The machine sits in a white room above a lavender distillery in Grasse, and it has never smelled anything in its life. It is an automated compounding rig, a gleaming frame of glass vials and fine steel needles, connected to a model trained on more than two hundred thousand perfume formulas collected over fifty years. Given a brief, perhaps a fig tree after rain or a cold leather glove, it proposes a formula. The rig then weighs and blends a sample in about four minutes. A human perfumer smells it and decides.
The system belongs to Brume Analytica, a fragrance-technology company founded in 2021 by a chemist and a former search engineer. It sells its services to independent perfume houses that cannot afford a full team of noses. In the past year, around thirty fragrances developed with the model have reached the market, several of them quietly successful. The company is careful to describe its product as an assistant rather than an author, and careful, too, about who gets the credit.
A nose for probability
What the model does well is explore. A human perfumer, trained over a decade, develops habits and preferences, favourite materials and familiar accords. The machine has none of these. It will suggest pairing a green galbanum note with smoky birch tar and a trace of salted caramel, combinations a person might dismiss before trying. Most of its proposals are dreadful. Perhaps one in forty is interesting. Occasionally one is remarkable, and those are the ones that end up in bottles.
Hélène Castagnier-Roux, a perfumer with thirty years in the industry who now works alongside the system three days a week, describes the experience as like working with a brilliant, tireless student who has read every book and never been outside. She rejects most of what it produces, reworks the rest by hand and considers the final fragrances entirely her own. Her younger colleagues, she notes, are more relaxed about sharing authorship with software.
It does not know what a memory is. It only knows what other people have bottled. My job is to bring the memory back in. — Hélène Castagnier-Roux
The traditional houses are uneasy. Perfumery has long guarded its mystique, the idea of the solitary nose composing in a quiet room, and the arrival of algorithms threatens both the romance and the economics. Critics argue that a model trained on past successes will inevitably produce variations on what already sells, flattening the art into statistically safe choices. Supporters respond that the same accusation has been levelled at human perfumers for decades.
For now, the most interesting results are the ones nobody expected. One of Brume Analytica's recent formulas, a strange, cool accord of wet stone, cucumber skin and old books, was rejected by its commissioning client and later adopted by a small house in Kyoto, where it has become a cult favourite. Its label credits both the perfumer and the machine. Customers, it seems, do not mind at all.
