Osmo says its artificial-intelligence models have reached “human-level olfactory intelligence”: they can detect and distinguish scents about as well as an average person and create fragrances at the level of a junior perfume designer. The startup is already working with the fragrance industry, and says its model recently outperformed an in-house perfumer at one of the world’s largest consumer-goods companies. Its founder, Alex Wiltschko, believes the technology could eventually support applications ranging from threat detection to disease diagnosis at airports.
Building an AI model for smell
Wiltschko spun out Osmo in 2023 from Google Brain, the search giant’s former artificial-intelligence research lab. The company is pursuing a strategy used by AI startups in other fields: feeding large amounts of data into a difficult problem and observing what the models can learn.
Smell presents a particular data challenge, Wiltschko says. Before Osmo, existing databases contained fewer than 200,000 human descriptions of smells, by his count. The chemical sensors needed to gather additional data are expensive, while the fragrance industry has traditionally been secretive about its methods for studying olfaction.
Much of Osmo’s work has focused on creating a dataset containing more than 10 million human labels and 800 million “odor tokens.” These tokens describe the specific molecules and concentrations that produce particular odors.
Osmo operates spectrometers and other chemical sensors around the clock, breaking down the molecular structure of materials ranging from fruits and flowers to food products and organs. About 70 trained data labelers in the United States and abroad smell samples and describe them using a standardized scent vocabulary.
Mapping the relationships between odors
Mapping scents in this way has revealed some unexpected similarities. Just as changing the order of words can substantially alter a sentence’s meaning, changing the proportions of scent molecules can significantly affect how people perceive an odor.
“The odor tokens that define the smell of a newborn baby overlap significantly with the odor tokens found on a plate of freshly cooked spaghetti,” Wiltschko says. He specifically referred to Barilla spaghetti, a detail that could raise questions for people working at the pasta company.
Printing custom fragrances
Osmo also creates scents using a scent-printing robot the size of a school bus. The machine mixes ingredients according to precise formulas to produce samples for the company’s research and fragrances for paying customers.
Customers can design and print scents to order using prompts similar to those used with ChatGPT. During a demonstration in Lisbon for the Web Summit, Wiltschko opened a text box on his laptop and was asked to generate a scent resembling a Portuguese custard tart for men aged 18 to 23.
After several minutes, the system produced three potential scent profiles matching the requested outline—creamy, coconut and fruity—along with the molecules needed to create them. Had the order been completed, the resulting signature fragrance, described as “l’eau de custard,” would have been ready in about a week.
Source: www.technologyreview.com











