pith. sign in

arxiv: 2604.20206 · v1 · submitted 2026-04-22 · 💻 cs.CE

Predicting food taste with bound-driven optimization

Pagkratis Tagkopoulos , Dimitris Sfondilis , Ilias Tagkopoulos , Tarek Zohdi This is my paper

Pith reviewed 2026-05-09 23:15 UTC · model grok-4.3

classification 💻 cs.CE
keywords taste predictionfood formulationcomposite materialsHashin-Shtrikman boundshybrid modelinginverse designsensory attributeschemistry proxies
0
0 comments X

The pith

Treating recipes as composite materials and adding eight chemistry proxies removes systematic underprediction of taste from ingredients.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether food taste can be predicted from ingredient lists by adapting bounds originally used for the stiffness of mixed solids. These bounds supply a simple additive starting point but consistently fall below measured sweetness, sourness, umami, and saltiness because they omit the non-additive changes that occur during cooking. The authors therefore append eight interpretable proxy variables that stand in for Maillard reactions, caramelization, hydrolysis, and similar processes. The resulting ten-feature model removes the underprediction bias and lowers mean absolute error by 27 to 62 percent on four taste dimensions, matching the accuracy of a black-box model that uses more than one hundred per-ingredient inputs. The same framework is then used to solve the inverse problem of recovering ingredient combinations that meet a target taste profile while obeying recipe constraints.

Core claim

Recipes are modeled as composite materials whose taste contributions obey Hashin-Shtrikman and Reuss-Voigt bounds; these bounds supply an additive baseline that underestimates panel-measured taste in 77 percent of cases. Augmenting the baseline with eight chemistry-proxy features that encode processing effects eliminates the bias. The ten-feature hybrid model reduces mean absolute error by 27-62 percent for sweetness, sourness, umami, and saltiness and reaches accuracy comparable to Lasso regression on 115 per-ingredient features. Constrained differential-evolution optimization then recovers ingredient formulations that match prescribed taste targets.

What carries the argument

Hashin-Shtrikman upper bound and Reuss-Voigt bounds applied to ingredient taste values, then augmented by eight chemistry-proxy features that represent non-additive processing mechanisms.

If this is right

  • Taste prediction becomes feasible with a small, interpretable feature set instead of hundreds of per-ingredient variables.
  • The systematic underprediction bias that affects pure additive models is removed for sweetness, sourness, umami, and saltiness.
  • Ingredient combinations that achieve a target taste profile can be recovered by constrained optimization.
  • The same bound-plus-proxy structure can be applied to other sensory attributes once suitable chemistry proxies are defined.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the proxies transfer across cuisines, the method could support rapid reformulation of products to meet new nutritional or allergen constraints.
  • The material-science analogy suggests that other emergent properties in complex mixtures, such as aroma or mouthfeel, might also be bounded and then corrected with a handful of interaction terms.
  • The approach supplies a transparent prior that could reduce the data requirements for machine-learning models in food design.
  • Testing on larger, more diverse recipe collections would reveal whether the current eight proxies remain sufficient or whether additional terms are needed for different processing regimes.

Load-bearing premise

The eight chemistry-proxy features capture the dominant non-additive taste changes caused by cooking without needing recipe-specific recalibration.

What would settle it

Collect panel ratings for a fresh set of 20-30 recipes that undergo documented processing steps, then check whether the hybrid model's predictions remain within the training-set error bounds.

Figures

Figures reproduced from arXiv: 2604.20206 by Dimitris Sfondilis, Ilias Tagkopoulos, Pagkratis Tagkopoulos, Tarek Zohdi.

Figure 1
Figure 1. Figure 1: Overview. Left: 70 recipes decomposed into 209 ingredients with five taste dimen [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Dataset characteristics. (A) Ingredient usage frequency across 70 recipes (ingredi [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Recipe clustering. (A) t-SNE embedding of 70 recipes (RV-weighted taste vectors, [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Model evaluation. (A) HS bound predicted vs. actual taste scores (PCC = 0 [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 3 minor

Summary. The manuscript applies Hashin-Shtrikman (HS) and Reuss-Voigt bounds from composite-materials theory to predict five taste attributes from ingredient lists in a 70-recipe dataset with 209 ingredient-level references. It reports that the bounds systematically under-predict (77% of observations exceed the HS upper bound) and attributes the gap to processing chemistry. The authors augment the HS baseline with eight chemistry-proxy features to form a 10-feature hybrid model that eliminates bias and reduces MAE by 27-62% on sweetness, sourness, umami and saltiness, reaching performance comparable to Lasso regression on 115 per-ingredient features. They also demonstrate constrained inverse design via differential evolution.

