Skip to main content
PHOTOGRAPH: GETTY IMAGES

This Startup Is Using AI to Unearth New Smells

Google Research spinout Osmo wants to find substitutes for hard-to-source aromas. The tech could inspire new perfumes—and help combat mosquito-borne diseases.

ALEX WILTSCHKO OPENS a black plastic suitcase and pulls out about 60 glass vials. Each contains a different scent. One smells starchy with soft floral notes, like jasmine rice cooking. Another brings to mind ocean air and the white rind of a watermelon. One is like saffron with hints of leather and black tea. The next is the pungent aroma of fig leaves, boxwood, and basil. The most surprising one has the tang of a Thai chili pepper without the nostril-burning heat.

The molecules wafting into my nose are nothing like I’ve ever smelled before. In fact, I’m one of only a handful of people who have ever smelled them. And yet, before any person had sniffed them, a computer model predicted how they’d smell to us.

Wiltschko has been obsessed with scents since he was a teenager, and for the past several years he has been developing software at Google Research to predict the scent of molecules based on their structure alone. The vials he’s invited me to smell are the basis of his new startup, Osmo, a spinout of Google Research based in Cambridge, Massachusetts. With $60 million in an initial funding round led by New York-based Lux Capital and GV (Google Ventures), Osmo aims to create the next generation of aroma molecules for perfumes, shampoos, lotions, candles, and other everyday products.

The $30 billion global fragrance industry relies on raw ingredients that are becoming increasingly difficult or controversial to source. Supplies of flowers popular in perfumery are dwindling because of extreme weather driven by climate change. Species like sandalwood trees are endangered from overharvesting. Other ingredients, like saffron or vetiver, are vulnerable to supply chain disruptions due to geopolitical turmoil. Some brands still use musk and other odors sourced from animals, which presents ethical issues, since it means they must be captured or killed. Meanwhile, some synthetic alternatives, such as lilial, which smells like lily of the valley, are facing regulatory bans for safety reasons.

Chemists at fragrance companies have figured out how to replicate some natural aromas, but it’s still a largely manual process, and many scents don’t have synthetic substitutes. “We need to be building replacements. Otherwise, we’re going to have to continue to harvest these plants and animals from our ecosystem,” says Wiltschko, cofounder and CEO of Osmo, who headed the digital olfaction team while he was at Google Research. “There’s a huge opportunity to build safe and sustainable and renewable ingredients that don’t require that we harvest life.

In the near term, the company wants to design molecules for the flavor and fragrance industry that are potent, allergen-free, and biodegradable. “We see Osmo as a rational design business model where people want a very specific odor and we design the chemicals, just like you would design a drug in a biotech or pharma company and then be able to license those,” says Josh Wolfe, a managing partner at Lux Capital and cofounder of Osmo. In the long term, the company wants to give computers a sense of smell—to “digitize” scent—although that concept is less far along and faces some uphill technical challenges.

The olfactory system isn’t as well understood as our other senses, but that’s because it’s arguably more complex, says Joel Mainland, an olfactory neuroscientist at the Monell Chemical Senses Center in Philadelphia who collaborated with Wiltschko’s olfaction team at Google Research but isn’t involved in Osmo.

The ability to detect smells—baking bread, grass after rain, cigarette smoke, or your grandmother’s perfume—starts when these scent molecules float through the air, enter your nose, and bind to odor receptors, which relay information to the brain through the olfactory nerve. The human nose has about 400 types of receptors, or special sensor proteins. By comparison, the eye uses three types to produce vision, and we taste with about 40 kinds of receptors.

That complexity makes it harder to categorize scents than other perceptual experiences. Color can be represented with a gradient known as a color wheel, and sounds by the frequency of their waves. Nothing similar exists for odors. “Right now, we need some way to understand how odors are related to each other,” Mainland says. “We don’t have a good way to organize smells.”

So Wiltschko’s Google team worked to build what they call an “odor map”—a way of categorizing scents so molecules that smell alike are clustered together. But instead of relying on human noses to make these distinctions, they used artificial intelligence.

