Miniaturized Spectrometer Enables On-Chip Hyperspectral Sensing

For years, spectrometers used to analyze the chemical composition of materials in medicine, food, and environmental monitoring have relied on large, expensive instruments that spread light into a rainbow and measure each color along a long optical path. This need for a physically extended path has made traditional devices bulky and difficult to integrate into portable systems. A recent study from the University of California Davis, reported in Advanced Photonics, describes an effort to reduce a lab-grade spectrometer to a device with a footprint comparable to a grain of sand, creating a spectrometer-on-a-chip suitable for compact platforms.

Instead of spatially separating the colors of light, the new design takes a reconstructive approach that avoids conventional dispersion optics. The chip uses 16 distinct silicon detectors, each engineered to respond differently to incoming light, so that the full spectrum is encoded across their collective outputs. This is likened to handing a mixed drink to several specialized sensors, with each one sampling a different aspect of the mixture. The challenge of recovering the “original recipe” from these mixed signals is addressed by artificial intelligence, which forms the second core component of the system.

The work centers on two primary technological advances. First, the researchers modified standard silicon photodiodes with photon-trapping surface textures (PTSTs). While silicon typically performs well in the visible range, it struggles with near-infrared light up to 1100 nm, a region that is important in applications such as biomedical imaging because it can penetrate human tissue more deeply than visible wavelengths. The PTST structures cause near-infrared photons to scatter within the thin silicon layer rather than pass straight through, which increases the chance of absorption and broadens the spectral sensitivity of the device. In addition, the architecture uses high-speed sensors that can measure photon lifetime, providing temporal resolution that reveals short-lived light–matter interactions that conventional instruments cannot detect.

Second, a fully connected neural network is trained on thousands of examples to learn the relationship between the noisy, encoded outputs of the 16 detectors and the original light spectrum. By solving this “inverse problem,” the AI reconstructs the spectrum with a resolution of about 8 nm, eliminating the need for bulky optics. The resulting chip combines a small 0.4 square millimeter footprint with high sensitivity and strong resistance to electrical noise, maintaining clear signals in conditions typical of portable, low-cost electronics. By extending silicon’s responsiveness into the near-infrared and enhancing performance through machine learning, the work outlines a path toward integrated, real-time hyperspectral sensing for diverse applications.