Spectral-driven bidirectional design framework for polarization-selective structural colors

The design of structural color metasurfaces typically relies on computationally expensive full-wave electromagnetic simulations and empirical parameter tuning, making it challenging to achieve efficient optimization in high-dimensional parameter spaces. We propose a forward-inverse neural network closed-loop framework for polarization-selective structural color design. In the forward modeling stage, a spectrally resolved multilayer perceptron is developed to explicitly learn the continuous mapping between structural parameters and reflection spectra, thereby enabling the joint optimization of spectral fidelity and perceptual color consistency. In the inverse design stage, a mixture density network (MDN) is introduced to model the multimodal distribution of structural parameters conditioned on a given target performance under different polarization conditions. The results demonstrate that the proposed method achieves excellent performance. After optimization, approximately 97% of the samples satisfy ΔE ≤ 1, with good spectral consistency maintained, thereby validating an effective balance between perceptual accuracy and physical fidelity. This work provides an efficient and reliable solution for the rapid inverse design of structural color, with potential applications in polarization-controlled displays and optical information encoding.