倒傳遞類神經網路為目前所知精度較高，且最常被採用作為建構模型之前饋式網路;但其最大之缺點為訓練時間較長，較不易達成即時性之需求。為解決此問題，本研究改採簡易循環式網路。此法可大幅提昇網路運作效率，且不致影響預測精度。參數篩選為資料前處理之重要一環，可有效提昇預測精度及運作效率。然而，為使虛擬量測模型能保有機台原有的特性，所有機台之重要參數均不可輕易刪除。為兼顧精度之提昇，且不刪除所有機台之任一重要參數，本研究提出一套提昇虛擬量測精度之方法與架構，將影響預測精度之參數進行篩選，並以鈍化之方式取代刪除的效用。此作法除能有效提昇預測精度外，亦可避免因刪除參數所造成在虛擬量測運作階段，該參數無法在系統上顯示任何異常資訊，以致機台發生故障時，無法針對該參數進行調機動作之問題，如此可避免或減少因機台異常造成晶圓報廢之損失。本研究採用二個案例製程進行參數鈍化實驗，其結果均顯示參數鈍化對於精度提昇效果比刪除參數更加明顯。為驗證鈍化實驗對於精度提昇確有其顯著效果，本研究利用Statistica 6.0統計軟體進行虛擬量測預測精度可信度檢定，結果顯示在顯著水準α為0.05時，二個案例製程計三種參數鈍化實驗之p-value均小於0.05，表示該鈍化實驗具有顯著性。

Back propagation neural network (BPNN) is one of the best known feedforward neural networks for high precision performance and model establishment. However, the disadvantage is that BPNN requires a tremendous period of training time; thus it cannot fulfill the real-time implementation requirements. To solve the problem mentioned above, in this work, the simple recurrent neural network (SRN) is adopted, which greatly reduces the training time and shows no impact to conjecture precision.
Parameter sifting is essential to data preprocess, which can improve conjecture precision and operating efficiency. However, in order to make the conjecture model keep track of all the variations of equipment properties, any equipment parameter cannot be deleted without thorough considerations. For these two concerns, a parameter-blunting method for virtual-metrology precision enhancement is proposed. By applying this method, those immaterial parameters will be blunted instead of sifting. As such, the effect of parameter sifting remains while maintaining monitoring all of the equipment parameter variations. Two illustrative examples are included in this work. Both of the examples show that the parameter-blunting method is better than the parameter-sifting method for enhancing virtual metrology precision. Statistica 6.0 is adopted to perform tests of significance for the proposed parameter-blunting method concerning the enhancement of virtual metrology precision. With the significance level α being 0.05, the p-values of three parameter-blunting experiments in each of the two examples are all smaller than 0.05. This reveals that the parameter-blunting method has certain significance indeed.

[1]A. C. Diebold,“Overview of Metrology Requirements based on the 1994 National Technology Roadmap for Semiconductor,” in Proc. of IEEE/SEMI Advanced Semiconductor Manufacturing Conference and Workshop, pp. 50-60, November 1995.
[2]2003 International Technology Roadmap for Semiconductors (ITRS), December 2003. http://public.itrs.net/.
[3]Y.-C. Chang and F.-T. Cheng, “Application Development of Virtual Metrology in Semiconductor Industry,” in Pro. of the 31st Annual Conference of the IEEE Industrial Electronics (IECON 2005), Raleigh, North Carolina, USA, pp. 124-129, November 2005.
[4]Y.-C. Su, M.-H. Hung, F.-T. Cheng and Y.-T. Chen, “A Processing Quality Prognostics Scheme for Plasma Sputtering in TFT-LCD Manufacturing,” IEEE Transactions on Semiconductor Manufacturing , vol. 19, issue 2, pp. 183-194, 2006.
[5]D. E. Rumelhart, G. E. Hinton and R. J. Williams, “Learning Representations by Back-propagating Errors,” Nature, vol. 323, pp. 533 -536, 1986.
[6]J. L. Elman, “Finding Structure in Time,” Cognitive Science, vol. 14, pp.179-211, 1990.
[7]S. Huang, and H. Zhang , “Neural Network in Manufacturing:A Survey, “ IEEE/CPMT International Electronics Manufacturing Technology Symposium, pp.177-190, 1993.
[8]葉怡成「應用類神經網路」，儒林圖書公司，1997.
[9]葉怡成「類神經網路模式應用與實作」，儒林圖書公司，2003.
[10]S. Haykin, “Neural Networks, A Comprehensive Foundation,”Prentice Hall, NJ:07458, 1999.
[11]羅華強「類神經網路，MATLAB的應用」，清蔚科技，2001.
[12]Hogg and Tanis, “Probability and Statistical Inference,” Prentice Hall, NJ:07458, 1997.
[13]羅正忠、張鼎張「半導體製程技術導論」.
[14]Matlab 7.0, The MathWork Corporation. http://www.mathworks.com/
[15]Statistica 6.0, The StatSoft Corporation. http://www.statsoft.com
[16]A. C. Tsakoumis, S. S. Vladov, V. M. Mladenov, “Electric Load Forecasting With Multilayer Perceptron And Elman Neural Network,” in Proc. of Neural Network Applications in Electrical Engineering, pp. 87 - 90, September 2002.
[17]B. Kim et al., “Prediction of plasma processes using neural network and genetic algorithm,” Solid State Electronics, vol. 49, pp. 1576-1580, 2005.
[18]X. Shi, Y. C. Liang, X. Xu, “An improved Elman model and recurrent back-propagation control neural networks.” Journal of Software, 2003,14(6): 1110~1119.
[19]Yao Yu, Yu Ge, Gao Fu-Xiang, “A Neural Network Approach for Misuse and Anomaly Intrusion Detection.” Wuhan University Journal of Natural Sciences Vol. 10 No 2, pp. 115-118, 2005.
[20]林嶸銓「通用型虛擬量測機制與系統」，國立成功大學製造工程研究所碩士論文，2005.
[21]張斐章、張麗秋「類神經網路」，東華書局，2004.
[22]陳順宇、鄭碧娥 「統計學」，四版，華泰書局，2004.
[23]張智星「MATLAB程式設計與應用」，清蔚科技，2004.