A new paper, Predicting food taste with bound-driven optimization, combines food science, mathematical modeling, and machine learning to investigate taste prediction at the formulation level.
The authors analyzed a curated dataset of 70 multi-ingredient foods with trained-panel sensory ground truth. Their starting question was: if we know the taste characteristics of a recipe’s ingredients and their proportions, how much can we predict about the finished product?
To test this, the authors treated recipes as composite systems and used additive models as a baseline. These models showed positive correlations with observed taste, but they also systematically underpredicted it. Across the dataset, 77% of observed taste values exceeded the Hashin–Shtrikman upper bound, with exceedance particularly pronounced for saltiness, sweetness, umami, and sourness.
The results showed a clear gap between the composition-only additive baseline and the observed taste of the finished foods.
The pattern was consistent with known effects of processing chemistry. The researchers therefore introduced eight chemistry-informed proxy features encoding the potential for mechanisms including evaporative concentration, Maillard chemistry, caramelization, protein hydrolysis, and nucleotide synergy.
Adding these features to a hybrid model substantially improved prediction. For sweetness, sourness, umami, and saltiness, mean absolute error decreased by 27–62% relative to the Hashin–Shtrikman baseline. Within the study dataset, the hybrid model also removed the systematic bias observed in the additive baseline while remaining interpretable and using far fewer features than the per-ingredient Lasso model.
Can the model work in reverse?
The authors also explored an inverse problem with direct relevance to food R&D. Using constrained optimization, the researchers tested three reformulation scenarios: reducing salt in pea soup, reducing sugar in a chocolate-hazelnut spread, and increasing umami while reducing sweetness in ketchup. In each case, the model adjusted ingredient proportions within formulation constraints to move the predicted sensory profile toward the target.
These examples demonstrate computational reformulation rather than bench-validated formulations; prospective sensory testing would be an important next step.
The research points toward a broader opportunity in food product development. By combining food science, computational modeling, experimental data, and process understanding, predictive approaches could help R&D teams explore larger formulation spaces, identify trade-offs earlier, and make more informed decisions about which formulations to test physically.
Read the paper: Predicting food taste with bound-driven optimization, published in Current Research in Food Science.