The Shape of Smell

We have good theories of color and sound. Light is a wavelength: longer is redder, shorter is bluer, and the eye's three cone types let you place almost any color on a tidy three-dimensional solid. Pitch is a frequency: double it and you land on the same note an octave up. In both senses the physics of the stimulus maps cleanly onto the structure of the experience — you can do arithmetic on it, and the arithmetic predicts the percept.

Smell has no such map. A molecule's odor is not its weight, not its size, not any single number you can read off its structure.

Take carvone. There is one carvone — one formula, one connectivity — but it occurs in two forms that are mirror images of each other, like a left hand and a right. By every flat description a chemist would write down they are identical. They are not identical to the nose: one form smells unmistakably of spearmint, the other of caraway. Same atoms, same bonds, same everything you can draw — two different smells. And it runs the other way too: molecules that look nothing alike on paper can land on nearly the same odor, while moving a single atom on a pleasant, rose-like molecule can tip it into something sulphurous.

This isn't for lack of effort; it's a property of the system. The nose carries roughly 400 receptor types (vision has three). Each responds to many molecules, each molecule trips many receptors, and the brain reads the combination — a 400-way chord, not a single note. There is no clean axis to ride. Odor is high-dimensional and stubbornly combinatorial.

So here is the uncomfortable fact at the center of the field: molecular structure underdetermines odor. You can train a model to predict a molecule's smell-words from its structure and it will do okay — genuinely useful, well above chance — but it hits a ceiling, and the ceiling is real, not a failure of engineering. The carvones are that ceiling made vivid: no structure-only model can ever separate two things it sees as the same molecule. Part of what a thing smells like simply isn't in the picture.

If not structure, then what?

If structure doesn't pin down smell, what carries information about it? One answer I find quietly profound: how things are used together.

A perfumer reaches for bergamot, neroli, a thread of musk; a cook pairs tomato with basil and olive oil. These pairings aren't random. Across millions of perfumes and millions of recipes, materials that go together — that recur in the same accords, that show up in the same dishes — trace out a structure. Embed each material by its company (the way you'd embed a word by the words it tends to appear near) and you get a geometry where things that smell and behave alike land close together. No receptor theory, no labels — just usage.

What's striking is that this usage-geometry is, in part, domain-general. Build a map of how fragrance materials are arranged by their usage, then a separate map of how food ingredients are arranged by theirs — two unrelated corpora, perfumery and cooking, with nothing shared in their construction — and for the molecules that live in both worlds, the maps agree. Take linalool, the soft floral-citrus note of lavender and bergamot. In the fragrance map it sits among the gentle florals a perfumer treats as nearly interchangeable; in the recipe-built map the same molecule sits beside coriander, basil, and citrus — the herbs a cook reaches for in the same gesture. Different worlds, no shared wiring, and the neighborhood travels with the molecule.

The agreement is modest — a shared relative ordering, not a rigid overlay — but it holds up across independent food datasets, including a peer-reviewed one. There seems to be a partly shared geometry of smell sitting underneath both the perfumer's palette and the kitchen. That's the kind of result that makes you sit up: it suggests the structure in odor space isn't an artifact of one industry's habits, but something closer to a property of how aromatic matter gets experienced — showing up wherever humans organize smells for use.

Nobody smells a molecule

There's a second shift worth naming. Almost all of machine olfaction works on single molecules. But nobody smells a molecule. A perfumer works with materials — bergamot oil, a rose accord, a captive isolate — each a small ecosystem of dozens or hundreds of compounds. The material is the unit of creative, economic, and regulatory decision, and most of the interesting structure lives there.

So do most of the real problems. The sharpest is substitution. Oakmoss — the damp, forest-floor backbone of the classic chypre — gets restricted because two of its constituents (atranol and chloroatranol) are potent allergens, and overnight thousands of formulas need something that holds the accord together without it. "Find me a compliant material that keeps the smell" is a material-level question, and the molecule-level maps don't answer it well. It's also the level that has been comparatively unmapped. Putting naturals, isolates, accords and their constituent molecules in one space — bridged to cuisine on one side and to the body's own volatile chemistry on the other — is most of what this project has been.

What's knowable, honestly

It would be easy to oversell this. I'd rather bound it.

• You cannot read odor off structure. Inverting the map to invent a brand-new molecule and predict its smell isn't honest — the underdetermination forbids it. The right tools retrieve and recombine known materials; they don't conjure new ones.
• The cross-domain agreement is real but modest — a shared relative ordering, not a precise isomorphism. It would be easy, and wrong, to draw it as a clean overlay.
• Usage-geometry encodes how things are used, which overlaps with but isn't the same as how things smell. Co-used is not co-smelling, and keeping that line honest matters more than the headline.

What you're left with is a map you can navigate: find a material's neighborhood, cross from a fragrance note to its culinary cousin, see which regions are dense and which are empty — while staying clear that the map is partial and the territory keeps some of its secrets.

That, to me, is the appeal. Smell is the sense that most resists being turned into a number, and yet it isn't structureless. There's a shape to it — emergent from use, partly shared across the worlds that use it, and still mostly uncharted. Worth mapping. And worth being honest about how much of it we've actually mapped.