Dynamic multifunctional metasurfaces: an inverse design deep learning approach

Abstract

Optical metasurfaces (OMs) offer unprecedented control over electromagnetic waves, enabling advanced optical multiplexing. The emergence of deep learning has opened new avenues for designing OMs. However, existing deep learning methods for OMs primarily focus on forward design, which limits their design capabilities, lacks global optimization, and relies on prior knowledge. Additionally, most OMs are static, with fixed functionalities once processed. To overcome these limitations, we propose an inverse design deep learning method for dynamic OMs. Our approach comprises a forward prediction network and an inverse retrieval network. The forward prediction network establishes a mapping between meta-unit structure parameters and reflectance spectra. The inverse retrieval network generates a library of meta-unit structure parameters based on target requirements, enabling end-to-end design of OMs. By incorporating the dynamic tunability of the phase change material Sb2Te3 with inverse design deep learning, we achieve the design and verification of dynamic multifunctional OMs. Our results demonstrate OMs with multiple information channels and encryption capabilities that can realize multiple physical field optical modulation functions. When Sb2Te3 is in the amorphous state, near-field nano-printing based on meta-unit amplitude modulation is achieved for X-polarized incident light, while holographic imaging based on meta-unit phase modulation is realized for circularly polarized light. In the crystalline state, the encrypted information remains secure even with the correct polarization input, achieving double encryption. This research points towards ultra-compact, high-capacity, and highly secure information storage approaches.