Data-driven Modeling for Magma Density in the Continuous Crystallization Process

Crystallization processes have been widely used for separation in many fields, such as food, pharmaceuticals, and chemicals. The crystallization process is a highly nonlinear system, owing to complex crystallization dynamics; therefore, it is difficult to model the process to control the crystal product quality. In this study, a data-driven neural network was implemented to predict the magma density of the continuous crystallization process that produces maleic acid crystals from the mother liquor. Three neural network algorithms, namely deep neural network, long short-term memory, and gated recurrent unit (GRU), were applied for magma density prediction. Process variables, such as the feed flow rate, pressure, and steam flow rate were defined as input, while magma density, the most important control variable in continuous crystallization, was defined as an output variable. The grid search method was used to select suitable hyperparameters for each method, and the predictive accuracy of the models was compared with the root mean square error (RMSE). The GRU-based model afforded the best prediction accuracy among the applied models, with an RMSE of 2.04. Consequently, the developed predictive model can be used as a proper control strategy.