Food Product Design: A Hybrid Machine Learning and Mechanistic Modeling Approach

At present, food products are designed by trial and error and the sensorial ratings are determined by a tasting panel. To expedite the development of new food products, a hybrid machine learning and mechanistic modeling approach is proposed. Sensorial ratings are predicted using a machine learning model trained with historical data for the food under consideration. The approach starts by identifying a set of food ingredient candidates and the key operating conditions in food processing based on heuristics, databases, etc. Food characteristics such as color, crispness, and flavors are related to these ingredients and processing conditions (which are design variables) using mechanistic models. The desired food characteristics are optimized by varying the design variables to obtain the highest sensorial ratings. To solve this gray-box optimization problem, a genetic algorithm is utilized where the design constraints (representing the desired food characteristics) are handled as penalty functions. A chocolate chip cookie example is provided to illustrate the applicability of the hybrid modeling framework and solution strategy.