Mapping smell: vector embeddings in perfumery

Colour has coordinates. Any shade you can see can be written as a position in a colour space, and two colours that look alike sit close together in it. Smell has never had that. We describe it with words, and the words are slippery: one person's "fresh" is another person's "soapy", and "woody" covers everything from pencil shavings to incense.

Over the last decade, researchers have started to give smell a map. The tool that makes it possible is the vector embedding, and it is one of the foundations of how we work at Atlas. This piece explains what an embedding is, what the published research has shown, and where the approach still falls short.

What an embedding is

An embedding turns something messy, like a word, an image or a molecule, into a list of numbers. That list is a point in a space with many dimensions. The trick is in how the numbers are learned: the model is trained so that things which behave alike end up near each other. Distance in the space starts to mean something. Two points close together should smell alike. A direction through the space can correspond to a quality, such as moving from green towards floral.

Once smell lives in a space like that, questions a perfumer answers by experience become questions you can also compute. What sits between these two materials? What is the nearest thing to this reference that costs less, or that passes a stricter regulatory limit? Which materials fill the gap in this accord?

From descriptors to structure

The early work tried to predict smell straight from chemistry. In 2013, Snitz and colleagues showed that the perceptual similarity of odour mixtures could be predicted from the structure of their molecules [1]. In 2017, the DREAM Olfaction Prediction Challenge brought together teams worldwide to predict how people rate odour molecules from chemical features alone, and the best models predicted perceptual attributes well enough to establish that structure carries real information about smell [2].

Those models learned from hand built chemical descriptors. The next step was to let a neural network learn its own representation. In 2019, Sanchez-Lengeling and colleagues trained graph neural networks, which read a molecule as a graph of atoms and bonds, to predict odour descriptors, and showed the learned representation carried over to related tasks [3].

A principal odour map

The clearest result so far came in 2023. Lee and colleagues trained a graph neural network on about 5,000 molecules labelled with 55 odour descriptors, drawn from the GoodScents and Leffingwell perfumery databases. The embedding it learned, which they called a principal odour map, was then tested prospectively on 320 molecules the model had never seen, rated by a trained human panel [4].

• The model's predictions were closer to the panel consensus than the median panelist for 53 percent of the molecules tested.
• The map preserved known relationships between smells, and generalised to other olfactory tasks it was not trained for.
• The authors used it to chart roughly 500,000 potential odour molecules that nobody had yet smelled.

The comparison they drew is telling: the map aims to do for smell what colour spaces do for sight.

Open data makes it possible

None of this works without data, and olfactory data has historically been scattered across papers, catalogues and private archives. Pyrfume, an open source project released under the MIT licence, gathers dozens of published olfactory datasets into consistent, machine readable form [5]. It is one of the large public resources we build on, alongside the public perfumery databases the research above relied on.

Mixtures are the hard part

A perfume is not a molecule. It is a mixture of dozens or hundreds of them, and mixtures do not behave like simple sums. Some materials mask others, some lift them, and the balance shifts as the fragrance evaporates. Research on embedding whole mixtures is newer. In 2025, Tom and colleagues extended the principal odour map to mixtures with a model called POMMix, which combines molecule level embeddings to predict how similar two blends will smell [6]. It is promising, and it is also a reminder of how much is still open.

How we use it

At Atlas, embeddings are one layer in a larger system that our own team has built and continues to develop. In broad terms:

• A brief, whether words, a reference fragrance or a mood, is placed into the embedding space.
• The materials that live in that neighbourhood become candidates.
• Our own mathematical models then balance those candidates against the constraints that make a fragrance real: cost, IFRA limits, how long each material lasts and how the blend will perform in the product it goes into.
• A person smells the result, edits it, and the outcome feeds back into the system.

The embedding makes the search fast and broad. The constraints keep it honest. The evaluation keeps it good.

Where it falls short

Smell is personal and cultural. The descriptors most datasets use were written largely from a Western vantage point, and many of the materials people across Southeast Asia know best are thinly represented. A map is only as good as the experiences it was drawn from, which is one reason we care about building for this region specifically.

An embedding also cannot tell you whether a fragrance is beautiful, whether it suits a brand, or whether a customer will buy it twice. Those remain human judgements. The value of the map is not that it replaces a perfumer. It is that it lets far more people get to the point where a perfumer's judgement matters.