Electronic Nose Chip Detects Lab-Tested Food Spoilage With 99% Accuracy

Forget Sniffing the Milk: Researchers Built a Tiny Chip That Does It for You

Smell is one of the oldest tools humans have for spotting when food may be going bad. But human noses are inconsistent, easily fooled, and can’t send a warning when the chicken in the fridge has turned. Now, a team of researchers has built a chip smaller than a postage stamp that can, under laboratory conditions, detect the chemical signatures associated with spoiled food, distinguish between different types of nuts, and tell the difference between 16 different foods with an accuracy rate of 92.6%.

Described in the journal Science Advances, the device is about the size of a small square of fingernail, just 7.5 millimeters on each side. It works by detecting the invisible chemical clouds that every food item releases into the air, then pairs that detection with artificial intelligence to make sense of what it’s smelling. It is an electronic nose that does not get tired or sick, though it still has to be trained on examples before it can recognize what it is smelling.

What Sets This Chip Apart From Earlier Electronic Noses

Most existing electronic nose devices use only a handful of sensors, typically between two and ten, and those sensors often react to the same chemicals in similar ways. This new chip uses 16 completely different sensing materials packed onto a single platform, so each one picks up on different chemical signals, producing a richer fingerprint for each food.

At the heart of the device are microscopic transistors coated with materials that react when certain airborne chemicals land on them. Researchers used carbon nanotubes, incredibly thin cylinders of carbon that are highly sensitive to chemicals even at room temperature.

That last part matters more than it might seem: most existing gas sensors need to be heated to very high temperatures to work properly, which limits what other materials can be used alongside them. Carbon nanotubes sidestep that problem entirely. Each sensor received a different coating, from electrically conductive plastics to metal-based powders, deposited through a process the researchers designed to be compatible with automated manufacturing systems.

How the Electronic Nose Chip Was Trained to Identify Food

When a food item is placed in a sealed flask and air is passed over it and into the device chamber, the chip’s 16 sensors respond simultaneously, each one shifting its electrical current by a different amount depending on which chemicals it encounters. Those shifting electrical signals become a kind of scent signature, which is then fed into a machine learning model, a type of artificial intelligence, trained to match signatures to specific foods.

Training the AI was a careful process. Each food was exposed to the chip 20 times in timed pulses: 95 seconds of exposure followed by a 185-second recovery window. The first pulse from each session was discarded to account for startup inconsistencies, leaving 19 usable readings per food. Those readings were divided into training, validation, and testing groups so the AI could learn on some data and be evaluated on data it had never seen before.

Foods tested covered a wide range: raw chicken, boiled eggs, whole milk, four types of nuts (walnut, hazelnut, cashew, and peanut), and freeze-dried fruits including blueberries, strawberries, and bananas. Spoilage tests tracked chicken, eggs, and milk at fresh, 24-hour, and 48-hour stages, all left out at room temperature between measurements.

Where the Chip Struggled and Where It Performed Best

Across all 16 foods, the system landed at 92.6% accuracy. Toughest calls were between hazelnuts and peanuts, which share enough chemical overlap that even a 16-sensor AI nose gets confused. Similarly, 48-hour spoiled boiled egg and spoiled raw chicken were occasionally mixed up, likely because both produce similar breakdown chemicals as they decay.

When the researchers trained a separate, more focused model exclusively on the spoilage data, accuracy climbed to 99.0%. A nut-only model brought hazelnut accuracy from 77.4% to 92.0% and peanut accuracy from 37.2% to 80.7%. A more specialized model that isn’t trying to distinguish 16 foods at once performs dramatically better at the specific jobs that matter most.

Nut classification was also tested as a possible future food-labeling or screening use case. Peanut and tree nut allergies can be life-threatening, so the researchers flagged this as a high-stakes area worth pursuing. According to the paper, this appears to be the first demonstration of a gas sensor array used to detect and distinguish different nut samples by their airborne chemical signatures. It is not an allergy-safety tool yet: the chip distinguished scent profiles from nut samples under lab conditions, not trace allergen contamination in a kitchen or a factory.

For now, this remains a laboratory prototype. Real-world messiness, from humidity to temperature swings to competing smells, still stands between this chip and a grocery store shelf. But the architecture is in place, and what it managed in a controlled setting is hard to dismiss.

---