Linear and non-linear multivariate calibration methods for multicomponent kinetic determinations in cases of severe non-linearity

Abstract

Both p-aminophenol ( p-AP) and p-phenylenediamine ( p-PDA) undergo oxidative coupling reactions in the presence of octacyanomolibdate(V) as oxidant. The kinetics of these reactions can be followed spectrophotometrically. When they happen to be in the same reaction mixture, relevant characteristics are (1) related to reaction conditions; the reagent must be in deficit and (2) related to the reaction products; ‘cross-coupling’ reactions are also involved. All this gives rise to systems with severe non-linearities. This is the reason why they have been used here to test the predictive ability of several linear and non-linear algorithms in cases of special complexity. Principal component regression (PCR), partial least squares regression (PLS) and three different forms of applying artificial neural networks (ANNs), with and without reduction of variables, have been tested. All the models were applied to the same calibration and validation sets. Typical prediction errors ranged between 2 and 5%. Not very different results were found when independent accuracy and precision studies were accomplished, regardless of whether linear or non-linear calibration methods were used. This is probably due to differences in the spectroscopic as well as the kinetic behaviour of both analytes. These differences allowed the construction of ANN calibration models with only 30 original variables (kinetic profiles of 10 data points at three selected different wavelengths).