Comparison of prediction power between theoretical and neural-network models in ion-interaction chromatography

Abstract

The separation by ion-interaction chromatography (IIC) of metal complexes having single and double charges has been studied in order to compare the prediction power of soft (neural-network) and hard modelling (IIC equation). The two approaches have been used to model the retention behaviour as a function of the composition of the mobile phase. With ion-interaction mobile phases, the parameters involved included the concentrations of ion-interaction reagent, organic modifier and ionic strength. From a set of 69 experimental design points (the different mobile phase compositions at which capacity factors are measured), one test set of ten design points and ten training sets, containing from 59 to 11 design points, have been extracted. Chromatographic and chemometric considerations for the selection of the data sets and minimum number of observations required have been discussed. The study showed that the IIC equation predicted more accurately when few experimental data were available, while a similar prediction power was obtained with both models when the number of data was more than 17. Nevertheless the neural-network accounted for a greater versatility without the need to develop an equation.