- 學位論文
以模糊自適應共振理論網路建構晶圓針測圖與缺陷圖之關聯模式分析
Using Fuzzy ART to Construct the Correlation Model of Wafer Bin Map and Defect Map
摘要
半導體產業乃為資本與技術高度密集的產業，且具有高成本與上市時間的壓力，如何提升良率皆是半導體產業不容小覷的課題。在實際產線上，晶圓圖乃提供追查晶片發生異常的重要線索，建構辨識故障系統可找出可能產生低良率的原因，目前國內晶圓廠對於晶圓圖的圖像辨識大多是以人工目視的方式來進行，但由於晶圓圖分佈的情況雜亂，且可能因為工程師本身的經驗和主觀判斷因素造成辨識時間過長與準確性的不足，影響問題解決的效率。本研究考慮不同缺陷差異組合影響的程度以及過濾雜訊等資料前置處理，並結合類神經網路中模糊自我適應共振理論網路(Fuzzy ART)，將晶圓針測圖(Bin Map)與晶圓缺陷圖(Defect Map)做討論，建構各Bin值與缺陷樣版圖的辨識，並做出合理的分類的系統，提供工程師在觀察晶圓針測圖時，可找出哪一製程階段出的缺陷，同時得知還有哪些晶方處於相同情況，進而回推到錯誤率高的製程點做改善，本研究樣本為75筆141顆規格的晶方，其中55筆做為訓練，20筆為測試樣本，而實證分析結果顯示，辨識正確率達90%，系統可以成功的區分出隨機性故障圖樣，並且消除系統性故障圖樣上的雜訊，提供有用的資訊幫助提高診斷的效率。
並列摘要
Semiconductor industry is an industry that requires both capital and technology, it faces not only pressure from high cost and timing for entering markets, but also has a major issue regarding to its yield. As for the production line, wafer bin map provides an essential clue for identifying any abnormality in wafers. It builds up a system which points out the causes that might be accounted for having such low yield. Nowadays, the majority method for recognizing wafer bin map in the fab is done through naked human eyes. However, the distribution of wafer bin map, as well as the engineer’s personal experiences or subjective estimations, could create problems of overtime-identifying and lack of accuracy which lead to results of low efficiency when it comes to problem solving. In order to discuss the correlation between the wafer bin map and defect map, this research conducts a methodology by using Fuzzy ART algorithm. During the analyzed model, firstly, this research recognizes number of wafer bin maps and defect maps. While speculating the wafer bin map, this model allows the engineers to find out not only which the manufacturing process is mistaken but also the particular die that appears to be abnormal. Since the engineers are able to trace back the mistake where it’s being made, they can of course improve the procedure by correcting the error. The sample size of wafers to this research is 75, and each of them contains 141 dies. While 55 wafer maps are used for training, the other 20 are for testing. The analytical result shows that there is 90% of accuracy. The system is able to successfully differentiate the random malfunction, also it provides supportive information for improving the efficiency while analyzing.