Predicting food taste with bound-driven optimization
Pagkratis Tagkopoulos, Dimitris Sfondilis, Ilias Tagkopoulos, Tarek Zohdi · 2026
Read Predicting food taste with bound-driven optimization by Pagkratis Tagkopoulos, Dimitris Sfondilis, Ilias Tagkopoulos, Tarek Zohdi (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-20206.txt.
Introduction
Tagkopoulos and three colleagues predict food taste from ingredients by treating recipes as composite materials. Hashin-Shtrikman and Reuss-Voigt bounds, borrowed from elastic moduli, cover five tastes across 70 recipes with 209 ingredient references and panel ground truth. The bounds under-predict: 77 percent of actual values exceed the upper bound, from 26 percent for bitterness to 97 percent for saltiness. Eight chemistry-proxy features cut mean absolute error by 27 to 62 percent on four tastes, using only ten interpretable features. Inverse design then recovers formulations matching target taste profiles.
Catalog record
| CREATOR | Pagkratis Tagkopoulos, Dimitris Sfondilis, Ilias Tagkopoulos, Tarek Zohdi |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 34093 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.20206
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Predicting food taste with bound-driven optimization Pagkratis Tagkopoulos, Dimitris Sfondilis, Ilias Tagkopoulos, Tarek Zohdi Abstract The prediction of sensory attributes from ingredient-level formulations is an emerging challenge at the intersection of food science and artificial intelligence. We address the fundamental question of whether the taste of a food can be predicted from its ingredients by treating recipes as composite materials. We apply Hashin–Shtrikman (HS) and Reuss–Voigt (RV) bounds, techniques originally developed for elastic moduli, to predict five taste dimensions (sweetness, sourness, bitterness, umami, saltiness) on a curated dataset of 70 recipes decomposed into 209 ingredient-level taste references with trained-panel ground truth. The bounds provided an additive baseline but systematically under-predict perceived taste: 77% of actual taste values exceeded the HS upper bound, with the exceedance rate ranging from 26% (bitterness) to 97% (saltiness). We traced this gap to specific processing chemistry (Maillard reactions, caramelization, evaporative concentration, protein hydrolysis, and nucleotide synergy) and introduced a hybrid model that augments the HS baseline with eight chemistry-proxy features encoding these mechanisms. Our results show that our interpretable hybrid model eliminates the systematic bias and reduces mean absolute error by 27–62% for sweetness, sourness, umami, and saltiness while using only 10 interpretable features, achieving performance comparable to a black-box Lasso regression on 115 per-ingredient features. We further demonstrate constrained inverse design via Differential Evolution, recovering ingredient formulations that match target taste profiles subject to compositional bounds. Our work demonstrates how key chemical processes during food preparation can inform and augment physics-based and machine learning models, providing a quantitative fingerprint of processing chemistry’s contribution to taste perception and paving the way for model-driven food formulation with targeted sensory characteristics. Keywords: sensory prediction, composite material bounds, Hashin–Shtrikman, machine learning, food formulation, taste modeling, inverse design 1 Introduction Human taste—the perception of chemical stimuli by taste receptor cells—is typically described through five basic modalities: sweet, sour, salty, bitter, and umami [ 6 ] . Each corresponds to distinct classes of tastants: sugars for sweetness, acids for sourness, sodium salts for saltiness, plant alkaloids for bitterness, and glutamate derivatives for umami. The human mouth contains thousands of taste receptor cells distributed across taste buds, each capable of detecting and integrating multiple taste stimuli [ 7 ] . This sensory input gives rise to a highly complex and dynamic taste space, which becomes even more intricate when combined with olfaction, texture, temperature, and auditory cues—factors that together constitute the broader perception of flavor [ 11 ] . Flavor perception is thus inherently multi-sensory, shaped by chemical and cognitive processes [ 18 ] . A central question in computational gastronomy is whether the taste profile of a recipe can be predicted from its ingredient composition and their proportions. Conversely, the inverse problem asks which ingredients and in what proportions should be combined to achieve a desired sensory profile. Both problems have been the subject of recent research, although most work focuses on single molecules rather than composite foods. Models may predict a single taste modality from molecular structure [ 8 ] , predict the full taste profile via multi-objective optimization [ 2 ] , or fine-tune large language models on flavor databases [ 30 ] . Prediction of food taste as a composite