Hydroxyl and acid number prediction in polyester resins by near infrared spectroscopy and artificial neural networks

Back-propagation artificial neural networks (BP-ANN) are applied for modeling hydroxyl number and acid value of a set of 62 samples of polyester resins from their near infrared (NIR) spectra. The results are compared to the classical calibration approaches, i.e. principal component regression (PCR) and partial least squares (PLS). The set of available samples is split into: (i) a training set, for models calculation; (ii) a test set, for setting the correct number of latent variables in PCR and PLS and for selecting the end point of the training phase of BP-ANN; (iii) a "production set" of samples, which are predicted to evaluate the models predictive ability. This approach guarantees that the predictive ability of the models is evaluated by genuine predictions. BP-ANN resulted always better than the classical PCR and PLS, from the point of view of the predictive ability. The study of the breakdown number of experiments to include in the training set showed instead that this factor does influence PCR and PLS at a lesser degree than what happens for BP-ANN. The latter approach requires a larger number of experiments for obtaining good results. The choice of optimal training sets is efficiently performed by Kohonen self-organizing maps (SOMs). It can be concluded that FT-NIR spectroscopy and BP-ANN models can be properly employed for monitoring the polyesterification of dicarboxylic acids with diols by predicting the acid and hydroxyl numbers directly along the process line.