Phy-ChemNODE: an end-to-end physics-constrained autoencoder-NeuralODE framework for learning stiff chemical kinetics of hydrocarbon fuels

Predictive computational fluid dynamics (CFD) simulations of reacting flows in energy conversion systems are accompanied by a major computational bottleneck of solving a stiff system of coupled ordinary differential equations (ODEs) associated with detailed fuel chemistry. This issue is exacerbated with the complexity of fuel chemistry as the number of reactive scalars and chemical reactions increase. In this work, a physics-constrained Autoencoder (AE)-NeuralODE framework, termed as PhyChemNODE, is developed for data-driven modeling and temporal emulation of stiff chemical kinetics for complex hydrocarbon fuels, wherein a non-linear AE is employed for dimensionality reduction of the thermochemical state and the NODE learns temporal dynamics of the system in the low-dimensional latent space obtained from the AE. Both the AE and NODE are trained together in an end-to-end manner. We further enhance the approach by incorporating elemental mass conservation constraints directly into the loss function during model training. This ensures that total mass as well as individual elemental species masses are conserved in an a-posteriori manner. Demonstration studies are performed for methane combustion kinetics (32 species, 266 chemical reactions) over a wide thermodynamic and composition space at high pressure. Effects of various model hyperparameters, such as relative weighting of different terms in the loss function and dimensionality of the AE latent space, on the accuracy of Phy-ChemNODE are assessed. The physics-based constraints are shown to improve both training efficiency and physical consistency of the data-driven model. Further, a-posteriori autoregressive inference tests demonstrate that Phy-ChemNODE leads to reduced temporal stiffness in the latent space, and achieves 1-3 orders of magnitude speedup relative to the detailed kinetic mechanism depending on the type of ODE solver (implicit or explicit) used for numerical integration, while ensuring prediction fidelity.