應用多變量指數加權移動平均管制圖監控晶圓表面缺陷數與缺陷群聚問題

Abstract

積體電路製造業廠商常用缺陷點數管制圖（c-chart）來監控晶圓表面之缺陷點（defect），但隨著晶圓面積逐漸增大，缺陷點開始出現群聚(clustering)現象，而c-chart只能管制到晶圓之總缺陷點數的問題，因此出現群聚現象之缺陷點數會產生過多之假警報（false alarms）。此外，晶圓表面若出現太多之缺陷點數或嚴重之群聚現象，都表示晶圓之製程可能出了問題，因此利用管制圖來監控晶圓品質時，需同時監控缺陷點數及群聚嚴重程度，而過去中、外文獻有針對此問題提出了一些修正管制圖，如尼曼A型分配（Neyman Type-A distribution）及Hotelling’s T2多變量管制圖等，但是在缺陷群聚和敏感度部分還是有一些缺點。由於多變量指數加權移動平均（Multivariate Exponentially Weighted Moving Average, MEWMA）管制圖可以利用一張管制圖來同時監控多個變數，並且當變數發生微小偏差時，此管制圖會有很好的敏感度。因此，本研究之主要目的是構建一個有效監控晶圓缺陷點與群聚現象之MEWMA管制流程。針對晶圓表面出現的缺陷點數及能反應晶圓缺陷群聚強弱之群聚指標（cluster index）兩個變數，構建MEWMA管制圖以監控晶圓品質。對於超出管制上限的失控點，本研究利用T²統計量分解法，將各別變數的T²值分解出來，比較兩個變數的大小，以判斷資料的失控原因屬於哪個變數。本研究最後以模擬之八吋晶圓缺陷資料來觀察所提方法之可行性與有效性，並以新竹科學園區某晶圓製造廠之實例說明如何應用本研究方法。

The c-chart is widely used in the integrated circuits (IC) manufacturers to monitor the wafer defects. However, the clustering of defects on a wafer due to the complicated manufacturing process becomes more evident with increased wafer size. The defect clustering phenomenon causes c-chart invalid, and the false alarms often appeared. The modified c-chart such as Neyman Type-A distribution and Hotelling’s T2 multivariate control chart still have some the disadvantages for defect clustering phenomenon and sensitivity. In addition, if there are too many defect counts or the serious clustering phenomenon on wafer surface that expressed the process possibly had problems; therefore, the control chart should monitors defect counts and defect clustering phenomenon simultaneously. The objective of this study is to develop a Multivariate Exponentially Weighted Moving Average (MEWMA) process control method for monitoring wafer defect with clustering phenomenon simultaneously. This study employed MEWMA control chart using the defect counts and cluster index as two quality characteristics. The proposed method employs the technique of decomposition of T2 to determine which quality characteristic (or the interaction of both quality characteristics) causes the process to out of control. The simulated data and a real case study are also presented, the results indicate that the proposed method is more effective than that of previous studies.