Predicting food taste with bound-driven optimization

This paper treats a recipe the way a materials scientist treats a composite: a mixture whose overall properties are bounded by the properties of its parts. The authors applied classical mixing bounds from material science to 70 recipes and compared the predicted sweet, salty, sour, bitter and umami intensity against how people actually rated them.

The bounds alone consistently underestimated perceived taste. The paper therefore builds a hybrid model that starts from the bound and adds eight chemistry-proxy features. With only ten interpretable features it performs on par with much more complex regression methods, and cuts prediction error by 27 to 62% across taste dimensions.