藉由半導體的製程相關技術於過去30多年的快速進展之下，許多高科技產品才得以問世，由此可見半導體產業在工業發展過程當中，著實扮演著相當吃重的角色。然而，半導體是一種高技術層次、高投資成本、高度標準化等高門檻特色的產業，因此，各家半導體廠無不致力於提高獲利能力，在眾多方法論當中，良率管理 (Yield Management) 是最受重視而相繼投入研究的；簡單來說，良率可以定義成產線所產出之良品佔所有投入生產總數的百分比，而其中伴隨著製造所產生的龐大的眾多資料，透過系統化的方式予以交叉分析以預測出良率 (Yield Prediction) ，務求最終能達到改善良率 (Yield Enhancement) 、提高獲利能力的目的。 承上所述，本研究特地選定幾種目前在半導體領域裡常被研究和應用的良率模式 (Yield Model) ，利用相同的缺陷檢測 (Defect Inspection) 資料來評估各個良率模式的優劣，其中亦包含了『個案公司』自行發展的良率模式，其特色在於利用每一顆晶粒於缺陷檢測後的分類結果，建構出簡單而能廣泛套用於所有缺陷檢測站點 (Defect Inspection Step) 的良率模式，藉由這樣的評估分析，本研究最後得以提供相關的決策規則以提供『個案公司』管理階層參考。 經過上述的評估分析之後發現，『個案公司』自行發展的良率模式有相當不俗的預測能力表現，約莫有58%的缺陷檢測站點在評估後顯示，『個案公司』自行發展的良率模式有既精確又穩定的預測能力，甚至僅從「精確度」的角度來分析時，100%的缺陷檢測站點均顯示『個案公司』自行發展的良率模式有最精確的預測能力。 Thanks to the rapid progress of semiconductor technology for the past 30 several years, many high-tech products could be hence developed. Therefore, we could hereby understand how important the role that semiconductor plays in the industrial development. However, semi- conductor is a field with high thresholds in technology, cost and standardization, therefore, all semiconductor companies aim for increasing their profitability without any exception. Among so many methodologies, Yield Management is highly emphasized and researched. In brief, Yield could be defined as a percentage of good chips over all chips in production line. In order to gain Yield enhancement and increase the profitability, we must predict Yield by analyzing the huge amount of data generated during the manufacturing process systematically. In this thesis, several popular Yield Models in semiconductor field were evaluated by the same Defect Inspection data to understand which one is the best for each Defect Inspection Step, including those Yield Models developed by the Case Company. The Case Company intended to develop the simple Yield Models based on the classifications of the tested chips, and tried to apply such Models on all Defect Inspection Steps. After evaluation, this thesis could contribute several decision rules to the management level of the Case Company to help make decisions. According to the above evaluation, we found those Yield Models developed by the Case Company all have pretty good performance in Yield prediction. There were about 58% of the Defect Inspection Steps revealed that the Yield Models developed by the Case Company had both precise and stable results in prediction. The prediction results became even better, 100%, if we only considered precision to evaluate those Yield Models.
