Smell Can Actually Travel Over the Internet! Scientists Build an AI Nose That Detects and Sends Scents Remotely.

The scent can actually travel through the network! Scientists create an AI olfactory system that can identify and detect remotely..

I was scrolling late at night and came across a food video; the high‑definition visuals and spatial audio satisfied my audiovisual experience, but something still felt missing. Smartphones have been evolving for over twenty years, with the screen, speaker, and vibration motor each taking over vision, hearing, and touch, yet to this day we still cannot have a phone that can “smell” flavors.

If a phone could emit scents, the thing that wakes you up in the morning could be more than an alarm—it could be the aroma of freshly brewed coffee or toasted bread. Recently, this delightful idea has taken another small step toward reality. In late October, researchers from The Ohio State University and the Zhijiang Laboratory published a paper in Science Advances showcasing an AI‑assisted olfactory interface called Scentronic (meaning “electronic scent”).

The system uses an electronic nose on one end to identify odors, and a wearable device on the other end can spray the corresponding fragrance thousands of miles away.With the help of AI, researchers have linked odor detection, network transmission, and wearable release into a complete automated chain.

From Detection to Release

The corresponding author of this study is Jinghua Li, an assistant professor in the Department of Materials Science and Engineering at Ohio State University, whose lab has long focused on wearable and implantable flexible biosensors. Scentronic continues this line of work, shifting the sensing target from bodily fluids to odor molecules in the air.

The system’s transmitter is an electronic nose, which works by using a set of gas sensors, each with its own preference, to mimic the olfactory receptors in the nasal cavity. Although a single sensor has limited resolution, the combined response pattern can distinguish different odors.

Scentronic uses six commercial gas sensors; when volatile molecules such as ethanol or limonene adsorb onto the heated metal‑oxide surface inside the sensor, the sensor’s resistance changes, and the six signals together form a “fingerprint” of a scent.

The researchers selected five everyday foods—orange, cake, coffee, beer, and red wine—collecting 50 samples of each. They trained a classification model using the k‑nearest neighbors algorithm (a classic machine‑learning method that classifies new samples based on similarity to known samples), achieving an identification accuracy of over 98%.

After recognition is completed, the system wirelessly transmits a category label such as “coffee” to the terminal receiving device. The team made the receiver in three forms: a wristband, a necklace, and a handheld device, each internally storing diluted versions of five fragrances. Upon receiving a signal to release a certain scent, the built‑in piezoelectric ceramic in the receiver vibrates, atomizing the fragrance into a fine mist and controlling the intensity of the odor by adjusting the on/off ratio of the vibration.

From receiving the signal to spraying the scented mist, the entire system’s response time is about 0.6 seconds.However, before that, when faced with the complex aromas of food, the electronic nose needs several seconds to dozens of seconds to complete identification.

Whether people can recognize these remotely delivered scents is another test. The study recruited 14 healthy volunteers for testing; after becoming familiar with the corresponding odors, their correct identification rate for a single scent released by the device was about 73%, and about 71% when two scents were mixed.

The researchers also created a “digital cup” equipped with five scent channels and five taste channels, the latter delivering flavored liquid via electromagnetic pumps. Users wear a Meta Quest 3 mixed-reality headset, select a beer or coffee in the virtual scene, and the cup simultaneously releases the corresponding aroma and flavor, allowing them to drink while smelling.

(Source: Science Advances)

The team envisions that this device could be used for remote olfactory testing, helping physicians assess smell changes in patients after viral infection or with neurodegenerative diseases.

Pre‑storing scents and then releasing them directionally is not a new idea. The 1960 film “Scent of Mystery” used “Smell‑O‑Vision” technology, triggering a central unit from signals on the film to deliver about 30 odors through ducts to each seat.

The 1996 computer game Leisure Suit Larry: Love for Sail! came with a scratch‑off card featuring nine scent zones. At certain points in the story, a number flashes on the screen, prompting the player to rub the corresponding square to release the fragrance.

What sets Scentronic apart is the source of the trigger signal. The cues on film and the numbers in games are pre‑written scripts by the creators, whereas Scentronic’s commands come from an electronic nose’s real‑time assessment of actual objects on site.The four stages—perception, classification, transmission, and release—are automatically closed into a loop by the machine, and the receiver is reduced to a size that can be worn on the body.

From Selection to Formulation

Only transmitting tags without transmitting formulas determines Scentronic’s expansion method: each time a new scent is added, the electronic nose must undergo targeted training, and an additional fragrance cartridge must be installed in the receiver. Yang Chunyu (Chunyu Yang), a Ph.D. student from Li Jinghua’s team and the first author of the paper, said in a media interview that in the visual domain, the three primary pixels—red, green, and blue—can be combined to produce virtually unlimited colors, but olfaction lacks such simple basic units; everyday smells are usually composed of a large mixture of molecules with vastly different structures.

Some teams have already begun trying to bypass this major limitation by attempting to create a larger variety of scents from a few basic odorants. In April of this year, researchers from the MIT Media Lab and New York University posted the AromaGen system on a preprint server. Working with olfactory experts, the researchers selected 12 basic odorants and packaged them into a neck‑worn device.

The user describes a food with text, images, or voice; the multimodal large language model uses the odor knowledge learned during training to calculate the mixing ratios of 12 basic odorants and hands them to the device for sequential release. If the user feels something is off, they can request modifications in natural language, and the model adjusts the ratios accordingly.In human tests, after several iterations, the median similarity score between the generated scent and the real food aroma reaches 8 out of 10, and participants’ rating of the “artificial feel” drops to a level comparable to real food.

Some teams choose to return to a more fundamental question, investigating whether a standard coordinate system exists in odor space.

In 2023, Google’s research team and other institutions published a paper in Science that used graph neural networks to map molecular structures to human odor perception, constructing a “Primary Odor Map” (POM). For 400 previously undescribed odor molecules, the model’s descriptions were comparable to those of trained human evaluators.

The core team of this research spun off a startup called Osmo, which refers to the odor map as the “RGB” of the olfactory domain. However, some olfactory researchers argue that this analogy is overly optimistic; the map currently solves mainly “what a single molecule smells like,” whereas everyday odors are almost always complex mixtures, and the perception of a mixture is not a simple additive sum of its components. There is still a long way to go before we can simulate all odors.

In its current form, Scentronic uses only six metal‑oxide sensors that cost a few dollars each, paired with a classic classification algorithm, to achieve very high recognition accuracy on fixed categories. This is already sufficient for scenarios that require rapid odor discrimination, such as food batch comparison and remote olfactory screening.

The researchers also said the next step is to test the system in environments outside the lab and to expand the range of recognizable odors.

Marianna Obrist, a professor at University College London who studies multimodal interaction, called the work a valuable advance, while noting that there is still a large gap between identifying and emitting a specific odor from a limited odor library and recording and replaying olfactory experiences the way we do with sound and images.

Returning to the two scenarios mentioned at the beginning, the bedside coffee alarm can be realized through a pre-installed solution; the real challenge is conveying the taste from food videos.

The aromas of these foods usually come from specific combinations of hundreds of molecules; only by studying the basic ways to combine scents can digital olfaction become as rich as images and sound.

Reference:

https://www.science.org/doi/10.1126/sciadv.aee9266

https://www.reuters.com/science/new-scentronic-technology-can-detect-transmit-aromas-remotely-2026-10-23/

https://www.newscientist.com/article/2595189-digital-scents-could-add-a-new-dimension-to-virtual-worlds/

https://arxiv.org/abs/2604.01650

Operation/Typesetting: He Chenlong

Note: Cover/featured image generated with AI assistance