Agentic AI 驅動的超穎介面逆向設計方法

Abstract

超穎透鏡(Metalens)是一種極薄、多功能、高設計自由度的光學元件，但其設計依賴具備奈米光子學、電磁模擬與最佳化經驗之研究人員，設計流程通常耗時、計算成本高，且在多焦點或自由形狀光場等高自由度任務中容易受限於人為目標函數設計。因此我們提出一套結合 Agentic AI、電磁場代理模型與可微分反向設計之超穎透鏡自動化設計框架，能夠將使用者以自然語言描述之光學需求轉換為可執行之設計目標，並完成奈米柱結構最佳化。 本研究以深度學習電磁場代理模型取代重複性的全波模擬，並採用結合 U-Net 與 Transformer 之模型架構學習奈米柱結構至複數電磁近場之映射關係。為提升模型之物理解釋性，Maxwell 方程式被加入至損失函數中，使模型能維持電磁場的物理一致性。同時也將代理模型與角頻譜法結合，建立端到端可微分設計鏈，使遠場 Figure of Merit 的梯度能直接反向傳播至奈米柱半徑參數，實現以最終光場性能為目標的Metalens inverse design。 此外，本研究導入 Agentic AI 作為目標函數生成與設計修正模組。系統可解析自然語言需求，自動產生結構化 FoM specification，包含目標焦點、ROI 條件、背景抑制、強度平衡與二維目標圖案等設計項目，降低人工撰寫 loss function 與反覆試錯的需求。同時我們也將設計好的Metalens與傳統設計方法比較，發現在多焦點設計中，我們的方法能產生更集中的焦點、更均勻的能量分佈與較佳的側瓣抑制效果。透過五聚焦、雙聚焦與環形光場案例，本研究驗證了所提出框架在不同自然語言設計任務下之自動化能力。 本研究提供了一種結合專家代理、物理約束代理模型與可微分最佳化的超穎光學設計策略，為未來高自由度、多功能 Metalens 與 Metasurface 自動化設計的解決方案。

Metalens is an ultrathin, multifunctional optical element with a high degree of design freedom. However, its design process still relies heavily on researchers with expertise in nanophotonics, electromagnetic simulation, and optimization. Conventional design workflows are usually time-consuming and computationally expensive, and they are often limited by manually defined objective functions when dealing with high-degree-of-freedom tasks such as multi-focus generation or free-form optical field design. Therefore, this study proposes an automated Metalens design framework that integrates Agentic AI, an electromagnetic field surrogate model, and differentiable inverse design. The proposed framework can convert optical requirements described by users in natural language into executable design objectives and optimize the corresponding nanopillar structures. In this study, a deep-learning-based electromagnetic field surrogate model is used to replace repetitive full-wave simulations. A hybrid architecture combining U-Net and Transformer is adopted to learn the mapping relationship from nanopillar structures to complex electromagnetic near fields. To improve the physical interpretability of the model, Maxwell’s equations are incorporated into the loss function, allowing the model to maintain the physical consistency of electromagnetic fields. In addition, the surrogate model is combined with the Angular Spectrum Method to establish an end-to-end differentiable design pipeline. Through this pipeline, the gradient of the far-field Figure of Merit can be directly backpropagated to the nanopillar radius parameters, enabling Metalens inverse design based on the final optical field performance. Furthermore, Agentic AI is introduced as a module for objective-function generation and design refinement. The system can parse natural-language requirements and automatically generate a structured FoM specification, including target focal points, ROI conditions, background suppression, intensity balance, and two-dimensional target patterns. This reduces the need for manually writing loss functions and repeated trial-and-error. The designed Metalenses are also compared with those obtained using conventional design methods. In multi-focus design, the proposed method produces more concentrated focal spots, more uniform energy distribution, and better side-lobe suppression. Through five-focus, dual-focus, and ring-shaped optical field cases, this study verifies the automation capability of the proposed framework under different natural-language design tasks. This study provides a Metasurface optical design strategy that combines expert agents, physics-constrained surrogate modeling, and differentiable optimization, offering a potential solution for the automated design of high-degree-of-freedom and multifunctional Metalenses and Metasurfaces.