Surrogate-Assisted PINNs with Hard Constraints for Heterogeneous Catalytic Reactor Modeling

We propose a hard-constrained PINN framework for efficient catalytic reactor modeling that guarantees atom conservation through a dedicated neural network layer.By choosing the weights of this layer based on the concept of key species, we replace multiple output-nodes with physical constraints, while simultaneously preserving positivity and improving training stability.Further, we include a detailed micro kinetic description of the surface chemistry through a physically plausible Global Reaction Neural Network surrogate.Applied to a CO$_2$ methanation reactor, our approach achieves 1000× speed-up over conventional solvers while maintaining physical fidelity.