Neural network modeling of reactive ion etching using principal component analysis of optical emission spectroscopy data

Abstract

In this paper, neural networks trained by the error back-propagation algorithm are used to build models of etch rate, uniformity, selectivity and anisotropy as a function of optical emission spectroscopy (OES) data in a reactive ion etching process. The material etched is benzocyclobutene (BCB), a low-k dielectric polymer, which is etched in an SFs and 9 plasma in parallel plate system. Neural network training data are obtained from a multi-way principal component analysis (MPCA)of the OES data. These data are acquired from a z4 factorial experiment designed to characterize etch process variation with controllable input factors consisting of the two gas flows, RF power and chamber pressure. Evaluation of the trained neural networks is performed in terms of root mean square (RMS) error, and less than 3% prediction errors are achieved.

A novel algorithm applied to filter spam e-mails for iPhone

“…In this paper, the simulation results are compared with DT, backpropagation network (BPN), and support vector machine (SVM). BPN is the most widely used neural network model, and its network behavior is determined on the basis of inputoutput learning pairs [32,33]. SVM is a learning system proposed by Vapnik that uses a hypothesis space of linear function in a high-dimensional feature space [34].…”

Estimation and Control in Semiconductor Etch: Practice and Possibilities

“…Hong et al [72] compared the use of PCA and ANNs for feature extraction from OES data, with a further ANN used to model the reduced data. However, 226 "relevant" wavelengths are initially chosen from the 2048, prior to compression.…”

“…Himmel and May [65] found ANNs to be superior to quadratic response surfaces, which might be expected since complexity is more limited in the quadratic case. Hong et al [72] compared ANNs and PCA for data reduction in selectivity prediction, concluding that the ANN reduction was significantly better.…”

Multiple-inputs Dual-outputs Process Characterization and Optimization of HDP-CVD SiO2Deposition

“…For the purpose of process characterization, a single output neural network is generally recommended for building a model rather than two outputs model due to the complexity and convergence point of view [8][9][10]. For the NN modeling and GA recipe synthesis, OBOrNNs, a neural network simulator developed by Integrated Circuits-Computer Integrated Manufacturing Lab.…”

“…at Georgia Institute of Technology, was used. Detailed information about neural networks for semiconductor process modeling can be found in [10][11].…”