深度学习辅助电磁超表面设计:算法演进及应用OA

To address the challenges in electromagnetic metasurface design,such as the high dimensionality of design space,the heavy computational cost of full-wave simulations,and the optimization difficulty caused by the coupling of electromagnetic,thermal,and device-level constraints,the methodological evolution and application progress of artificial intelligence-assisted metasurface design are reviewed,aiming to provide references for the efficient design and practical implementation of electromagnetic metasurfaces.From a methodological perspective,existing studies can be broadly categorized into five core paradigms:fast forward prediction based on supervised surrogate models;inverse solving and multi-solution generation using cascaded/bidirectional frameworks and invertible networks;structural prior learning and global exploration enabled by generative models such as Variational Autoencoders(VAE)and diffusion models;physics-guided learning for improved physical consistency and cross-condition generalization;and closed-loop automation driven by reinforcement learning,Large Language Model(LLM),and multi-agent systems through a generation-verification-feedback process.From an application perspective,related studies mainly focus on static meta-optical tasks,reconfigurable/programmable metasurface design,and robust design and engineering closed loops under multiphysics coupling.For representative scenarios such as structural color design,thermal radiation regulation,programmable beam shaping,and freeform shape design,the applicability of different algorithms under system-level metrics,multi-constraint conditions,and multi-solution spaces is further summarized,and the pathways toward multiphysics collaborative optimization and trustworthy validation for real-world systems are discussed.