Multivariate analysis of chemical data using multilayer perceptrons

The successful implementation of multilayer perceptron network models in the analysis of nonlinear, multivariate data involves a complex strategy which involves a number of interdependent design features. The analysis of nonlinear, multivariate data structures that are typically encountered in chemical analysis deserves special consideration because of the expense of the acquisition of data, which is typically obtained through laboratory experimentation, and because an understanding of the physical nature of the data permits the application of a priori knowledge in the development of a modeling strategy. Such a strategy is developed and applied to simulated and experimental data in this work. Multilayer perceptron network models are developed using an understanding of chemical systems and of the network design process. The successful development of network models demands a thorough understanding of their mathematical function, which is available in the engineering and computer science literature, while understanding of the nature of chemical data structures is a central part of the field of chemometrics. In this work, the optimization of the network weights via multiparameter, nonlinear optimization methods is integrated with the strategies of multivariate chemical analysis via latent variable methods commonly found in the chemometrics literature. The studies contained herein involve the use of multilayer perceptron networks and some of the more conventional chemometric methods in the modeling of nonlinear multivariate calibration, and pattern recognition of spectroscopic data. The advantages and disadvantages of modeling with multilayer perceptrons as compared with nonlinear biased regression methods derived from Partial Least Squares and Principal Components Regression, and classification discriminants based on linear and quadratic discriminant analysis are examined.