Neural Network-Assisted End-to-End Inverse Design for Polarimetric Microspectrometer

Metasurfaces represent a cutting-edge platform for the miniaturization of spectral and polarimetric sensing applications. However, achieving multi-channel control has been a significant challenge for conventional design methodologies. This work introduces an end-to-end multi-channel inverse design framework capable of directly optimizing metasurface geometry for precise optical field manipulation. A cascaded metasurface system can be conceptualized as an optical diffractive neural network. By developing a deep neural network that correlates the optical response of meta-atoms with their structural properties, a differentiable pipeline is established to drive metasurface designs toward multi-functional requirements. Leveraging this approach, a polarization-independent microspectrometer is designed, which is unaffected by the polarization state of incident light and theoretically has a spectral resolution down to 2 nanometers across the visible spectrum. Furthermore, a polarization-dependent microspectrometer is demonstrated to simultaneously capture both spectral and polarization information from a single exposure. This device maintains the same high spectral resolution and also provides precise determination of the polarization angle. Such polarization-sensitive microspectrometer has potential applications in the detection of chiral substances. The results will accelerate the progress of meta-optics, with implications for spectro-polarimetric detection, multi-target holographic displays, parallel optical information processing, and optical computing.