半導體化學氣相沉積厚度預測之研究

Abstract

近年來倒傳遞類神經網路(back propagation neural network，BPNN)被提出來應用於半導體的良率預測。因為此網路於訓練時，需調整的網路參數甚多，較為複雜。後來有學者提出較為簡單的輻狀基底函數類神經網路(radial basis function neural network，RBFNN)來預測半導體良率，僅需調整一個參數且預測準確率比倒傳遞類神經網路更佳。 本論文為提出比輻狀基底函數類神經網路準確性更佳的反傳遞模糊類神經網路(counterpropagation fuzzy neural network，CFNN)及廣義迴歸類神經網路(general regression neural network，GRNN)並以預測半導體化學氣相沉積(chemical vapor deposition，CVD)厚度為例，來比較這三種網路。實驗結果證明，反傳遞模糊類神經網路和廣義迴歸類神經網路之準確性均比輻狀基底函數類神經網路好且更適合應用於化學氣相沉積厚度的預測。

Applying back propagation neural network(BPNN) to predict the yield in semiconductor has bean proposed in the recently years. Because too many network parameters need to be adjusted in the training phase, it is more complicated. Afterward a scholar purposes more simple radial basis function neural network(RBFNN) to predict the yield in semiconductor. Its adjustable parameter is only one and predict accuracy achieves higher than back propagation neural network. The thesis proposes couterpropagation fuzzy neural network(CFNN) and general regression neural network(GRNN) which achieve higher prediction accuracy than radial basis function neural network and compare with the three networks by the example to predict chemical vapor deposition(CVD) thickness. The experiment proves that CFNN and GRNN achieve higher prediction accuracy than RBFNN and more applied suitable to predict CVD thickness.