Researchers at the University of California, Davis have built a spectrometer-on-a-chip that occupies just 0.4 square millimeters — roughly the size of a grain of sand — and reconstructs light spectra to about 8 nm resolution without prisms or gratings. The team, led by M. Saif Islam and including Ahasan Ahamed, Htet Myat, Amita Rawat and Lisa N. McPhillips, published the work in Advanced Photonics (2026, vol. 8, issue 01, DOI: 10.1117/1.AP.8.1.016008).

Conventional spectrometers split light into its component colors with prisms or diffraction gratings, then measure the intensity of each wavelength. Because the light must travel a relatively long path before detection, the instruments stay bulky and resist miniaturization. The UC Davis design removes that path almost entirely.

How does the chip replace the optics?

The device uses 16 silicon detectors, each engineered to respond slightly differently to incoming light. No single detector isolates a color. Instead, each one captures an encoded signal that carries partial spectral information, and a fully connected neural network trained on thousands of examples learns the mapping from those noisy signals back to the original spectrum.

The researchers describe this as solving an "inverse problem": the system infers the spectrum computationally rather than measuring each wavelength separately. That trade — detector simplicity against model complexity — is what allows the 0.4 mm² footprint.

The takeaway: machine learning does the work that a long optical path used to do, and the resolution penalty is about 8 nm.

Why the near-infrared extension matters

The second result is material. Standard silicon photodiodes detect visible light well but lose sensitivity in the near-infrared, up to about 1100 nm, and NIR light matters because it penetrates human tissue more deeply than visible light — the basis of biomedical imaging.

The team modified the photodiode surface with photon-trapping surface textures (PTSTs). These textures scatter incoming NIR photons repeatedly inside the thin silicon layer instead of letting them pass straight through, which raises the probability of absorption. The result is a silicon sensor sensitive across a much wider spectral range than standard silicon allows, without changing the base material or the manufacturing platform.

The takeaway: the infrared extension comes from surface texturing on ordinary silicon, not from exotic materials.

What about speed and noise?

The chip also carries high-speed sensors that measure photon lifetime with high temporal precision. According to the researchers, this lets the device capture ultrafast light-matter interactions that traditional spectrometers may miss entirely, a capability that normally requires far larger and more expensive systems.

The design also holds up electrically. The AI-assisted architecture preserves clear signal quality even in noisy environments, which the team flags as a major challenge for portable, low-cost electronics. High sensitivity and noise resistance together are the conditions a wearable or handheld sensor must meet before it leaves the lab.

The takeaway: the chip pairs spectral reconstruction with timing measurement, and it survives the electrical noise of cheap hardware.

Where could it land?

The researchers point to compact, real-time hyperspectral sensing as the end goal. Named applications include portable medical diagnostics, wearable health monitors, environmental remote sensing and food quality analysis — the same markets benchtop spectrometers serve today in disease diagnosis, food inspection and pollution monitoring.

All performance figures come from the published paper; the application list is the team's own roadmap rather than shipping product. If the 8 nm resolution and NIR sensitivity hold up outside the laboratory, the work points toward hyperspectral sensing becoming a standard block in consumer and medical silicon rather than a standalone instrument category.