Thermodynamics-Guided Neural Network Modeling of a Crystallization Process

Melt crystallization is a promising separation technique that produces ultra-high-purity products while consuming less energy and generating lower CO2 emissions than conventional methods. However, accurately modeling melt crystallization is challenging due to significant non-idealities and complex phase equilibria in multicomponent systems. This study develops and evaluates two neural network-based surrogate models for acrylic acid melt crystallization: a stand-alone (black-box) model and a thermodynamically guided (hybrid) model. The hybrid model incorporates UNIQUAC-based solid–liquid equilibrium constraints into the learning process. This framework combines first-principles thermodynamic knowledge—particularly activity coefficient calculations and mass balance equations—with multi-output regression to predict key process variables. Both models are rigorously tested for interpolation and extrapolation, with the hybrid approach demonstrating superior accuracy even under operating conditions significantly outside the training domain. Further analysis reveals the critical importance of accurate solid–liquid equilibrium (SLE) data for thermodynamic parameterization. A final case study illustrates how the hybrid approach can quickly explore feasible operating regions while adhering to strict product purity targets. These findings confirm that integrating mechanistic constraints into neural networks significantly enhances predictive accuracy, especially when processes deviate from nominal conditions, providing a practical framework for designing and optimizing industrial-scale melt crystallization processes.