本篇論文中，我們著重在兩個部分一個是相似性模型的分析，另一個是相似性叢聚分析應用的部分。先討論相似性分析的部分，選擇實驗的對象為不同良率以及不同diesize數量下的合成晶圓，定義相似性判定公式以及其相似性數值與良率在常態分佈下之模型。 再來討論相似性叢聚分析應用的部分，選擇實驗的對象真實晶圓的錯誤樣態做分析，透過叢聚分析(K-means)將其分成數群，找出每一群具有的特徵，我們所使用的真實晶圓為台積電所提供的WM-811K晶圓資料庫，其中的錯誤樣態可分為以下九類，Center、Donut、Scratch、Edge-Ring、Edge-Loc、Loc、Near-Full、Random、None。 最後分別計算在不同晶圓大小的效能比，隨著晶圓越大，所需要的計算時間也越多。 最後綜合此兩分析方法，運用了相似性公式及叢聚分析來判斷合成晶圓以及真實晶圓的錯誤樣態，從而瞭解其特性，最後可以判斷該產品具有什?特徵。 ;In this paper, we focus on two parts, one is the analysis of similarity model and the other is the application of similarity clustering analysis. In this paper, we first discuss the similarity analysis. The experimental subjects are synthetic wafers with different yields and different numbers of diesize, and we define the similarity determination formula and the model of similarity values and yields under the normal distribution. The wafers we used are from the WM-811K wafer database provided by TSMC, and the error patterns can be classified into the following nine categories: Center, Donut, Scratch, Edge-Ring, Edge-Loc, Loc, Near-Full, Random, and None. Finally, the performance ratios for different wafer sizes were calculated separately, and the larger the wafer, the more computation time was required. Finally, the similarity formula and cluster analysis are used to determine the error patterns of synthetic wafers and real wafers to understand their characteristics and finally to determine the characteristics of the product.
