Research

AI-generated text

Osmo builds AI models to recognize and design scents

Osmo is combining human scent descriptions with chemical measurements to train AI models that can identify and create odors.

Osmo builds AI models to recognize and design scents

In a laboratory near Newark Airport in New Jersey, around 30 experienced human sniffers describe the odors of different materials. Their impressions, ranging from nutty and peach-like to marine, are matched with chemical analyses to create training data for an artificial intelligence model developed by Osmo. Founder Alex Wiltschko’s goal is to give computers the ability to sense smells.

Wiltschko says the model has reached “human-level olfactory intelligence” in recent months. Osmo defines that as roughly matching an average person at recognizing and distinguishing odors, and a junior perfume designer at creating fragrances. The company also says that last month its model outperformed the in-house perfumers at a major consumer goods company.

Gathering data about smells

Wiltschko founded Osmo in 2023, building on his work at Google Brain, Google’s former artificial intelligence research laboratory. He says a shortage of suitable data was a major obstacle: by his count, existing databases previously contained fewer than 200,000 human descriptions of smells. Chemical sensors are expensive, and the fragrance industry shares little information about its own research.

Osmo therefore built its own dataset, with more than 10 million human labels and 800 million so-called scent tokens. The tokens describe the molecules and concentrations that produce particular odors. Using spectrometers and other chemical sensors running around the clock, the company examines the molecular makeup of items including fruit, flowers, food and organs. Around 70 trained data labelers in the United States and other countries describe samples using a standardized fragrance vocabulary.

Changing the proportions of molecules can substantially alter a perceived smell. Wiltschko says, for example, that there is considerable overlap between the tokens describing the smell of a newborn baby and those for a freshly cooked plate of Barilla spaghetti.

Designing scents and finding uses

Osmo also uses a school-bus-sized scent-printing robot to mix samples from precise formulas for research and paying customers. The system can design a fragrance from a ChatGPT-style text prompt. At a demonstration in Lisbon, it suggested three possible scent profiles and their molecular ingredients for a fragrance inspired by a Portuguese custard tart and aimed at men aged 18 to 23. A sample, if ordered, could have been ready in about a week.

The technology has already been used to create fragrances sold at retail, including for Procter & Gamble. The Gates Foundation invested $5 million when Osmo launched and later provided a further $3.5 million grant to assess molecules that repel disease-carrying insects. Osmo identified eight substances it found more repellent than DEET.

The company has also worked with the online clothing marketplace StockX on detecting counterfeit sneakers through smell-related differences in the chemical composition of shoe materials. Wiltschko cites airport threat detection and disease diagnosis as possible future applications.

Questions and remaining obstacles

Avery Gilbert, a smell scientist and author of What the Nose Knows, has argued that some of Osmo’s marketing claims are undocumented. He has also criticized the company’s patent filings for continuing to rely on human sensory panels and perfumers. An Osmo representative did not respond before the source article was published.

Michael Schmuker, a former professor of neural computation at the University of Hertfordshire who previously consulted for Osmo, regards the technology as an achievement. He and colleagues reverse-engineered an open-source version of the company’s 2023 model, but it lacks the extensive data Osmo has gathered since.

Schmuker cautions, however, that context is important to human olfaction: the same molecule can evoke a different association in a Camembert cheese than in a used diaper. That may matter less in a controlled fragrance laboratory, but it could be consequential in more complex applications.

For the technology to detect food spoilage on a production line or identify tumors by smell—as dogs have been trained to do in pilot studies—both sensors and models need to improve considerably, Schmuker says. Wiltschko acknowledges that medical diagnosis is an expensive and difficult goal; for now, Osmo is focusing on building the foundations and exploring potential applications.