Neural networks as a tool for modelling the retention behaviour of dihydropyridines in micellar liquid chromatography

Artificial neural networks have been applied for modelling the retention behaviour of a group of dihydropyridines in Micellar Liquid Chromatography and their capabilities have been compared with classical statistical methods (multiple regression). Some parameters concerning the architecture of the nets have been evaluated. Thus, type of transfer function in the hidden layer, number of neurons in the hidden layer and the use of recurrent networks and transformations of the output data have been optimized. These studies have been carried out with the retention data for 27 dihydropyridines in an octyl silica column using micellar mobile phases containing hexadecyl trimethylammonium bromide (CTAB) or sodium dodecyl sulphate (SDS) as surfactants and n-propanol or n-butanol as organic modifiers. From the results obtained some considerations can be drawn. The selection of the transfer function is very important to get good results, the logarithmoid function being the best one among the different functions studied. Also, the least errors were obtained when three nodes in the hidden layer and the reciprocal of retention factors (as the output variable) were employed.