Scholarlink

<p>As a paradigm of 2D artificial composite electromagnetic media, metasurface architectures exhibit cross-scale synergistic regulation properties. Their core mechanism manifests as multi-dimensional electromagnetic parameter coupling across physical domains, encompassing nonlinear response processes including wavefront reconstruction (amplitude modulation), phase transition (wavefront manipulation), and polarization dimension conversion. However, system implementation requires not only interdisciplinary knowledge integration but also emergent property modeling based on complex system theory, involving nonlinear iterative computations along topological optimization paths. Deep learning, as a core implementation pathway for computational intelligence, establishes nonlinear mapping models through high-dimensional feature deconstruction. Recent extensions of this framework into metasurface inverse design demonstrate progressive innovation. Existing neural network architectures transform physical space mapping into cross-modal projection systems, achieving inverse modeling of topological optimization through hypersurface correlations between electromagnetic response spectra and subwavelength structures. However, this method lacks freedom in the structural design of metasurfaces, as it often limits the optimization process to a set of pre-designed choices, making it difficult to create completely new metasurface designs. In this study, we designed a deep learning scheme with variational auto-encoders (VAE) as the core, which demonstrated excellent innovative capabilities and was able to effectively generate unprecedented graphic patterns. Compared with traditional neural networks, this VAE encoder-based method has unique advantages in the reverse design process. It not only significantly improves design accuracy but also performs outstandingly in generalization performance, bringing new breakthroughs and possibilities for the generation and design of graphic patterns, and effectively promoting the development and progress of related fields.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Reverse Design of Electromagnetic Response for Metasurface Based on Deep Learning

  • 期刊论文

  • 发表日期：

    2025年9月12日

查看全文
  • Xuan Xu,
  • Huizhen Feng,
  • Manna Gu,
  • Chenxia Li,
  • Ying Tian

摘要

As a paradigm of 2D artificial composite electromagnetic media, metasurface architectures exhibit cross-scale synergistic regulation properties. Their core mechanism manifests as multi-dimensional electromagnetic parameter coupling across physical domains, encompassing nonlinear response processes including wavefront reconstruction (amplitude modulation), phase transition (wavefront manipulation), and polarization dimension conversion. However, system implementation requires not only interdisciplinary knowledge integration but also emergent property modeling based on complex system theory, involving nonlinear iterative computations along topological optimization paths. Deep learning, as a core implementation pathway for computational intelligence, establishes nonlinear mapping models through high-dimensional feature deconstruction. Recent extensions of this framework into metasurface inverse design demonstrate progressive innovation. Existing neural network architectures transform physical space mapping into cross-modal projection systems, achieving inverse modeling of topological optimization through hypersurface correlations between electromagnetic response spectra and subwavelength structures. However, this method lacks freedom in the structural design of metasurfaces, as it often limits the optimization process to a set of pre-designed choices, making it difficult to create completely new metasurface designs. In this study, we designed a deep learning scheme with variational auto-encoders (VAE) as the core, which demonstrated excellent innovative capabilities and was able to effectively generate unprecedented graphic patterns. Compared with traditional neural networks, this VAE encoder-based method has unique advantages in the reverse design process. It not only significantly improves design accuracy but also performs outstandingly in generalization performance, bringing new breakthroughs and possibilities for the generation and design of graphic patterns, and effectively promoting the development and progress of related fields.