基於模糊支持向量迴歸之半導體CVD製程厚度預測

FSVR-based Thickness Prediction for CVD Process

許晉榮(Jin-Rung Shiu)

指導教授 ： 劉益宏

中原大學/工學院/光機電及資電控產業研發碩士專班/碩士(2009年)

https://doi.org/10.6840/cycu200900507

摘要

在先進的半導體製造(Semiconductor manufacturing)中，生產製程上之晶圓皆須進行監控以維持生產機台高穩定性和高良率，確保所有機台在製程上有良好的穩定性且合乎品質規格的產出。目前作法是在機台中放置三片監控晶圓(monitor wafers)，並藉由量測機台去量测監控晶圓來代表產品晶圓的品質，然而，如此不但增加製程的循環時間(cycle time)，且監控晶圓在尚未量測之前，無法即時察覺出機台發生性能漂移的問題。這可能導致生產出有瑕疵的晶圓，因此仍產出之產品被作廢，造成極高的成本損失。
本文以模糊支持向量迴歸(Fuzzy Support Vector Regression, FSVR)演算法來對半導體CVD機台所製成的晶圓沉積厚度來做預測。在訓練FSVR時，每一筆資料的重要性皆不同， 本論文依照資料產生時間的不同來給予該筆資料不同的權重值。權重值越大，代表該筆資料在訓練時就越重要。因此，本論文建構了一個模糊歸屬函數(Fuzzy Membership Function)，其可依照輸入資料的產生時間來給予此筆資料一個權重值。再利用量測晶圓的厚度資料來訓練此FSVR模型。訓練完的FSVR模型，就能用來預測真實的晶圓厚度，且不需實際量測就能控制晶圓的品質。考慮到生產機台型號的不同，本文也會對不同的製程分別建立適合的預測模型。此外，特徵篩選對晶圓厚度的預測精確度也會有相當大的影響。因此，本論文也針對有無特徵篩選來進行實驗與比較，用來找出最佳的特徵組合，以得到良好的預測精度。實驗結果證明，本論文所提出的FSVR預測模型不但比傳統的SVR模型好，而且可以達到小於0.5%的MAPE值。實驗資料皆由實際DRAM廠中所取得。

關鍵字

模糊支持向量迴歸 ； 化學氣相沉積 ； 特徵篩選 ； 歸屬函數

並列摘要

In advanced semiconductor manufacturing, in-process wafers need to be monitored periodically in order to achieve high stability and high yield rates, by which the products that satisfy the specifications can be obtained. Till now, the general method is that a monitoring wafer is inserted to three pieces of the process at a specific time for measuring the quality, However, it not only increases the cycle time, but also the sudden drift of the equipment can not be observed in real time. This may result in a large number of defective wafers and they will therefore be discarded, leading to a high production-cost loss.
This thesis adopts a FSVR (Fuzzy Support Vector Regression) algorithm to predict the deposit thickness of each wafer manufactured by IC CVD equipments. The importance of each data is different in the training of the FSVR model. Each data is assigned different weight according to the production time of the data. The larger the weight value, the more important the data in the training stage. Therefore this thesis proposes a fuzzy membership function, which can assign the data a weight value according to the production time of this data. Furthermore, the FSVR model is trained on the thickness data of monitored wafers. The trained FSVR mode can therefore be used to estimate the true thickness of wafers. In consideration of various production equipments, this thesis would also construct different prediction models associated with different processes. In addition, feature selection would also influence the prediction accuracy greatly. Therefore thesis also try to find the optimal feature combination by comparing the experimental results. Experimental results show that the proposed FSVR model not only performs better than the traditional SVR model, but also achieves a MAPE value of less than 0.5%.The experiment data were acquired from a real DRAM fab.

並列關鍵字

Membership function ； Fuzzy Support Vector Regression ； Chemical Vapor Deposition(CVD) ； Feature selection

參考文獻

[9] 顧尚芳，“生產系統中利用製程不良率評估設備預防維護之研究,” 私立中原大學工業工程學系碩士論文，2003.

[24] 張逸凡，“支撐向量機在即時河川水位預報之應用,” 國立成功大學水利及海洋工程研究所碩士論文，2006.

[29] 馬晨榮，”應用暫態製程參數於半導體晶圓品質預與晶圓品質量測簡化”私立中原大學機械工程研究所碩士論文，2008年.

[2] Y.-J. Chang, Y. Kang, C.-L. Hsu, C.-T. Chang, and T. Y. Chan, “Virtual Metrology Technique for Semiconductor Manufacturing,” in Proc. 2006 International Joint Conference on Neural Networks (IJCNN’06), pp. 5289-5293, July 2006.

[11] Y.-J. Chang, Y. Kang, C.-L. Hsu, C.-T. Chang, and T. Y. Chan, “Virtual Metrology Technique for Semiconductor Manufacturing,” in Proc. 2006 International Joint Conf. Neural Networks, pp.5289-5293, 2006.