Aplicação de redes neurais artificiais no sinal FBRM para monitorar a cristalização de Paracetamol

Abstract

Obtaining high added value products in the chemical industry is directly related to separation operations carried out to achieve the specifications required by regulatory agencies. The crystallization, one of the most used separation processes, has its performance directly related to the control of properties such as kinetics, particle size distribution (PSD), shape and polymorphy. Therefore, the Food and Drug Administration (USA), through the Process Analytical Technologies (PAT) Guide, prioritize the small-scale development of online and in-line techniques to obtain real time data, aiming to reach process optimization, control and scalability. In this context, this work proposes the application of the FBRM equipment, considering the high amount of measured data in real time, the capability to obtain reliable primary data of chord length distribution (CLD) and the fact that the technique avoids external disturbances. However, CLD cannot be converted directly to PSD. Recent works suggest that artificial neural networks (ANN) can be used to solve this problem, although they present performance problems with different morphologies and network optimization. In this work, ANN were trained using different training methods (Levenberg-Marquardt, Bayesian Regularization and Scaled Conjugate Gradient) and tested to ensure the non-occurrence of overffiting and overtraining to obtaining optimized networks. A melhor rede foi escolhida para a sequência do trabalho, apresentando desvio máximo de 15% em relação a DTC obtida em Malvern. Subsequently, monitoring tests and external experiments were carried out to ensure the reliability of the following test: the monitoring of an isothermal test. Hence, four crystallization assays (A, B, C and D), which differ in terms of the number of crystal counts, average growth rate, DTC, DCC, supersaturation over time, were monitored through the best ANN obtained. The ANN precision also allows to predict experiment’s kinetic parameters through Moments Method and to make inferences about the process, being compared with documented experiments. Thereby, the nucleation kinetic parameters (n from 1,15 to 2,05 and ln(kn) from 25,50 to 28,89) and the growth kinetic parameters (g from 1,6 to 2,00 and ln(kg) from -8,55 to -12,22) were obtained, approaching experiments described by other authors under similar conditions.