Physics-informed learning of chemical reactor systems using decoupling-coupling training framework

It is known that physics-informed learning become a new learning philosophy that has been applied in many scientific domains. However, this approach often struggles to achieve optimal performance in addressing the issue of multiphysics coupling. Here, for the first time, we extend this approach to modeling chemical reactor systems. We design a new decoupling-coupling training framework, which consists of decoupling pre-training and multiphysics coupling training steps. With decoupling pre-training, the complex physical domain is decomposed into subdomains of fluid flow, heat transfer, and mass transfer combined with reaction kinetics. Each subdomain is represented by a specialized neural network that can provide a coarse but reasonable distribution of network parameters for initializing the sub-networks for the subsequent multiphysics coupling training. The capabilities of this approach, in comparison with the traditional CFD simulation, are demonstrated through an example of a plate reactor system with a heating cylinder.