基於粒子群演算法之最佳化支持向量迴歸
及其在半導體製程診斷應用

在半導體製造(semiconductor manufacturing)中，必須維持晶圓生產機台的穩定性，以及產品晶圓的良率，確保所有產品都能符合品質規格中。目前檢測品質規格的作法是在生產機台中放置三片監控晶圓(monitor wafer)，當製程完成時，利用量測機台對監控晶圓做量測，代表此批製程產品之產品品質，但此方式會增加製程的循環時間，且當監控晶圓尚未被量測之前，無法得知機台是否發生異常或漂移問題。可能生產出有瑕疵的晶圓，導致整批晶圓報廢，增加生產成本。本論文提出一套線上診斷系統來解決上述問題。本系統結合支持向量迴歸(Support Vector Regression, SVR)與支持向量資料描述(Support Vector Data Description, SVDD)作為系統核心，SVR的作用為預測晶圓品質，而SVDD用來評估新近製程資料與歷史資料的相似度，若差異過大，則啟動實際量測，並經由重新訓練指標判斷需重新訓練SVR或重新訓練SVDD，藉此提升線上診斷系統預測準確度。在系統實際運作前，必須將歷史製程資料訓練出SVR最佳模型，而訓練SVR最佳模型必須找尋最佳參數組合，傳統方法是利用格子搜尋法，但其方法缺點是非常耗時，因此本文提出利用啟發式演算法來找尋建模最佳參數組合，其利用2種啟發式演算法與傳統參數搜尋法來做參數搜尋比較，分別為粒子群演算法(Particle Swarm Optimization, PSO)、禁忌搜尋法(TABU search)以及格子搜尋法；由實際晶圓廠提供之資料經由實驗結果顯示，利用PSO能夠大幅縮短最佳參數的搜尋時間，並驗證本論文提出之線上診斷系統搭配PSO參數搜尋，對於新進資料能夠有快速且精確的預測的效果。

關鍵字

線上診斷系統；支持向量迴歸；禁忌搜尋法；粒子群演算法；半導體製造；支持向量資料描述

並列摘要

In semiconductor manufacturing, it is necessary to maintain the stability of wafer machine and wafer yield. To make sure all the wafer production can be qualified in the specification, the current way of inspection is to put three monitor wafers in a wafer machine before and after. When the manufacture procedure achieve, we can know the quality of all the other wafers by measurement equipment to measure the monitor wafers. However, the process increases the cycle time and the problems of the machine cannot be detected in real time until the monitor wafer measurement is done, which means there is risk of mass wafers scrap and costing up. This thesis proposes an e-diagnostics system to solve the problems described above. The e-diagnostics system incorporates Support Vector Regression (SVR) and Support Vector Data Description (SVDD). SVR is used to predict wafer quality, while SVDD is to compare new process data with historic database. When the new data unusual, it will execute in-situ measurement. And through retrain index, we can figure out whether SVR or SVDD needs to be retrained, which makes e-diagnostics system work better. Before the system is operated in the production, the collection of historic database is helpful to find out optimum parameters and to train the optimum SVR model. The traditional grid searching technique is too time-consuming to apply. Therefore, the thesis proposes heuristic algorithm as a way to search for the optimum parameters. The thesis compares the efficacy of three kinds of parameter search methods, including Particle Swarm Optimization (PSO), TABU search, and Grid search. The experiments using the data from wafer companies show that PSO will greatly shorten the required time for optimum parameters searching and prove that the incorporation of e-diagnostics system and PSO will enhance fast and precise prediction for the new data.

並列關鍵字

TABU search ； Particle Swarm Optimization ； Suppor ； SVDD ； semiconductor manufacturing ； e-diagnostics system

參考文獻

[12] 吳泰熙，張智欽，“以禁忌演算法則解決推銷員旅行問題，” 大業學報，vol.2，no.1，pp.87-99，1997。

[16] 林伯勳，“最佳化方法於大地工程上之應用，” 中興工程期刊，no.103，pp.13-24，2008。

[45] 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程研究所碩士論文，2007年。

[46] 林佑駿， “應用遞增式支持向量機於TFT-LCD陣列電路工程閘電極光罩線中瑕疵辨識，” 私立中原大學機械工程研究所碩士論文,2007年。

[4] A. C. Diebold, “Overview of metrology requirements based on the 1994 national technology roadmap for semiconductor,” In Proceedings of IEEE/SEMI Advanced Semiconductor Manufacturing Conference and Workshop, pp.50-60, 1995.

國際替代計量

基於粒子群演算法之最佳化支持向量迴歸及其在半導體製程診斷應用

全文下載

主題瀏覽