在2011年德國政府提出工業4.0高科技計畫後，世界各國為了保持其優勢皆開始以智慧工廠為目標輔導企業轉型；其中虛擬量測的技術在半導體領域中扮演著重要的角色，虛擬量測可幫助提升產品良率、降低量測成本、提高產出率並且將產品量測從抽樣測量提升為全面測量；而虛擬量測在業界已有一定程度的發展，各公司皆已發展出各自的架構，但現行使用的模型為人工設定的數學模型居多，人工設定需花費大量時間，在相關參數的選擇上亦需倚賴製程人員的專業知識選擇，相對而言容易造成人為誤差，並且無法針對複雜的量測參數進行預測。 故本研究提出以現有架構為基準，並以深度學習模型取代現有人工設定模型的改良做法，以減少人力花費，並降低模型對製程人員專業知識的倚賴程度，且能支援複雜的量測參數預測，本研究以Pure NN結合Rule Based Improved NN加上Weight NN的深度學習模型架構進行訓練，深度學習數學模型平均誤差值為0.044相較於人工設定數學模型的平均誤差值0.05，有更準確的預測結果，證明以深度學習模型取代目前人工設定數學模型的可行性。 ;After the German government proposed the Industry 4.0 high-tech project in 2011, all countries in the world began to guide the transformation of enterprises with the goal of smart factories in order to maintain their advantages. The technology of Virtual metrology plays an important role in the semiconductor field, and Virtual metrology can help improve product yield, reduce measurement cost, increase output rate and upgrade product measurement from sampling measurement to comprehensive measurement; Virtual metrology has developed to a certain extent in the industry, and each company has developed its own architecture, but the currently used models are mostly mathematical models that are manually set. It takes a lot of time to manually set the parameters. The selection of relevant parameters also depends on the selection of the expertise of the process personnel. It is relatively easy to cause personal error and cannot be complicated. And the mathematical models can’t predict complicated parameters. Therefore, this study proposes an improved approach based on the existing architecture and replaces the existing manual setting models with deep learning models to reduce labor costs reduce and the model′s reliance on process expertise, and support complex measurement parameter prediction. In this study, Pure NN is combined with Rule Based Improved NN plus Weight NN′s deep learning model architectures. The average error value of deep learning mathematical model is 0.044 compared with the average error value of manually set mathematical model of 0.05, which has higher accurate prediction. As a result, it proves that the feasibility of the manual setting mathematical models is replaced by deep learning models.
