The Shape of Smell

We have useful maps of color and sound. Wavelength gives us an ordering for spectral colors, while the eye's three cone types help explain their relationships and mixtures. Pitch is tied to frequency: double it and you move an octave up. Perception complicates both, but there is a physical structure we can work with.

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. It occurs in two forms with the same molecular formula and connectivity, but their three-dimensional arrangements are mirror images of each other, like a left hand and a right. One smells of spearmint, the other of caraway. A representation that records only which atoms are connected will treat them as identical, even though their spatial arrangement matters to the receptors in the nose. And it runs the other way too: molecules with quite different structures can have similar odors, while a small structural change can alter the character considerably.

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.

The challenge is to find which features of molecular structure predict the experience. Structure-based models can make useful predictions, including for previously uncharacterized odorants. But a model's representation has to preserve the relevant distinctions, and its prediction still concerns a particular task: descriptors, intensity, similarity, or pleasantness. A good answer to one is not automatically a complete account of how something smells.

If not structure, then what?

What other information can help us organize smell? One answer: 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.

More striking is that this usage-geometry is partly 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, built from different records of how materials are used) 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.

From molecules to materials

The second shift is in the unit itself. Much of machine olfaction works on single molecules. A perfumer also works with materials: bergamot oil, a rose absolute, an accord, a pure aromachemical. Some are complex mixtures; others are individual compounds. The material is the unit of creative, economic, and regulatory decision, and much of the practical problem 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

What the map can do depends on what went into it.

• Predicting an unfamiliar molecule's odor and recommending a known perfumery material are different tasks. This usage map is designed for the second: retrieving and recombining known materials, with their documented properties available for inspection.
• 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.

You can navigate the map: find a material's neighborhood, cross from a fragrance note to its culinary cousin, see which regions are dense and which are empty, without confusing a useful neighborhood with a complete description of its materials.

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. It is worth mapping without pretending we have mapped more than we have.