Significance. If the reported error reductions survive rigorous validation, the work provides a novel, interpretable bridge between physics-based bounds and food-chemistry mechanisms for sensory prediction. Strengths include the use of established composite bounds as a transparent baseline, the small number of interpretable features, and the inverse-design demonstration. These elements could support formulation optimization if the gains prove generalizable beyond the current dataset.

major comments (3)
  1. Abstract: the central performance claims (27-62% MAE reduction and bias elimination for four tastes) are presented without any description of cross-validation procedure, feature-selection or construction details for the eight chemistry-proxy features, or statistical significance testing of the improvements. With N=70 this information is load-bearing for assessing whether the gains reflect general mechanisms or dataset-specific fitting.
  2. Abstract and hybrid-model description: the claim that the eight proxies (Maillard, caramelization, hydrolysis, synergy, etc.) suffice to capture dominant non-additive effects without overfitting or requiring new-dataset calibration is untested. No ablation study, external test set, or sensitivity analysis is reported, leaving open the possibility that the proxies absorb recipe-specific processing variance rather than transferable chemistry.
  3. Abstract: the statement that the 10-feature hybrid model achieves performance 'comparable' to Lasso on 115 features lacks the exact MAE values, regularization details, or confirmation that both models were evaluated under identical cross-validation folds, making the interpretability advantage difficult to quantify.
minor comments (3)
  1. Abstract: the 77% overall HS-exceedance rate is given with a per-taste range (26% bitterness to 97% saltiness); a supplementary table or figure breaking down exceedance statistics by taste dimension would improve clarity.
  2. The manuscript does not specify how the HS upper-bound formula is adapted from elastic moduli to scalar taste intensities, nor whether any scaling or normalization is applied to the ingredient taste references.
  3. The inverse-design section would benefit from explicit statement of the compositional constraints enforced and the success rate of recovering formulations that match target profiles within a stated tolerance.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive and detailed comments, which have helped us improve the clarity and rigor of the manuscript. We address each major comment point-by-point below, indicating the specific revisions made. All changes preserve the original scientific claims while enhancing transparency.

read point-by-point responses

  1. Referee: Abstract: the central performance claims (27-62% MAE reduction and bias elimination for four tastes) are presented without any description of cross-validation procedure, feature-selection or construction details for the eight chemistry-proxy features, or statistical significance testing of the improvements. With N=70 this information is load-bearing for assessing whether the gains reflect general mechanisms or dataset-specific fitting.

    Authors: We agree that these details are essential for evaluating robustness given the dataset size. In the revised manuscript we have updated the abstract to reference the 5-fold cross-validation procedure applied to all reported results. A new subsection in Methods now details the construction of each of the eight chemistry-proxy features (e.g., Maillard proxy derived from reducing-sugar and amino-acid concentrations, caramelization proxy from temperature-time integrals, hydrolysis proxy from protein content and pH, etc.). We also added Wilcoxon signed-rank tests confirming that the MAE reductions are statistically significant (p < 0.01) for sweetness, sourness, umami, and saltiness. These additions directly address the concern without altering the reported performance numbers. revision: yes

  2. Referee: Abstract and hybrid-model description: the claim that the eight proxies (Maillard, caramelization, hydrolysis, synergy, etc.) suffice to capture dominant non-additive effects without overfitting or requiring new-dataset calibration is untested. No ablation study, external test set, or sensitivity analysis is reported, leaving open the possibility that the proxies absorb recipe-specific processing variance rather than transferable chemistry.

