Developing ANN-Kriging hybrid model based on process parameters for prediction of mean residence time distribution in twin-screw wet granulation

Abstract

Artificial neural network (ANN) modelling is applied to predict the mean residence time of pharmaceutical formulation in a twin-screw granulator. Process parameters including feed flow rate, screw speed, and liquid to solid ratio are correlated with the obtained values of mean residence time to build a predictive tool. In order to improve the ANN predictive capability, a kriging interpolation approach is utilised and both ANN models (before and after kriging) are compared. Experimental data is obtained for wet granulation of microcrystalline cellulose using a bench-scale 12 mm twin-screw granulator. In addition, the effect of screw configurations on mean residence time is investigated by the developed ANN. The ANN model is made of two hidden layers with 2 linear nodes in each layer, and the linear system of equations is derived for the improved ANN model. The results revealed that the developed model was capable of predicting the mean residence time in the granulator more accurately after applying kriging interpolation, with an R2 value of about 0.92 for both training and validation. ANN model after kriging shows a dramatic improvement of R2 by 4% and 22% in training and validating phases, respectively. Also, the RMSE was improved by 40% and 61.5% in training and validating phases, respectively. Furthermore, this improvement was reflected in the contour profiles of the ANN models before and after kriging interpolation, where the model that uses the interpolated data points shows a smoother contour profiles and wider prediction areas. Screw configuration has the most significant effect on the residence time of granules inside the granulator where adding more kneading zones results in a substantial increase in the mean residence time compared to other process parameters.