在過去，類神經網路已經被成功地應用於建立預測模型。本論文說明如何建立一個類神經網路，來預測生產微波主動元件之磊晶製程的電流增益。本研究利用實驗設計之反應曲面法，來決定類神經網路之連續型重要參數。為了進一步提升類神經網路之預測績效，我們提出以分割資料子集之方式，來訓練多個類神經網路，產生數個預測模型來進行預測。本研究是以平均絕對百分比誤差作為績效指標，利用實際之製程數據來建立類神經網路並驗證其績效。根據經驗及專業知識，我們以電力（power）來作為分割資料子集之基礎。研究結果顯示，利用本研究所提出之分割資料子集法來訓練類神經網路，可以有效地提升預測之績效。

In recent years, the photoelectric technology has been successfully and widely applied to wireless communication, fiber-optical, solar power, display and illumination industry. The function of photoelectric products is relied on the device performance of compound semiconductor epitaxy wafer with quality and yield that are dominated by substrate and epitaxy technology mainly. The destroyed test with high cost and sampling risk is implemented to assure the current gain of epitaxy wafer.This study proposes a prediction model for current gain based on artificial neural networks (ANNs). Response surface methodology was employed to optimize the parameters of ANN prediction model. In order to enhance the prediction accuracy, the training dataset was partitioned off into three subsets and trained by three separate ANN models. Real data collected from manufacturing process are provided in this study to demonstrate the superiority of the proposed ANN prediction model. The results indicate that the proposed ANN prediction model performs better than traditional regression analysis method. Neural networks trained by partitioned data subsets can predict better than the neural network trained by the whole dataset.