Implementation and Assessment of Chemical Kinetics Neural ODEs for Combustion Simulations

A data-driven methodology based on neural ordinary differential equations (NODE) is introduced for computationally efficient integration of chemical kinetics stiff ODEs in combustion simulations. The methodology consists of a deep neural network representing the derivatives at the hidden states, integration of which using an ODE solver yields the time evolution of thermochemical variables. A novel approach is established to train the NODE on data generated using a canonical constant pressure homogeneous hydrogen-air reactor. The trained NODE is then evaluated for accuracy and performance on the same reactor. Furthermore, it is applied to a more realistic case of pairwise mixing stirred reactor (PMSR), characterized by complex coupling of chemistry and mixing. The NODE is shown to effectively reduce numerical stiffness, enabling the use of explicit ODE solvers for the integration. It also improves multistep prediction accuracy and robustness, enhancing generalizability to realistic flames. The PMSR results show that, compared to direct integration of detailed kinetics, the NODE can achieve a significant computation time speed up for comparable accuracy. This warrants further extension and application of this approach for large-scale turbulent combustion simulations.