Quantitative analysis of paper coatings using artificial neural networks

This paper describes a neural network approach to the quantitative analysis of paper coatings. Infrared spectra of samples of coated paper were used as input data for the different types of artificial neural networks. Unsupervised learning was applied to obtain the clustering of samples with respect to similarities in the spectra. The self-organizing artificial neural network of Kohonen type produced a visual representation of the discovered groupings on a two-dimensional plane. Such mapping provided the expert a possibility to analyze the mutual arrangement of samples and to predict the properties of the test samples using their relative position with respect to existing clusters. Supervised learning with a multilayer feedforward network was used to construct the non-linear models that relate the spectral information and concentrations of three basic components of paper coating - styrene, butadiene, and carbonate. These models were used for prediction of concentrations of paper coating components for the test data set. The results of modeling demonstrate that accuracy of classification and prediction is better than those obtained with traditional methods like principal component analysis or partial least squares (from 4% to 2% for different components). According to our experience, the modeling with artificial neural networks is intuitively clear for the expert. This method allows to construct complex multivariable and multiresponse models in unified style. Causal relationships between inputs and outputs can be analyzed and explained.