A race is on to reverse-engineer the nose, and more broadly, the entire olfactory system, which is incredibly complex, still not fully understood, and extremely underutilized by technology, compared to sight and sound. Quite apart from the novel possibilities opened up by being able to “design” scent, or give robots “a sense of smell”, the results could reshape fields like environmental monitoring, as Envirotec discovers.
Our experience of scent is obviously a powerful and emotionally-loaded one. The immediacy of the sensation can seem to disguise or downplay what turns out to be a toweringly complex process of exquisitely fine-tuned chemical recognition.
But bridging the gap between these two worlds – the chemical and the experiential – turns out to be just the kind of task to which AI can be usefully turned. And this undertaking – sometimes referred to as “the digitization of smell” – is starting to yield groundbreaking possibilities, such as the ability to design entirely new scents.
In March, Californian startup firm Osmo launched Generation,1 billed as the world’s first AI-powered fragrance house, and seemingly the first attempt to offer commercial opportunities based on this emerging capability. The promise has been that it will allow chemists to specify chemicals and their scents somewhat precisely, without the need to perform thousands of laboratory tests. There is also interest in using it to help design-out toxic or unsustainable chemicals from products.
Fascinating possibilities also come into view when, for example, you can specify a smell like “burnt popcorn” and the system will suggest suitable molecules. Osmo’s publicity for Generation declares that it gives brands the power “to create fragrances with more precision and creativity than ever before”.
The firm has released very little detail about the underlying technology, dubbed Olfactory Intelligence (OI), although the groundwork for its approach appears to have been laid when some of Osmo’s team, including CEO Alex Wiltshko, worked at Google Brain (now part of Google DeepMind), and papers published at that time might be assumed to offer a nod to the approach taken.
Ahead by a nose
Such fundamental groundwork included the development of a machine learning model – the “Principal Odour Map” – that can predict the smell of something based on its chemical structure, the focus of a 2023 paper in Science.2 The study was a significant milestone in achieving performance that is on par with, even to some extent beyond, the capabilities of trained human sniffers.
Utilizing machine learning, the model itself was trained using datasets pairing molecules with descriptors provided by such human experts – labels like “woody”, “minty” and so on. This kind of data is available from research fields like fragrance chemistry and food science, as well as academic datasets like GoodScents. Also fed into the model were data about molecular structures.
The result is a system that, when you feed it data about a molecule, it will predict its likely smell (in terms of the aforementioned descriptors). And it does so with an accuracy that meets or exceeds that of trained human sniffers: “On a prospective validation set of 400 out-of-sample odorants, the model-generated odour profile more closely matched the trained panel mean than did the median panellist,” says the paper.
One seemingly remarkable aspect of the work is the fact that it achieves the accuracy it does without being grounded in any knowledge of the actual human biology underlying smell. Study of the latter therefore seems to open up further exciting avenues by which important leaps might be made in the future.
Conk chemistry
The sheer variety of chemicals involved in producing odour is “enormous”, as one researcher recently puts it.2 And the way in which this chemistry is sensed by our nose and brain has been yielding its secrets recently, albeit slowly. A decisive biochemical agent in the process seems to be the olfactory receptor (OR), responsible for “recognising” particular smells. The OR is a complex protein molecule residing in the nasal epithelium, in the vicinity of the olfactory sensory neurons. There are around 400 different ORs, whose functioning accounts for our ability to perceive the great many substances that possess smell (how many? One recent estimate is about 20 billion).4
These ORs “recognise” specific odorous molecules by binding to them, an event that the olfactory nerve is able to translate into an electrical signal, the medium of the nervous system. So, different smells – whether fruity, minty, citrus or whatever – will be recognised by different types of OR.
How the ORs do this is a key question. Studying them seems to be particularly challenging, as biomolecules go, but this interaction is becoming better understood. For example, 2023 saw the publication of the first protein structure of a human OR bound to an odourant.5
At this stage, knowledge about human ORs is very limited. Of the 400 or so types believed present in human noses, only about a fifth of them have been linked to specific odour molecules. There also appears to be mystery surrounding the action of these ORs. One theory is that a single OR might engage with different odourants in distinct and different ways, a factor that may account for some of the puzzling complexities in how molecular structures correspond to particular smells – the fact that very similar molecules can possess drastically different smells while quite different molecules can smell similarly.
The perceptual mapping models, the work that precedes the formation of startups like Osmo, appears to be agnostic to such biological specifics. But, presumably, our growing knowledge of how we smell will enable further, unexplored possibilities for digitizing smell.
So far this new capability – to predict human smell perception from molecular attributes – is just getting started commercially. Extrapolating its potential to fields like envirnomental monitoring suggests possibilities such as the ability to predict the odour of a faclity before it has even been built, or that odours in a region like a city could be mapped in great detail, or that early warning of dangerous leaks could become a much more exact science.
Notes
[1] https://www.osmo.ai/blog/osmo-launches-generation-worlds-first-ai-powered-fragrance-house
[2] “A Principal Odor Map Unifies Diverse Tasks in Human Olfactory Perception”, Science, 2023
[3] Comment attributed to Aashish Manglik, a biochemist at the University of California, quoted in “The Most Mysterious Sense: Cracking the Odour Code” by Kerri Smith, in Nature, 5 September 2024, p26.
[4] “Digital smell has arrived. Are we ready for Stinkygram?”, Salon, January 2025
[5] “The Most Mysterious Sense: Cracking the Odour Code” by Kerri Smith, in Nature, 5 September 2024, p26.
