Scientists develop artificial intelligence with ‘sense of smell’

A group of scientists, including a team from KU Leuven, has developed algorithms that are partly able to predict the odor of molecules based on their chemical structure. Of the nineteen scents used, the algorithms correctly predicted eight.

In a competition between 22 teams of computer scientists, the aim was to develop algorithms that are best able to accurately predict people’s odor perceptions. In this study, 49 people had to ‘label’ a total of 476 fragrances with one of the 19 different odor indications available, such as ‘fish’, ‘garlic’, ‘sweet’, ‘sour’ and ‘burnt’. The intensity of the fragrances were also assessed and how pleasant the smells were experienced by the test subjects. This writes the journal Science.

These results were used by computer scientists to create algorithms to predict how an average person would rate a scent. The scientists got two-thirds of the volunteers’ predictions, the chemical structure of the molecules used, and 4,800 descriptions of each molecule. With the latter one can think of the atoms and how they are arranged in the molecule. This resulted in a dataset with more than two million data points on the basis of which the algorithms were developed.

The Arizona State University team scored the best in predicting how all the volunteers together would rate the different scents on average. This team’s algorithms had the best prediction of how the human test subjects rated the fragrances; in eight of nineteen odor indications used, the models were right. Team leader Richard Gerkin believes that the developed models can help fragrance producers better give their perfumes a certain scent.

However, there is some criticism of this research. Avedry Gilbert, a biological psychologist from the American company Synesthetics and an experienced man in the fragrance industry, finds the new research especially useful because it is a large database. However, because the study was based on nineteen odor indications, Gilbert considers the study too limited. He points out that many more fragrance categories were used in similar studies.

The tricky part about predicting the odor of molecules is the fact that some molecules that are very similar chemically can be judged very differently by the human sense of smell. At the same time, molecules with very different chemical structures can be perceived by humans as almost identical in smell.