應用統計學習方法於晶圓瑕疪之分類

Abstract

對於有經驗的半導體工程師而言，利用晶圓瑕疪點的分佈，可判別某一段製程是否發生狀況，此為重要的一項製程診斷技術。本文利用圖形識別方法取代工程師對晶圓瑕疪圖作分類，以增加工作效能和良率。本文所要識別的晶圓瑕疪圖共有五類；利用觀察晶圓瑕疪點的幾何特性，共開發出八種特徵，其中三種特徵須設定控制參數。和一般圖型識別系統不同的是，本文利用Fisher判別式選定最佳參數，實驗結果驗証此方法可顯著提高正確率。本研究計含135個模擬瑕疪樣本，因此選用較不須大量訓練資料的分類方法，分別為：決策樹、樸素的貝氏分類器、K-最鄰近分類器。由交互驗証實驗結果發現，樸素的貝氏分類器與K-最鄰近分類器皆可達到96%以上的正確辦識率，而決策樹僅有89%的辦識率。

By examining and inspecting the defect patterns on trouble wafers, experienced semiconductor engineers can usually identify the candidate manufacturing processes that caused the problem. This is of the important diagnosis methods in IC manufacturing industry to enhance fabrication efficiency and yield. In this thesis, the pattern recognition approach is used to classify defect patterns automatically instead of human inspection. In our study, five different defect patterns have been identified and categorized by human experts. After reviewing the geometric characteristics of these detect patterns, we extracted eight different features, three of them containing control parameters. To achieve optimal parameter setting, we adopted Fisher’s linear discriminant criterion, which is not common in conventional pattern recognition systems. And the experimental results showed that it did improve the system performance effectively. Our study included 135 simulated samples, implying that the classification models should not require massive training data. Thus we chose the following methods: the decision tree, naïve Bayesian classifier, and K-nearest-neighbor classifier. The results of cross validation showed that both naïve Bayesian and K-nearest-neighbor classifiers achieved extremely high accuracy, above 96%, while the decision tree only reached 89% of accuracy.