AI has been used in fragrance work in two documented ways: software that suggests perfume formulas from past formulas and sales data, and research models that predict how a single molecule smells from its structure. Symrise and IBM Research announced the first on October 24, 2018, and a team publishing in Science in September 2023 reported the second. Neither source shows a system that designs an ideal scent for one particular person, which is how the phrase is often used in marketing.
Key takeaways
- Symrise and IBM Research described Philyra on October 24, 2018 as a tool that proposes formulas for perfumers to refine, with a first project of two fragrances for a Brazilian brand planned for 2019.
- A 2023 Science paper trained a graph neural network on about 5,000 molecules and rated odor quality on 55 labels, and reported that it matched an average panel rating better than the median panelist did.
- The odor map predicts the smell of single molecules by their structure. It does not include concentration, so it cannot predict how strong a smell is.
- A fragrance made with AI still follows ordinary cosmetic rules, including Health Canada's fragrance allergen labeling, which began on April 12, 2026.
Two different jobs for AI in perfumery
A perfume is a blend of dozens of materials, and AI can enter at different points. One point is the formula: given thousands of past recipes, a model can suggest new combinations. Another is the molecule: given a chemical structure, a model can predict what a human nose would report. The two tasks use different data and answer different questions, and the table below sets them side by side.
| Approach | Input | Output | Not covered by the source |
|---|---|---|---|
| Philyra (Symrise and IBM Research, announced 2018) | Formula data bank, fragrance family data, historical data | Suggested formulas and complementary components | Any published test of how wearers rated the results |
| Principal odor map (Science, 2023) | Molecular structure | Ratings on 55 odor labels for one molecule | Blends, intensity at a given concentration, one person's taste |
What Philyra was announced to do
In the October 24, 2018 announcement, Symrise said it was developing, with IBM Research, a method that uses AI to create perfumes from digital fragrance models. The system draws on a large data bank of fragrance formulas, data on fragrance families such as fruity, oriental, and flowery, and historical data, and it looks for patterns that predict novel combinations. Symrise described the perfumer as working with an "AI apprentice," and said its first project, two fragrances for the Brazilian company O Boticário, would reach the market in 2019.
The announcement came from the company that sells the service, so it is a statement of intent and design rather than an independent evaluation. It says fragrance design demands more precision than other IBM models because small deviations in amounts can ruin a perfume, and it describes a perfumer adjusting the AI's suggestions by hand to emphasize a note and improve how long the scent lasts on skin.
The odor map and what it measured
The research paper, "A principal odor map unifies diverse tasks in olfactory perception," appeared in Science, volume 381, issue 6661, pages 999 to 1006, in September 2023, with Brian K. Lee as first author. The authors built a graph neural network on a reference set of about 5,000 molecules, each labeled for odor, drawn from the GoodScents and Leffingwell flavor and fragrance databases. The layer just before the output serves as the odor map, a coordinate system in which molecules that smell alike sit close together.
The test was prospective: a trained panel rated 400 new odorants, and the model's odor profile came closer to the panel average than the median individual panelist's did. The preprint version of the paper adds details. Panelists rated each molecule on 55 odor labels, 80 of the 400 molecules were dropped from the final evaluation (for low intensity, contamination, and other reasons), and the model beat the median panelist on 53 percent of the remaining molecules.
Limits the authors stated
The preprint lists several practical limits. Concentration changes how a molecule smells, but it is not part of the map, so the map cannot predict how strong a smell will be above the detection threshold. The strongest results were for organic molecules, and the authors did not extend predictions to halides or molecules with novel elements because of missing safety data. High accuracy is expected only for chemical motifs represented in the training set.
The comparison target also matters. The model was scored against the average of a trained panel, not against a given individual. Panelists were invited to join only if their own ratings were repeatable, so the benchmark describes a consensus nose. A model that predicts consensus ratings for single molecules does not by itself predict which finished blend a particular person will like.
Rules that still apply to an AI-assisted scent
In Canada, Health Canada's industry guide for the labelling of cosmetics says that, as of April 12, 2026, new and existing cosmetics must disclose 24 fragrance allergens when they exceed 0.01 percent in rinse-off products or 0.001 percent in leave-on products, listed outside the general term "parfum." The guide adds that new cosmetics must disclose fragrance allergens from a second list from August 1, 2026, and existing products from August 1, 2028. The method used to create a formula does not change these duties.
In the United States, the Food and Drug Administration lists fragrance allergen labeling among the regulations the Modernization of Cosmetics Regulation Act of 2022 requires it to establish. The agency's page, current as of August 10, 2026, gives no rule status for that item.
This article gives general information. It is not medical or safety advice, and a person with fragrance allergies or skin reactions should consult a health professional and read product labels.
FAQ
Can AI make a perfume for one person?
The sources reviewed here do not document it. They describe formula suggestion for perfumers and odor prediction for single molecules, and neither includes a published test of individual wearers.
Did AI replace perfumers?
Symrise's 2018 announcement describes the system as an apprentice that proposes formulas, with the perfumer making final adjustments. That is the company's own account.
Why does an odor model need human panels?
Odor quality is described in words by people, so the researchers scored the model against ratings from a trained panel. According to the preprint, each molecule was rated by at least 15 panelists, with two replicates.
Are AI-designed fragrances regulated differently?
No difference appears in the rules checked here. Ingredient and labeling duties attach to the product, whatever tool produced the formula.
Related reading
- How vectors represent data in AI
- Whether AI can be creative
- How AI programs analyze large data sets
- The math behind AI models
More in AI Basics.
Last verified: October 2026
Sources
- Breaking new fragrance ground with artificial intelligence: IBM Research and Symrise are working together, Symrise, October 24, 2018
- A principal odor map unifies diverse tasks in human olfactory perception (preprint), Lee et al., bioRxiv, 2022
- A principal odor map unifies diverse tasks in olfactory perception, Lee et al., Science 381(6661), 2023
- Industry guide for the labelling of cosmetics, Health Canada, 2026
- Modernization of Cosmetics Regulation Act of 2022 (MoCRA), U.S. Food and Drug Administration, content current as of August 10, 2026