應用模糊類神經網路於銅膜化學機械研磨之批次製程控制

Abstract

化學機械研磨(Chemical Mechanical Polishing, CMP)是一個極為複雜的製程,此系統大致包含化學,機械以及物理三個研究領域[1]，論文[1]中所探討的是CMP的整個研磨機制,以及推導CMP Model,並以實驗驗證其Model的準確性。然而在現代半導體廠房裡,製程控制(Process Control)應用在CMP製程上時，一般都是使用R2R(Run-to-Run)製程控制的方式來做移除率(Remove Rate, RR)或不均勻度(Non-Uniformity)的控制。 本研究使用的控制方法主要是參考了論文[2]裡所提之誤差零超越量的觀念並結合了論文[3]所提出的輻射基底函數式模糊類神經網路(Radial-Basis Function Based Fuzzy Neural Network)來預測CMP製程，最後提出一零誤差追蹤模糊類神經網路控制器，以達成移除率(Remove Rate, RR)的控制，本研究所用之類神經網路並不完全和論文[3]所提的一樣，本研究之網路不需要將輸入正規化,有別於論文[3]所提之輻射基底函數式模糊類神經網路。模擬時並與傳統的EWMA Run-to-Run製程控制法做一效能之比較，最後以實驗來驗證模擬之結果。

Chemical Mechanical Polishing (CMP) is a very complex process and the process involved many areas of studying [1]. They are Chemistry, Mechanics, Physics, etc. The research of [1] explores the removal mechanism of the CMP process and it also derives the model of CMP process. It verified the accuracy of the CMP model in experiment. However, the so called “Process Control” applying to the CMP process in the semiconductor fabrication today uses the method of Run-to-Run process control to control the “Remove Rate (RR)” or the “Non-Uniformity (NU)” of the CMP process. The design of controller of my research is designed mainly from two papers [2][3]. One of them is the “Zero Error Tracking” concept from paper [2] and the other is the “Radial-Basis Function Based Fuzzy Neural Network” from paper [3]. From the concept in paper [2], the error will converge to zero. On the other hand, the “Radial-Basis Function Based Fuzzy Neural Network” can predict the CMP process. Combing both papers [2][3], I finally derive a new controller. It is called “Zero Error Tracking Fuzzy Neural Network Controller”. In my research, the purpose of the controller is to control the Remove Rate of CMP process. The “Fuzzy Neural Network” used in my research is not the same as that in paper [3]. In my research, the input data does not need to be normalized between -1 and 1. In the simulation, I compared the effectiveness between the traditional EWMA (Exponentially Weighted Moving Average) Run-to-Run process controller and the one I designed. Finally I confirm my simulation results by experiment in NDL (National Nano Device Laboratories).