    Authors: We acknowledge the value of ablation and sensitivity analyses for demonstrating that the proxies capture transferable mechanisms rather than dataset-specific artifacts. The revised manuscript now includes an ablation study that removes each proxy in turn and reports the resulting MAE increases, showing that every feature contributes measurably. Bootstrap-based sensitivity analysis has also been added to quantify feature stability. However, no suitable external dataset with matched panel ratings and ingredient-level processing metadata is publicly available, so external validation remains infeasible at present. The proxies are nevertheless grounded in established food-chemistry literature, which supports their intended generality. revision: partial

  3. Referee: Abstract: the statement that the 10-feature hybrid model achieves performance 'comparable' to Lasso on 115 features lacks the exact MAE values, regularization details, or confirmation that both models were evaluated under identical cross-validation folds, making the interpretability advantage difficult to quantify.

    Authors: We have replaced the qualitative term 'comparable' in the abstract with quantitative statements and added a new comparison table (Table 3) that reports exact MAE values for both models under identical 5-fold CV splits. The Lasso baseline used nested cross-validation to select the regularization parameter (alpha range 0.001–1.0, final value 0.1). The table shows the hybrid model achieves MAEs within 5–12 % of the Lasso model across the four tastes while employing only 10 features. This quantifies the interpretability advantage and confirms that both models were evaluated under the same protocol. revision: yes

standing simulated objections not resolved
  • Absence of an external test set for validating the hybrid model's generalizability beyond the 70-recipe dataset.

Circularity Check

0 steps flagged

No significant circularity; derivation relies on external bounds and independent feature engineering

full rationale

The paper applies Hashin-Shtrikman and Reuss-Voigt bounds from established composite materials literature as an additive baseline for taste prediction. It empirically observes that 77% of ground-truth values exceed the HS upper bound, attributes this to specific processing mechanisms, and introduces eight new chemistry-proxy features to augment the baseline in a hybrid model. No equation, prediction, or result in the provided text reduces by construction to a fitted parameter or self-referential definition drawn from the same evaluation data. The performance claims (MAE reduction, comparability to Lasso) are empirical outcomes of fitting the hybrid model, not tautological derivations. No self-citations are load-bearing, no uniqueness theorems are invoked, and no ansatz is smuggled via prior work. The central chain is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that taste can be decomposed into additive ingredient contributions plus a small number of global chemical-process corrections; the eight proxy features are introduced without independent validation that they are the minimal or optimal set.

free parameters (1)
  • eight chemistry-proxy features
    Hand-crafted encodings of Maillard, caramelization, hydrolysis, and synergy; their exact functional forms and any scaling constants are chosen to close the bound gap.
axioms (1)
  • domain assumption Taste perception can be approximated by linear superposition of ingredient-level references plus a small set of process corrections.
    Invoked when the HS/RV bounds are treated as the additive baseline before augmentation.

pith-pipeline@v0.9.0 · 5581 in / 1424 out tokens · 29810 ms · 2026-05-09T23:15:46.459104+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages

  1. [1]

    The receptors and cells for mammalian taste.Nature, 444(7117):288–294, 2006

    Jayaram Chandrashekar, Mark A Hoon, Nicholas JP Ryba, and Charles S Zuker. The receptors and cells for mammalian taste.Nature, 444(7117):288–294, 2006

    work page 2006

  2. [2]

    The cell biology of taste.The Journal of cell biology, 190(3):285, 2010

    Nirupa Chaudhari and Stephen D Roper. The cell biology of taste.The Journal of cell biology, 190(3):285, 2010

    work page 2010

  3. [3]

    A systematic review of data and models for predicting food flavor and texture.Current Research in Food Science, page 101127, 2025

    Michael Gunning and Ilias Tagkopoulos. A systematic review of data and models for predicting food flavor and texture.Current Research in Food Science, page 101127, 2025

    work page 2025

  4. [4]

    Scribner, New York, revised edition, 2004

    Harold McGee.On Food and Cooking: The Science and Lore of the Kitchen. Scribner, New York, revised edition, 2004

    work page 2004

  5. [5]

    Decoding the taste of peptides: Structure, interac- tions with taste receptors, bioactivities, and applications.Sustainable Food Proteins, 3(2):e70009, 2025

    Nesma Elhadad and Jianping Wu. Decoding the taste of peptides: Structure, interac- tions with taste receptors, bioactivities, and applications.Sustainable Food Proteins, 3(2):e70009, 2025

    work page 2025

  6. [6]

    Predicting multiple taste sensations with a multiobjec- tive machine learning method.npj Science of Food, 8(1):47, 2024

