結合物理知識與機器學習之應用：廠務真空壓力控制

Application of Physics-Informed Machine Learning in Facility Vacuum Pressure Control

真空系統為半導體製程與先進封裝之關鍵基礎設施，其壓力穩定度直接影響製程均勻性、污染控制與缺陷形成。在共享前級管路／歧管（shared foreline／manifold）架構之sub-fab系統中，多腔體耦合運作易放大壓力瞬態，特別是在材料加熱或固化過程中所產生之放氣（outgassing）尖峰，將導致壓力回升、回復時間延長與泵浦hunting，進而提高微孔洞（void）缺陷風險並縮窄製程視窗。
本研究提出一套物理知識輔助（physics-informed）之AI真空壓力控制架構，結合集中參數動態模型、導通度（conductance）限制與有效抽速約束，並透過時間序列模型預測壓力變化與放氣事件。進一步採用混合式控制策略（AI bias +模型預測控制，MPC），於安全邊界內調節節流閥與泵浦轉速，實現前饋補償以抑制壓力尖峰。
此外，系統整合異常診斷與預知保養功能，可辨識漏氣、管路阻塞與泵浦劣化等狀態，並透過shadow mode進行漸進式導入，以降低實際部署風險。實驗結果顯示，相較於傳統PID控制，所提方法可顯著降低壓力波動（ΔP）、壓力變化速率（|dP／dt|）及回復時間，並同步降低能耗與void風險代理指標。
本研究建立一套結合物理模型、AI預測與最佳化控制之整合框架，可有效提升共享真空系統之穩定性，並為先進封裝製程之缺陷風險控制提供具體可行之技術途徑。

Vacuum systems are critical infrastructures in semiconductor manufacturing and advanced packaging, where pressure stability directly impacts process uniformity, contamination control, and defect formation. In sub-fab configurations with shared foreline/manifold architectures, multi-chamber coupling amplifies pressure transients, particularly during outgassing bursts induced by heating or curing processes. These transient gas loads propagate through the manifold, resulting in pressure excursions, prolonged recovery time, and pump hunting, which in turn increase the risk of void formation and narrow the process window.
This study proposes a physics-informed AI-based vacuum pressure control framework that integrates lumped-parameter dynamic models, conductance limitations, and effective pumping constraints. A time-series prediction model is employed to forecast short-term pressure evolution and outgassing events. A hybrid control strategy combining AI bias and model predictive control (MPC) is then applied to regulate throttle valves and pump speeds within safety constraints, enabling feedforward compensation to suppress pressure transients.
The framework further incorporates anomaly detection and predictive maintenance capabilities to identify system abnormalities such as leaks, clogging, and pump degradation. A staged deployment strategy using shadow mode is introduced to ensure safe and gradual system integration. Experimental results demonstrate that, compared with conventional PID control, the proposed method significantly reduces pressure fluctuation (ΔP), pressure variation rate (|dP/dt|), and recovery time, while also lowering energy consumption and void risk proxies.
This work establishes an integrated framework combining physical modeling, AI prediction, and optimal control, providing a practical solution for enhancing pressure stability and mitigating defect risks in shared vacuum systems for advanced semiconductor packaging applications.