How Electronic Noses Work: From Scentronic's Scent Messages to a Student E-Nose

Ohio State's Scentronic sends smells over the internet by sending only a word. This guide explains how: the tin-dioxide physics inside an MQ gas sensor, why six sensors together form a fingerprint one sensor cannot, how a small k-NN classifier turns that pattern into a scent name, why only a label travels over MQTT, and how a wristband turns the label back into mist — plus the practical protocol for recording your own fingerprints and the honest limits of the whole idea.

Written by Projectech16 min readPublished
For Engineering students building sensor, IoT or wearable projects who want to understand how gas-sensor arrays recognise smells and how a scent travels as a label. Topics: MQ gas sensors, tin dioxide (SnO2), ESP32, k-NN classification, MQTT, ultrasonic atomizer
Technical illustration of a ring of six gas sensors around a sample cup, with vapour patterns flowing to a small wristband device that releases a fine mist
Illustration generated for this guide.
In this guide

How Electronic Noses Work: From Scentronic's Scent Messages to a Student E-Nose

A friend can send you a photo of their coffee in a second. The smell stays in the kitchen — phones carry pictures, sound and video, but nothing for the nose. On 16 November 2026, researchers at Ohio State University reported a system called Scentronic that closes that gap: an electronic nose names a food from its fumes, and a wearable puffs the matching scent to a friend in another city. Only the word travels. This guide explains the physics and the engineering behind that trick — how a gas sensor actually senses, why six sensors together can name a smell when one sensor cannot, how a classifier turns a pattern into a word, and how a wristband turns that word back into a smell — and what a student build does differently, honestly.

What a smell actually is

A smell is chemistry in the air. When coffee beans sit in a cup, hundreds of volatile molecules escape from them — alcohols, aldehydes, acids, sulphur compounds — and drift through the air. Your nose detects them with about 400 types of olfactory receptor, each responding to many molecules, and your brain reads the pattern across all of them. No single receptor identifies coffee on its own; the pattern does.

That is the first idea an electronic nose copies. A single gas sensor is not a nose. It is a detector for a family of molecules, and it responds to many things at once. The trick is the same one biology uses: several different sensors, each with a different sensitivity profile, read the same air together. The pattern across them is the fingerprint, and the fingerprint is what gets recognised.

Scentronic: the news that made smells sendable

The Scentronic system was built by a team at Ohio State University — first author Chunyu Yang, Jinghua Li's group, with Zhejiang Laboratory — and published in Science Advances, reported on 16 November 2026. All the figures in this section are reported figures from that coverage, not measurements from any student build.

The sniffing end is a ring of six gas sensors around a spot where the food sits. Each sensor is a heated chip of tin dioxide. Vapours from the food change how easily electricity flows through the sensor, and the six together produce a pattern. The team trained a pattern-matching program on it: they recorded each of five foods — beer, cake, coffee, orange and red wine — 50 times, for 200 seconds at a stretch, making 250 recordings. On recordings the program had not seen before, the best version named all five foods correctly. Coffee gave the faintest signal, and the authors wrote that many of the compounds behind its aroma may not be the kind these sensors detect well.

Once the program decides on coffee, that word is all that travels — to a program on a cloud server and on to the wearable, a trip that took about 65 thousandths of a second in the team's tests. Recognising the food takes longer, from several seconds to a few tens of seconds.

The wearable is built around a thin disk that vibrates about 110,000 times a second. A cotton pad feeds it scented liquid from a small chamber, and the vibration pushes that liquid through tiny holes as a fine mist. A wristband version holds two smells; a necklace holds five, lined up, as the authors wrote, "like pearls on a string". Run flat out, a cartridge empties in under five minutes, and at the lowest setting it lasts more than 35 minutes. After an hour of continuous use, the warmest spot on the device measured 96.4 °F (35.8 °C), below body temperature.

