(333d) Super Learner Based Recurrent Neural Network Modelling Twin-Screw Granulation

Continuous manufacturing has gained significant attention in recent years due to the advantages it offers over traditional batch manufacturing in the pharmaceutical industry. In pharmaceutical manufacturing process, several processes such as mixing, drying, granulation, milling and coating are employed. Granulation, specifically, is an important process in pharmaceutical formulation, excipient active pharmaceutical ingredient (API) coming with powder are formulated to the granules. Twin-screw granulation (TSG) is one such granulation technique that has become increasingly popular in the production of solid oral dosage forms. TSG is a complex process that involves the mixing, wetting, and granulation of powders in a twin screw extruder. The process is influenced by many factors such as screw configuration, screw speed, feed rate, feed material property and process conditions, which makes it difficult to control and optimize using traditional methods.

Recurrent neural networks (RNNs) are a type of neural network that can handle sequential data and capture the dependencies between variables over time, compared to the traditional feedforward neural networks, RNNs have cycle connection in the architecture that allow them to take into account the time series data. In TSG, the variables can change rapidly and interact with each other in complex ways, by modelling the temporal dynamics of the process, RNNs can provide more accurate predictions of the granule qualities, such as the particle size.

Methods:

Twin screw granulation simulations were carried out with Siemens gPROMS Formulated Product, in which the feed flowrate, liquid flowrate, screw configuration and feed solid material size are altered. In order to enhance variety and dynamics in the TSG simulation, the pseudorandom binary sequences (PRBS) signals were utilized for the feed and liquid flowrates within the condition ranges, meanwhile, different numbers of kneading elements were employed to vary the screw configurations while two distinct material sizes were used.

Two recurrent neural networks, namely the Layer Recurrent Neural Network (LRNN) and the Nonlinear Autoregressive with Exogenous Inputs Neural Network (NARXNN), were employed in this study. Both networks contained one input layer, one hidden layer, and one output layer. In the LRNN, the hidden layers are connected in a feedback loop, thereby enabling the network to retain a memory of the previous inputs and computations. On the other hand, the NARXNN's output at each time step depends not only on the inputs at the current time step but also on the inputs and outputs from previous time steps. To obtain the optimal prediction results, the number of nodes in the hidden layer and the time delay steps had to be identified for both RNNs. Figure 1 illustrates the diagrams of the two recurrent neural networks.

To improve the predictive accuracy of the two RNNs, the Super Learner algorithm was employed. This algorithm is a machine learning technique that uses an ensemble of multiple base models to create a more accurate and resilient predictive model. The algorithm employs an ensemble learning strategy in conjunction with cross-validation to combine the predictions of multiple base models, resulting in a final prediction that surpasses the individual model predictions in terms of accuracy.

Finally, the developed super-learner based RNN model is utilized to predict the particles size and porosity in twin-screw granulation. To evaluate the prediction accuracy of the model, a comparison is made between predicted values with the gPROMS simulated values, and the model accuracy is shown by root-mean-squared error (RMSE), mean absolute percentage error (MAPE) and coefficient of determination .