    Lampros Androutsos, Lorenzo Pallante, Agorakis Bompotas, Filip Stojceski, Gianvito Grasso, Dario Piga, Giacomo Di Benedetto, Christos Alexakos, Athanasios Kalogeras, Konstantinos Theofilatos, et al. Predicting multiple taste sensations with a multiobjec- tive machine learning method.npj Science of Food, 8(1):47, 2024

    work page 2024

  7. [7]

    A chemical language model for molecular taste prediction.npj Science of Food, 9(1):122, 2025

    Yoel Zimmermann, Leif Sieben, Henrik Seng, Philipp Pestlin, and Franz G¨ orlich. A chemical language model for molecular taste prediction.npj Science of Food, 9(1):122, 2025

    work page 2025

  8. [8]

    Ai for food: accelerating and democratizing discovery and innovation.npj Science of Food, 9(1):82, 2025

    Ellen Kuhl. Ai for food: accelerating and democratizing discovery and innovation.npj Science of Food, 9(1):82, 2025

    work page 2025

  9. [9]

    Predicting and improving complex beer flavor through machine learning.Nature Communications, 15(1):2368, 2024

    Michiel Schreurs, Supinya Piampongsant, Miguel Roncoroni, Lloyd Cool, Beatriz Herrera-Malaver, Christophe Vanderaa, Florian A Theßeling, Lukasz Kreft, Alexan- der Botzki, Philippe Malcorps, et al. Predicting and improving complex beer flavor through machine learning.Nature Communications, 15(1):2368, 2024. 10

    work page 2024

  10. [10]

    Flavor network and the principles of food pairing.Scientific Reports, 1:196, 2011

    Yong-Yeol Ahn, Sebastian E Ahnert, James P Bagrow, and Albert-L´ aszl´ o Barab´ asi. Flavor network and the principles of food pairing.Scientific Reports, 1:196, 2011

    work page 2011

  11. [11]

    Compu- tational methods for food reformulation: A review.Journal of Food Engineering, 349:111478, 2023

    Mahdi Ghasemi-Varnamkhasti, Seyed Hasan Yoosefian, and Jorge Lozano. Compu- tational methods for food reformulation: A review.Journal of Food Engineering, 349:111478, 2023

    work page 2023

  12. [12]

    A variational approach to the theory of the elastic behaviour of multiphase materials.Journal of the Mechanics and Physics of Solids, 11(2):127–140, 1963

    Zvi Hashin and Shmuel Shtrikman. A variational approach to the theory of the elastic behaviour of multiphase materials.Journal of the Mechanics and Physics of Solids, 11(2):127–140, 1963

    work page 1963

  13. [13]

    Taste, fat and texture database—taste values dutch foods.DANS, Wageningen, 10, 2020

    Monica Mars, C de Graaf, PS Teo, and AWB van Langeveld. Taste, fat and texture database—taste values dutch foods.DANS, Wageningen, 10, 2020

    work page 2020

  14. [14]

    Lehrbuch der kristallphysik

    Woldemar Voigt. Lehrbuch der kristallphysik. Teubner, Leipzig, 1928

    work page 1928

  15. [15]

    Berechnung der Fließgrenze von Mischkristallen auf Grund der Plastizit¨ atsbedingung f¨ ur Einkristalle.Zeitschrift f¨ ur angewandte Mathematik und Mechanik, 9(1):49–58, 1929

    Andras Reuss. Berechnung der Fließgrenze von Mischkristallen auf Grund der Plastizit¨ atsbedingung f¨ ur Einkristalle.Zeitschrift f¨ ur angewandte Mathematik und Mechanik, 9(1):49–58, 1929

    work page 1929

  16. [16]

    Springer, Berlin, 2nd edition, 2005

    Tarek I Zohdi and Peter Wriggers.Introduction to Computational Micromechanics. Springer, Berlin, 2nd edition, 2005

    work page 2005

  17. [17]

    Dehydrated foods: chemistry of browning reactions in model systems.Journal of Agricultural and Food Chemistry, 1(15):928–943, 1953

    John Edward Hodge. Dehydrated foods: chemistry of browning reactions in model systems.Journal of Agricultural and Food Chemistry, 1(15):928–943, 1953

    work page 1953

  18. [18]

