Reduced order model for stiffness removal in chemically reacting flow simulations

The computation of finite-rate chemical kinetics forms the bottleneck of modern reacting flow simulations due to the inherent stiffness of the system. The operator splitting approach has been introduced to eliminate the time step requirements due to stiffness. However, integrating the operator split chemistry ordinary differential equations (ODEs) is computationally intensive and limits the overall performance of the modern reacting flow simulations. In this work, we introduce a machine learning-based autoencoder approach to eliminate the stiffness of the system. Specifically, the operator split chemistry ODEs are transformed into a latent space using autoencoders and integrated using the neural ordinary differential equation (neural ODE) approach. This approach results in elimination of the stiffness and thus a relaxed time step requirement. Furthermore, the dynamics of the reduced variables in the latent space, obtained from autoencoders and principal component analysis, are compared for two different chemistry mechanisms: i) H2-air and ii)CH4-air. The results obtained from the constant pressure batch reactor integration with H2 and CH4 mixtures show a good agreement with Cantera.