According to a report in Science magazine on the 1st, a newly designed machine learning model by British and American scientists has reached a level comparable to human olfactory perception, capable of describing the odor of chemical substances in language. Researchers have used it to "depict" odor maps corresponding to hundreds of chemical structures, such as "fruit flavor" or "grass flavor". This guide map can help researchers design new synthetic odors and may provide new insights on how the human brain interprets odors, marking another step towards digitizing odors. To explore the connection between the structure of chemical substances and odors, Google Research Institute derived startup Osmo collaborated with the Monell Chemical Sensory Research Center in the United States, the University of Reading in the United Kingdom, and Arizona State University in the United States to design a neural network system that can match one or more descriptive words out of 55 with descriptions of odors. The team trained AI using industry datasets, which included the odors of approximately 5000 known odorants. AI also analyzed the chemical structure of each odor to determine the relationship between structure and aroma. The system identified the correlation between specific patterns in the structure of approximately 250 chemical substances and specific odors. Researchers combined this relevant information into the main odor map (POM). When AI predicts the odor of new molecules, it can refer to this figure. To compare the olfactory levels of POM and human noses, 15 human volunteers matched specific odors with the same set of descriptive vocabulary used by AI. Next, the researchers collected hundreds of odorants that do not exist in nature but are familiar and descriptive to humans. They asked volunteers to describe 323 molecules and AI to predict the odor of each new molecule based on its chemical structure. As a result, AI's guess is very close to the average response given by humans and closer to the correct answer. Specifically, the model performed better than the average of group members in 53% of the tested molecules. (New News Agency)