Chapter 13: Designing Flavors
“A good cook is like a sorceress who dispenses happiness.” — Elsa Schiaparelli, Shocking Life (1954)
RUNNING EXAMPLE — THE NEAREST PLANT POINT
Dev Anand has a beaker of beef on his bench. Not beef — a vial of the headspace above a seared ribeye, captured an hour ago and run through the gas chromatograph in the next room, now reduced to a printout: two hundred and forty volatile compounds with their concentrations, and beneath it the sensory panel’s scores on thirty descriptors. This is the target. Somewhere in front of Dev, in the rows of amber bottles that line his workstation at a flavor house in Cincinnati, is the closest plant-derived approximation to it that exists — he just has not built it yet.
The problem is concrete and it is geometric. The beef sits at a point $\mathbf{f}^\star$ in flavor space: a particular balance of roasted, fatty, bloody-metallic, umami, and sulfurous-savory notes, the signature of the Maillard reaction running on muscle and fat. Dev cannot use any animal compound. He has, instead, a palette: yeast extract (umami, a little meaty), beet juice (the blood note, and the color), coconut and sunflower fats (the mouthfeel), methional and 2-methyl-3-furanthiol synthesized from plant feedstocks (the cooked-meat thiols), a Maillard reaction precursor of plant sugars and amino acids he can heat to order. Each bottle is a vector. He is allowed to add them in nonnegative amounts, up to a cost ceiling and a “clean label” constraint that forbids certain molecules entirely.
He is not searching the bottles at random. He is solving for the point on the convex hull of his palette that lands nearest to $\mathbf{f}^\star$ — and where it cannot land near enough, deciding which dimensions of the beef to sacrifice. That decision, repeated a thousand times across the industry, is why the plant burger you ate last year tasted closer to beef than the one from five years before. The gap is a distance in flavor space, and it is shrinking along measurable axes.
Engineering in a Perceptual Space
Everything in this book up to now has been description: flavor is a vector, pairing is an operation, cooking is a transformation, a cuisine is a distribution. Description is the easier half. This chapter turns the framework around and asks the engineer’s question. Given the geometry, can we design? Given a point in flavor space we want to reach — the taste of beef, a cola that is sweet but not bitter, a children’s antibiotic a child will actually swallow — can we compute the ingredients that get us there, rather than discovering them by the centuries of trial that produced every traditional cuisine?
The answer, increasingly, is yes, and an entire industry has been built on it without ever using the vocabulary. The world’s flavor and fragrance houses — Givaudan, IFF (International Flavors & Fragrances, which absorbed Firmenich in 2021), Symrise, Takasago — employ thousands of chemists whose daily work is, formally, optimization in a few tens of dimensions. They have target points (a client’s brief: “a cleaner mango, less green, more ripe”), a palette of basis vectors (tens of thousands of approved aroma chemicals and naturals), constraints (cost, regulatory status, allergen labeling, heat stability, solubility in the final matrix), and an objective (perceptual distance to the target, scored by trained panels). They do not write down a metric tensor. But the structure of what they do is exactly the structure this book has been building, and stating it geometrically does two things: it explains why the work is hard in the specific ways it is hard, and it tells us when a target is reachable and when no combination of the palette can get there.
[Modeling Axiom.] Flavor design is constrained optimization in $\mathcal{F}$. A design problem is a quadruple $(\mathbf{f}^\star, B, \mathcal{C}, d)$: a target point $\mathbf{f}^\star \in \mathcal{F}$; a palette $B = \{\mathbf{b}_1, \dots, \mathbf{b}_m\}$ of available ingredient vectors; a constraint set $\mathcal{C}$ on the admissible coefficients; and the flavor metric $d$ of Chapter 3, which says what “close” means. A design is a coefficient vector $\mathbf{x} \in \mathcal{C}$, and the designed flavor is $\mathbf{f}(\mathbf{x}) = \sum_i x_i \mathbf{b}_i$. The problem is to minimize $d(\mathbf{f}(\mathbf{x}), \mathbf{f}^\star)$ over $\mathbf{x} \in \mathcal{C}$. Every section of this chapter is a special case of this one problem.
