應用資料探勘技術於半導體業製程良率改善之研究

Abstract

追求高良率 (Yield) 為晶圓製造業內最基本也最重要的工作任務，高良率意味著以同樣的生產成本，業者能夠產生相對低良率更有效用的產出。因此，在晶圓代工業，一家擁有高良率的公司，也代表著具有高度的競爭力。 本研究整合資料探勘演算法中的分群技術 (Clustering)，首先將複雜度高的製程資料由分群使得具有相似特徵的資料聚類，接著再使用分類技術 (Classification) 中的決策樹推論法則 (Decision Tree Induction)，將分群的結果作出群聚規則的推論，建構一完整之晶圓代工製程良率分析資料探勘架構。分群過程中使用SOM (Self-Organizing Map) ，K-Means以及 SOM整合K-Means之二階段分群等三種分群法，並使用業界之實體資料作為實驗的依據，以業界所發生之實際案例作為設計實驗，以驗證本研究所提出之資料探勘架構內之各項演算法之效果，最後，再根據領域專家的意見評比定義出最適合半導體製程資料探勘架構的分群演算法。 本研究之最終目標為提出此一有效之晶圓代工業製程良率分析之資料探勘架構，並找出適合半導體製程資料探勘的分群演算法，作為晶圓代工業者於改善良率時用以增加效率，分析故障因子時的另一種可行的分析手法，以期望達到最終能夠快速找出造成低良率之故障因子，進而改善良率的效果。

Achieving high yield is the most essential issue in the Semiconductor Manufacturing Industry. A high yield company can have more output using the same productions cost relatively to the low yield company. Therefore, in the Semiconductor Manufacturing Industry, a company capable of producing the wafer with high yield means to have high competitiveness. In this study, we integrated the clustering technique and the decision tree induction of classification technique to construct a data mining approach in the yield analysis of Semiconductor Manufacturing process. First, we used clustering technique to group data with similar characteristics, and then we used decision tree induction to induct the rules of the clustering results. In clustering, we used three clustering methods, including SOM (Self-Organizing Map), K-Means, and SOM combined with K-Means, to apply on the real data in the industry. At the same time, we used the real cases in the industry as the basis of the experiment design to verify the effectiveness of each method. The final goal of this study is to propose an effective data mining approach for the yield analysis of Semiconductor Manufacturing process, and to find out the most suitable clustering method for analyzing low yield root causes.