Machine learning (ML) methods have been widely applied to analyze design of experiments (DOE) data in such industries as chemical, mechanical, and pharmaceuticals, yet receive limited attention in the food industry, especially for recipe optimization.

To address this, we explored ML and sequential learning for recipe formulation, aiming to optimize product quality. We combined DOE with ML to select optimal combinations of one, two, or three ingredients from 12  candidates, adjusting ingredient dosages based on the number of combined ingredients: 0-1 for single ingredients, 0-0.5 for pairs, and 0-0.33 for triplets. After assessing the main effects of all 12 ingredients, we narrowed the focus to five key ingredients. A full factorial design was applied to two-ingredient combinations, alongside collecting one data point at maximum dosage for each triplet. Three promising combinations were further analyzed using a space-filling design to explore the full parameter space.

Subsequently, ML models were developed to predict product quality, with sequential learning guiding additional experiments to refine the model for one specific combination. This approach identified the optimal mixture with fewer than 100 lab experiments, demonstrating the efficiency of combining ML, sequential learning, and DOE in reducing experimental efforts while identifying high-performing ingredient mixtures.