That last claim is worth taking seriously, because it unifies what look like five unrelated commercial activities. Building a meat analogue, masking a drug’s bitterness, designing a flavor from scratch, formulating to a cost ceiling, and tuning a recipe to one person’s palate are not five problems. They are one problem under five different settings of $(\mathbf{f}^\star, B, \mathcal{C}, d)$ — five places to stand in the same optimization. The chapter walks through them in turn.
13.1 Flavor Houses as Navigators of Flavor Space
The flavor industry is the empirical proof that flavor space is navigable, because it navigates the space for profit, daily, at scale. A flavor house receives a brief — a beverage company wants a “wild strawberry” note that survives pasteurization and a year on the shelf — and returns, weeks later, a formula: a list of compounds and weights that, dosed into the client’s matrix, produces the requested percept. Between the brief and the formula is a search, and understanding the search geometrically explains both the industry’s competence and its structure.
[Empirical.] The palette is enormous but bounded. Roughly 3,000 aroma chemicals are in common commercial use, drawn from a regulatory-approved universe of some thousands more (the FEMA GRAS program in the United States, the standard reference being Reineccius, Flavor Chemistry and Technology (2006), catalogs on the order of a few thousand flavor substances; IOFI and EU registers overlap heavily). These are the basis vectors. Crucially, they are not the ten thousand-plus volatiles found in nature; they are the subset that is synthesizable at scale, stable, safe, and affordable. The palette is therefore a finite, curated frame spanning a subspace of the full olfactory manifold of Chapter 1 — and the first fact of industrial flavor design is that the reachable subspace is smaller than the perceptual space it sits inside.
This immediately predicts the industry’s central difficulty. A natural flavor — a real strawberry — is a point $\mathbf{f}^\star$ that lies on the olfactory manifold but generally not in the span of the commercial palette. The designer cannot reproduce it; they can only reach the nearest point in the palette’s span (or, more precisely, its admissible convex region, §13.4). The residual — the distance from that nearest point to the true target — is what every taster means by “it tastes artificial.” Artificiality is not a chemical property of synthetic molecules; many are molecule-for-molecule identical to natural ones. It is a geometric property of the formula: the perceptible distance between where the palette could land and where the real food sits. [Empirical.] This is why “natural” extracts taste rounder: they carry the dozens of low-concentration background volatiles that the designer, working with twenty or thirty dominant character-impact compounds, omits — and those omitted dimensions are exactly the residual.
What a flavorist possesses, then, is a learned chart of this subspace. A senior perfumer or flavorist has internalized, over a decade or two, where each of a few thousand materials sits and which directions they move a formula — the same internalized metric the line cook of Chapter 1 used over her forty containers, scaled to thousands of bottles and made explicit in a formula sheet. The flavor houses are, structurally, repositories of the metric of Chapter 3: not the abstract tensor, but a vast empirical sampling of which moves in palette-space produce which perceptual displacements — the proprietary analogue of the public compound-and-pairing networks that Ahn et al. (2011) assembled from shared volatiles. Their competitive advantage is a map, and the map is why the search is fast. This is the same claim the series makes about expertise everywhere — that mastery is a learned metric, not a lookup table — and the flavor house is its purest commercial instance.
13.2 Flavor Masking as Projection
Some of the most valuable flavor design is not addition but removal. A great many active pharmaceutical ingredients are intensely bitter — the macrolide antibiotics, acetaminophen, ibuprofen, ranitidine, many antiretrovirals — and bitterness is a deal-breaker for pediatric and geriatric medicines, where a refused dose is a failed therapy. Mennella and colleagues have documented this directly in children, whose heightened bitter sensitivity (Drewnowski & Gomez-Carneros 2000) makes the masking problem sharper than for adults. The task handed to the flavor chemist is sharp: remove the bitter, change nothing else that matters. The molecule’s efficacy lives in its pharmacology and must be untouched; only its flavor projection may be altered. Stated this way, masking is the cleanest geometric operation in the chapter.