whole—from ingredient-level properties to recipe-level perception—remains an open problem with far-reaching applications in food design [ 15 , 11 , 23 , 1 , 9 ] . In this study, we ask: can composite-material theory bracket the taste of a food? We apply Hashin–Shtrikman and Reuss–Voigt bounds [ 12 ] —originally developed to estimate effective elastic properties of multiphase materials—to predict recipe taste from ingredient taste scores and mass fractions. We quantify where the analogy holds and where it breaks, trace the failures to specific food-chemistry mechanisms, and introduce a hybrid model that combines the physics-informed baseline with chemistry-informed corrections. We evaluate all models against a curated dataset of 70 recipes with trained-panel ground truth on five taste dimensions ( Figure 1 ), and demonstrate constrained inverse design via evolutionary optimization. 2 Methods 2.1 Dataset Ground-truth taste scores (sweetness, sourness, bitterness, umami, saltiness, 0–100 Spectrum™ scale) for 627 Dutch food items were drawn from the trained-panel SVT database of Mars et al. [ 17 ] . From this database, 126 single-ingredient items served directly as ingredient-level taste references, and 83 additional ingredients were scored from published sources (USDA FoodData Central, NEVO 2021, peer-reviewed sensory studies). Seventy multi-ingredient food items were decomposed into constituent ingredients with mass fractions assigned via a four-tier evidence hierarchy (label-declared, NEVO recipe documentation, food-science literature, arithmetic residual), with all fractions summing to 1.000\pm 0.001 per recipe. Each recipe carries a composition confidence rating (High: 22, Moderate: 42, Low: 6). Fat sensation, the sixth SVT dimension, was excluded because it reflects textural properties not consistently modeled by compositional approaches. The complete data pipeline is documented in the Supplementary Information, and all data are provided in Supplementary Data File 1. 2.2 Forward prediction: composite-material bounds We treat each recipe as a heterogeneous N -phase composite, where ingredient i has taste score T_{i} and mass fraction v_{i} ( \sum_{i}v_{i}=1 ). Each of the five taste dimensions is modeled independently. Reuss–Voigt (RV) bounds. The Voigt bound [ 28 ] (arithmetic weighted average) and Reuss bound [ 22 ] (harmonic weighted average) give the widest theoretical brackets on the effective property T^{*} : T^{*}_{\text{Reuss}}=\left[\sum_{i=1}^{N}\frac{v_{i}}{T_{i}}\right]^{-1}\leq\;T^{*}\;\leq\sum_{i=1}^{N}v_{i}\,T_{i}=T^{*}_{\text{Voigt}} (1) When any T_{i}=0 (e.g., sugar has Sour =0 ), we regularize T_{i}\to\max(T_{i},\varepsilon) with \varepsilon=0.01 . Hashin–Shtrikman (HS) bounds. For N isotropic phases, the HS bounds provide tighter limits than RV [ 12 , 31 ] . Defining the auxiliary function A(T_{0})=\left[\sum_{i=1}^{N}\frac{v_{i}}{T_{i}+2\,T_{0}}\right]^{-1}-2\,T_{0} (2) the multi-phase HS bounds are T^{*-}=A(T_{\min})\;\leq\;T^{*}\;\leq\;A(T_{\max})=T^{*+} (3) where T_{\min}=\min_{i}T_{i} and T_{\max}=\max_{i}T_{i} . The midpoint (T^{*-}+T^{*+})/2 serves as the point estimate. The factor “2” in Eq. ( 2 ) arises from the spherical inclusion geometry in three dimensions, and it equals (d-1) in d dimensions. We adopt d=3 following the original derivation. We emphasize that this is a phenomenological analogy: taste intensity is not a physical modulus, and d serves as a shape parameter controlling bound tightness rather than a literal spatial dimension. The robustness analysis (Supplementary Section S9.1) confirms that the processing gap persists across d\in\{2,3,5,10,50\} , supporting the utility of the analogy despite its heuristic nature. For N=2 , Eq. ( 3 ) reduces to the classical two-phase result [ 12 ] . Derivations are given in Supplementary Section S3. 2.3 Chemistry-informed hybrid model The HS bounds capture additive mixing but not the processing chemistry that amplifies taste during cooking: Maillard reactions [ 13 ] generate umami-active peptides and sweet furanones from amino acids and sugars [ 3 ] . Caramelization creates sweet and bitter compounds from thermal sugar decomposition. Evaporative water loss concentrates dissolved tastants (particularly NaCl). Protein hydrolysis releases free glutamate, and nucleotide synergy (IMP \times glutamate) amplifies umami perception non-linearly [ 4 ] . We augment the HS baseline with a learned correction over K=8 chemistry-proxy features \phi_{k} that encode these mechanisms from the ingredient list alone (no process parameters required): \hat{T}_{j}^{*}=\underbrace{\frac{T_{j}^{*-}+T_{j}^{*+}}{2}}_{\text{HS midpoint}}+\underbrace{\beta_{j,0}\,T_{j}^{\text{Voigt}}+\sum_{k=1}^{K}\beta_{j,k}\,\phi_{k}}_{\text{chemistry correction}} (4) where T_{j}^{\text{Voigt}} is the Voigt bound for taste dimension j , and the features \phi_{k} are: protein fraction, sugar fraction, their product (Maillard potential), salt fraction, water fraction, concentration factor 1/(1-v_{\text{water}}) ,