prompting industry research into the flavor profiles of raw milk. The flavor quality of raw milk varies across different dairy farm brands. Consequently, there is a need for inexpensive and scalable flavor assessment methods to enable efficient dairy grading and sensory prediction. To address this need, the present study analyzed raw milk samples obtained from typical farms in China. The PLS-DA models developed from volatile flavor data and electronic nose data both achieved relatively high classification accuracy for Class I and Class III raw milk; however, their classification accuracy for Class II samples was comparatively low. Feature analysis performed on these models identified the specific compounds that affect flavor and acceptance. We innovatively integrated machine learning with chemical data to predict sensory attributes in advance. The present study iteratively optimized 19 models, with the coefficients of determination (R²) for the relevant models reaching up to 0.90. Gradually increasing the training set size from 30 to 80 moderately improved sensory prediction performance. Therefore, this study demonstrates that machine learning can effectively reveal the relationships among chemical components of food, flavor properties, and consumer perception, thereby providing a foundation for the development of innovative milk varieties with more distinctive and personalized flavor characteristics.