Cambridge's $10 Convolutional Spectrometer Chip Targets Wearable Biomarker Sensing

A team at the University of Cambridge and its spin-out GlitterinTech has built a packaged near-infrared spectrometer that costs about $10. In the team's reported tests it matched or beat benchtop instruments on several measurements. The paper, published in Nature Photonics (volume 20, issue 6, pages 664–672) and online since 15 April 2026, is being recirculated this month. The pitch is a wearable that reads blood chemistry from light alone. The data support the optics far more strongly than they support the medicine.

01Doing the transform in the optics

Miniature spectrometers usually trade away range, resolution or accuracy. According to the authors, miniaturized versions are often limited to coarse identification of signature peaks and are inadequate for metrology. Dispersive designs need path length to separate wavelengths. Fourier-transform designs need moving parts. Computational "reconstructive" designs sample the spectrum through many scrambled filters and then solve an ill-conditioned inverse problem, which makes them sensitive to noise and calibration drift.

The Cambridge device, which the authors call a convolutional spectrometer, avoids that inverse problem. It applies the convolution theorem directly to incoming light using repeating optical components (unbalanced Mach-Zehnder interferometers or micro-ring resonators), with Fast Fourier transforms recovering the spectrum.

The implemented version cascades four stages of unbalanced Mach-Zehnder interferometers (MZIs) on the photonic chip. The authors describe that as a trade-off between resolution, footprint and control complexity. An unbalanced MZI has a periodic spectral response. Its free spectral range depends on the group index and the arm-length difference. The stages have different periods, so their combined response has its own composite period, which the paper's figure sets at 420 cm⁻¹.

Applying proportional phase modulation to the stages shifts the overlaid system response across the spectral domain. The photodetector readings at each shift are therefore a circular convolution of the incident spectrum with that response. Because convolution becomes multiplication in the Fourier domain, recovering the spectrum is a linear step: a fast Fourier transform. That linearity explains the claimed noise tolerance and the modest computing load.

The scaling argument is that the response is periodic, so the team says bandwidth can in theory be extended almost without limit. Cascading more stages, meanwhile, can raise resolution exponentially. Both remain theoretical. The packaged unit has four stages and one bandwidth.

02What was actually built

The chip uses a silicon nitride platform. Silicon nitride is transparent across 1,200–1,700 nm and can be made in established photonic foundry processes. The paper says curved directional couplers are used to support the ultra-wide bandwidth. The reported specifications are:

• Range: 5,900 to 8,300 cm⁻¹, about 1,200 to 1,700 nm, a 2,400 cm⁻¹ bandwidth.
• Package: 4.8 × 6.2 × 0.6 cm³, 16.5 g, and a total cost described as US$10-level. The paper's cost breakdown is in its Supplementary Table 1.
• Speed: under 0.4 seconds for sampling and local processing.
• Temperature: operation from −20 °C to 80 °C, per the university's account of the work.

A footnote on "wrist-sized": 4.8 × 6.2 cm is a prototype module, not a smartwatch component. Integrating it into a watch case with a light source, skin-contact optics and a battery is still future engineering. The paper's own phrasing is that the work shows the potential for "portable and/or wearable" metrology.

For comparison, other groups have used cascaded-MZI reconstructive designs to reach very fine resolution. One earlier programmable-circuit design reported under 10 pm resolution over more than 200 nm of bandwidth using 729 sampling channels. The Cambridge design is aiming at a different point: moderate resolution, very wide range, simple reconstruction and low cost.

03Bench results

On solids, the device classified pharmaceuticals, coffee, flour and tea with a 100% success rate. The abstract also lists plastics. For liquids, the team measured aqueous solutions of ethanol, lactate and glucose and organic solutions of ethylene glycol in isopropanol across concentrations from 0.1% to 40%, with 80 repetitions per concentration. A pretrained support-vector regression model tracked glucose solution in real time with a system-level time resolution of 4.9 s. The team reports concentration accuracy of about 0.01%, ahead of commercial benchtop spectrometers in those experiments.

Two caveats apply to these numbers. A 100% classification rate on a handful of material classes shows the spectra are separable. It says little about generalization to the variability of real production lines. And "better than benchtop" is relative to the specific instruments and protocols the authors chose.

04Biomarkers and the clinical gap

The healthcare demonstrations are the headline. The abstract reports non-invasive mean errors of 2.49% for skin moisture, 1.70 mg/dL for blood alcohol, 0.81 mmol/L for blood lactate and 0.36 mmol/L for blood glucose. That glucose error is roughly 6.5 mg/dL by standard unit conversion, which is my arithmetic, not a figure from the paper. The Cambridge release says the implications range from hydration assessment and intoxication alerts onward.

The physics explains why this is hard. As the authors note, blood lactate and glucose exist at low concentrations and have overlapping absorption bands. Independent work on tissue spectroscopy finds that in the 1,400–1,700 nm band glucose absorbs, but the scattering change still dominates the glucose signal. Reliable models must separate glucose from scattering, hydration and temperature effects. A wide 1,200–1,700 nm window gives a regression model more spectral features to work with, which is the argument for this chip.

The study also shows how far this is from a medical product:

• Glucose tracking: The paper reports single-participant long-term tracking of daily glucose fluctuations. It also shows glucose variation in three participants during oral glucose tolerance tests.
• Transferability: The authors state that, at the current data scale, neither the multi-participant nor the single-participant glucose model transfers to new participants. They attribute this to interindividual variability. A device that needs per-person calibration against reference measurements is not yet a needle-free replacement for a fingerstick or a continuous glucose monitor.
• Metrics: Mean absolute error is not the same as the error-grid analysis and mean absolute relative difference used in regulatory glucose-monitor evaluation. The paper does not provide regulatory-grade accuracy.
• Conflicts of interest: The preprint says the research was mainly supported by GlitterinTech Limited. The published paper declares a granted Chinese patent held by GlitterinTech on the convolutional spectrometer.

05What comes next

The nearer-term uses are probably less demanding than glucose. Skin hydration, alcohol and lactate tracking, food and agricultural screening, plastics sorting and pharmaceutical identity checks tolerate looser accuracy, or can be calibrated against known materials. Cheap spectrometers like this could also be embedded across production lines for continuous quality control.

The open questions are practical. They include larger multi-participant datasets with proper cross-validation across people, skin tones and days. They include power draw for the phase shifters and electronics in a battery device. And they include whether four-stage resolution suffices for the narrow spectral features of the biomarkers that matter. Professor Qixiang Cheng, who led the project, frames the aim as making spectrometry as ubiquitous as temperature or motion sensing. The paper shows the optical and computational architecture works at low cost, but the clinical case has not yet been made.