Modelling High Shear Granulation using Artificial Neural Networks

High shear granulation is a commonly used technique for the size enlargement of powders in the pharmaceutical industry. It is a very complex process that is difficult to model due to the number of interacting material properties and processing conditions that control the granule properties. In this thesis an attempt to model the high shear granulation process in a Glatt/Powrex FM-VG-01 lab scale mixer/granulator using Artificial Neural Networks (ANNs) has been made. ANNs are well suited to modelling tasks that contain 'noisy' and interacting data. The properties of granules prepared with three different grades of calcium carbonate, one grade of magnesium carbonate, one grade of powdered cellulose and Eudragit RPSO with binder solutions of either polyvinylpyrrolidone or hydroxypropyl methylcellulose were investigated using ANN models. All the powders were chosen as they are insoluble in water and exhibit a range of surface energetic properties, surface areas, flow characteristics and particle sizes. Granulation experiments were carried out using variable processing conditions with all the powder/binder combinations. The porosity, friability, the mean granule size by sieving, a measure of the width of the granule size distribution by sieving, flow properties and binder content variability were all determined after preparing and drying the granules. The data for each granule property were used in ANN models and their ability to predict the outcome of previously unseen experimental data was evaluated. Overall the ANNs showed some promise in modelling the porosity and the width of the granule size distribution, and were particularly accurate in predictions of the binder content variability. There was some agreement with direct observations made with simple graphical analysis of the data and with literature data. However, it was found that caution was required when using ANNs as a tool to model relationships in complex data as critical input variables may need to be removed to improve the quality of fit. This is due to the 'black-box' nature of ANNs. Modelling of the other granule properties were not considered accurate enough to be of practical use