半導體的研發是我國經濟發展及提昇的重點科技之一,但由於產品製造之複雜度的提昇,使得晶圓經常有不良的情形發生。目前半導體產業對於不良晶圓的檢測大部份仍是以人工目視的方式來進行,此方法因無法符合快速、精確的要求,而有失檢測標準的精確性及一致性,生產成本也因此而無法降低。為解決此問題本文應用區間SOM(Interval Self-Organizing Map)神經網路來進行晶格(die)缺陷的檢測工作。在特徵值擷取方面,是採用紋路分析技術之灰階值共生矩陣法,擷取不良晶格影像的熵、均勻性、對比性、相關性及區域同質性等五種紋路特徵值,再利用區間SOM神經網路聚類的特性來進行檢測。 良率(yield)是衡量半導體製程的重要指標,為得到最精確的良率,廠商必須根據產品的不同而變動良率模式。本文另將利用BPN神經網路,來學習Poisson、Seeds及Murphy等三種良率模式,以期對多樣化產品的不同製程,所產生之可能不同晶格缺陷分佈的統計分配,能快速預測出晶圓的製程良率值。
Semiconductor research and development is one of high techniques, however, due to the fact that the sophisticated process of semiconductor manufacturing is increased so that there always generates defected semiconductor wafers. At present, the defect inspection almost still is a tedious manual work, which cannot meet the goal of agile manufacturing, the reduction of cost. This thesis develops an interval SOM neural network capable of self-learning uncertain data of defect image of a die and then self-identifying the further defected dies of a wafer. Five kinds of texture features, entropy, energy, contrast, correlation and local homogeneity, are extracted from a die image and are computed to form gray level co-occurrence matrices presented to the interval SOM. This interval SOM can then cluster the defects automatically. Another BPN neural network tandem with the interval SOM is then applied to predict the yield based on the defect area and density of the wafer detected by the coupled interval SOM. One amazing characteristic of the neural prediction is that the statistical distribution of defects can be ignored.