Spectrometers have traditionally treated light as a two-stage problem. First, an optical system encodes the incoming light by separating or modulating its wavelengths. Then a detector records the resulting signal, and a computer reconstructs the spectrum through numerical algorithms. That arrangement has enabled powerful scientific instruments, but it also creates a bottleneck: the raw measurements must leave the sensor before they can become useful information. A new study published in Nature Sensors describes a device designed to remove that separation. Instead of sending encoded photocurrent data to an external processor for interpretation, the sensor performs both spectral encoding and decoding at the point where light is detected.
The researchers call this approach in-sensor spectral sensing, and its central idea is to program electrical operations directly into the sensor’s photocurrent readout. Photocurrent is the flow of electrical charge generated when incoming photons interact with a light-sensitive material. In a conventional computational spectrometer, the detector collects signals that represent a mixture of wavelengths, while a processor later applies a mathematical reconstruction. In the new system, the readout itself is programmed to carry out operations that would normally be performed off-chip. The sensory terminal therefore does not merely acquire measurements; it transforms those measurements into wavelength-resolved information before transmission. This reduces the need to move large volumes of intermediate data and allows the output to be organized around the information an application actually needs.
According to the study by J. Wang, B. Ouyang, L. Fan and colleagues, the device can directly reconstruct optical spectra spanning wavelengths from 600 to 1,000 nanometres. This range covers much of the visible red region and extends into the near-infrared, where materials often display distinctive absorption and reflection patterns. The reported peak signal-to-noise ratio reaches 20.1 decibels at an average integration time of 1.05 seconds per scan. Signal-to-noise ratio is a key measure of spectral quality because it indicates how clearly a genuine optical signal can be distinguished from electronic fluctuations, environmental interference and other sources of uncertainty. A higher value generally allows more reliable identification of subtle differences between wavelengths, although performance also depends on illumination, material properties and measurement conditions.
The device’s operation can be understood as a form of hardware-level spectral inference. Incoming light produces photocurrent responses that contain information distributed across the measured wavelength range. Rather than recording every response as an undecoded data stream, the programmed scan operations process the current in a sequence designed to recover the spectrum. The resulting readout channels correspond to distinct wavelength channels, meaning that each channel is associated with a particular portion of the optical signal. This arrangement gives the sensor a direct relationship between electrical output and spectral information. The approach does not eliminate the physical challenge of distinguishing wavelengths; instead, it embeds part of the mathematical decoding process into the sensing hardware, bringing computation closer to the photon-to-electron conversion event.
That architecture also enables multispectral imaging. In an ordinary camera, each pixel is typically optimized for measuring brightness across a broad band, while colour or spectral information may require filters, multiple exposures or extensive computational processing. A sensor whose readout channels represent different wavelength bands can capture spatial and spectral information together. Different regions of an image can therefore be compared not only by their brightness but also by how they reflect or absorb selected portions of the spectrum. Such capabilities are relevant to material inspection, environmental monitoring, biomedical analysis and industrial quality control. The study’s results suggest that in-sensor decoding could make these systems more compact and reduce the latency associated with transferring and processing raw sensory data.
The researchers also demonstrate a more targeted use of the technology: spectroscopic recognition of ink. In this application, the sensor is not required to reconstruct a complete spectrum before a decision can be made. Instead, the photocurrent scan operations are programmed to extract the spectral features most useful for distinguishing inks. This is an example of application-specific sensing, in which the hardware is configured to produce a meaningful answer rather than a large general-purpose dataset. The sensor can consequently output information related to the task itself, such as whether a document shows signs of falsification, without requiring an external computer to first generate a full spectral record.
The reported document-authentication demonstration detects falsification in 0.12 seconds, a subsecond result that highlights the potential impact of moving decoding into the sensor. Document inks that appear visually identical under ordinary illumination can differ in their spectral behaviour, particularly outside the narrow range of human colour perception. A spectroscopic system can examine those differences and reveal whether printed or written regions were produced with matching materials. By programming the device to focus on the relevant signatures, the system reduces both response time and sensory data load. Instead of continuously streaming detailed spectral measurements for later analysis, it can deliver a compact, application-relevant result at the point of measurement.
This reduction in data movement is important because sensing systems are increasingly being deployed in settings where power, bandwidth and processing capacity are limited. Cameras and spectrometers used in mobile devices, robots, manufacturing lines or distributed monitoring networks may generate more information than can be efficiently transmitted. Off-chip processing introduces additional energy consumption and can create delays between acquisition and interpretation. In-sensor computation addresses that problem by performing selected operations locally, potentially allowing downstream electronics to receive only the reconstructed spectrum, a set of spectral channels or a classification outcome. The approach may therefore support faster feedback and more efficient hardware, although practical systems will still need to address calibration, device-to-device variation, illumination changes and the flexibility required when applications evolve.
The work points toward a broader shift in how sensory hardware is designed. For decades, sensors have primarily been treated as front-end components whose purpose is to preserve as much raw information as possible for later computation. The new strategy treats the sensor as an active information-processing terminal, capable of combining measurement with interpretation. Its demonstrated wavelength range, spectral reconstruction quality, multispectral readout and rapid ink-recognition operation show how the same platform can support both general spectral analysis and specialized recognition tasks. The researchers’ results do not make conventional spectrometers obsolete, particularly where maximum resolution, broad adaptability or detailed archival data are required. But they show that decoding need not always be a separate computational stage. By integrating programmed photocurrent operations with photodetection, spectral sensors could become faster, leaner and more directly connected to the decisions they are built to make.
Subject of Research: In-sensor spectral sensing and hardware-based spectral decoding for multispectral imaging and spectroscopic ink recognition.
Article Title: In-sensor spectral decoding for efficient spectroscopic recognition
Article References: Wang, J., Ouyang, B., Fan, L. et al. “In-sensor spectral decoding for efficient spectroscopic recognition.” Nature Sensors 1, 691–700 (2026). https://doi.org/10.1038/s44460-026-00085-5
Image Credits: AI Generated
DOI: 10.1038/s44460-026-00085-5
Keywords: in-sensor computing, spectral sensing, spectral decoding, computational spectrometry, multispectral imaging, photocurrent, near-infrared sensing, spectroscopic ink recognition, document authentication, sensor technology
