化學機械研磨之虛擬量測與製程控制

Abstract

本論文主要目的為針對半導體廠內之化學機械研磨機台，設計一套Run-to-Run控制系統，以不斷調變各區域的研磨壓力，使研磨後之晶圓保持在理想的平坦度內。 首先，透過類神經網路建立輸出預測模型，利用此模型做為虛擬感測器，即可在未量測的情形下預測晶圓研磨後之輪廓，並將此輸出值供Run-to-Run控制器回授使用。Run-to-Run控制器則是先透過實驗設計法得到機台輸入與輸出間之非線性迴歸模型，並同時將迴歸模型之殘差擬合時間序列模型，以補迴歸模型之不足；結合此兩模型，再利用遞迴最小平方法，由線上量測資料不斷地隨機台狀況同時更新此兩模型。以透過更新後的模型為預測器，由最小變異控制器計算出下一片製程之輸入參數，使得輸入參數改變量最小的情況下，使輸出靠近所期望的目標值，完成此一製程控制系統。

In this thesis, a run-to-run control system is designed for chemical-mechanical polishing process. By continuously adjusting the pressure of different polishing zones, the nonuniformity can be constrained in the desired specification. First, the output prediction model is established by artificial neural network, and is used as a virtual sensor. It can estimate the current output for the run-to-run controller when there is no measurement. About the run-to-run controller, we first get the nonlinear input-output regression model of the machine by design of experiments, and also use the residual of the model to fit another time-series model for the compensation of the mismatch of the regression model, then, combine the two models. Moreover, we use the recursive least squares to continuously update the model with on-line measurement data in order to follow the situation of the machine. By regarding the updated model as a predictor, we can calculate the recipe for the next run by the minimum variance controller, with the rule of minimizing the change of each input and approaching the output to the target, and the process control system is accomplished.