The created system turned out to be able to recognize odors based on the structure of molecules, but there is still a lot of work to be done.
For decades, perfumers and scientists have been trying to predict the relationship between the structure of a molecule and its smell. It is easy to find out what color the light will be by determining its wavelength. But in the case of odors, everything is much more complicated, it is impossible to predict how a molecule will smell according to its structure. However, researchers at the Google Brain Team are hopeful that artificial intelligence can change that. In an article published on Arxiv, they explain how they train AI to recognize smells.
The researchers created a data set of nearly 5,000 molecules identified by perfumers, who tagged the molecules with descriptions like "oily", "tropical" or "weak". About two-thirds of this data set was used to train AI based on neural networks, whose task was to identify the connections of molecules with the received descriptions. The researchers then used the remaining scents for testing. And it worked. The system was able to predict the smells of molecules based on their structures.
But it is still too early to celebrate victory, the science of smell is not so complicated by chance. Needless to say, two people can describe the same smell in different ways, such as "woody" or "earthy". Sometimes molecules have the same atoms and bonds, but due to their mirror arrangement, they have completely different smells. A typical example is the smells of cumin and mint. Things get even more complicated when you start combining scents.
However, Google researchers are confident that their work will be an important step in this direction, eventually having a significant impact on chemistry, physiology, neuroscience and other sciences. However, they are not alone, research in this direction is being conducted by many scientists around the world. So, British scientists recently tried to recreate with the help of AI the smell of a long-extinct flower; IBM is experimenting with perfumes created with the help of machine intelligence; and in Russia, they are trying to adapt AI to detect potentially lethal gas mixtures.