AI described smells better than the typical trained panelist

Researchers at Google Research built a graph neural network trained on thousands of molecules labeled with odor descriptors in perfumery reference databases, then tested it on 400 new molecules rated by a trained human sensory panel.

On 400 new odorants, the model's descriptions, words like woody or fruity, matched the panel's average rating more closely than the typical individual panelist did. The same map also beat standard chemistry-based models on several other odor prediction tasks.

The map the model built placed chemically different molecules near each other when people perceived them as similar, capturing how odors are perceived rather than how chemists classify them. The system still trails the full panel's combined judgment, and describing a smell in words remains an imperfect stand-in for smelling it directly.

A machine out-described the people trained to do it.