Digital Design of Cooling Crystallization Processes Using a Machine Learning-Based Strategy

Abstract

This study introduces a novel machine learning-based digital design strategy for predicting the final yield and particle size distribution (PSD) in cooling crystallization processes. The proposed approach integrates an artificial neural network (ANN) to correlate operating conditions and product qualities with a generative adversarial network-based model of the crystallization process (CrystGAN) to generate new data, thereby reducing the dependence on extensive real data sets. After training the CrystGAN model with only 30 data sets, the generated yield values showed an RMSE of 6.428 and an MAE of 4.622 compared to real values. The coupled CrystGAN-ANN model exhibited predictive performance, with R² values greater than 0.93 for all Dv predictions. This method was demonstrated using a simulated data set of the l-glutamic acid crystallization process and validated using a new experimental data set of the p-toluenesulfonamide crystallization process. The average prediction error of the proposed approach was within 5% for both the product PSD and yield. The proposed strategy significantly reduces the experimental effort and design complexity required compared with traditional mechanistic model development and validation, offering a robust, scalable, and efficient framework for crystallization process design and optimization in industrial applications.