Researchers have opened the “black box” of artificial olfaction, using explainable AI (XAI) to visualise for the first time how an electronic nose actually identifies different smells.
The breakthrough, led by Japan’s National Institute for Materials Science (NIMS), advances AI-assisted sensors from simple classification to a more comprehensible process by revealing which receptor materials respond to specific molecules.
Artificial olfaction technology mimics the human sense of smell using chemical sensor arrays and AI to classify odours. However, its practical use has been limited by the discrimination accuracy of the sensors.
Previously, AI models could identify odours, but it was not understood how they were doing it or which sensor materials were most important for the task.
In the study, published in ACS Applied Materials & Interfaces, the team measured the responses of 94 different odorant molecules using a nanomechanical sensor equipped with 14 different receptor materials. They then used a Convolutional Neural Network (CNN) to classify the odours and an XAI technique called Score-CAM to visualise the CNN’s decision-making process.
“Importance maps”
The XAI successfully created “importance maps” showing which parts of the sensor signal the AI relied on to make its identifications. The analysis revealed that the key sensor responses varied depending on the molecule.
For example, the AI relied heavily on receptor materials containing aromatic rings when identifying aromatic molecules, confirming a clear structure-activity relationship. The CNN model achieved prediction accuracies exceeding 80 per cent for classification tasks, such as identifying molecules with or without oxygen atoms.
The researchers state this approach will enable the efficient selection and development of optimal receptor materials tailored for specific target odours. The findings are expected to accelerate the practical application of artificial olfaction in fields like food safety and medical diagnosis, and may also advance the understanding of human olfactory mechanisms.
