Image Courtesy of Unsplash.
Scientists have long known that the chemical structure of a molecule influences its smell. However, it’s still unclear how tiny structural changes in a molecule’s structure can turn a sweet, delicate scent into a fishy stench.
Enter artificial intelligence (AI). Researchers from the start-up Osmo based in Cambridge, Massachusetts trained a type of AI system called a neural network to predict a compound’s odor based on its structure. The scientists instructed the system to assign descriptors from a list of fifty-five, such as “grassy” or “fruity,” to a scent. The AI system then generated an odor map by screening roughly five thousand well-studied molecules.
To test the validity of this map, a panel of fifteen trained study participants sniffed a set of compounds with undocumented scents. Their answers were averaged to account for genetic differences, personal experiences, and preferences. The researchers found that the AI system achieved results comparable to the human assessments for fifty-three percent of the molecules.
This new odor map could be a helpful reference tool when designing new scents in the food or perfume industries. But it doesn’t seem to reveal much about how humans interpret smell. Odor descriptors are quite subjective, and it’s unclear if averaging the answers of a group of people is the best way to obtain a “correct” description of a smell.
Most smells in the real world come from a mixture of compounds. The next frontier for the AI system may be to chemically describe the complex smoky smell of freshly brewed coffee or the aromatic sweetness of a new perfume.