Sixteen volunteers held their noses 20 cm from a device while orange rose through ten strength levels. Fourteen volunteers then identified single scents from the device about 73 percent of the time, named both parts of a two-scent mixture about 71 percent of the time, and ranked five strengths of an alcohol scent about 59 percent of the time. The authors focus on medicine: an adjustable device could let a clinician raise an odour step by step over the internet until a patient notices it, finding the smell threshold without an office visit. Scentronic is a laboratory prototype, and its first author said it will take several years, potentially longer, before such devices become commercially available.

Inside a tin-dioxide gas sensor

The sensors in Scentronic and in most student e-nose builds are metal-oxide semiconductor sensors — the MQ family that fills every electronics parts list. Understanding one means understanding the physics of a heated ceramic film.

Inside each MQ sensor is a tiny alumina ceramic tube coated with tin dioxide (SnO2), with a heater coil running through it. The heater holds the film at a few hundred degrees Celsius, and that heat is not incidental — it is the whole operating principle. At that temperature, oxygen molecules from the air stick to the tin-dioxide surface and steal electrons from it, forming a layer of negatively charged oxygen ions. That layer makes the film resistive: with few free electrons, current struggles to cross it.

When a vapour reaches the surface — alcohol from a beer, terpenes from orange peel, the volatile compounds of coffee — those molecules react with the adsorbed oxygen, return electrons to the film, and the resistance falls. The sensor's circuit turns that resistance change into a voltage the microcontroller can read. Remove the vapour and the oxygen layer rebuilds, and the resistance climbs back. That is the whole measurement: a resistance that rises and falls with the chemistry of the air.

Three consequences follow, and every one of them matters for a student build:

Cross-sensitivity is the rule, not the exception. An MQ-3 is sold as an alcohol sensor, but its tin-dioxide film also responds to smoke, benzene and other vapours. The Hanwei MQ-3 datasheet gives its detection range as 0.05–10 mg/L of alcohol, but nothing in the device can tell alcohol from another reducing gas. A single sensor reports "something is in the air", never "this is coffee".

Heat costs power. The MQ-3 heater runs at 5 V and draws about 150 mA (its heater resistance is 33 Ω ±5% per the datasheet, so 5 V / 33 Ω ≈ 150 mA). Six heated sensors together draw about 900 mA, which is why an e-nose sender needs a 5 V 2 A supply and cannot run from a laptop USB port. An under-powered heater rail is the single most common cause of a drifting fingerprint.

Warm-up and drift are real. The datasheet specifies over 24 hours of preheat for full stability; practical builds settle in roughly 10–20 minutes, but readings taken cold are meaningless, and the clean-air baseline drifts with the room's temperature and humidity. A serious build re-records its baseline at the start of every session.

Why one sensor is a gas alarm and six sensors are a nose

If each sensor responds to many vapours, how can six of them name a scent? Because they respond differently. An MQ-3 is most sensitive to alcohol; an MQ-135 to ammonia and benzene; an MQ-2 to smoke and LPG; an MQ-7 to carbon monoxide; an MQ-8 to hydrogen; an MQ-136 to hydrogen sulphide. The same puff of coffee vapour pushes each sensor's resistance by a different amount, so the six readings together form a pattern — six numbers that rise and fall in a shape that is different for coffee, orange, lemon, mint and vanilla.

That pattern is the fingerprint. One sensor gives a single number, which can only be compared against a threshold: above this line, alarm. Six sensors give a six-dimensional point, and points for the same scent cluster together while points for different scents sit apart. The classifier's job is to measure which cluster a new point belongs to.

This is also why the sensor ring matters physically. In Scentronic, the six sensors sit in a ring around the sample so they all breathe the same air at the same time. A student build that reads its sensors one after another from different corners of a room is measuring different air, and its fingerprints will not repeat. Same air, same moment, same temperature: that is what makes a fingerprint comparable from one recording to the next.

From fingerprint to a name: how the classifier works

A student e-nose does not need a neural network. The honest, explainable classifier for five scents is k-nearest-neighbours (k-NN), and it is small enough to run on an ESP32.

