Google Researchers Successfully Create A Smell-Recognizing AI

For a long while, scientists and perfumers have put many efforts on figuring out how a molecule's structure and its unique scent relates to each other. In fact, they can identify a creature's color depending on the wavelength of light, but they still can't identify a scent based on just a specific molecule. However, some researchers from Google are hoping to change that with the help of the AI. According to a publication cited from Arxiv, Google Brain Team has explained the way they train AI to solve the amazing query.

At least 5,000 molecules that perfumers have identified, including "buttery", "tropical", "weedy" labels, are put together into a special data set. In this way, researchers made use of more than half the data set to teach their graph neural network then coupled every molecule with a descriptor. The rest data, in the meantime, is used for testing purposes. The AI model finally passed the test thanks to its algorithms that give predictions about smells while analyzing their structures.

Wired points out that some caveats cast doubt on the science of smell. For example, two starters might give different descriptions on the same scent, such as "earthy" and "woody." In addition, molecules sometimes share the same chemical bond, but they own completely different smells because they're simply arranged as mirror images. One example is the duo of spearmint and caraway, called a chiral pair. Even more, when scents are combined, things will get more complicated.

As for Google researchers, the stage of training AI to identify scents based on their molecules is a fundamental step ahead of all the others. So far, we can apply it more effectively on other fields like chemistry and our understanding of sensory neuroscience and human nutrition, especially when it comes to synthetic fragrance manufacturing.