Whenever you say a color name, you’re referring to specific properties of light waves. Sounds work the same way, but with properties of compression waves. But what about smell? With all of the different scented chemicals out there and their complex interactions, it’s been impossible to create a simple scale to describe the odors or noses detect.
Scientists understand that the way things smell, be they like garlic or piss, comes from the way receptors in our body interact with the structures of different odiferous molecules. Still, given a certain molecule or combination, it’s difficult to accurately determine what smell we’re going to get. An international team of scientists realized this and, using a large dataset, hosted a contest to creating a smell-to-molecule structure algorithm. The scientist hope the research will revolutionize the way humans understand smell.
“Today, what they do to create new smells is use specialists trained for years and years,” or hundreds of people testing thousands of odors, study author Pablo Meyer from IBM’s Thomas J. Watson Research Center told Gizmodo. “I think that’s going to change.
The research began with a dataset that Rockefeller University researchers collected on how 49 people—many recruited from Craigslist—perceived the scents of almost 500 different chemicals. After each sniff test, participants were asked to select a descriptor, from a list of 19, that best matched the odor hitting their noses. This produced a shitload of data, which 22 teams crunched with machine learning algorithms to create optimal models of scent. Through the IBM-run crowdsourced effort, the teams developed a straightforward system: in goes the molecule with its structural and chemical information, out comes the name of the scent. The organizers even hid a set of 69 (nice) from the participants as a sort-of answer key. The better the algorithm could link these molecules to their correct scent, the better it was.
The models found that various properties of molecules correlated with various smells. Molecules that contain sulfur, for instance, yielded burnt and garlicky smells. Molecules that looked similar to vanillin, which gives vanilla beans its smell, left a bakery-like scent, according to the research that will be published this week in the journal Science. Molecules that looked most like the cancer drug paclitaxel, or a chemical called phenyl acetate, smelled pleasant.
Meyer said that some might oppose the selection of only 19 scent descriptors, which I’ve listed at the end of the story for your convenience, but argued that this was a good place to start. Eventually it may be possible to “go further into the compelexity and look at how people describe wine, perfume or coffee, he said. “We also think it’s possible to predict mixes,” or create a sort of color wheel like you’d see in grade school art class, “but no one has ever tried this before. We had to start somewhere.” The paper also notes that this model still has a long way to go—it predicted certain smells with much better accuracy than others, and had difficulty predicting smells like “acid,” “cold,” “warm,” and “urinous.” I personally do not have difficulty determining the smell of urine.
So, one day we might learn about and be able to mix smells like paint, based on the shapes and properties of molecules rather than just our own personal opinions. Because who doesn’t want a future where you sit down in chemistry class, look at a molecule and shout, “Hey, professor, that molecule smells like piss!”
For the dedicated readers who made it this far: the 19 smells used in the study were grass, bakery, sweet, fruit, spices, cold, wood, chemical, decayed, flower, sour, acid, musky, garlic, warm, burnt, fish, sweaty and urinous.