Scholarlink

<p>This paper proposes an AlphaGo-inspired algorithm for automated antenna generation. Analogous to how AlphaGo navigates the discrete move-space of a board game, our framework treats antenna design as a sequential decision-making process. With an innovated topology generation rule, a Monte Carlo tree search-based optimization is applied to guide the moving steps to approach good designs. During the process, a machine learning surrogate model is trained and continuously refined to predict antenna performance from design topology, enabling an online optimization loop with progressively improving accuracy. Finally, postprocessing is applied to fine-tune continuous geometric parameters, which is necessary in most cases to meet the final design specifications. This approach reduces the need for manually crafted initial antenna topologies within the demonstrated class of single-layer planar designs, while maintaining sufficient flexibility to explore design spaces for complex objectives. Compared with current popular machine learning-based methods, the proposed method achieves effective designs with the number of evaluations on the 10<sup>3</sup> scale, and the resulting geometries are potentially more fabrication-friendly. Its effectiveness and efficiency are validated through multiple representative antenna design examples.</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

AlphaGo-driven generative machine learning framework for inverse topology synthesis and optimization of planar antennas

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月9日

查看全文
  • Xiaobo Wang,
  • Suomin Cui,
  • Moein Nazari,
  • Gang Kang,
  • Jian Liu

摘要

This paper proposes an AlphaGo-inspired algorithm for automated antenna generation. Analogous to how AlphaGo navigates the discrete move-space of a board game, our framework treats antenna design as a sequential decision-making process. With an innovated topology generation rule, a Monte Carlo tree search-based optimization is applied to guide the moving steps to approach good designs. During the process, a machine learning surrogate model is trained and continuously refined to predict antenna performance from design topology, enabling an online optimization loop with progressively improving accuracy. Finally, postprocessing is applied to fine-tune continuous geometric parameters, which is necessary in most cases to meet the final design specifications. This approach reduces the need for manually crafted initial antenna topologies within the demonstrated class of single-layer planar designs, while maintaining sufficient flexibility to explore design spaces for complex objectives. Compared with current popular machine learning-based methods, the proposed method achieves effective designs with the number of evaluations on the 103 scale, and the resulting geometries are potentially more fabrication-friendly. Its effectiveness and efficiency are validated through multiple representative antenna design examples.