我國半導體是以生產製造為主,如何降低生產成本、增加設備效能、減少製程風險及提昇製程良率,為提升其國際競爭力最重要的因素。目前的半導體廠大部分仍是採用統計製程管制技術來監控製程參數,以例行測機及預防保養等方式提高製程穩定性,改善製程良率。對於無預警式的系統良率損失來說,當製程中發生異常時,會造成大量數目的報廢,嚴重地影響到成本與產能。 本篇論文以化學機械研磨(Chemical Mechanical Polishing,CMP)製程結合虛擬量測(Virtual Metrology, VM)的概念,再搭配類神經網路與模擬控制,進行趨勢預測以與配方調整預測,整合這兩方面對於響應的影響來做量测的預測,以達到改善製程及提昇設備效能的目標。 在虛擬量測的系統架構下,結合趨勢狀態與配方響應的預測結果,在模擬資料的部分,從半導體廠中化學機械研磨設備取得,取針對影響研磨速率最重要的研磨平台轉速、研磨頭轉數、製程時間以及研磨時之下壓力量,作為神經網路的輸入,每秒的研磨量作為輸出,以網路學習方式來建立類神經網路系統模型,並以實際預測兩台CMP機台參數來驗證平均絕對誤差在0.944Å/sec與0.967Å/sec,而整體虛擬量測系統的預測結果的準確性令人滿意。
Semiconductor manufacturing is the leading industry in Taiwan. In this competitive market, reducing fabrication cost, increasing overall equipment efficiency, and promoting yield and throughput are the key factors to promote the competitiveness. Currently, statistical process control (SPC) is the most commonly used method for the monitoring of manufacturing processes in the semiconductor industry. In addition, routine test and preventive maintenance of process tools are also performed for keeping process stable. Any abnormal process will lead to wafer scraps and yield loss. This thesis is concerned with the virtual metrology for CMP process. A hybrid system, consisting of piecewise linear neural network for estimating the trend of system dynamics and fuzzy neural network for calculating the possible change on the process outcome due to the adjustment of recipe, was developed. The aim of virtual metrology is to improve process quality and to promote equipment efficiency. Four important process parameters including platen speed, revolutions of polishing head, polishing time, and polishing downforce were collected from CMP tools as the input of neural network. Removal rate was the output of network. The average absolute errors of virtual metrology for two CMP tools were 0.944Å/sec and 0.967Å/sec, respectively. The system was proven to have good performance for CMP process.