利用不平衡資料取樣策略改善晶圓缺陷群聚分類模型

Abstract

在製造晶圓的過程中，晶圓表面難免會產生缺陷點（defects），這些缺陷點會造成晶圓的良率損失(yield loss)。此外，這些缺陷點也會隨著晶圓的尺寸擴大逐漸產生群聚現象，形成多種類別之缺陷點群聚圖案，如：牛眼、半月、環狀、底部、隨機等圖案，每種圖案又是由不同之製程失誤所造成。因此，如能建構出一個準確的晶圓缺陷點分類模型，就能快速找出改善製程失誤之原因加以改善，就能有效提升晶圓良率，這對半導體廠而言，是一個重要議題。由於在實際之晶圓缺陷點群聚圖案資料通常屬於不平衡資料(imbalanced data)，即某一類之缺陷點圖案數量遠多於或少於另一類圖案之數量，若直接使用不平衡資料來建構分類模型，則不論使用何種分類方法(如：統計方法或類神經方法等)，通常都會出現分類模型之整體分類準確率雖然相當高，但少數類別之分類準確率卻過低的情況。現有文獻在建構晶圓缺陷點圖案分類模型時，大都是針對各類別圖案資料筆數相同的情況，使得所建分類模型對少數類別圖案之預測準確率偏低。因此，本研究之主要目的是利用實驗設計(Design of Experiments；DOE)及雙反應曲面法(Dual Response Surface Methodology；DRS)，針對晶圓缺陷群聚圖案屬不平衡資料的情況，提出一個重新取樣策略(re-sampling strategy)來分別減少及增生各種缺陷群聚圖案的資料量，再建構分類模型，以有效提升多類別缺陷點群聚圖案之不平衡資料中少數缺陷群聚圖案之分類準確率，進而快速找到造成此類缺陷圖案之製程失誤原因以及時改善。本研究最後利用模擬之晶圓資料，驗證了本研究方法確實有效。

In the manufacturing process of wafers, defects on the wafer surface are inevitable. The total number of defects and the defect clustering phenomena on wafers will cause yield loss. The patterns of defect clustering become more apparent as the wafer size increases. The common patterns of defect clustering are: Bull, Bottom, Chord, Circle, Random, etc., and each pattern is usually caused by a specific process error. An effective classification model of defect patterns can identify the defect patterns accurately. According, the process errors can be corrected. Therefore, it is an important issue to construct an accurate classification model for wafer defect patterns. Because the real-world data of defect patterns are usually imbalanced, that is, the number of wafers for some defect patterns is much less than the number of other patterns, Therefore, direct use of imbalanced data to construct a classification model for defect patterns, no matter what classification method (e.g. statistical methods or artificial neural network methods) is used, the accurate classification rate is low for minor wafer patterns, That is, although the overall classification accuracy rate of the classification model is quite high, the classification accuracy of a category with few data may be very low. Previous studies on constructing the wafer-pattern classification model only considered the balanced data. In this study, Design of Experiment (DOE) and Dual Response Surface Methodology (DRS) are utilized to develop an optimal re-sampling strategy for imbalanced defect-pattern data and the classification model is constructed accordingly. Finally, the simulated defect data of wafers are utilized to verify that the re-sampling strategy proposed by this study is effective in improving the accuracy rate of classifying the minor defect clustering patterns.