Of the five senses, we've already seen significant advancements in sight and hearing through artificial intelligence. But what about smell?

One intriguing company, led by Alex Wiltschko, is making waves by attempting to give computers a sense of smell. This article delves into how AI is transforming the way computers understand and interpret scents, and how this groundbreaking technology can benefit us. 

The Obsession With Smell

Alex Wiltschko, the CEO and co-founder of Osmo, has always been fascinated by the sense of smell. His lifelong passion for understanding olfaction led him to earn a bachelor's degree in neuroscience from the University of Michigan and subsequently, a Ph.D. in olfactory neuroscience from Harvard University. His academic journey was a prelude to his groundbreaking work at Google Research, where he spearheaded efforts to use machine learning for scent prediction.

The Genesis of Osmo

In 2022, after five fruitful years at Google Research, Wiltschko co-founded Osmo, a startup that leverages artificial intelligence to digitize scent. With backing from Lux Capital and Google Ventures, Osmo aims to improve human health and happiness by enabling computers to “generate smells like we generate images and sounds,” according to its mission statement.

Why Give Computers the Ability to Smell?

Medical Advancements

One of the most compelling reasons for digitizing smell is its potential in medical diagnostics. Smell carries crucial information that could help in early disease detection. Wiltschko mentions, “We’ve known that smell contains information we can use to detect disease. But computers can’t speak that language and can’t interpret that data yet.” This breakthrough could revolutionize medical diagnostics, making early detection of diseases like cancer more feasible.

Safer Consumer Products

In the near term, Osmo aims to create safer, more sustainable aroma molecules for everyday products like perfumes, shampoos, and insect repellents. These products traditionally rely on fragrances developed by a small number of secretive companies. Osmo seeks to offer better, safer alternatives that are non-toxic and non-irritating.

Developing the Principal Odor Map

The Challenge

During his tenure at Google, Wiltschko’s team faced the daunting task of developing a "principal odor map" through machine learning. Molecules are notoriously complex, and a slight alteration in their structure can result in drastically different scents. For instance, moving one bond in a molecule can change its smell from that of roses to rotten eggs.

The Solution

By training their AI model on a dataset of 5,000 aroma molecules across various odor categories (such as floral, fruity, or minty), the team enabled the model to identify patterns in the molecular structures. These patterns allowed the AI to accurately predict the odors of other molecules, achieving what Wiltschko describes as "superhuman" predictive capabilities.

Building Osmo’s AI Model from Scratch

The Data Dilemma

Unlike AI chatbots that can be trained on expansive data from the internet, a similar digital library for scents didn't exist. This necessitated the creation of a new kind of data. Initially, partnerships with fragrance industry companies were explored, but their datasets proved inadequate.

Creating New Data

Osmo acquired thousands of molecules and had master perfumers describe their scents. This data was then fed into graph neural networks (GNNs), sophisticated algorithms that analyze relationships between data points. Think of it as a social network where people are connected through friendships, but in this case, the nodes and connections represent atoms and bonds within molecules.

GNNs in Action

The GNNs helped Osmo’s AI model understand molecular structures and their corresponding odors. This pioneering effort was crucial for developing a reliable AI system that can interpret and generate scents.

The Future of Osmo