System Identification Methods for Plasma Enhanced Chemical Vapor Deposition

Abstract

The introduction of new in situ sensing creates the possibility of directly controlling critical process variables in plasma enhanced chemical vapor deposition systems (PECVD). The complexity of this process makes it necessary to develop empirical models of the system dynamics. This paper describes the experimental and numerical procedures for identifying both transfer function and recurrent neural net multi-input multi-output models of PECVD dynamics. The models are compared in terms of the resulting mean-squared-prediction-error for verification data sets. The neural net models are shown to be mildly superior in cases where the input signals vary widely, stimulating the nonlinear response of the system.