APPLICATION OF ARTIFICIAL NEURAL NETWORKS IN THE OPTIMIZATION OF HPLC MOBILE-PHASE PARAMETERS

The prediction capability of forward feed neural networks was tested on computer generated capacity factors. The capacity factors were simulated from equations reflecting the contribution of mobile phase changes in pH organic modifier concentration, and ion-pair concentration. Simulated data allows an appropriate experimental design which assures the training of the network does not involve memorization but guarantees the network we generalize. The use of different mathematical forms to calculate the behaviour of capacity factor with changes in pH, methanol concentration and ion-pair concentration permitted us to explore the capability of neural networks to fit a variety of curves. Each of the independent variables were studied separately, and then in combination. The effect of variable transformation played a very important role in effective training of the network. The neural network output equations were used to formulate nonlinear regression problem and the behaviour of this model was compared to the neural network system. When the neural network systems had only sufficient processing units needed to solve the problem, nonlinear regression models and neural networks arrived at identical solutions. When the network contained excessive neurons, nonlinear regression techniques were unstable, having high intraparameter correlations and showing matrix singularity.