For sight, the colors we see relate to wavelengths of light; for sound, the pitch relates to the frequency of sound waves. The sense of smell does not have such a simple, mathematical explanation. A person sniffing a perfume can easily tell whether it smells floral, tropical, or earthy, but the chemists who make perfumes cannot determine what a new perfume will smell like based on its chemical composition alone. Some molecules with similar shapes smell different, while others with different shapes smell the same. This makes it challenging to design new pleasing scents for products.
A recent study has shown that artificial intelligence (AI) may be the key to solving this challenge. The researchers used an AI technique called graph neural networks to analyze a database of 5000 molecules with known scents, identifying the complex patterns that link molecular structure to odor. After this training, the researchers asked the AI to guess the scent of 400 new compounds based only on their chemical compositions. They also asked human participants to smell these compounds and compared their descriptors (like “smoky” or “powdery”) to the AI’s guess. For over half of the test compounds, the AI was better at identifying the scent than the average human participant.
The researchers have used their AI to create a database of scents for 500,000 new compounds that have never been smelled before, increasing the number of compounds with known scents by a factor of 100. These findings pave the way for the development of new, better-smelling consumer products, and are also an important step in understanding the biology of smell.
This study was led by Brian Lee, a Research Software Engineer at Google Brain, and Emily Mayhew, an Assistant Professor in the Department of Food Science and Human Nutrition at Michigan State University.
Managing Correspondent: Emily Pass
Press Article: AI rivals the human nose when it comes to naming smells (News from Science)
Original Journal Article: A principal odor map unifies diverse tasks in olfactory perception (Science)
Image Credit: Pexels/Pavel Danilyuk