    Springer, Berlin, 4th edition, 2009

    Hans-Dieter Belitz, Werner Grosch, and Peter Schieberle.Food Chemistry. Springer, Berlin, 4th edition, 2009

    work page 2009

  19. [19]

    Paul A.S. Breslin. An evolutionary perspective on food and human taste.Current Biology, 23(9):R409–R418, 2013

    work page 2013

  20. [20]

    Regression shrinkage and selection via the lasso.Journal of the Royal Statistical Society: Series B (Methodological), 58(1):267–288, 1996

    Robert Tibshirani. Regression shrinkage and selection via the lasso.Journal of the Royal Statistical Society: Series B (Methodological), 58(1):267–288, 1996

    work page 1996

  21. [21]

    Differential evolution – a simple and efficient heuris- tic for global optimization over continuous spaces.Journal of Global Optimization, 11(4):341–359, 1997

    Rainer Storn and Kenneth Price. Differential evolution – a simple and efficient heuris- tic for global optimization over continuous spaces.Journal of Global Optimization, 11(4):341–359, 1997

    work page 1997

  22. [22]

    SciPy 1.0: fundamental algorithms for scientific computing in Python.Nature Methods, 17(3):261–272, 2020

    Pauli Virtanen, Ralf Gommers, Travis E Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, et al. SciPy 1.0: fundamental algorithms for scientific computing in Python.Nature Methods, 17(3):261–272, 2020

    work page 2020

  23. [23]

    Nolden and John E

    Alissa A. Nolden and John E. Hayes. Perceptual and temporal variation in bitterness and savoriness of culinary herbs and spices.Nutrients, 9(9):1001, 2017

    work page 2017

  24. [24]

    The synergistic taste effect of monosodium glutamate and disodium 5’-inosinate.Journal of Food Science, 36(6):846–849, 1971

    Shizuko Yamaguchi. The synergistic taste effect of monosodium glutamate and disodium 5’-inosinate.Journal of Food Science, 36(6):846–849, 1971. 11

    work page 1971

  25. [25]

    Perception of sweetness and bitterness in different vehicles.Perception & psychophysics, 54(6):751–758, 1993

    Amalia M Calvi˜ no, Mar´ ıa Rosa Garc´ ıa-Medina, J Enrique Cometto-Muniz, M´ onica B Rodr´ ıguez, and C´ atedra de Fisiolog´ ıa. Perception of sweetness and bitterness in different vehicles.Perception & psychophysics, 54(6):751–758, 1993

    work page 1993

  26. [26]

    Modification of the bitterness of caffeine.Food Quality and Preference, 15(3):259–269, 2004

    Russell SJ Keast and Paul AS Breslin. Modification of the bitterness of caffeine.Food Quality and Preference, 15(3):259–269, 2004

    work page 2004

  27. [27]

    Visualizing data using t-SNE.Journal of Machine Learning Research, 9:2579–2605, 2008

    Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE.Journal of Machine Learning Research, 9:2579–2605, 2008

    work page 2008

  28. [28]

    hdbscan: Hierarchical density based clustering.Journal of Open Source Software, 2(11):205, 2017

    Leland McInnes, John Healy, and Steve Astels. hdbscan: Hierarchical density based clustering.Journal of Open Source Software, 2(11):205, 2017

    work page 2017

  29. [29]

    Dynamic instrumental and sensory meth- ods used to link aroma release and aroma perception: A review.Molecules, 28(17):6308, 2023

    Jean-Luc Le Qu´ er´ e and Rachel Schoumacker. Dynamic instrumental and sensory meth- ods used to link aroma release and aroma perception: A review.Molecules, 28(17):6308, 2023

    work page 2023

  30. [30]

    Taste mixture interactions: suppression, additivity, and the predominance of sweetness

    Barry G Green, Juyun Lim, Floor Osterhoff, Karen Blacher, and Danielle Nachtigal. Taste mixture interactions: suppression, additivity, and the predominance of sweetness. Physiology & behavior, 101(5):731–737, 2010

    work page 2010

  31. [31]

    Orig” and “Opt

    Fabian Pedregosa, Ga¨ el Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in Python.Journal of Machine Learning Research, 12:2825–2830, 2011. 12 Figure 1: Overview. Left: 70 recipes decomposed into 209 ingredients with fiv...

    work page 2011