Vision and hearing have been digitised for decades. We have pixels for images and frequencies for sound. Smell has no such simple code. Two molecules that look almost identical on paper can smell completely different, while two very different molecules can smell nearly the same. That is why scent has been called the last frontier for artificial intelligence.
The problem: structure does not equal smell
A perfumer learns, over years, that a certain family of molecules tends to smell musky, or that a particular structure gives a green, cut-grass effect. But the rules have many exceptions. For a computer, the challenge is to learn those patterns from examples rather than rules.
How machine learning tackles it
Researchers train models on large datasets in which molecules are paired with descriptions written by trained human panels: words like fruity, woody, floral, sulphurous. The model learns which structural features predict which descriptors. In 2023, a team including researchers from Google and the start-up Osmo published work in the journal Science describing a “principal odour map” that predicted how molecules would be described, in many cases about as reliably as an individual trained panellist.
From single molecules to finished perfumes
Real perfumes contain dozens or hundreds of materials, and the way they interact is even harder to predict than single molecules. This is where a second kind of AI helps: models that work at the level of notes and accords rather than chemistry. By learning from thousands of documented perfumes, their note pyramids and the way people describe them, these models can say that one composition is “close to” another, even without simulating every molecule.
What this means in practice
- Discovery: matching a customer's favourite perfume to the closest available oil.
- Formulation support: suggesting materials a perfumer might explore to push a blend in a direction.
- Quality and consistency: spotting unusual deviations between batches.
Where Haveli AI fits
Our scent finder works at the note-and-accord level. It compares what you describe, or the perfume you name, with the profiles of more than 6000 oils in our library and ranks the closest matches. The final check is still done by a human nose on our perfumery bench in Mumbai. Try Haveli AI and see how close it gets to your favourite scent.