應用主成份分析於晶圓瑕疵圖形辨識

Abstract

對於有經驗的半導體工程師而言，藉由晶圓瑕疵分佈而成的圖樣，可判別某一段製程是否發生狀況，而不同的製程錯誤將會造成不同的晶圓圖樣。若能建立迅速且一致的自動化系統有效地分類晶圓瑕疵圖，達到錯誤診斷的目標，必能降低時間與人力,增加晶圓製程的作業效率。 本論文結合影像處理、統計方法與資料挖礦，發展出一自動化的瑕疵圖樣辨識器，共可辨識隨機型、中央型、邊緣型、四足型、啞鈴型、矩陣型、刮線型、環狀型、區塊型九種不同類型。在設計自動瑕疵辨識器過程中，我們針對晶圓尺寸、位置之問題加以善，輔以主軸成分分析之一貫精神，結合k最鄰近分類與樸素貝式分類辨識出大部分類別，其準確率皆可達92%以上。

For experienced semiconductor engineers, the pattern formed by the wafer defect distribution can be used to judge if a certain stage of the manufacturing process is problematic. Accordingly, different problematic stages will result in distinct wafer defect patterns. In case, an efficient and consistent automatic system can be established to effectively classify wafer defect graphs to achieve the goal of diagnosing defects. Then, the expense of time and human resources can be decreased while the efficiency of manufacturing wafers can be raised. This thesis integrates image process, statistics, and data mining to develop an automatic defect pattern identification machine. Nine different types can be identified: random, center, edge, four buttons, dumbbell, matrix, scratch, ring, or local type. In the process of designing the automatic defect pattern identification machine, the problems of wafer size and location have been resolved. Besides, based on the essence of the Principal Components Analysis, K- Nearest Neighbor Classification and Naïve Bayesian Classification have been integrated to identify most types. The accuracy of over 92% can be achieved.