在先進的半導體晶圓製造廠,線上生產製程之機台皆需維持高穩定性和高良率。目 前廠商在產品檢測的做法是在機台中放置監控晶圓,並藉由量測機台量测監控晶圓來維 持產品晶圓的品質。此種監控方式的缺點是在製程中發生異常時,並無法即時得知生產 機台出現錯誤。先前已有相關研究使用虛擬量測(Virtual Metrology, VM)方法,利用生產 機台環境參數之穩態資料來預測產品品質,然而,僅使用製程參數之穩態資料對於預測 的幫助有限,因此,本論文不但使用以往之穩態資料,更進一步加入暫態資料(transient production equipment data)於徑向基底函數類神經網路(Radial Basis Function Networks, RBFN)的虛擬量測系統,以提升虛擬量測精度,來克服上述問題。 此外,為了保有虛擬量測對線上新進資料的預測準確度,本論文也使用支持向量資 料描述(Support Vector Data Description, SVDD)來提供VM一個線上學習之機制,藉由判 斷新進製程資料與舊有歷史製程資料的相似程度來給予信心值,可協助虛擬量測系統判 定是否需要即時重新訓練。 此外,目前半導體廠是針對監控晶圓上數個固定位置之量測點進行厚度量測。整體 量測過程將耗時約15分鐘,每一個量測點平均耗費20~30秒。本論文針對減少實際量測 點問題,利用數學趨近方法將晶圓上可被取代之固定量測點去除,可使機台量測更少的 點,而量測結果又幾乎驅近於原始結果。本研究不但可替廠商縮減量測時間,於量測所 耗費的資源也可有效減少。
In advanced semiconductor manufacturing, both on-line stability and yield for production equipments need to be maintained. In current practice, the method for product inspection is to place the monitoring wafers into the equipments so that the wafer quality can be maintained by measuring the quality of the monitoring wafers. However, such a method cannot feedback the error information about the production equipments in real time when some anomalies occur. To solve this, the virtual metrology (VM) method has been proposed. However, the previous VM methods only considered the steady-state data of process parameters as the inputs, of which the VM accuracy will be limited. To improve the VM accuracy, in this thesis, not only the steady-state process data, but also the transient-state process data will be used as the input for the proposed RBFNN (radial basis function neural network) based VM system. In addition, a support vector data description (SVDD) based on-line learning mechanism (OLM) is also proposed to maintain the VM accuracy for the newly coming process data. OLM is able to assign a confidence value to a newly coming data by measuring the similarity between the new data and the historical data (training data). Then, OLM will determine whether the original VM system needs to be retrained based on the measured confidence value. Additionally, the mean thickness of a wafer is obtained by measuring and averaging the thicknesses of 17 measurement points with different positions on the wafer via metrology equipment. In real practice, it will take about 15 minutes to accomplish the actual metrology task for each wafer. That is, the measurement time for each measurement point is around 20-30 seconds, which is time consuming and increases the cycle time. Therefore, how to reduce the number of actual measurement points is a key issue for a semiconductor manufacturer. To solve this problem, this thesis proposes a curve-fitting based method to achieve the goal of measurement-point reduction. Experimental results show that the proposed method can reduce the number significantly while keeping almost the same measuring average thickness.