Forget flower sniffing robots – the future of artificial intelligence just took a huge step towards building computers with an actual sense of smell.
In an exciting breakthrough, researchers at Google AI have developed an AI system capable of predicting how molecules will smell to humans. This “electronic nose” allows computers to detect scents and odor chemicals just by analyzing their molecular structures.
Previous attempts at digitizing smell have fallen flat. But by representing odor molecules as intricate maps of their atomic connections, Google’s AI has learned to identify key chemical features associated with scents like “floral” or “musky”.
After training on a dataset of 5000 known smelly molecules, the e-nose AI was let loose to sniff 400 mystery chemicals. Incredibly, its predictions matched or even exceeded the performance of human scent experts. For the first time, a computer can smell a molecule never encountered before and imagine how it might smell to us.
This odor map beats other chemical description techniques at predicting not just scents but also properties like intensity, pleasantness and similarity. It’s like having a complete periodic table arranging smells rather than elements.
So what’s next for this pioneering work? The lead researchers behind this digital odor breakthrough have now launched a startup called Osmo (http://osmo.ai) to commercialize their nose AI.
Osmo co-founder Alex Wiltschko explains, “What gets us really excited is using this technology to help address real-world problems – detecting gas leaks, monitoring health, etc.”
It’s a heartening example of bench-to-market AI research translating into impactful products. By digitizing the forgotten sense of smell, innovators like Osmo are opening up intriguing possibilities to enhance human life and health.
Even in today’s visual world, maybe there are some things our noses know that our eyes don’t. This new AI gives computers the power to sniff out the world like we do – and reach insights impossible to see but easy to smell.
Research Article: https://www.science.org/doi/10.1126/science.ade4401