Comparison of partial least squares regression and multi-layer neural networks for quantification of nonlinear systems and application to gas phase Fourier transform infrared spectra

The performance of back-propagation artificial neural networks (NN) and partial least squares (PLS) regression for the calibration of linear and nonlinear systems has been investigated by using six types of synthetic data. Three PLS methods, conventional linear-PLS and two nonlinear-PLS methods, have been used in the study. In all but one of the synthetic data types, the band intensities varied nonlinearly with concentration. These five data types were designed to represent the effect of band shifts with increasing concentration, a nonlinear relationship between peak height and concentration, or a combination of both types of nonlinearities. The results showed that NNs perform better than PLS for all the nonlinear datasets. When a band shift is the major reason for the nonlinearity, the relative performance of NNs and PLS depends on the overlap of the absorption bands. If there is no band overlap, neither NN nor PLS can calibrate the data accurately but the results could be improved by convolving the spectral features with a Gaussian broadening function. The results indicate that a combination of peak position shift and peak height change is the most difficult nonlinearity to calibrate. NN and PLS were also used to determine the concentration of CHCl₃ in pure component and mixtures of CHCl₃ and CH₂Cl₂ using their Fourier transform infrared (FT-IR) spectra, a dataset that has been proved nonlinear in high concentrations due to the nonlinear response of the detector. The best results for the experimental data were obtained by applying one hidden layer NN to the mean-centered absorbance spectra.