New AI Learned to Smell — And It Can Name Mixtures

Almost nothing in nature smells like just one thing. Coffee, smoke, fresh bread — each is a cloud of hundreds of different molecules hitting your nose at once. Your brain sorts that mess out instantly. Computers have never been good at it. So a team of researchers built an AI that works more like a nose than a calculator. It mimics how single molecules switch on smell receptors, then blends those signals the way your brain does. In tests, it correctly identified mixtures 92.2% of the time.

Why is this so hard? Because smells change with strength. The same perfume smells different up close than across a room — some ingredients suddenly shout, while others go silent. It's like a band where one loud instrument drowns out the rest depending on how loud the whole song is. The new model accounts for that, giving each molecule's signal a weight based on how strongly it's actually present.

The team's clever trick was borrowing knowledge. They compared the "response curves" of well-understood chemicals to unfamiliar mixtures, letting the AI transfer what it already knew — a bit like tasting a new dish and recognizing it by comparing it to meals you've had before. That shortcut is what pushed accuracy so high, and it means the system doesn't need to be trained on every possible smell in advance.

The honest catch: 92.2% still means roughly one in thirteen answers is wrong, and this is a research result, not a product you can buy. It was also tested on data the researchers assembled, not on the chaotic real world. Still, the goal is bigger than perfume. The authors say their method could plug into robots and other "embodied" AI — machines that sense and act in the physical world — to help them judge food safety, detect hazards, or assist in medical screening.