They began by feeding machine-learning software a data set of 5,000 scent molecules available from fragrance catalogs—all odors that have been commonly used and are well described. For example, is the scent fruity, buttery, woody? From this training set, the software began to note associations between the chemical structure of each odor molecule and how a human would describe it, building out a high-dimensional map of odors that grouped molecules based on these characteristics. “It sounds like a simple problem, but little tiny changes in a molecule’s structure can move it from smelling like roses to rotten eggs,” Wiltschko says. For instance, the chemical bonds or number of carbon atoms in a molecule can affect its odor.

Then they gave the software a more mysterious data set to parse: 400 molecules that had been designed by scientists but never produced, so their odors remained undescribed. They asked the model to predict what each molecule would smell like to people—based solely on its structure.

To test how well these predictions stacked up, Mainland and his colleagues at Monell asked a panel of 15 volunteers to sniff each odor and assign it labels: floral, minty, smoky, and so on. The panelists didn’t always agree with each other; olfaction is more subjective than many other senses. But for 53 percent of the scents, the model’s predictions were closer to the panel average than to any one volunteer.

The team considered that a success, although the system has some limitations, says Wiltschko. For instance, two molecules can be mirror images of each other but smell different. “The smells aren’t always radically different, but they are subtly different, and our neural network is completely blind to that,” he says.

The team posted its findings to the preprint server bioRxiv in September, and the paper is currently being peer-reviewed at a scientific journal.

“One thing we want to do in olfactory science is understand how it is that humans perceive odors,” says Krishnan Padmanabhan, an olfactory neuroscientist at the University of Rochester School of Medicine who isn’t involved in Osmo. He says the group’s odor map points to a way to do that. “It’s really striking what they were able to accomplish.”

The glass vials Wiltschko had me smell contained the same scents those Monell panelists had sniffed. He says Osmo is in active talks with several fragrance companies to license some of them.

Some novel scents are more commercially viable than others, says Christophe Laudamiel, a French master perfumer who serves as an adviser to the company, and who guided me over Zoom as I smelled the different odors. (There are just 600 perfumers in the world, according to New York-based International Flavors & Fragrances, one of the major companies that concocts new scents.) For example, there are very few molecules available that smell like the ocean, Laudamiel says, so a new sea scent would be highly desirable. He’s not sure how the fragrance industry would use the chili pepper one, but he could see it being used for food flavoring.

“The industry is very small, and there are only a few companies that have embarked on finding new molecules,” he says. “It takes a lot of serendipity to find a new molecule with a new scent.”

And the failure rate is high. Not only do those molecules have to smell good, they also have to be safe and biodegradable. Companies might test a thousand molecules a year just to bring a few to market that check all these boxes. When Wiltschko sent him the molecules Osmo had created, Laudamiel said: “You realize you’ve created an alternate universe of perfume ingredients.”

THERE IS ANOTHER problem Wiltschko thinks Osmo’s technology can solve: creating a better mosquito repellent.

 Mosquito-borne diseases like malaria and dengue fever are responsible for more than 700,000 deaths annually, according to the World Health Organization. Female mosquitoes feed on human blood and are attracted to the smell of skin. Most chemical repellents, including DEET, which is considered the gold standard, work by confusing mosquitoes’ olfactory signals, preventing them from finding their next target.

But DEET has some drawbacks. It has to be used at high concentrations, it can degrade plastic, and it can cause skin irritation. It’s also possible that mosquitoes could develop resistance to DEET, as they have to other chemicals, says Chris Potter, a neuroscientist at Johns Hopkins who studies the mosquito olfactory system. “I think there is a good reason to look for additional repellents,” says Potter, who isn’t involved with Osmo. “We always need to have a backup.”

In 2020, the Environmental Protection Agency approved the first new repellent in 11 years—a naturally occurring chemical called nootkatone, which gives grapefruit its characteristic scent. But Wiltschko and his team at Google thought they could use their machine-learning system to find new ones.

