Modeling of propane thermal cracking process via physics-informed neural networks with process-consistent constraints

Propane thermal cracking provides the primary route for ethylene production. Developing an accurate model is important for optimizing its operation. However, coupled heat and mass transfer, intricate chemical reactions and coking dynamics make modeling challenging. This paper develops a tailored physics-informed neural network (PINN) by incorporating the process physics into the data-driven modeling framework. Two process-consistent constraints, i.e. the monotonicity constraints and coupling constraints are developed through process characteristics analysis, converted into loss functions and embedded in the training objective. Monotonicity constraints ensure the model complies with the fundamental process characteristics, while coupling constraints leverage the ethylene-to-propylene production ratio to characterize reaction depth and further reflect the intervariable dependencies. This approach effectively captures the complex behavior of propane thermal cracking, demonstrating robust extrapolation capabilities and high predictive accuracy. Simulation results show that, even under data-limited conditions, the proposed method significantly reduces maximum prediction errors across the entire operating range, maintaining them lower than 4 %, 2 % and 4 % for the coking rate, ethylene flowrate and propylene flowrate, respectively. This study pioneers a novel PINN framework for process modeling of thermal cracking, pointing to a promising direction for integrating physics information into the modeling of complex industrial processes.