[Modeling Axiom.] Masking is suppression along a subspace. Write the drug’s flavor vector as $\mathbf{f}_{\text{drug}}$, with a large component along the bitter axis (in the taste block of Chapter 1) and a fixed, untouchable pharmacological vector $\mathbf{p}$ orthogonal to $\mathcal{F}$ entirely — efficacy is not a flavor coordinate. The masking goal is to find an additive formula $\mathbf{m}$ such that the perceived flavor $\Pi_{\perp}(\mathbf{f}_{\text{drug}} + \mathbf{m})$, projected off the bitter axis, is acceptable, while $\mathbf{p}$ is unchanged. The ideal mask is a projection: it should zero out, or sharply attenuate, the bitter coordinate while leaving the rest of the flavor vector — and all of the pharmacology — fixed.
There are three geometrically distinct ways to achieve this, and the distinction is not pedantic — it is the difference between approaches that work and approaches that fail. The first is physical removal of the coordinate at the source: microencapsulation or ion-exchange-resin complexation that prevents the bitter molecule from ever reaching the T2R receptors, so the bitter component is literally not generated. Geometrically this is the true projection — the bitter axis is set to zero before perception. The second is receptor-level antagonism: a few compounds (sodium salts, certain lipoproteins, the much-studied adenosine monophosphate) bind or block T2R bitter receptors, attenuating the bitter response without removing the molecule. This rescales the bitter coordinate toward zero — also close to a projection, but only along the specific receptors the antagonist hits, and the ~25 T2R receptors of Chapter 1 mean a single antagonist rarely covers them all.
[Empirical.] The third and most common method — additive overshadowing, dumping in sweetness and a strong flavor (cherry, bubblegum) — is not a projection, and its inferiority is geometric. It does not reduce the bitter coordinate; it adds large components on other axes so the bitter is a smaller fraction of the whole. But the bitter vector is still there, and worse, the sweet-suppresses-bitter interaction is one of the off-diagonal metric couplings of Chapter 1 §1.1 — it bends perception but does not delete the dimension. The masked drug therefore reads as “sweet with a bitter back-end,” because in a curved flavor metric you cannot hide a coordinate by piling up others; the bitter resurfaces wherever the suppressing interaction is weak (notably in the aftertaste, where sweetness clears first). The lesson generalizes: true masking projects out a dimension; cheap masking merely lengthens the vector in other directions, and the metric of Chapter 3 guarantees the unwanted dimension leaks back.
13.3 Plant-Based Meat as Flavor Mimicry
The most public flavor-design problem of the decade is the one on Dev Anand’s bench: reproduce the flavor of meat using only plants. It is mimicry in the exact sense of the framework — find an ingredient combination whose flavor vector lands as close as possible to a target $\mathbf{f}^\star$ that the available palette cannot, in principle, reach exactly, because the target was generated by chemistry (heme iron, animal fat, the specific Maillard products of muscle) that the palette excludes by definition. This is the nearest plant point problem, and it is the running example because it shows every feature of constrained design at once.
[Empirical.] The decisive insight of the modern plant-meat industry was geometric, even when its discoverers did not say so. Earlier veggie burgers tried to mimic meat’s whole-recipe gestalt and landed nowhere near it. The breakthrough — associated with Impossible Foods’ work on leghemoglobin — was to identify the character-impact dimension of cooked beef and target it directly. Cooked-meat flavor is dominated by the reaction of heme iron with fats and amino acids during cooking; heme is the catalyst that drives the meaty Maillard pathway (the chemistry reviewed in Mottram 1998). Soy leghemoglobin, a heme protein from the root nodules of legumes, supplies the same catalytic heme from a plant source. The move was not “make something meaty”; it was “find the single axis along which beef is farthest from existing plant food, and acquire a basis vector that points along it.” That is principal-direction reasoning applied to a design target.
