Abstract Near‐infrared (NIR) meta‐optics often suffer from a limited field of view (FOV) and bandwidth in compact designs. While metasurfaces offer high design freedom in the lateral dimension, computational methods provide more flexibility for achieving precise control over the thickness dimension, offering a superior approach to reach the physical limits of the system. Here, a computational imaging method based on physics‐constrained embedded training, which integrates intrinsic physical constraints, including chromatic dispersion and fabrication tolerances, into the forward imaging simulation and neural network training, is proposed. This approach is demonstrated through the implementation of a NIR camera featuring a 6.8‐mm‐diameter monochromatic metalens, which, when coupled with a nonlinear activation‐free generative adversarial network, achieves broadband achromatic imaging with a FOV of 78° across the 800–1000 nm spectral range. The camera successfully captures real‐world scenes with high fidelity, enabling applications like vein detection.
Zhou et al. (2026) studied this question.