應用自組性演算法建構具不同缺陷點群聚圖案之晶圓良率預測模型

Abstract

晶圓的良率(yield)是積體電路製造商用以提升產能及獲利的重要指標，影響良率的因素很多，主要因素為晶圓上缺陷點(defects)的總數以及缺陷點的群聚(clustering)嚴重程度。近年來晶圓製程逐漸進步，晶圓面積不斷增大，缺陷點的群聚現象也越來越明顯，導致傳統的卜瓦松(Poisson)良率模型不再能準確地預測晶圓良率。針對此問題，中外文獻提出一些複合卜瓦松良率模型(compound Poisson yield models)，或應用迴歸分析或類神經網路等方法來建構良率模型，但這些良率模型依舊各有不完善之處。 因此本研究之主要目的為針對晶圓上缺陷點的群聚圖案(patterns)另建構一個良率預測模型，以有效提升晶圓良率之預測準確度。本研究所提出的方法共分成兩個階段進行：第一階段為應用自組性演算法(Group Method of Data Handling, GMDH)建立缺陷點群聚圖案之判別模型，本研究所考慮之群聚現象有隨機、牛眼、底部、新月及環狀等常出現之晶圓缺陷點圖案；第二階段為依照不同的群聚圖案分別應用自組性演算法建立良率預測模型。本研究最後利用模擬之晶圓資料驗證了本研究所建構之良率模型確實有效。

Wafer yield is an important indicator of increasing production capacity and profit for semi-conductor manufacturers. There are several factors that influence the wafer yield, the defects on wafers and the degree of defect clustering are two major factors among them. Over the past few years, the techniques of wafers have been advanced. The surface areas of wafers are increased, but the degree of defect clustering become apparent accordingly. This leads to the traditional Poisson model not being able to accurately predict the wafer yield. Several studies have developed compound Poison yield models or applied regression analysis and neural networks to construct yield models. However, these models still have some drawbacks. Therefore, the object of this study is to develop a two-stage procedure to construct yield models for various defect patterns in order to increase the prediction accuracy. The first stage is to apply Group Method of Data Handling (GMDH) to construct a discriminate model to discriminate various defect patterns. The second stage is to construct yield models for various defect patterns using Group Method of Data Handling (GMDH) respectively. Finally, this study applied simulated yield data to verify the effectiveness of the proposed method.