AI 輔助貼片天線設計AI-Assisted Patch Antenna Design

電磁模擬器不可微分，且自由拓樸的金屬圖樣搜索空間極大（25×25 二值像素 = 2⁶²⁵）。本研究以梯度估計網路（GEN）將模擬轉為可優化，並提出三項機制解決局部解、金屬孤島與樣本效率問題。EM simulators are non-differentiable and the free-topology search space is enormous (a 25×25 binary grid = 2⁶²⁵). A Gradient Estimation Network makes simulation optimizable, and three mechanisms address local optima, metal islands, and sample efficiency.

電磁模擬器不可微分，且自由拓樸的金屬圖樣搜索空間極大（25×25 二值像素 = 2⁶²⁵ 種組合）。以梯度估計網路（GEN）將模擬轉為可優化雖可行，但純梯度法有三大缺陷：收斂停滯於局部解、產生破壞電流路徑的金屬孤島、以及冷啟動下樣本效率低落。後處理修補孤島屬被動策略，易破壞有效金屬集群。 EM simulators are non-differentiable, and the free-topology search space is enormous (a 25×25 binary grid = 2⁶²⁵ combinations). A Gradient Estimation Network (GEN) makes simulation optimizable, but a purely gradient-driven approach has three defects: convergence stalls in local optima, disconnected metal islands that break the resonant current path, and poor sample efficiency under cold-start. Post-hoc island repair is passive and can destroy electromagnetically useful metal.

• 生成器 + 可微代理模型（GEN）：皆為 MLP；代理模型每輪由回放緩衝區線上即時更新，無需預蒐集資料集。Generator + differentiable surrogate (GEN): both MLPs; the surrogate is updated online each iteration from a replay buffer — no pre-collected dataset.
• ACP 自適應循環策略：以主動式高原偵測 + 自適應重啟取代固定週期排程，並將學習率與二值化溫度 τ 綁在同一退火曲線，於單一階段內兼顧探索與邊界銳化。ACP (Adaptive Cyclical Policy): plateau detection + adaptive restart replaces fixed-period scheduling; couples learning rate with a binarization temperature τ on one annealing curve, balancing exploration and boundary-sharpening in a single stage.
• SC Loss 圖譜連通損失：以圖拉普拉斯的代數連通度（Fiedler 值 λ₂）倒數作為端到端可微的拓樸懲罰，於訓練期主動消除孤島，免去後處理。SC Loss (Spectral Connectivity Loss): uses the reciprocal of the graph Laplacian's algebraic connectivity (Fiedler value λ₂) as a differentiable topology penalty, eliminating islands during training instead of post-processing.
• DLF 動態損失過濾：經驗回放機制，以歷史最佳損失為自適應門檻，門檻隨模型提升收緊，在不增加模擬次數下提升樣本資訊密度。DLF (Dynamic Loss Filter): experience replay whose elite threshold tracks the running-best loss, raising sample information density without extra EM simulations.
• 可微二值化 + 物理熱啟動：溫度控制 sigmoid 搭配直通估計（STE）；以傳輸線理論推算 28 GHz 初始幾何取代隨機初始化。Differentiable binarization + physics warm-start: temperature-controlled sigmoid with a straight-through estimator; a transmission-line-theory initial geometry replaces random init.

頻段Band

28 GHz（5G NR n257 毫米波）28 GHz (5G NR n257, mmWave)

基板Substrate

Rogers RO4003C（ε_r 3.55, h 0.508 mm）

設計區Design region

5×5 mm · 25×25 像素5×5 mm · 25×25 pixels

模擬器Simulator

Ansys HFSS 2023 R2

運算Compute

純 CPU（i9-14900 · 128 GB · 無 GPU）· PyTorch 2.5CPU-only (i9-14900 · 128 GB · no GPU) · PyTorch 2.5

• ACP&SC&DLF 全系統：響應損失壓至 0.99 dB，同時饋入連通性 R_feed = 61.98%，電磁性能與可製造性雙優。Full ACP&SC&DLF: response loss down to 0.99 dB with feed reachability R_feed = 61.98% — best on both performance and manufacturability.
• 連通性：SC Loss 將饋入連通性由約 18.89% 提升至 60%+（高權重可達 98%）。Connectivity: SC Loss raises feed reachability from ~18.89% to 60%+ (up to 98% at high weight).
• 收斂：DLF 使最小損失較基準改善逾 50%。Convergence: DLF cuts the minimum loss by over 50% vs. baseline.
• 泛用性：同一框架不調超參即生成合規的雙埠濾波器，驗證規格無關性。Generality: the same framework, untuned, also generates a spec-compliant dual-port filter.