MODELING OPTICAL EMISSION SPECTROSCOPY DATA GENERATED DURING REACTIVE ION ETCHING USING AN AUTOENCODER NEURAL NETWORK

摘要

Optical emission spectroscopy (OES) is used to monitor plasma emission intensity in reactive ion etching (RIE) systems as a means for determining process endpoint. While OES is an excellent tool for monitoring plasma emission intensity, its primary problem is the large dimensionality of the spectroscopic data. To alleviate this concern, the use of an autoencoder neural network (AENN) is proposed as a means of feature extraction to reduce the dimensionality of OES data. AENNs are trained with OES data generated from factorial experiment designed to characterize RIE process variation during the etching of benzocyclobutene (BCB) in an SF_6/O_2 plasma with controllable input factors consisting of the two gas flows, RF power, and chamber pressure. Models of etch rate, uniformity, selectivity, and anisotropy are constructed. For each response, AENN with 7 hidden neurons is able to capture the relevant variation with scaled RMS errors less than 6%. Subsequently, multi-layer perceptron NNs are used to model the responses using only the compressed OES data. These models exhibit scaled RMS prediction errors as low as 1.49%.