- 學位論文
應用類神經網路模擬預測電子迴旋共振式蝕刻 機台蝕刻深度與蝕刻良率系統
The Prediction of Etching Depth and Yield of ECR-RIE Using Neural Network
摘要
本研究是應用倒傳遞( Back-Propagation )類神經網路，建立了一套能夠模擬電子迴旋共振式蝕刻機台(ECR-RIE)之蝕刻深度預測系統，在製程資料的收集上，係自半導體晶圓製造廠中，蝕刻部門所使用的ECR-RIE機台中所取得，其資料包括了該蝕刻製程所使用的氣體種類及流量、腔體( chamber )壓力、腔體溫度、線圈電流量、磁控管功率與射頻功率等資料，並將其歸納整理後所得到的14筆資料數據作為神經網路的輸入，再將蝕刻製程所量測到的蝕刻深度相關數據，作為神經網路的輸出，其中包括蝕刻深度之最大值、最小值、平均值等3筆資料數據，以倒傳遞的網路學習方式來建立類神經網路系統模型；實驗中總共蒐集了99筆機台製程資料，將其中的65筆作為神經網路訓練樣本，並對訓練樣本進行1000次的訓練，訓練完成之網路以訓練樣本數據先對其作檢測，確認網路的架構正確無誤後，再將訓練完畢的類神經網路模型以另34筆測製程資料做樣本檢測，研究中證實此系統能有效預測ECR-RIE製程之蝕刻深度品質狀況；最後將此神經網路模型，結合蝕刻率的實驗，當製程轉換時或蝕刻條件變更時，可作為製程微調的建議,可以提供製程工程師系統化的估測蝕刻深度方法，亦可作為診斷與預警上的依據，當機台參數的漂移可能令蝕刻深度超出規格時，即可提出警示，並提供工程師執行參數修正的建議及參考方向。
並列摘要
In this research, back-propagation neural network was applied to setting up a prediction system for Etching depth which was able to simulate electron cyclotron resonance reactive ion etching. In the semiconductor wafer fab, we received data files in the etching process including gases、gas flow、pressure( chamber )、temperature(chamber)、coil electric current、magnetron power and reflect power, etc. Then 14 data files were classified as the input of the neural networks while 3 data files, including maximum、minimum、average, were classified as the output of the neural networks. Neural network system model was set up based on back-propagation learning networks. In the experiment, the researcher took 65 data files out of 99 in the Etching process as the training samples of the neural networks. The samples which were trained for 1000 times were inspected. After assuring the frame networks, the researcher inspected the trained neural network system models and 34 data files in the etching process. The research proved that the neural network system was able to predict the quality of Etching depth in the process of ECR-RIE. By combining the neural network system models and the Etching rate, it helps make adjustment in the Etching process and it helps engineers predict the etch depth in a varied condition. It serves as a basis in making diagnoses and issuing an alert. As one of the parameters shits beyond scale in the etching process, the neural networks would issue an alert. It helps engineers to correct the parameter and offers referential data.