Connect with us

Tech

Led by a former Google researcher, this company wants to give computers a sense of smell—here’s how it works

Published

on

Led by a former Google researcher, this company wants to give computers a sense of smell—here’s how it works

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

Alex Wiltschko is the CEO and co-founder of Osmo, a startup that uses artificial intelligence technology to help computers “generate smells like we generate images and sounds,” per the company’s website.

Wiltschko has been “obsessed with smell” for a long time, he tells CNBC Make It. “It’s been my passion to try to understand smell. It’s this very powerful emotional sense, yet we know so little about it.”

That’s why he earned a bachelor’s degree in neuroscience from the University of Michigan and studied olfactory neuroscience at Harvard University, where he earned his Ph.D. in 2016.

The next year, he went on to become a research scientist at Google Research, where he spent five years leading a team that used machine learning to help computers predict how different molecules smell based on their structure.

While Osmo began as a research project during Wiltschko’s days at Google, he went on to co-launch it as a separate startup in 2022 with support from Lux Capital and Google Ventures.

As Osmo’s CEO, he says the startup’s mission is to “improve human health and happiness” by digitizing humans’ sense of smell.

Here’s why Wiltschko believes humans can benefit from giving computers the ability to process scent, how Osmo developed its unique technology and what he hopes the technology will be able to achieve in the future.

How AI scent detection and creation can help humans

The big question is, Why give computers the ability to smell at all? One of the main reasons Wiltschko cites is that it’s critical in helping medical professionals to detect diseases.

“We’ve known that smell contains information we can use to detect disease,” he says. “But computers can’t speak that language and can’t interpret that data yet.”

DON’T MISS: How to be more successful with your money

While that’s his long-term goal for the company, in the near term he wants Osmo to make safer, more sustainable aroma molecules for fragrances in everyday products like perfume, shampoo, insect repellant and laundry detergent.

“Those products will usually have fragrance in them designed by a very small number of secretive companies,” Wiltschko says. “We think we can do better with them by building better and safer ingredients that aren’t toxic … and don’t irritate your skin or your eyes.”

How Osmo uses AI to digitize scent

During his time at Google Research, Wiltschko’s team used machine learning software to develop a “principal odor map.” To do this, his team trained their AI model on a dataset of 5,000 aroma molecules across various odor categories such as floral, fruity or minty.

Osmo used thousands of aroma molecules to help train its AI model to accurately predict how a molecule would smell based on its structure.

Osmo

Wiltschko found that molecules can be tricky for computers to analyze due to their complex structures.

“The reason why it’s so challenging is because you can move one tiny thing around in that molecule, like one bond, and the scent of the molecule goes from roses to rotten egg,” he says.

But thanks to advances in AI technology, the model was able to pick up on patterns in the different structures of the molecules and use that knowledge to accurately predict the odor of other molecules.

“It was superhuman in its ability to predict what things smelled like,” he says.

Building Osmo’s AI model from the ground up

While large language models, known as AI chatbots, can be trained on data from “the entire internet,” a similar digital library of information on scents wasn’t as readily available when they began building their AI model, Wiltschko says.

“The one thing we realized was that we couldn’t use anybody else’s data,” he says. “We actually spent about a year working with companies in the fragrance industry that had what they thought were great datasets, which we found were not.”

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

They obtained thousands of molecules and descriptions of their scents according to master perfumers. They then fed that data into graph neural networks (GNNs), which fall under the umbrella of machine learning and use powerful algorithms to detect 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.

Wiltchko’s team could then use the GNNs to help their AI model understand atoms, the bonds that connect them and how that molecular structure determines its odor, he says.

What’s next for Osmo

Continue Reading