天線設計自動化方法：機器學習與最佳化研究

Abstract

天線設計是一個複雜的電磁逆問題，傳統上通常透過基於經驗的啟發式方法或耗時的參數化掃描來解決。本論文旨在實現天線設計流程的自動化和一般化，並比較和討論不同的方法。 研究首先探討並實現以離線機器學習代理人模型微調天線的幾何參數。我們的模型可以實時的根據環境的變化優化操作在2.45 GHz之PIFA天線。類似的離線模型在天線領域有許多其他應用，我們也將對其進行分析討論。其次，我們也嘗試利用像素化形式表示平面天線，從而實現對材料分佈的離散控制。機器學習的引入顯著的加速了優化的流程(MLAO)。這類採用MLAO架構的研究近來蓬勃發展，我們同樣大量分析並討論其方法的優劣。再者，為了應對可擴展性的挑戰，本文介紹了一種基於梯度的最佳化方法，利用帶有懲罰的固體各向同性材料 (SIMP) 方法和伴隨方法進行高效的靈敏度分析。伴隨方法顯著加快了收斂速度，即使在高維設計空間中也是如此。 論文最後討論了新興方法，例如強化學習、生成模型和基於物理的神經網絡等，這些方法或可透過實現即時自適應、更廣泛的設計空間探索和更低的模擬成本來進一步增強天線設計，具有很大的潛力。總而言之，本論文對天線設計自動化進行了深入且廣泛的探討，並統整出一個全面的理解架構，為未來數據驅動和基於物理的設計範式的整合奠定了基礎。

Antenna design is essentially an inverse electromagnetic problem, traditionally addressed through empirical heuristic methods or time-consuming parametric sweeps. This thesis aims to automate and generalize the antenna design process while comparing and discussing various approaches. The study first explores and implements the fine-tuning of antennas using an offline machine learning surrogate model. Our model can optimize a PIFA antenna operating at 2.45 GHz in real-time based on environmental changes. Similar offline model frameworks have numerous other applications in the antenna domain, which will also be analyzed and discussed. Secondly, we attempt to represent planar antennas in a pixelated form, enabling discrete control over material distribution. The introduction of machine learning significantly accelerates the optimization process. Research adopting the MLAO framework has recently flourished, and we extensively analyze and discuss the strengths and weaknesses of these methods. Furthermore, to address scalability challenges, this thesis introduces a well-developed gradient-based optimization approach, utilizing the Solid Isotropic Material with Penalization (SIMP) method and the adjoint method for efficient sensitivity analysis. The adjoint method markedly enhances convergence speed, even in high-dimensional design spaces. Finally, the thesis discusses emerging methods, such as reinforcement learning, generative models, and physics-based neural networks, which hold significant potential to further enhance antenna design by enabling real-time adaptation, broader design space exploration, and reduced simulation costs. In summary, this thesis provides a thorough and comprehensive exploration of automated antenna design, establishing a robust framework for understanding and laying the foundation for the integration of future data-driven and physics-based antenna design paradigms.