Researchers at the Google Brain Team have trained an artificial intelligence (AI) system to recognise smells. The team started by creating a data set of almost 5000 molecules identified by perfumers and labelled with descriptors such as 'buttery', 'tropical' or 'weedy'. It then developed a system to translate the structure of molecules into graph representations. A large part of the data set was later used to train a graph neural network (GNN) to associate molecules with the descriptors they are often associated with. (GNNs are deep neural networks designed to operate on graphs (abstract data types) as input.) The rest of the data set was used to test the AI's ability to match molecules and descriptors of smells. According to Google, in addition to predicting odour descriptors, GNNs can also be applied to other olfaction tasks, such as classifying new or refined odour descriptors using only limited data. While noting that smell is the most elusive sense in the realm of machine learning, Google believes that its research opens the door to future work in this area, that could lead to designing new olfactory molecules that are cheaper, digitising scent, and maybe even allowing those without a sense a smell to finally enjoy this sense.