In-sensor spectral decoding for efficient spectroscopic recognition

Conventional spectrometers rely on optical encoding and computational decoding to resolve spectral information. Although computational spectrometers merge spectral encoding with photodetection, the decoding still depends on computationally intensive off-chip processing. Here we introduce in-sensor spectral sensing that unifies both encoding and decoding at the sensory terminal. By programming photocurrent scan operations directly into the device, the sensor performs in-sensor decoding and eliminates off-chip spectral processing. The device directly reconstructs light spectra from 600 to 1,000 nm, achieving a 20.1 dB peak signal-to-noise ratio at a 1.05 s average integration time per scan. Each readout channel corresponds to a distinct wavelength channel, enabling multispectral imaging. We further program the decoding operations for spectroscopic ink recognition, allowing the sensor to output application-relevant semantics without reconstructing a full spectrum. The operations facilitate subsecond (0.12 s) document falsification detection and demonstrate how moving decoding to the sensory terminal reduces sensory data load for efficient spectral sensing.