First, they needed a large data set of scent molecules so they could train their model to recognize the correlations between a compound’s structure and its effectiveness as a repellent. But they could only find a few dozen mosquito repellents described in recent scientific literature. So Wiltschko tracked down a US government report from the 1940s, when scientists tested around 19,000 compounds for their effectiveness. That effort ranked these compounds according to how well they worked and led to the discovery of DEET. Wiltschko and his team digitized the data set, then trained their algorithms on it.

As they had in the fragrance experiment, they provided their model with 400 novel molecules that had not been tested for their repellency. In this case, they asked the model to predict which would work, based only on each one’s chemical structure. From these, they chose 317 for screening with a standard lab test. It showed that more than 10 of them had a repellency similar to or greater than DEET and other chemicals currently in use.

The team published its findings in a preprint on bioRxiv, but the paper has not yet been peer-reviewed. Next, Wiltschko says Osmo plans to test those molecules for skin safety and biodegradability.

Potter is impressed with the team’s method. “It unleashed this data that we’ve been sitting on for so long,” he says. “Now we have this great list to work from. It’s worth taking a closer look at these chemicals.”

WILTSCHKO AND WOLFE see bespoke scent molecules and new repellents as just the beginning—they’re on a mission to give computers a sense of smell. They think AI can get us closer to digital olfaction by predicting what odors smell like and how they relate to other scents. “The long-term vision is ‘Shazam’ for smell,” as Wolfe puts it. Just as you can use an app to identify the song that’s playing on the radio, Wolfe thinks you should be able to capture, save, and transmit scents with your phone.

But that’s a difficult problem. While a phone is built to transmit sound, it’s not built to transmit chemicals. Such a device would have to collect scent molecules, convert them into a digital signature, and pass the signal to someone else’s phone or computer where it would be decoded. Then they’d need some kind of chemical-releasing device to convert that signal back into an inhalable scent.

And Osmo hasn’t yet given specifics about how it would approach digitizing smell, although Wiltschko has laid out the basic idea. “You need three parts: a sensor, a map, and a printer. The sensor takes the physical world and converts atoms to bits. The map helps you interpret, store, compress, and transmit the bits. In color, these are technologies like RGB and JPEG. Then, you need to be able to turn the bits back to atoms,” Wiltschko says. “We think the time is now to begin putting these all together.”

The company hasn’t built any sensors yet for capturing smells into digital signals—or devices to “print” smells, for that matter—but Wiltschko says they are collaborating with outside researchers to do so. Wiltschko calls the problem “ridiculously hard” and says it will take years.

In fact, people have been trying for decades. DigiScents iSmell, a USB-connected cartridge for desktop computers, launched in 1999. It was supposed to encode and then play back scent data collected online, and WIRED claimed it would “launch the next Web revolution.” But the company shut down in 2001 due to a lack of funding.

In 2014, Vapor Communications launched the oPhone, a device that connected to an iPhone or iPad to allow users to send scents with messages. In 2016, the company also introduced a “scent speaker” called Cyrano that allowed people to play sequences of scents, like a playlist for odors. Neither of those products is still on the market.

Most recently, tech startup Feelreal tried to incorporate scent into a virtual reality headset, but it ran into a regulatory snag with the Food and Drug Administration because the agency considered it a vaping product. The headset has yet to make it to consumers.

Why try to digitize smell at all? For Wiltschko and Wolfe, it’s because smells have the exceptional ability to trigger memories. “We have not been able to capture what is arguably evolutionarily our most salient sense, which is our sense of smell,” Wolfe says. “We evolved that sense to be able to ward off danger, detect loved ones, smell rotting food, and enjoy the beauty of the world, and those are difficult things to share with people unless somebody else is there.”

For now, they will start with trying to shake up the fragrance industry by recreating existing scents and unearthing new ones. As Wiltschko puts the scent vials back into the suitcase, I consider whether I’d want to wear any of the Osmo scents I sampled. One was rose with hyacinth and fresh greenery. My usual perfume, a rose scent made by a French brand, has been steadily increasing in price, and its aroma has changed over the years. Once, when I was in the store, I asked why: The issue was the rose supply chain.

Perhaps if I could find an alternative, I’d switch.

Source: This Startup Is Using AI to Unearth New Smells | WIRED