But one axis is not the whole vector, and the residual structure is instructive. Beef’s flavor vector has, at minimum, a roasted-Maillard component, a fatty/lactone component (beef tallow has a characteristic ruminant note from branched-chain fatty acids), a bloody-metallic component (heme again, uncooked), an umami component (free glutamate, inosine monophosphate — the IMP that animals carry and plants largely do not), and a sulfurous-savory component (the meaty thiols 2-methyl-3-furanthiol and methional). Dev’s palette can reach some of these cleanly: yeast extract and fermentation supply glutamate; coconut and cocoa-butter fractions plus synthesized lactones approximate the fat; heme and reaction flavors cover the roasted and metallic axes. Others resist. [Empirical.] The IMP/GMP nucleotide umami — and the salt-glutamate-nucleotide synergy, a superlinear interaction (Chapter 4 §4.6) in which glutamate plus a 5’-ribonucleotide produces umami far exceeding the sum — is hard to source and to stabilize from plants, which is one reason plant meats lean on added yeast-derived nucleotides and still read as “savory but not quite deep.”
THE NEAREST PLANT POINT, REVISITED — WHICH DIMENSION TO SACRIFICE
Dev’s optimizer has converged, and the news is mixed. His formula lands close to $\mathbf{f}^\star$ on the roasted, fatty, and front-of-palate umami axes — within the panel’s discrimination threshold, the geometric definition of “indistinguishable.” But there are two stubborn residuals. The bloody-metallic note is slightly high (his heme dose, tuned for the cooked Maillard pathway, overshoots raw), and the long-finish umami depth is low (the missing nucleotide synergy). He cannot fix both: lowering heme to correct the metallic note also weakens the Maillard drive that the roasted axis depends on. The axes are coupled through a single ingredient, exactly the curvature problem of Chapter 3.
So he makes a perceptual judgment that is really a choice of metric. Diners weight the front-of-palate, first-bite roasted-fatty experience far more heavily than the finish; the metric $d$ should be anisotropic, stretched along the dimensions eaters notice. He accepts the thin finish and protects the first bite. The “nearest plant point” was never the Euclidean nearest. It was nearest in the perceptually weighted metric — and knowing which way that metric is stretched is the whole craft.
The plant-meat problem also clarifies what “perfect mimicry” would and would not require. The framework predicts that the analogue need not match the chemical vector of beef — an impossible target with a plant palette — but only its perceptual vector after the projection from chemistry to perception (Chapter 2). Because that projection is many-to-one, distinct chemical formulas can land on the same perceptual point: there exists, in principle, a plant formula perceptually indistinguishable from beef even though no plant molecule equals an animal one. Whether the admissible plant palette spans enough of the perceptual neighborhood of $\mathbf{f}^\star$ to contain such a point is an empirical question, and the honest answer today is “along most axes yes, along the deep-umami and the precise fat-melt axes not yet.” The gap is a set of named, shrinking residuals — which is a far more useful thing to know than “it doesn’t quite taste like beef.”
13.4 The Inverse Problem
We can now state the design problem in its general form and solve it. Chapters 7 and 8 treated the forward problem: given ingredients and operations, what flavor results? Design is the inverse: given a target flavor, what ingredients produce it? Inverse problems are generically harder than forward ones — they may have no solution, or many — but the geometry of flavor makes this particular inverse problem unusually tractable, because in the relevant regime it is convex.
[Modeling Axiom.] The inverse flavor problem is a constrained least-distance program. Let $B \in \mathbb{R}^{n \times m}$ be the matrix whose columns are the $m$ palette vectors in the $n$-dimensional flavor space, $\mathbf{x} \in \mathbb{R}^m$ the nonnegative dose vector, and $M$ the positive-definite matrix of the flavor metric (the local metric tensor of Chapter 3, $d(\mathbf{u},\mathbf{v})^2 = (\mathbf{u}-\mathbf{v})^\top M (\mathbf{u}-\mathbf{v})$). The design problem is
$$ \min_{\mathbf{x}}\; (B\mathbf{x} - \mathbf{f}^\star)^\top M\, (B\mathbf{x} - \mathbf{f}^\star) \quad \text{subject to}\quad \mathbf{x} \ge 0,\;\; \mathbf{c}^\top \mathbf{x} \le \kappa,\;\; \mathbf{x} \in \mathcal{C}. $$
The objective is the squared perceptual distance to the target; $\mathbf{x} \ge 0$ because you cannot add a negative amount of an ingredient; $\mathbf{c}^\top\mathbf{x} \le \kappa$ is the cost ceiling ($\mathbf{c}$ the per-unit costs, $\kappa$ the budget); and $\mathcal{C}$ collects the regulatory, allergen, solubility, and clean-label constraints. With $M$ positive-definite and the constraints linear, this is a convex quadratic program — the workhorse of convex optimization, solvable to global optimality in milliseconds.
