NONLINEAR MULTIVARIATE CALIBRATION USING PRINCIPAL COMPONENTS REGRESSION AND ARTIFICIAL NEURAL NETWORKS

This manuscript describes methods for detecting and modeling nonlinear regions of spectral response in multivariate, multicomponent spectroscopic assays. Simulated data and experimental UV/visible data were used to study the capability of multivariate linear models to approximate nonlinear response. The sources of real and apparent nonlinearity simulated included nonlinear instrument response functions (e.g. stray light), concentration-dependent wavelength shifts, and concentration dependent absorption bandwidth changes. A weighting algorithm was devised to reduce the influence of nonlinear spectral regions in principal component regression (PCR) calibrations, thereby improving the performance of multivariate linear calibration models. Second-order calibration methods using quadratic principal component scores and nonlinear calibration methods using artificial neural networks were compared to unweighted and weighted linear calibration methods. Orthogonal transformation of the input variables was used to significantly improve neural network training speed and reduce calibration error. Some conditions where second-order and nonlinear calibration techniques outperform linear calibration techniques have been identified and are described.