Enabling micro-kinetics based simulation of industrial packed-bed reactors by physics-enhanced neural networks

Multiscale CFD simulations can provide insights into the coupling of the surface catalytic mechanism and reactor scale transport phenomena. To address the high computational cost of detailed micro-kinetic models, we have applied for the first time artificial neural networks (NNs) as surrogate models for micro-kinetic rate evaluations, with the aim of accelerating particle-resolved CFD simulations of a catalytic reactor. To evaluate the efficacy of the proposed strategy, a methane steam reforming packed bed reactor was selected as a benchmark case. A global reaction neural network with embedded thermodynamic and stoichiometric information has been implemented as a surrogate for a UBI-QEP micro-kinetic model available in literature. Two distinct test cases have been employed. The first targets a lab scale reactor, enabling either the full evaluation of the micro-kinetic scheme or the novel NN accelerated approach in the CFD simulations. The comparison between the two strategies showed deviations in the computed mole fractions of less than 1 % across a wide range of operating conditions along with a 19-fold total simulation and 63-fold chemistry speed-up. Consequently, the total simulation time of the benchmark was reduced from 114 h to 6 h, with only 29.2 % or 1.75 h of the computational cost being allocated to the chemistry sub-step of the solver. Therefore, source term evaluations are no longer the bottleneck of reactive CFD. Given the obtained excellent accuracy and speed-up, we applied the NN-accelerated micro-kinetics to the CFD simulation of an industrial scale packed-bed reactor, discretized with 44 M cells (54 % solid phase). To the best of our knowledge, this represents the largest reactive simulation with micro-kinetic level of detail. Overall, these results pave the way for the scale-up of multiscale simulations to industrially relevant scales.