Artificial Intelligence could assist in preserving historical scents that are in danger of disappearing

Artificial intelligence can assemble formulas to recreate perfumes based on their chemical composition. One day, a single sample may be used to recreate rare scents that are at risk of being lost, such as incense from culturally specific rituals or forest scents that change as temperatures rise.

Idelfonso Nogueira Researchers at the Norwegian University of Science and Technology profiled two existing fragrances and determined their scent families (subjective words such as “spicy” and “musky” commonly used to describe perfumes); They classified them by a so-called “odor value” scale. About how strong certain smells are. For example, one of our fragrances received the highest odor value for ‘coumarin’, a group of scents similar to vanilla. The other received the highest odor value for the scent family “alcohol.”

To train the neural network, the researchers used a database of known molecules associated with specific fragrance notes. The AI ​​learned how to generate a set of molecules that match the odor score of each scent family in the sample fragrance.

But simply producing those molecules isn’t enough to recreate the desired scent, Nogueira says. That’s because the way we perceive smells is influenced by the physical and chemical processes that molecules go through when they interact with the air and skin. Immediately after spraying, the “top note” of a perfume is most noticeable, but it disappears within minutes as the molecules evaporate, and the “base note” can remain for several days. To address this, the team selected molecules produced by AI that evaporate under conditions similar to the original fragrance.

Finally, they again used AI to minimize the discrepancy between the odor value of the original mixture and the odor value of the AI-generated mixture. Their ultimate recipe for one of the fragrances showed a slight deviation regarding its “coumarin” and “sharp” notes, but the other appeared to be a very accurate replica.

Predicting the smell of chemicals is notoriously difficult, so the researchers used a limited number of molecules in their training data. But the process could become even more accurate if the database could be expanded to include more, more complex molecules, Nogueira says. He suggests that the perfume industry could use his AI to create recipes that create cheaper, more sustainable versions of fragrances.

richard gerkin Arizona State University and OsmoThe startup, which aims to teach computers how to generate smells the way AI does for images, says that combining AI with physics and chemistry is the strength of this approach, and that it understands how smells are generated. He says that this is because it can explain subtle points that are often overlooked, such as whether the image evaporates into water. But the effectiveness of this process still needs to be confirmed in human studies, he says.

Nogelia and his colleagues are already almost there. In a few weeks, he plans to travel to his colleague’s lab in Ljubljana, Slovenia, to experience the AI-generated scents for himself. “I’m really looking forward to smelling it,” he says.