本論文提出以GRNN(General Regression Neural Network) 的方法來建構一種預測良率之模式,良率模式可用以預測產品之良率以及管理良率。傳統的良率模式裡,如Poisson 或 Murphy’s 等等之良率模式,模式裡之母體均設定具固定之分配,然而此假設並不一定適用於每一種半導體的製程。這些傳統上模式中的缺失將會導致良率預測的不精準及不可靠。在GRNN演算法裡,不需要事先假設良率與各因子之間的函數關係,也不需要事先知道母體的分配,利用由歷史資料中自我學習之原理,所有的關係式均可在GRNN的網路架構中尋求解答。。本研究中,影響良率之因子包括晶粒面積、晶圓的晶粒數、缺陷數將考慮在GRNN模式中,並以holdout方法來決定GRNN中的平滑參數值,可更有效地建構良率學習模式。最後以模擬的方法討論GRNN建構良率模式的效果,結果發現,GRNN方法所建構之良率模式能準確的預估良率,並且較部份傳統的良率模式(Seed’s及Murphy模式)更為精確、穩健。
We can use yield models to predict and manage yield. In traditional yield models, we assume population to be a certain distribution. However, it may not be always correctly. In this thesis, we use a GRNN(General Regression Neural Network) method to construct a yield model. In GRNN model, it’s unnecessary to know the distribution forms of population neither to know the function form between yield model factors. All of the relation functions are include in the network structure. In this paper, we use holdout method to determinate the smoothing parameter of GRNN. And then, we discuss the outcome by simulation. Finally, we found, GRNN can predict the yield accurately.