自動化分類晶圓缺陷圖樣-使用混合圖形特徵辨識

Automatic Classification of Defect Maps - Using Hybrid Approach of Pattern Recognition

黃志凱

指導教授 ： 陳飛龍

國立清華大學/工學院/工業工程與工程管理學系/碩士(2006年)

摘要

近年來半導體產業展現出高變動、高成長的特性，其中最常被討論的議題就是如何有效的提升良率，而高良率亦是製造業最重要的獲利指標，在半導體製造業中，晶圓缺陷不但是良率的最大阻礙，亦會造成製造成本的升高，因此晶圓缺陷的偵測與缺陷圖樣的辨識便成為晶圓廠內最重要的議題，發展出有效的辨識方法及系統更是改善良率不可或缺的要件，因此本研究同時採用多種的圖樣辨識手法，提出一能夠有效率的辨識晶圓缺陷圖樣的方法論，在研究中將晶圓缺陷圖樣概分為直線型、曲線型、環形、區域型、放射型、重複型及晶片邊緣型，除了這些單一缺陷圖樣問題外，本研究亦提出多重缺陷圖樣問題的解決方法，以期此方法論能夠更切合實際應用領域，研究中定義出五種不同的多重圖樣關係，包含分離型、分支型、交錯型、包圍型及重疊型。
本研究所發展的方法論中，首先將所得的資料轉換成圖形型態，利用影像處理技術濾除非系統性的缺陷點，再將鄰近的缺陷點加以匯整並計算密度，形成不同密度的方格，以獲取缺陷圖樣結構之訊息，在辨識核心部份，本研究分別採用了模板比對、元素分離、特徵萃取及特徵選取等手法，以期能發揮不同手法的優點，進而達成高準確度的辨識，另外在樣本描述及樣本分類的階段中，將樣本轉化為特徵向量，並將其投射至特徵空間，利用模糊理論之分群方法，將樣本分類並賦予其特定的信心程度。
本研究收集了某半導體廠商實際的缺陷資料，建立系統以驗證辨識方法的之可行性，以實證結果而言，本研究所提出之辨識方法相較於人工辨識，具有相當的準確性亦可大幅縮短辨識時間。

關鍵字

特徵辨識 ； 晶圓缺陷 ； 良率 ； 模糊分群

並列摘要

The semiconductor industry historically has been an extremely dynamic and high-growth business. Yield improvement is the important element to ensure the profitability. Defects incumber yield and increase the cost of manufacturing, so the defect inspection and recognition are vital in wafer fabrication. The development of recognition methodology is also the essential ingredient of yield enhancement.
This research intends to propose a high-efficiency solution of the pattern-related defects recognition based on the hybrid approach of pattern recognition. In this research we classify pattern-related defects into different categories including line, curve, ring, local, radial, repeat, and die edge. In addition to single-pattern problems we discussed before, this research also intends to solve the problems of multi-pattern recognition such that the practical applications can further be improved. There are five types of relationship in multi-pattern situation defined in this research, including separation, branch, crossover, encirclement and overlap. After transforming coordinate data into image format and using image processing techniques to eliminate the unsystematic noise and set grids to get structural information for further analysis. The procedures of hybrid identification technique performed in this methodology including templates matching, primitive separation, feature extraction and feature selection. By setting feature vectors into feature space and applying FCM theory to clustering, it is expected to reach the goal of description and classification.
The proposed methodology is verified with real industrial data and the experimental results show the advantages of high accuracy and short executing time.

並列關鍵字

Pattern Recognition ； Defects ； Yield ； FCM

參考文獻

Andrzej, J. S. and Stephen, W., “A Pattern Recognition Based Method for IC failure Analysis,” IEEE Transactions Computer-Aided Design, 76-92(1985.)

Baxes, Gregory A., “Digital image processing: principles and applications,” John Wiley and Sons, 1994.

Breaux, L. and Singh, B., “Automatic defect classification system for patterned semiconductor wafers,” IEEE/UCS/SEMI International Symposium on Semiconductor Manufacturing, 68–73(1995).

Chen, Y. F. and Nazir, A. W., “A Knowledge-Based System for Multiple Pattern Recognition Paradigms Support” IEEE Transactions Computer-Aided Design, 1130-1135(1995).

Carpenter, Gail A., Grossberg, Stephen, “A Massively Parallel Architecture for a Self-Organizing Neural Pattern Recognition Machine,” Computer Vision Graphics Image Processing, 54-115(1987).