The student records a training set first: each scent goes in the cup for 20 recordings of 30 seconds, and the six-channel fingerprint of each recording is saved with its label. That gives 100 labelled points in six-dimensional space, 20 per scent. When a new sample arrives, the classifier measures the distance from the new fingerprint to every stored point, takes the k nearest neighbours (k = 3 is a common choice), and returns the scent most of those neighbours belong to, with a confidence based on how close and how unanimous they are.

Three things about this procedure are worth understanding, because they decide whether the build works:

The report prints a confusion matrix, not an accuracy claim. The training set is split: most recordings train the model, and a held-out test set the training never saw measures it. The confusion matrix shows which scents get confused with which — coffee mistaken for vanilla, say — and that matrix is the honest result. No accuracy is claimed before those runs.

A fingerprint is tied to the sample it was trained on. Scentronic's team used one brand of each food — the cake was a bag of King's Hawaiian sweet rolls — and whether the program would recognise a different coffee is untested. The same applies to a student build: a classifier trained on one brand of coffee beans may not recognise another. Retraining with new labelled recordings is the fix, and the guide for the build covers it.

Normalisation is what makes the comparison fair. Raw voltages depend on the heater, the supply and the room. Before comparing fingerprints, each channel is scaled — for example, each reading is divided by the clean-air baseline recorded that session — so the classifier compares shapes, not absolute levels. Without that step, a hot afternoon looks like a different scent.

Why only a word travels

The most counter-intuitive part of Scentronic is also its most important design decision: the smell itself never travels. What crosses the internet is a label — the word "coffee" — and the receiving device releases coffee scent from a liquid it already holds.

There are two reasons. The first is physics: you cannot send molecules through a wire, and no one has a way to synthesise a smell from a description at the receiving end. The second is information: a label is a few bytes, it arrives in about 65 ms (the reported figure for Scentronic's cloud hop), and it is unambiguous. The receiver's job is mechanical, not chemical: pick the right cartridge and puff.

A student build copies this exactly, at local scale. The sender publishes one word over MQTT to a broker on the student's laptop, and the wristband subscribes to the same topic. Local Wi-Fi is enough; there is no cloud link in the base build, and sending a smell to another city is an extension, not a starting point. The dashboard — the Scent Console — shows the whole chain on one screen: the live fingerprint bars, the label, the confidence and the transmission log, so a reviewer can see that the network carried a word and the wristband did the rest.

How the wristband makes mist

The receiving end is a small piece of physics of its own. Scentronic's wearable uses a thin piezoelectric disk that vibrates about 110,000 times a second. A cotton pad feeds scented liquid to the disk, and the vibration pushes the liquid through tiny holes as a fine mist. Switching the vibration on and off sets how much mist comes out, which is how the device offers ten strength levels.

A student build uses the same principle at hobby scale: a 5 V ultrasonic atomizer disc in the 110 kHz class, a cotton pad, and a cartridge of food-grade essential oil. Two cartridges fit the base-build wristband — coffee and orange — and a five-cartridge necklace is the extension that mirrors the paper's necklace. The mist is a water-based aerosol, not smoke, and the demo runs in a ventilated room with an allergy check first, because the wristband releases real scented mist.

One design decision deserves emphasis: the student build deliberately excludes the paper's beer and wine scents. Alcohol near a heated sensor and a wearable is a safety risk a student build must not take, so the five training scents are food-grade essential oils and real kitchen samples — coffee beans, orange peel, lemon peel, mint leaves and vanilla.

What the student build does differently

The student-scale version of Scentronic is a real project in its own right, and it is worth stating plainly how it differs from the research device it was inspired by:

  • Sensors: six MQ modules (MQ-3, MQ-135, MQ-2, MQ-7, MQ-8, MQ-136) read through a CD74HC4067 16-channel multiplexer into an ESP32, instead of the paper's six heated tin-dioxide chips.
  • Classifier: a small k-NN model running on the sender ESP32, instead of the paper's pattern-matching program.
  • Network: local MQTT to a laptop broker, instead of a cloud server.
  • Receiver: a wristband with two cartridges and a 110 kHz class atomizer disc, instead of the paper's wristband and five-smell necklace.
  • Scents: five food-grade scents (coffee, orange, lemon, mint, vanilla), with beer and wine deliberately excluded for safety.

