Melt index modeling with support vector machines, partial least squares, and artificial neural networks

This article presents the application of three black-box modeling methods to two industrial polymerization processes to predict the melt index, which is considered an important quality variable determining product specifications. The modeling methods covered in this study are support vector machines (SVMs; known as state-of-the-art modeling methods), partial least squares (PLS), and artificial neural networks (ANNs); the processes are styrene–acrylonitrile (SAN) and polypropylene (PP) polymerizations currently operated for commercial purposes in Korea. Brief outlines of the modeling procedure are presented for each method, followed by the procedures for training and validating the models. The SVM models yield the best prediction performances for both the SAN and PP polymerization processes. However, the ANN models fail to accurately predict the melt index when sufficient data are not available for model training in the PP polymerization process. The PLS models are not effective either when applied to the SAN polymerization process, for which the melt index has strong nonlinear functionality with the process variables. The good prediction performance that the SVM models show despite the insufficient data or strong process nonlinearity suggests that SVMs can be effectively used as alternative to PLS or ANNs for modeling the melt indices in other polymerization processes as well. © 2004 Wiley Periodicals, Inc. J Appl Polym Sci 95: 967–974, 2005