Spectrum generally refers to the electromagnetic wave spectrum in the wavelength range from ultraviolet to infrared bands, containing rich information about the interaction between matter and light waves. Spectrum is also called the "fingerprint" of matter. Spectral imaging can obtain three-dimensional data cubes containing the spectral information of each point in the two-dimensional image, which surpasses the perception ability of human eyes, and it thus has important application prospects in many fields such as disease diagnosis, precision agriculture, food safety, astronomical detection, and face recognition.

According to the methods of data acquisition, spectral imaging can be divided into four categories: point scanning type, line scanning type, wavelength scanning type, and snapshot type. Traditional spectral imaging technology generally adopts the mode of spatial scanning or wavelength scanning, which fails to obtain the real-time spectral information of each pixel in the field of vision. In recent years, new single-point spectrometers based on computational spectral reconstruction have made a breakthrough in miniaturization, but there is no report about snapshot spectral imaging based on the above scheme. There is no research scheme for snapshot spectral imaging that can achieve high spectral accuracy, high spatial resolution, and high imaging speed simultaneously.

For the spectral imaging scheme based on metasurfaces, different metasurface units with different structure parameters are designed to realize rich broadband modulation on the spectra of incident light at each spatial point. The modulated light signal is detected by the image sensor, and the spectral information of incident light is obtained by computational reconstruction. The number of metasurface units can be significantly smaller than that of wavelength channels, which effectively reduces the volume of a single microspectrometer. Spectral imaging can be realized through the periodic array of the computational spectrometer, which has the advantages of high design freedom, high integration density, and low-cost mass production.

In 2022, we reported the world's first real-time ultraspectral imaging chip based on regularly shaped metasurface units. The designed metasurface units contain five types: round hole, square hole, cross hole, and square and cross hole after 45 degrees of rotation (Fig. 4). The real-time ultraspectral imaging chip reduces the size of a single-point spectrometer to less than 100 μm and can obtain spectral information of more than 150000 spatial points in a single shot. In other words, more than 150000 (356×436) micro spectrometers are integrated on a chip with a size of 0.5 cm², and the operational wavelength band of each microspectrometer is 450-750 nm. The measured wavelength accuracy of monochromatic light is 0.04 nm, and the spectral resolution is up to 0.8 nm. In order to break through the design restriction of regular shapes, we propose a design method of freeform-shaped metasurface units. The freeform shapes are generated by grid partitioning, random distribution of grid values, filtering, and binarization. The corresponding design freedom is expanded by 2-3 orders of magnitude compared with that of regular shapes. Thanks to the expansion of design space, the performance of ultraspectral imaging chip based on freeform-shaped metasurface units is further improved, with a wavelength resolution up to 0.5 nm (Fig. 5). In terms of spectral image reconstruction algorithm, we propose to use ADMM-net, a deep unrolled neural network based on ADMM iterative algorithm, to realize fast spectral image reconstruction. A single reconstruction only takes 18 ms, and the reconstruction speed is improved by about 5 orders of magnitude compared with the traditional point-by-point iterative spectral reconstruction algorithm. We also discuss the application prospects of metasurface spectral imaging chips in the brain imaging of living rats, face anti-counterfeiting recognition, automatic driving, and other fields

We summarize the work related to metasurface spectral imaging chips from the basic principles, structural design, reconstruction algorithms, and potential applications. In the future, metasurface spectral imaging chips with the advantages of high precision, low cost, and mass production are expected to become the basis for the development of artificial intelligence and big data. Further optimization directions of metasurface spectral imaging chips include improving the spectral image reconstruction algorithm and reducing the angle sensitivity of metasurface units.