The full build — sender prototype, wristband receiver, firmware, Scent Console dashboard and the complete review kit — is published as Scentronic-Inspired E-Nose Project with Gas Sensor Array for Sending Smells Over the Internet. The project page carries the comparison table against Scentronic's reported figures and the live breath-analyzer page, the full specification table, and the blind smell-test procedure the student runs to validate the build.

Building your own fingerprint: the practical protocol

If you build an e-nose, the hardware is the easy half. The fingerprint is the project, and it is won or lost in the recording protocol:

  1. Warm up first. Run the sensors for at least 10 minutes before any recording, and respect the datasheet's over-24-hours preheat for full stability. Record nothing cold.
  2. Record a clean-air baseline every session. Empty cup, same room, same temperature. Every reading is relative to that baseline, because the baseline drifts with the room.
  3. Same air, same moment. All six sensors read the same sample at the same time, from the same cup, under the same light and temperature. Change any of those and the fingerprint changes.
  4. Record more than you think you need. Twenty recordings of 30 seconds per scent is a starting point, not a law. Keep a held-out test set the training never sees, and print the confusion matrix from it.
  5. Test with a blind smell test. A classmate wears the wristband in another room while the sender sniffs samples, and you record how often the classmate names the scent correctly. That hit rate — not the classifier's training accuracy — is the honest answer to whether the system sends a smell a person can recognise.
  6. Power the heaters properly. Six MQ heaters draw about 900 mA together. Use a 5 V 2 A supply. A sagging heater rail shows up as a fingerprint that drifts for no visible reason.

Frequently asked questions

Does the smell itself travel over the internet?
No — only the word travels. The sender publishes one label, such as "coffee", over MQTT, and the wristband releases coffee scent from a cartridge it already holds. That is exactly how Scentronic works: the paper's authors state that what crosses the internet is a label rather than the smell itself, and the device reproduces a category from a liquid chosen in advance.

Why can't one gas sensor identify a smell?
Because every MQ sensor responds to many vapours. An MQ-3 is most sensitive to alcohol, but it also reacts to smoke and benzene. A single reading can only be compared against a threshold, which makes it a gas alarm. Six sensors with different sensitivity profiles produce a pattern, and the pattern is what a classifier can name.

How accurate is a student e-nose?
No accuracy is claimed before the student's own runs. The classifier is trained on the student's own recordings, its confusion matrix comes from a held-out test set, and the blind smell test measures whether a person wearing the wristband actually recognises the scent. The report records those numbers honestly, on the student's own scents.

Is this Scentronic?
No. This guide explains the research system reported on 16 November 2026 and the student-scale build it inspired. The student build shares no parts, method or performance with the Ohio State device, and it reproduces a scent category from a pre-loaded liquid, never the chemistry of the original sample.

Can it send any smell?
No. The vocabulary is the scents it was trained on — five in the student build. A sixth scent means new recordings, a retrained model and a new cartridge. And some smells are deliberately excluded: the student build uses food-grade essential oils and kitchen samples only, never alcohol-based scents.

Could this become a remote smell test?
That is the medical use the Scentronic authors describe — an adjustable device that raises an odour step by step over the internet until a patient notices it. The student build demonstrates the sensing and transmission chain only. It is not a medical device and must not be used to assess anyone's sense of smell.

Limitations

An electronic nose is a pattern recogniser, not a chemical analyser. It cannot tell you what molecules are in the air, only that the pattern matches a scent it was trained on. Its fingerprints drift with temperature, humidity, heater power and sensor ageing, so a model trained on Monday may need recalibration on Friday. Its vocabulary is small — five scents in the student build, five foods in the paper — and adding a scent means new recordings and retraining. It reproduces a category, never a specific sample: the coffee the wristband releases smells like coffee in general, not like the particular beans the sender sniffed. And the research device it was inspired by is a laboratory prototype, years from commercial use. None of that makes the project less worthwhile; it makes the honesty about it the project.

More project guides

More in IoT & Embedded