A Physics-Constrained Neural Ordinary Differential Equations approach for Robust Data-Driven Modeling of Chemical Kinetics

The computational cost of solving for detailed chemistry is one of the major bottlenecks for predictive CFD modeling of turbulent reacting flows. Deep learning approaches have been explored to develop surrogate models for stiff chemical source terms, but are often unstable when coupled with CFD solvers. This is because these techniques minimize the error during training without guaranteeing successful integration with ODE solvers, resulting in undesirable error accumulation over time. In this regard, Owoyele and Pal (Energy & AI, 2021) proposed, for the first time, a robust deep learning approach based on neural ordinary differential equations (NODE), known as ChemNODE, wherein the chemical source terms predicted by the neural networks are integrated during training, and by computing the required derivatives, the neural network weights are adjusted accordingly to minimize the difference between the predicted and ground-truth solution. In this work, the framework is extended to incorporate constraint terms in the loss function (during training) to explicitly enforce element/species mass conservation. Further, the ChemNODE framework is integrated with CONVERGE CFD solver and validation studies are performed for constant pressure homogeneous autoignition of hydrogen-air mixtures over a range of composition and thermodynamic conditions. It is shown that incorporation of physics-informed constrains enhances the robustness and accuracy of ChemNODE.