晶圓在製造過程中難免會有一些缺陷，造成晶圓上的缺陷原因有很多種，因此，工程師透過晶圓針測結果產生的晶圓圖，觀察其錯誤晶粒分佈，進而分析製造過程中錯誤原因。現今主要分成九種的錯誤樣態，分別為Center、Donut、Scratch、Edge-Loc、Loc、Near-full、Edge-Ring、Random和none，這些錯誤樣態都是由工程師以肉眼觀察晶圓圖找出，然而去分類錯誤樣態沒有一致的定義，主要都是以工程師的主觀判定，所以會造成判定上的分歧。 本論文的目的是進行錯誤樣態的特徵參數提取，主要是對台積電所提供的WM-172K資料集，將其轉換成圖片和利用影像處理及聚類演算法去找到最大群聚，再進行區域及統計方面上做特徵參數的提取，提取後得到各個參數的分布結果，並且利用盒鬚圖找到每種錯誤樣態的重要特徵參數數值區間。 最後，我們利用特徵參數的數值區間進行錯誤樣態的標記，對於WM-172K的7種樣態，平均準確度可達74.69%。 ;Some inevitable defects exist during wafer manufacturing process. There are many reasons for these defects. Therefore, engineers observe the distribution of bad dies after testing and analyse the causes of errors in the manufacturing process. We often categorize defect patterns into nine types, which are Center, Donut, Scratch, Edge-Loc, Loc, Near-full, Edge-Ring, Random, and none. These patterns need to be automatically identified as fast as possible after wafer testing. In this paper, we define and extract many feature parameters of defect patterns on wafer maps of WM-172K dataset. We use image processing and clustering algorithms to find the largest cluster for each map. Then, the feature parameters are extracted for the regional and statistical attributes, respectively. The distribution results of each parameter are obtained. We use Box Plot to mark the numerical interval of each feature parameter for each defect pattern. Finally, using the numerical intervals of feature parameters to rate the scores of defect patterns of WM-172K, we can obtain the average accuracy of 74.69% for seven types of defect patterns.
