Phy-ChemNODE: A Physics-Enhanced Neural Ordinary Differential Equations Approach for Accelerating Stiff Chemical Kinetic Computations*

A first-of-its-kind neural ordinary differential equations (NODE) based approach, known as ChemNODE, was developed at Argonne National Laboratory (Owoyele and Pal, Energy & AI 2021) to accelerate detailed chemistry computations. In this framework, chemical source terms predicted by a neural network are integrated during training, and by computing the required derivatives, the neural network weights are optimized to minimize the difference between the predicted and ground-truth thermochemical state solutions. ChemNODE was enhanced by incorporating elemental mass conservation constraints directly into the loss function during training (Kumar et al. 2023, https://arxiv.org/abs/2312.00038). Both a-priori and a-posteriori tests (for hydrogen-air 0D autoignition) demonstrated that the enhanced Phy-ChemNODE framework not only improves physical consistency of the data-driven model, but also enables faster model training. A-posteriori studies performed by coupling Phy-ChemNODE with a CFD solver exhibited robustness and generalizability to unseen initial conditions from within (interpolative capability) as well as outside (extrapolative capability) the training regime. Lastly, Phy-ChemNODE was recently extended to hydrocarbon chemistry by incorporating a non-linear autoencoder (AE) for dimensionality reduction, wherein the NODE learns the temporal evolution of the dynamical system in a reduced-order latent space obtained from the AE. Both the AE and NODE were trained together in an end-to-end manner. Numerical studies performed for methane-oxygen chemistry (32 chemical species; 266 chemical reactions) demonstrated 10X speedup achieved by Phy-ChemNODE compared to solving for the full system of stiff ODEs, while preserving prediction fidelity.