Scientists at The Ohio State University and Zhejiang Laboratory have built a system that smells something in one place and releases a matching scent in another, using machine learning to connect the two. In a human test, 14 volunteers identified single scents from the wearable about 73 percent of the time. In lab tests, the e-nose's models classified five odors with 96 to 100 percent accuracy.
The study, called Scentronic, appeared in Science Advances. Ohio State published its news release on November 4, 2026. The numbers are easy to confuse, so this post keeps them apart and explains what the AI actually does.
TL;DR: what it is and how well it works
| Question | Answer | Source |
|---|---|---|
| What was built? | An AI-assisted e-nose (sender) and a wearable scent releaser called Scentronic (receiver) | Ohio State; New Atlas |
| Which smells? | Coffee, beer, red wine, orange, chocolate cake | Ohio State |
| Lab accuracy of the e-nose models | 96 to 100 percent, depending on the method | New Atlas |
| Human accuracy, single scents | About 73 percent | Ohio State |
| Human accuracy, paired scents | About 71 percent | Ohio State |
| Intensity ranking accuracy | Around 59 percent | New Atlas |
| Can it recreate any smell? | No, it releases preloaded fragrances for known categories | New Atlas |
| Who led it? | Jinghua Li is senior author; Ohio State lists Chunyu Yang among co-authors | Ohio State |
The sending end: a six-sensor e-nose
According to New Atlas, the e-nose is a circular circuit board with six metal-oxide semiconductor gas sensors, using tin dioxide as the main sensing material. In ordinary air, oxygen settles on the sensor surface and raises its electrical resistance. When volatile organic compounds from food or drink reach the surface, some react with the adsorbed oxygen and give electrons back, so resistance drops and the circuit registers a change in voltage.
Ethanol from alcoholic drinks, limonene from citrus and diacetyl from buttery aromas are among the compounds that trigger these reactions. Because each sensor responds differently, each source produces a distinct voltage pattern across the array. New Atlas calls it a chemical fingerprint, "albeit a fairly blurry one."
Where the AI comes in
A blurry fingerprint needs a classifier. The researchers recorded 250 labeled examples, 50 for each of five foods and drinks, and tracked sensor voltages over 200 seconds per recording. They trained and compared several machine-learning models to map those patterns to categories. Once a pattern is recognized, the system assigns a label such as coffee or orange, and the result travels through an internet-connected cloud server.
That is a small supervised classification problem, similar in spirit to the edge-sensor work we covered in Exowatt's fine-tuned model for alarm triage, where a compact model on a narrow task did the job. Here the narrow task is sniffing.
A tiny green core inside a sensor housing, illustrating a small machine-learning model classifying signals from a gas sensor array
The reported 96 to 100 percent accuracy is for the e-nose models in controlled tests, with one product per category. That does not mean the system recognizes coffee in general. It means it recognizes the specific samples it was trained on.
The receiving end: Scentronic
The wearable holds miniature chambers filled with prepared scented liquids, each connected to its own release channel. A cotton filter draws liquid toward a piezoelectric atomizer, a small disc that flexes when electricity is applied. New Atlas says it is driven at roughly 110,000 vibrations per second and forces liquid through microscopic openings as a mist.
An ESP32 microcontroller controls the atomizer with fast pulses and adjusts spray frequency to regulate the amount of scent. Several channels can run together to make mixtures.
The device does not create smells from nothing. When a "coffee" instruction arrives, it activates the channel loaded with a coffee-like fragrance. The AI picks the category, and the receiver follows programmed instructions.
The team made several forms. Ohio State describes a wristband you lift to your nose and a necklace with five odor channels aligned "like pearls on a string," which open only after a signal arrives. New Atlas also mentions a handheld five-channel version and an experimental cup that pairs five scent channels with five pumps that drip flavoring into water for a taste effect, shown with a mixed-reality headset.
The human test: 73 percent, not 100
When 14 volunteers identified scents from the handheld device, they were right about 73 percent of the time for single scents and about 71 percent for two-scent combinations, according to Ohio State and New Atlas. In a separate intensity-ranking test, accuracy was around 59 percent.
