High-accuracy forward prediction and inverse design of deep-learning-assisted reflecting color filters

Structural colors generated from the reflection of light at subwavelength dimensions have drawn a lot of research interest because of their ability to replace the conventional pigment-based color filters. In this work, a CMOS-compatible multilayer Si-SiO 2 -Si dielectric metasurface is proposed for structural color filtering, enabling vivid color generation with a substantially extended color gamut across the CIE 1931 chromaticity diagram. Unlike previously reported single-layer or material-limited designs, the proposed multilayer architecture provides enhanced color tunability and scalability, underscoring its potential for practical implementation in next-generation photonic and optoelectronic applications. Here, a deep neural network is developed to replace the traditional finite-difference time-domain (FDTD) simulation method, resulting in significantly reduced power consumption and less computational time (nearly 8 ms per epoch). The proposed forward deep neural network (FDNN) can predict the color with an accuracy of 99.8% from the structural input parameters. Furthermore, a tandem network, capable of showing an accuracy of over 99%, is constructed to address the issue of the non-uniqueness problem of the inverse deep neural network (IDNN). The proposed metasurface is a reflection-type and gives better performance in the visible range. Therefore, our research illustrates a simpler technique to produce dynamic structural color tuning and may have potential applications in dynamic display, anticounterfeiting, and micro-nanophotonic devices.