隨著晶圓製造成本的不斷增加，控制晶圓良率變得至關重要。達成此目標的其中一種有效方法是分析晶圓圖中的缺陷樣態，然而幾乎所有研究都集中在聚類分佈的缺陷樣態上，相反的，本文致力於分析反聚類分佈的缺陷樣態。我們提出了對這些缺陷樣態進行標記和分類的方法，我們將反聚類分佈缺陷樣態分為兩種類型：Grid及Sparse。在Grid缺陷樣態中，缺陷晶片在晶圓上呈現規則分佈，而在Sparse缺陷樣態中，缺陷晶片在晶圓上呈現離散式分布。 我們的方法由三個階段組成。 在第一階段，我們對晶圓圖進行預處理，去除可能干擾Grid缺陷樣態識別的群聚和線狀缺陷樣態。接下來，我們從晶圓圖中提取相關特徵參數並進行綜合分析來識別Grid缺陷樣態，其中可能包含多種類型的缺陷樣態。在第二階段，我們採用預先建立的隨機缺陷模型來確定晶圓圖中的缺陷是否表現出反聚類分佈。 隨機缺陷模型中包括三種類型：Score、Discrete die和Under 2-die 模型。通過應用這些模型並在晶圓圖上進行加權計算，我們可以分析出缺陷樣態的分佈類型。如果缺陷樣態表現出反聚類分佈，則它們會進入第三階段。最後，我們從晶圓圖中提取特徵參數並分析它們是否具有特殊的Grid或Sparse缺陷樣態。 為了驗證我們的方法，我們在真實晶圓資料集上進行了實驗，並實現了高準確度的辨識結果。實驗結果證明了我們的方法在辨識和分類反聚類分佈缺陷樣態的有效性和可靠性。 ;With the increasing cost of wafer manufacturing, it has become crucial to control wafer yield. Analyzing defect patterns of wafer maps is an effective method to achieve this goal. While many studies have focused on defect patterns with cluster distributions, this thesis addresses the analysis of anti-cluster distribution defect patterns. We propose a method to label and classify these patterns. We classify anti-cluster distribution defect patterns into two types: Grid and Sparse. For the Grid pattern, defect dies exhibit a regular distribution across the wafer, while for the Sparse pattern, defect dies are dispersed across the wafer. Our method consists of three stages. In the first stage, we preprocess the wafer maps to remove cluster and line defect patterns that may interfere with the identification of Grid patterns. Next, we extract relevant features from the wafer maps and perform comprehensive analysis to identify Grid patterns, which may contain multiple types of defect patterns. In the second stage, we employ pre-established random defect models to determine whether the defects on the wafer maps exhibit an anti-cluster distribution. The random defect models include three types: Score, Discrete die, and Under 2-die models. By applying these models and employing weighted calculations on the wafer maps, we can analyze and determine the distribution type of the defect patterns. If the patterns demonstrate an anti-cluster distribution, they progress to the next stage. Finally, we extract features from the wafer maps and analyze them to determine if they represent specific Grid or Sparse patterns. To validate our approach, we conducted experiments on real wafers and achieved a high accuracy of recognition. Experimental results demonstrate the effectiveness and reliability of our method in identifying and classifying anti-cluster distribution defect patterns.