Li said in the release that "we did see that some people are more sensitive to certain odorants than others," and that despite scents being easy to distinguish, intensity varied among people. The gap between the lab number and the human number is the honest headline: the sensing side works well on its own training set, and the delivery side is judged by people, who differ.
Why smell is a hard medium for AI
Cameras and microphones record physical quantities that follow well-understood laws, and a file can store them. New Atlas makes the point that a smell is chemistry: molecules floating through air, often in complex mixtures, whose identities and concentrations decide what you perceive. There is no straightforward equivalent of a photograph or audio file for an aroma.
That is why this project does not try to transmit the smell itself. It transmits a label. The sender reduces a messy chemical signal to one of five categories, and the receiver replays a stored stand-in. This is the same design trade seen across applied machine learning: when the full signal is too hard to capture, you classify it into a small vocabulary and rebuild an approximation from a library.
The trade has costs. A vocabulary of five means every smell outside the set is invisible, and a library of five liquids means every output is a substitute. Scaling it up means more sensors, more training data per smell and more channels, which is exactly the list of next steps the researchers give.
What it cannot do yet
- Five smells only. The e-nose recognizes five categories, and training used one product per category.
- Preloaded scents. Scentronic needs a matching fragrance loaded in advance for every smell it should produce.
- No discovery. It cannot sniff an unfamiliar perfume, work out its molecular recipe and recreate it. With five channels, the output is a recognizable stand-in.
- Small human study. Fourteen volunteers is a pilot, not a trial.
- Prototype. Ohio State calls it a prototype and says it will take time before the internet is "fully scratch-and-sniff capable."
New Atlas also points out that smell-releasing wearables already existed, including 2023 wireless VR interfaces from Hong Kong researchers. What is new here is combining electronic sensing, AI recognition, internet communication and scent release into one remote-sharing system.
Possible uses, and how far they are
Ohio State says possible uses include assistive technologies, virtual training environments and medical settings, such as reducing the need for frequent clinic visits and enabling timely personalized care. Li noted that loss of smell is related to many neurological diseases and infections, so the team wants to understand the olfactory system and then reconstruct it.
The researchers also wrote about improving perceived food quality for people with nutritional challenges, such as astronauts or soldiers eating field rations. New Atlas questioned how well the current five-scent design fits that case, and noted a more practical near-term use may be virtual reality, where a game could ship with a dedicated scent pack.
For online shopping, the obvious idea is sampling a perfume before ordering. That would require the right fragrance already loaded at the receiving end, so it is far from a general solution.
What to watch next
Ohio State lists next steps: testing ways to build the devices into everyday accessories like watches or smart jewelry. New Atlas adds more sensitive sensors, bigger training datasets, additional fragrance channels, better handling of complex mixtures, more rigorous human testing and standardized methods for assessing medical smells.
The work was supported by the National Science Foundation, the Center for Brain Injury Recovery and Discovery Pilot Award at Ohio State, and the Ohio State Alumni Grant for Graduate Research and Scholarship, per the release.
If you follow wearable AI more broadly, see our coverage of Amazon's Bee and ambient AI wearables and the AI toothbrush and on-device machine learning explainer. For another health-sensing example, read Google's TB screening from chest X-rays.
How to read claims like this one
- Ask which accuracy. Sensor classification on a lab dataset and people identifying scents are different measurements.
- Count the categories. Five known classes is a much easier problem than open-ended smell.
- Check the sample. One product per category and 14 volunteers limit what you can conclude.
- Separate the AI from the hardware. Here the AI labels; the chemistry of the receiver comes from preloaded liquids.
- Look for the replication. The paper is in Science Advances, but independent replication is still to come.
Related reading
- Exowatt's fine-tuned model for alarm triage on SageMaker
- Amazon Bee and AI wearables
- AI toothbrush and on-device machine learning
- Google TB screening with chest X-ray AI
- SageMaker SDK v3 for per-camera quality inspection models
Details are accurate as of November 5, 2026, based on Ohio State's release and New Atlas's report on the Science Advances paper.