應用資料探勘技術於半導體晶圓允收測試參數預測之研究

Application of Data Mining Techniques for Wafer Acceptance Test Prediction

呂學地(Hsueh-Ti, Bruce, Leu)

指導教授 ： 鄭春生

元智大學/工程學院/工業工程與管理學系/碩士(2009年)

摘要

晶圓允收測試 (wafer acceptance test, WAT) 設置的目的在於對晶圓進行電性的量測，以做為晶圓出貨給代工客戶的依據。WAT 各項參數均訂有明確的規格，假如晶圓未能通過參數規格，因客戶對超過規格的晶圓有低良率的疑慮，所以通常的處置方式，就是報廢。如果可以事先預測 WAT 參數值，那麼即便發生 WAT 參數異常，工程師也可以盡快進行問題解決步驟，以減少 WAT 參數異常批的發生。

本研究之主要目的是發展一個以類神經網路為基礎之預測模型，用來預測半導體廠之 WAT 參數值。我們以某半導體業者之實際 WAT 參數及晶圓製造量測數據，來驗證預測模型之正確性。

此研究利用平均絕對誤差百分比 (mean absolute percent error, MAPE) 作為主要之評估指標，比較類神經網路與傳統迴歸分析之預測正確性。我們採用之類神經網路包含徑向基函數網路與倒傳遞網路。研究結果顯示，三種預測模型之 MAPE 均低於 2.5%，優於公司內現行之預測方法，其中以倒傳遞網路所建立的預測模型具有最佳的預測效果，徑向基函數網路則又優於傳統迴歸分析。

關鍵字

類神經網路 ； 徑向基函數網路 ； 倒傳遞網路 ； 迴歸分析

並列摘要

Wafer acceptance test (WAT) results are the basis of shipping wafers to foundry customers. If wafers fail in the criteria of WAT, the customer will reject them due to low yield. However, if we can predict the results of WAT during wafer process, engineers could conduct problem solving process sooner the WAT abnormality resulted. As a result, the quantity of WAT abnormal lots will be effectively reduced.

The main purpose of this study is to develop a neural network-based prediction model for WAT. Real WAT data and wafer process measurement data are employed to verify our proposed prediction model.

Two ANN models are considered in this research, namely RBFN and BPN. The traditional regression analysis is used as a benchmark for comparison. The mean absolute percent error (MAPE) is used as the primary performance measure in this research. A comparative study shows that the MAPE values of three prediction models are less than 2.5% and better than current method used in the company. BPN has the best overall performance, followed by RBFN and regression model.

並列關鍵字

ANN ； RBFN ； BPN ； regression analysis

參考文獻

3.Bishop, C., “Neural networks for pattern recognition,” Oxford University Press, (1995).

5.Cha, I. and Kassam, S. A., ”Channel equalization using adaptive complex radial basis function network,” IEEE Journal on Selected Areas in Communications, 13, 122-131 (1995).

6.Cybenko, G., ”Approximation by superposition of a sigmoidal function,” Mathematics of Control, Signals, and Systems, 2, 303-314 (1989).

7.Davies, P. C., “Design issues in neural network development,” Neurovest Journal, 5, 21-25 (1994).

8.Eksioglu, M., Fernandez, J. E. and Twomey, J. M., “Prediction peak pinch strength: artificial neural networks vs. regression,” International Journal of Industrial Ergonomics, 18, 431-441 (1996).