Google researchers train AI to recognise smells based on molecular structure

Google researchers used a data set of about 5,000 molecules, identified and described by perfume makers, as a benchmark for their experiment

Artificial intelligence researchers from the Google Brain Team claim to have trained a graph neural network (GNN) to predict smells of molecules based on their structures.

The findings of the study are detailed in an academic paper published on Arxiv [PDF], in which the researchers explain how they trained a bot through machine-learning algorithms to accurately assess the structure of various molecules and predict their smell based on the structures.

We hope machine learning can eventually do for olfaction what it has already done for vision and hearing

According to the researchers, they used a data set of about 5,000 molecules identified by perfume makers. The perfumers had also labelled those molecules with different smell descriptors, such as "grassy", "buttery", "pungent", "tropical", and "weedy".

Nearly two-thirds of the data set was used to train the Google GNN to link molecules with their smell descriptors.

The team used the remaining data set to test whether the GNN could accurately predict the smell of molecules based on their structures. They claim that the AI correctly predicted the smells of molecules in the majority of cases.

However, there are some caveats. The first is the subjectivity involved in describing the smells. For instance, two individuals may describe the same smell as "earthy" or "woody".

Another major hurdle is chiral pairs of molecules, which have same atoms and bonds, but are mirror images of each other. Such chiral pairs have completely different smells. Spearmint and caraway are one example.

There are other hurdles as well, for example, describing a new smell obtained from mixing two or three different scents.

Despite all those hurdles in accurately linking a smell to a molecule, the researchers believe their work could help in improving human understanding of sensory neuroscience and synthetic fragrance.

"Based on these early results with graph neural networks for molecular properties, we hope machine learning can eventually do for olfaction what it has already done for vision and hearing," the researchers wrote in their paper [PDF].

Google is not the only company doing research in this particular field. Earlier this year, scientists demonstrated at an exhibit at London's Barbican Centre the capabilities of their machine learning algorithms to recreate the smell of an extinct flower.

In Russia, AI-enabled robots are being used to detect potentially toxic gas mixtures. IBM is also carrying out research to advance the field of AI-generated perfumes.