運用資料探勘技術分析半導體晶圓圖形

Abstract

在多變且快速的時代，電子產品日新月異，半導體晶圓代工廠必須不斷提升高階精密製程的能力，且快速提升晶片之產品良率(Yield)，以提升競爭力。但是晶圓製造的生產步驟非常多與複雜，有許多可能影響產品品質的因子，所以良率異常的分析就變的很困難卻又非常的重要。良率的異常大多是由生產的機台異常所造成，然而生產機台的類別與數量很多，故發生異常時並不容易馬上被發掘出來。以工程師的經驗，同類機台的異常通常會形成相似之良率異常的圖樣，故本研究以竹科某知名半導體廠作為實驗分析對象，並應用(1)倒傳遞網路(Backpropagation)的技術，將半導體晶片相似之良率異常(Fail Bin)的圖樣分群，(2)再利用決策樹(Classification by Decision Tree)分析尋找到關聯與可疑異常之機台，以達到鑑別異常機台與良率提升的目的。實驗結果顯示所應用之方法能有效針對竹科半導體大廠良率低之產品，找出可疑異常的機台。

Taiwan semiconductor companies must design high-level technology product and increase the wafer yield rate to enhance their competitive advantages. However, the wafer process is very complex with a lot of steps and factors to affect the yield rate. Accordingly, it is very difficult and important for wafer yield analysis to find out the root cause. According to engineers’ experiences, process equipment issues are the major factors in low yield. However, the root cause is difficult to identify because many equipments are involved in a wafer process. Base on engineers’ experience, the low yield wafers usually have some special pattern in CP yield test result. Thus, we can use the characteristics to identify the problematic equipment. This study focuses on wafer yield pattern analysis to identify the problematic equipment. The thesis uses the data of a Taiwan semiconductor manufacturing company to conduct the analysis and experiment by using the following methods: (1) Apply ‘Classification by Backpropagation’ to classify wafer fail bin pattern; and (2) use the ‘Classification by Decision Tree’ Induction to identify related problematic equipment. The experiment results show that applying the proposed methods can fast classification fail yield bin pattern and find out the problematic equipments.