The convexity is the gift, and it is worth saying why it holds and where it breaks. It holds because, in the linear-mixing regime, the designed flavor $B\mathbf{x}$ is a linear function of the doses and the distance is a convex quadratic — so there are no false local optima, and the nearest reachable point is found exactly. The reachable set $\{B\mathbf{x} : \mathbf{x} \in \mathcal{C}\}$ is a convex polytope (the nonnegative, budget-limited combinations of the palette), and the solution is the projection of $\mathbf{f}^\star$ onto that polytope in the metric $M$. When $\mathbf{f}^\star$ lies inside the polytope, the distance is zero — the target is reachable, and the QP finds a formula that hits it. When $\mathbf{f}^\star$ lies outside, the solution is the boundary point of the polytope nearest the target, and the residual distance is the irreducible “artificiality gap” of §13.1, now given a precise value rather than a colloquial intuition. Either way the optimizer also returns the active constraints — which is to say it tells the chemist exactly which limit (cost, an excluded molecule, an allergen cap) is preventing a closer match. That diagnostic is often worth more than the formula.
[Empirical.] Convexity breaks where the chemistry is nonlinear, and the framework names where: the superlinear and sublinear interactions of Chapter 4 §4.6 (umami synergy, sweet-bitter cancellation), and the threshold and saturation nonlinearities of the Weber-Fechner metric (Chapter 3) at high dose. In these regions $B\mathbf{x}$ is no longer linear in $\mathbf{x}$ and the QP is only a local model. The industrial practice — and the right one — is to solve the convex QP to get a near-optimal formula, then refine with a few panel-tested iterations that locally correct for the nonlinear terms, exactly as gradient descent with a good convex initialization. The geometry tells you that the hard part (avoiding bad local optima across a thousand-ingredient palette) is solved by convexity, and the residual nonlinear tuning is a low-dimensional local search a human can finish at the bench. [Theorem.] In the linear-mixing regime with positive-definite metric $M$ and convex admissible set $\mathcal{C}$, the inverse flavor problem has a unique global minimizer, and a formula achieves zero perceptual distance if and only if $\mathbf{f}^\star$ lies in the convex hull $\{B\mathbf{x} : \mathbf{x}\in\mathcal{C}\}$. The proof is immediate from strict convexity of the metric objective over a convex feasible set; the content is the reachability criterion, which converts “can we make this flavor?” into a membership test.
13.5 Personalized Flavor
The final design problem changes not the target or the palette but the metric — and in doing so it closes a loop opened in Chapter 2. We established there (§2.5) that individual genetic variation acts as a near-identity rotation $R$ of the receptor basis: a person’s receptor vector for a stimulus is $\mathbf{r}' = R\,\mathbf{r}$, so the PAV/AVI bitter polymorphism, the OR6A2 cilantro receptor, the OR7D4 androstenone receptor, each turn one palate’s frame relative to another’s. Two people taste the same dish at points related by a known rotation. Personalized flavor design is the reverse operation: knowing a person’s rotation, pre-rotate the dish so that it lands, in their rotated frame, where the designer intended.
