Deep neural network for color prediction from a dielectric metasurface and inverse design of corresponding color filter

We propose a dielectric nanostructure that produces high-resolution colors and a 162.65% larger color gamut in CIE compared to sRGB by modifying its structural parameters and polarization angle. However, designing a specific color by iteratively optimizing the geometric parameters and polarization is computationally expensive, and it is quite challenging to obtain millions of distinct hues. In this instance, a deep neural model is trained which precisely forecasts colors through a forward modeling network with an accuracy exceeding 99%. Additionally, it solves the nonuniqueness problem of the inverse design process so that it can accurately predict certain device parameters from the necessary colors with an error value of 1.18×10 -4 . Our main findings demonstrate that deep learning is an effective technique for accelerating the color design process and substantially boosting the efficiency of nanophotonic design. A neural network is integrated here to operate with five photonic parameters within the tunable color filter platform. This approach holds promise for both practical commercial applications and future photonics research.