Researchers send scent data to wearables, with 73% recognition in a small lab test

A smell detected in one place can trigger a wearable to release a recognizable scent elsewhere. In a small laboratory study, researchers at Ohio State University and Zhejiang Laboratory demonstrated an AI-assisted system for sharing odors over the internet. Fourteen volunteers identified its individual scents correctly about 73% of the time.

The system, called Scentronic, is described in a Science Advances paper. It connects an electronic nose that detects and analyzes odors with a receiver that releases scents. The demonstration is narrower than sending any arbitrary smell: the researchers say the receiver reconstructs odor categories using preloaded scent channels, rather than reproducing the original odor’s full chemical composition.

AI identifies the odor; a wearable supplies the scent

Smell presents a different engineering problem from a picture or a sound. It depends on sensing chemical compounds, which must then be represented as data and recreated at the destination. As Discover explains, the team uses machine-learning algorithms to classify volatile organic compounds—the small molecules associated with a smell—into odor categories. AI’s role is to recognize what the sensor encounters.

The receiver can be built into a bracelet or necklace and dispenses the reconstructed scent near the wearer’s nose. The researchers designed the sequence to run automatically and hands-free, without someone manually triggering each release. That distinguishes the approach from earlier setups described in the research coverage, which sometimes required a person to activate the scent at the receiving end.

Five familiar odors, tested alone and in pairs

The team trained and tested the system on coffee, beer, red wine, cake and orange. Its human evaluation involved 14 healthy volunteers, who were asked to identify the smells released by the device. Those results measure whether people recognized the recreated odors—not whether the receiver produced a chemically exact copy of what the sensor had detected.

Alongside the individual odors, participants received pairs such as coffee and orange or beer and cake. They correctly identified those combinations roughly 71% of the time. The paired tests therefore extended the demonstration beyond one scent at a time, while still drawing on the same five everyday odors used in the experiment.

Speed also depended on which part of the process was being measured. The system could respond to a smell-sensor signal in 0.6 seconds. Classifying more complex food odors, however, took several to tens of seconds. The faster figure should not be read as the time needed to recognize and recreate every possible food smell.

The laboratory is still the boundary

The authors caution that their experiments were relatively simple and included limited control conditions, which may introduce bias. They say further testing is needed outside a controlled laboratory. That leaves a material question about how well the system’s recognition and reconstruction would hold up beyond the conditions used for this demonstration.

The team’s planned next steps include making the tests more complex and improving control of potential volunteer biases. Those experiments would probe a different hurdle from simply getting a sensor and dispenser to communicate: whether people can reliably recognize the output when the evaluation becomes less constrained. For now, the reported evidence comes from a small, controlled test.