Ex-Google researcher’s startup uses AI to digitize smells

Of the five human senses, AI can already mimic sight and hearing, and one company wants to use the technology to digitize another one: smell.

Alex Wiltschko is the CEO and co-founder of Osmo, a startup that uses artificial intelligence technology to enable computers to “generate smells in the same way that humans generate images and sounds,” according to the company’s website.

Wiltschko has long been “obsessed with smell,” he told CNBC’s “Make It.” “My passion is trying to understand smell. Smell is a very powerful emotional sense, and yet we know so little about it.”

So he went on to earn a BA in Neuroscience from the University of Michigan and studied Olfactory Neuroscience at Harvard University, where he received his PhD in 2016.

The following year, he became a research scientist at Google Research, where he spent five years leading a team that used machine learning to help computers predict different smells based on the structure of molecules.

Osmo began as a research project during Wiltschko’s time at Google, but was co-launched as an independent startup in 2022 with backing from Lux Capital and Google Ventures.

Osmo’s CEO says his startup’s mission is to “improve human health and well-being” by digitizing the human sense of smell.

Below, Wiltschko explains why he thinks humanity would benefit from giving computers the ability to process smells, how Osmo developed its unique technology, and what he hopes it can achieve in the future.

The big question is, why give a computer a sense of smell? One of the main reasons Wiltschko gives is because the sense of smell is crucial to medical professionals detecting disease.

“We know that smell contains information that can be used to detect disease,” he says, “but computers can’t speak that language or interpret that data yet.”

Don’t Miss: How to be more successful with money

While that is his company’s long-term goal, in the short term he wants Osmo to be able to produce safer, more sustainable fragrance molecules to flavor everyday products like perfumes, shampoos, insect repellents and laundry detergents.

“These products typically contain fragrances developed by a small number of secretive companies,” Wiltschko says. “We believe we can make these products even better by developing better, safer ingredients that are non-toxic and won’t irritate the skin or eyes.”

While at Google Research, Wiltschko’s team used machine learning software to develop a “dominant odor map.” To do this, he and his team trained an AI model on a dataset of 5,000 aroma molecules across a range of odor categories, including floral, fruity, and minty.

Wiltschko found that molecules can be difficult to analyze with a computer because of their complex structures.

“The reason this is so difficult is that moving something small in a molecule, like one bond, can change the molecule’s smell from roses to rotten eggs,” he says.

However, advances in AI technology have enabled models to pick up patterns in different structures of molecules and use that knowledge to accurately predict the smell of other molecules.

“His ability to predict the smell of things was superhuman,” he says.

While large-scale language models known as AI chatbots can be trained on data from “the entire internet,” Wiltschko says that when he started building his AI models, a similar digital library of information about scents wasn’t readily available.

“What we realized was, we couldn’t use someone else’s data,” he says. “In fact, we worked with a company in the fragrance industry for about a year on what they thought was a great data set, but it wasn’t.”

That prompted Wiltschko and his team to build a “new kind of data,” he says.

They took thousands of molecules and their scent descriptions from master perfumers. They then fed that data into a Graph Neural Network (GNN), a branch of machine learning that uses powerful algorithms to discover and analyze relationships between data points. At this point in the process, think of a social network where you can see people and how they are connected by friendships.

Wiltschko’s team could use GNNs to help AI models understand atoms, the bonds that hold them together, and how that molecular structure determines smell, he says.

Ultimately, Osmo hopes to use its technology to digitize a scent in one place and recreate it exactly in another, making it possible to teleport scents, Wiltschko said.

“This is a way for the AI ​​model to prove to itself that it really understands what something smells like, because if it can’t recreate the exact smell of the original, it’s fooling itself,” he says.

Wiltschko also plans to continue working towards his long-term goal of using the technology to help detect disease earlier.

“Eventually we will be able to detect disease by smell, and we are developing that technology,” he says. “It won’t happen this year or anytime soon, but we are moving in that direction.”

Do you want to be more successful and confident with your money? take New online course from CNBC Make ItOur expert instructors will help you master your money and find practical strategies to increase savings, reduce debt, and grow assets in the way that works best for you.Achieve financial wellness: be happier, wealthier, and more financially secure“Start your journey to financial freedom today!” Use coupon code EARLYBIRD to get 30% off until September 2, 2024.

plus, Sign up for the CNBC Make It newsletter Learn tips and tricks for success in work, money and life.