Google Brain Team's AI masters smell using 5,000 molecule database
AI now categorizes odors like 'earthy' and 'pungent' from molecular structure.
Google's Brain Team has achieved a significant breakthrough in artificial intelligence by training a machine learning model to 'smell' — accurately categorizing different odors by analyzing their molecular structure. Using a database of 5,000 molecules labeled by perfume makers with descriptors such as 'earthy' and 'pungent', researchers fed two-thirds of the data into a graph neural network. The AI then correctly identified the remaining scents, overcoming long-standing challenges like the subjectivity of human odor descriptions and molecular mirror-image isomers that smell differently despite similar structures.
The team believes this work, published on the Google AI Blog, brings machine learning for olfaction to the same level as advances already seen in vision and hearing. The model could help discover new synthetic odorants, reducing the ecological impact of harvesting natural products, and may provide new insights into the biology of smell. Competitors like IBM are also pursuing smell-enabled AI for applications like gas leak detection and perfume creation. This milestone opens the door for AI to interpret and generate smells with the same precision it now applies to images and sounds.
- Google used a database of 5,000 molecules labeled by perfume makers with subjective descriptors like 'earthy' and 'pungent'.
- The AI was trained on two-thirds of the data using graph neural networks and successfully categorized the remaining scents.
- The model aims to advance artificial olfaction to match AI's success in vision and hearing, with potential for eco-friendly synthetic odorants.
Why It Matters
Enables AI to interpret smell, opening applications in synthetic chemistry, safety detection, and fragrance design.