利用類神經方法建構晶圓缺陷點群聚圖案之辨識系統

Abstract

如何提升晶圓的良率(yield)一直是半導體廠最關注的問題之一。而良率問題往往和晶圓上缺陷點數和缺陷點群聚現象息息相關，其中缺陷點群聚現象的產生主要是由於晶圓尺寸越作越大，製程越來越複雜而造成缺陷點出現群聚現象。因為製程問題的不同，群聚圖案就會有不同的形狀，因此製程工程師若能準確地判斷出晶圓上的群聚圖案，即可以迅速找到製程的問題來提升良率。目前有一些中外文獻利用類神經方法辨識晶圓缺陷點之群聚圖案，效果不錯，但這些文獻所提之方法在類神經的輸入變數擷取上往往需要花費很多時間。因此本研究的主要目的是利用類神經方法建構出一套簡單好用且辨識率高的晶圓缺陷點圖案判別系統，能夠簡單擷取類神經輸入變數以及有效辨識缺陷點群聚圖案。本研究將以模擬資料進行類神經網路方法的訓練，以找到表現最佳的類神經網路方法及其參數組合，最後再以新竹科學園區某半導體廠商之實際晶圓資料來驗證本研究之辨識系統的有效性及可行性。

Being a semiconductor manufacturer, knowing how to improve the yield of wafer production has been regarded as the focus. However the causes of yield problems have much to do with the total number of defects on a wafer and defects clustering phenomenon. As the wafer size increases, the wafer processes get complicated and the defects clustering phenomenon tends to be apparent. Different problems of wafer processes always make different clustering patterns, so process engineers could find the process problems rapidly to improve the yield by identifying the clustering patterns correctly. Some papers make use of Artificial Neural Network (ANN) to identify wafer defects clustering patterns and come to the acceptable effects. However, it costs much time while transferring wafer defects data into input variables of ANN. This study constructs a wafer defects identification system by ANN, which characterize well identification rate and the method for easily getting input variables of ANN. Simulation data is used for training ANN and then to find out the combination of parameters of the best performance. The Real wafer defects data verify the effectiveness and feasibility of the identification system.