[Modeling Axiom.] Personalization inverts the perceptual rotation. Let $\mathbf{f}^\star$ be the target percept the designer wants the individual to experience, and let $R$ be that individual’s receptor rotation relative to the reference basis (estimable from a genotype or a short psychophysical calibration). The reference-frame flavor that the individual will perceive as $\mathbf{f}^\star$ is the pre-image $R^{-1}\mathbf{f}^\star$. The personalized design problem is therefore the inverse problem of §13.4 with the target replaced by $R^{-1}\mathbf{f}^\star$: solve for a formula whose reference-basis flavor is $R^{-1}\mathbf{f}^\star$, so that after the individual’s frame rotates it by $R$, it arrives at the intended $\mathbf{f}^\star$. Personalization is design composed with the inverse of the palate’s rotation.
Concretely, and as Chapter 2 promised, this is how you season for a specific palate. For a PAV/PAV supertaster (the TAS2R38 haplotype mapped by Kim et al. 2003; “supertaster” is Bartoshuk’s term) whose bitter axis is rotated toward high gain, the dish should be designed with the bitter coordinate attenuated in the reference frame — less of the kale, more compensating fat and salt — so that after their high-gain bitter rotation it reads as the intended pleasant-bitter rather than punishing. For an AVI non-taster, the same target percept requires the opposite pre-rotation: more bitter character in the reference frame, because their low-gain frame will shrink it. [Empirical.] The effective-dimensionality result of §2.6 adds a second adjustment beyond rotation: a supertaster’s higher effective rank means more axes clear their noise floor, so a formula tuned for them can carry more simultaneous components without muddying, whereas a non-taster’s lower-rank frame wants fewer, bolder dimensions. Personalization is thus a rotation and a rank-matching — turn the frame and match the resolution.
This is the geometric content of personalized nutrition (Chapter 11 §11.6) and the natural endpoint of the chapter’s arc, because it returns flavor design to a single eater. The mass-market flavor house of §13.1 solves the inverse problem against a population-averaged metric — a formula that reads as balanced across the distribution of palates, the robust-fixed-point problem of Chapter 6. Personalization replaces that averaged metric with one person’s, and the same convex QP, run with $R^{-1}\mathbf{f}^\star$ as target and the individual’s $M$ as metric, returns their formula. [Conjecture.] As genotyping and rapid psychophysical calibration cheapen, the binding constraint on personalized flavor will not be the optimization — that is a solved convex program — but the measurement of $R$ and $M$ for a given eater with enough fidelity that the pre-rotation improves on the population average rather than chasing estimation noise. The mathematics is ready; the sensing is the frontier, and it is the same frontier the series flags wherever a beautiful inverse problem waits on a hard forward measurement.
From the Bench to the Open Problems
This chapter took the descriptive machinery of the first twelve chapters and ran it backward, and the running-backward turned out to be a single convex program seen from five angles. The flavor house navigates a learned chart of a palette subspace (§13.1); masking projects out an unwanted dimension rather than burying it (§13.2); plant-meat mimicry seeks the nearest reachable point to an unreachable animal target and chooses which residual to keep (§13.3); the inverse problem is a metric-weighted quadratic program whose solution doubles as a reachability test (§13.4); and personalization composes that program with the inverse of an individual’s receptor rotation (§13.5). The common thread is the thread of the whole series: once the right space and the right metric are in hand, design stops being a combinatorial search over ingredient lists and becomes optimization in a few tens of dimensions — fast, diagnosable, and, in the convex regime, provably optimal.
But every design problem in this chapter assumed a static target: a point $\mathbf{f}^\star$ to reach. Real flavor is a curve in time — top notes, middle, base — and a formula that hits the target point at first bite may trace the wrong path through the seconds that follow, which is why the masked drug’s bitterness returns in the aftertaste and why the plant burger’s thin finish matters more than its formula sheet admits. The next chapter, on the open problems of the theory, takes up that temporal geometry directly, along with the context-dependence and satiety dynamics that turn the metric itself into a moving target. Connecting forward to Volume 6: Geometric Cognition, those same dynamics — a percept that unfolds, depends on what preceded it, and habituates as you attend to it — are not special to flavor; they are the signature of any perceptual manifold read by a brain in time. The geometry was always real. The scalars — a single “tastiness” score, a single bitterness number, a single match to beef — were always insufficient.