Fifty samples, five categories, one result: 100% classification success. That is the figure behind a spectrometer chip from the University of Cambridge that costs about $10 and sorted plastics, medicines, coffee, flour and tea correctly in the classification experiment that gives the claim its force. The number is real, but it comes from a controlled experiment, and the same paper reports errors, small ones, in its other tests.
The tech news worth your time: 400,000+ professionals from Google, Apple and OpenAI read Techpresso, a daily email with the most interesting tech news and insights.
We may earn a commission if you sign up.
The chip is a few centimeters across and reads near-infrared light between 1,200 and 1,700 nanometers. It processes a measurement in under a second and keeps working from minus 20 to 80 degrees Celsius, according to the university. The pitch is a spectrometer cheap enough to put in a wearable or on a factory line, instead of a benchtop instrument that costs far more.
The Nature Photonics paper behind the $10 price
The work was published in Nature Photonics as “Optical convolutional spectrometer,” in volume 20, issue 6, pages 664 to 672 (DOI 10.1038/s41566-026-01891-6). It went online on April 15, 2026, and the university’s announcement followed on Oct. 22, which is why the news cycle is later than the paper. Chunhui Yao is the first author, sharing the co-first marker with Jie Ma, and Qixiang Cheng is the last author and the one listed for correspondence. The device was developed with GlitterinTech, a start-up founded by the same group.
In the University of Cambridge release, republished by ScienceDaily, Yao describes the approach this way: “The key insight was to go back to the mathematics.” The chip itself uses a silicon nitride photonic platform with onboard electronics, and the paper’s title points to a convolution-based design. Cheng said the group’s long-term goal is to make spectrometry as ordinary as temperature or motion sensing.
What “every test” covers: 50 samples and a held-out split
The classification experiment, as laid out on the paper’s page, used 50 samples in five categories: 10 plastics, 10 pharmaceuticals, 10 coffees, 12 flours and 8 grades of green tea. Each sample was measured 80 times, and the measurements were split two to one into training and test sets. A k-nearest-neighbours classifier then reached 100% success in every category. Foods, plastics and medicines were all identified without a miss.
That wording needs its boundary. The result is a perfect score on a held-out third of repeated readings drawn from the same 50 physical samples, which is a strong test of whether the chip can tell those materials apart but a narrower one than sorting unfamiliar items in a kitchen or a pharmacy. The sources read for this article report no results from samples outside the set, and a perfect rate on 50 items does not predict what happens with thousands of varied products, which is the sort of test a commercial buyer would want.
The categories differ in how hard they are. Ten plastics and ten pharmaceuticals are distinct materials, while 12 flours and 8 green tea grades are closely related foods that a human taster might confuse, so the food results carry much of the weight. The paper reports success per category and not only overall, which is why the result can be stated for foods, plastics and medicines separately.
Where the accuracy is not 100%
The other experiments report ordinary measurement error, so the perfect score belongs to classification alone. The solution experiments used 80 repetitions at each concentration. For four solutions, ethanol, lactate, glucose in water, and ethylene glycol in isopropanol, at concentrations from 0.1% to 40%, a support vector regression model gave an R-squared of 1.000 with mean absolute errors of 0.006% to 0.017%. The university says that matched or exceeded commercial benchtop instruments in those experiments, about 0.01% accuracy.
The human measurements are rougher, and they are the ones most likely to be pictured in a wearable. For noninvasive glucose, 27 participants gave a mean absolute error of 0.363 millimoles per liter. Blood alcohol, in 24 participants, was off by 1.70 milligrams per deciliter on average. Lactate in 48 participants and skin moisture in 26 participants showed errors of 0.814 millimoles per liter and 2.49%. The human tests measured skin moisture, blood alcohol, blood lactate and blood glucose under conditions meant to mimic physiological environments, and the release says further testing is needed before any medical product, and glucose tracking was shown over an extended period in one participant.
So the sensor that read foods, plastics and medicines without error is not the same claim as a sensor that can replace a glucose meter. The price is about $10 for a research device, not a retail product with a price tag attached. The next evidence to look for is performance on materials outside the original 50, and a clinical-style study of the human readings.
This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.
The tech news worth your time: Subscribe to Techpresso →
We may earn a commission if you sign up.
More from Morning Overview
- Long use of a common prostate pill is tied to a higher chance of glaucoma
- Ford is recalling 223,472 F-150 pickups because the fuel tank can leak or detach
- Hybrids have 15% fewer problems than gas cars, while EVs and plug-in hybrids have about 80% more, Consumer Reports finds
- Regulators cleared the first U.S. small modular reactor, 4 months early
