(59ap) Intelligent Size Characterization of Granules By Machine Learning Method

Machine learning methods such as artificial neural network (ANN) modeling, has been shown to be a powerful technique for a complex and nonlinear process such as powder granulation with a strong ability to learn and predict the process. In recent years, the interest in ANN modeling in different fields such as pharmaceutical industries have been increased and several successful applications of ANN in this field have been reported. In the present work, an artificial neural network (ANN) approach is applied to model granules size distribution (GSD) curves of pharmaceutical granules which were produced by wet granulation method. A powder mixture containing 44% microcrystalline cellulose, 44% -lactose monohydrate, and 3% Polyvinylpyrrolidone as a binder was investigated. All other components were pre-combined with PVP by using dry binder mixing method. The granules production was carried out in a Steer Integraal® with six heating zones controlled at temperatures between 50^oC and 120^oC. The size characteristics of the granules during the process were determined by granules size analyzer. A neural networks model was used to identify and predict granules size characteristics such as average granules size (D₅₀) and simulation of the whole distribution graphs based on each processing parameter.

Since a three-layer network can map any function to any desired accuracy, so feed forward three-layer neural networks have been applied in this research. In this network, the granulation process parameters including temperature , liquid to solid ration (L/S), screw speed as well as the particle size were considered as input values and the size distribution values at each particle size were considered as output. The number of nodes in the hidden layer is related to the complexity of the system and here, it was considered that this number be 2n+1, where n is the number of neurons in the input layer. Since the input layer neurons is 5, therefore, the network architecture used in this study is 5-10-1. In first step, 70% of input/output patterns were randomly selected from the experimental data for training the designed networks and the rest of data were used for testing the performance of the networks. The designed neural networks were trained by the error back propagation algorithm. Momentum term was also incorporated in updating the weights of the networks to enhance the convergence rate of the algorithm.

Results showed that developed ANNs model was capable of predicting the D₅₀ and span [(D₉₀–D₁₀)/D₅₀] of granules and anticipating the different influential parameters involved during granulation. The good performance of this model was confirmed by large correlation coefficients (>0.95) achieved by plotting all the experimentally obtained data for their corresponding ANN predicted values in different samples. Our works creates a paradigm for future studies of application of neural network modeling of other pharmaceutical granules size distribution with the proposed method.