應用類神經網路預測TFT-LCD生產良率

Using Artifical Neural Network for TFT-LCD Yield Prediction

陳仕偉(Shih-Wei Chen)

指導教授 ： 皮世明

中原大學/商學院/資訊管理研究所/碩士(2012年)

https://doi.org/10.6840/cycu201200348

摘要

台灣是國際上TFT-LCD(Thin-Film Transistor Liquid-Crystal Display，薄膜電晶體液晶顯示器)主要的製造和生產基地。由於TFT-LCD在生產設備和儀器的投資金額相當廣大，因此如何降低製造的時間，維持一定的品質，縮短產品的生命週期和加速新產品的開發，比競爭對手更早推出產品在市場上，是企業生存的必要條件。而良率（Yield）是反應此一關鍵產業中整體技術與企業獲利高低之綜合代表，所以在眾多衡量指標中，其重要性自不言而喻。簡單的說，良率可定義成產出之良品佔所有投入生產總數的百分比；而良率管理（Yield Management）是指針對從整個TFT-LCD製造過程中產生龐大資料所作之整合分析、良率改善與預測（Yield Prediction）等。由於生產投料、交貨排程以及工程問題界定釐清等方面皆須仰賴一有效的預測模式，因此良率預測已漸漸成為該產業中重要之議題。
如上述理由，本論文嘗試提出一有效且簡單之良率預測模式：即以TFT-LCD面板陣列資料為基礎，運用類神經倒傳遞網路（Artificial Neural Back-Propagation Network）所具有學習、平行運算與容錯等優點來發展出預測TFT-LCD良率的方法。而為了有效控制神經網路建構的複雜度，本研究將分別應用主成分分析法與逐步選取變數分析法來達到降低輸入變數維度之目的；另也將以傳統迴歸分析方法與以上結果作不同模式結果之比較，以驗證類神經倒傳遞網路所建構之良率預測模式準確度。
而經由本研究所蒐集到的資料進行實際驗證的結果顯示，以生產TFT-LCD陣列資料結合逐步選取變數與類神經倒傳遞網路的模式，確實可作為一有效的預測TFT-LCD生產良率的方法。

關鍵字

面板 ； 良率 ； 類神經網路

並列摘要

Taiwan is the primary manufactured and produced country in TFT-LCD(Thin-Film Transistor Liquid-Crystal Display). Because the investment needs lots of cost in equipments and facilities produced of TFT-LCD, how it reduces manufactured time, keeping the quality, it is the necessary maintained requirement for enterprises to exhibit the production in market sooner than competitors. Yield is the correlated sample to reflect the whole technology and profit obtain in enterprises, so it is significant within many measured indicators. Simply speaking, yield is defined the percentage of good productions in all produced amount, and yield management is refer to produce a great number of data to integrated analysis, yield improvement, and yield prediction in the whole TFT-LCD making process. However, it would be the effective predicted model to define and clarify some aspects such as production materials, delivery time, and making problems, so yield prediction is gradually become the significant topic in enterprises.
Through the above instructions, the study tries to present an effective and simple yield prediction model: it is based on TFT-LCD panel arrayed data, and operates Artificial Neural Back-Propagation Network having the characteristics such as learning, parallel computing and compatibility to develop the method of TFT-LCD yield prediction. In order to control the complexity of ANBPN architecture, the study will apply Principal Components Analysis and Stepwise Variable Selection Analysis to reduce input variable instability.
A real case was presented to demonstrate the methodology and the result revealed that by stepwise variable selection of BPN can provide an acceptable for TFT-LCD yield prediction.

並列關鍵字

TFT-LCD ； Yield ； Artifical Neural Network

參考文獻

1. Cunningham, J. A. (1990). The use and evaluation of yield models in integrated circuit manufacturing. Semiconductor Manufacturing, IEEE Transactions on, 3(2), 60-71.

2. Freeman, J. A., & Skapura, D. M. (1992). Neural networks: Algorithms, applications, and programming techniques.

4. Murphy, B. T. (1964). Cost-size optima of monolithic integrated circuits. Proceedings of the IEEE, 52(12), 1537-1545.

5. Price, J. (1970). A new look at yield of integrated circuits. Proceedings of the IEEE, 58(8), 1290-1291.

6. Rumelhart, D. E., Hintion, G. E., & Willians, R. J. (1986). Learning Representations by Back-Propagating errors. Nature, 323, 533-536.