基於支持向量迴歸及支持向量資料描述 之半導體CVD 製程虛擬量測研究

A Study on CVD Virtual Metrology System Based on Support Vector Regression and Support Vector Data Description

林昌誠(CHANG-CHENG LIN)

指導教授 ： 劉益宏

中原大學/工學院/機械工程研究所/碩士(2008年)

https://doi.org/10.6840/cycu200900545

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，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。
本論文提出一套虛擬量測(Virtual Metrology, VM)系統來克服這個問題。本虛擬量測
系統結合了兩種新的機器學習演算法。首先，此系統利用支持向量迴歸(support vector
regression, SVR)當作VM預測模型，其可以針對未訓練過之樣本提供不錯的擴充能力
(generalization performance)，所以可以達到較高之預測精度，此外本系統也具備了VM
線上學習之機制。此功能以支持向量資料描述(Support Vector Data Description, SVDD)演
算法來實現，它可以自動評估新輸入資料與舊有資料的相似度，若差異過大，則啟動線
上學習機制，重新訓練SVR與SVDD，進而提升虛擬量測系統線上預測精度。由實際12
吋晶圓廠所提供之資料所得到實驗結果顯示，本論文提出之虛擬量測系統預測精度優於
輻射基底類神經網路，並證明本論文提出的虛擬量測系統對於新輸入資料有較好的擴充
能力。

關鍵字

虛擬量測系統 ； 機器學習 ； 半導體製造 ； 支持向量迴歸 ； 支持向量資料描述

並列摘要

In advanced semiconductor manufacturing, the in-process wafers need to be monitored
periodically in order to obtain high stability and high yield rate, 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 and sudden drift of the equipment can not be
observed in real time by this way. This may result in defective wafers and they will therefore
be discarded, leading to a high production-cost loss.
To overcome this problem, this study proposes a virtual metrology (VM) system in
which two kinds of machine learning algorithms are embedded. First, the support vector
regression (SVR) is adopted as the predictor for wafer-thickness prediction. SVR can provide
good generalization performance for unseen samples, thus achieving better prediction
accuracy. In addition, this VM system also embodies an online learning mechanism, which is
implemented by using support vector data description (SVDD) algorithm. SVDD is able to
evaluate the prediction accuracy of a new input data. If the difference is too large, the online
learning mechanism will be activated to retrain the SVR predictor, thus enhacing the online
prediction accuracy of VM system. The experimental results, carried out on real data provided
by a 12-inch wafer fab, show that the proposed SVR-based VM system outperforms Radial
basis function Neural Network (RBFNN), and prove that the proposed VM system has better
generalization performance for new input data.

並列關鍵字

Support Vector Data Description (SVDD) ； Semiconductor manufacturing ； Support Vector Regression (SVR) ； Virtual Metrology (VM) ； Machine learning

參考文獻

Independent Component Analysis and Support Vector Data Description,” Proceedings of

[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

[27] 呂學信，“結合SVDD及SVM分類器於會員人臉確認之研究,” 私立中原大學機械工

[31] 馬晨榮, “應用暫態製程參數於半導體晶圓品質預測與晶圓品質量測簡化, ”私立中

[10] 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