Neural Network Inverse Model Controllers for Paracetamol Unseeded Batch Cooling Crystallization

Crystallization is a common process for purification and product design in the pharmaceutical industrial field. To develop an efficient crystallization process, the generated crystals must present a size distribution respecting the regulatory constraints on product quality. Therefore, a control system is needed to achieve this goal in a crystallization process. Neural network inverse model controllers (NNIMCs) are an efficient control strategy previously used for controlling some chemical processes. Moreover, they are able to calculate the control action faster than the classic model predictive controller (MPC). In this work, a nonlinear model predictive controller (NMPC) was initially applied to a paracetamol batch crystallization process in ethanol. The goal of the NMPC was to maintain the mass and crystal size in the targets by manipulating the temperature. Then, the NMPC was used to simulate controlled batches of the paracetamol crystallization, which were used to generate data for developing NNIMCs. Multilayer perceptron (MLP), regular recurrent neural network (RNN), gated recurrent unit (GRU), and long short-term memory (LSTM) networks were trained to predict the optimum temperature value that maintains the controlled variables in the targets. The controllers’ performance was investigated for different targets, model mismatches, and noisy data. The five controllers could efficiently maintain the mass and crystal size in the targets. The NNIMCs presented lower computational costs and imposed fewer temperature changes than the NMPC when accounting for noise. Furthermore, the NNIMC based on MLP performed best in dealing with model mismatches, being the most efficient controller studied. Therefore, NNIMCs showed a robust and efficient performance for controlling the crystallization process and have the potential to be used to control real crystallization processes.