Physically Consistent Neural Network Surrogates of Catalytic Reactors with Detailed Surface Kinetics

Abstract

Motivated by their potential to unlock real-time control of catalytic devices and model-based process design under full mechanistic detail, we developed physics-informed neural networks (PINNs) that consider detailed surface kinetics through an auxiliary kinetic surrogate model and enforce atom conservation as a hard constraint. In particular, we inform the PINNs about detailed surface kinetics by coupling with an auxiliary neural network that determines the surface reaction rates. This allows meshless predictions of composition profiles and removes the most time-consuming step from model training. Further, we propose a dedicated neural network layer that imposes atom conservation as a hard constraint. This guarantees the physical consistency of predicted reactor compositions and thus reduces the amount of data required for model training, in our case, by a factor of 10. We demonstrate that our new framework provides accurate and consistent concentration profiles of a CO₂ methanation reactor more than 1000 times faster than established direct integration schemes.