Stiffness-reduced neural ode models for data-driven reduced-order modeling of combustion chemical kinetics

A novel methodology for data-driven reduced-order modeling of stiff ODE systems was developed. A combination of a reversible nonlinear mapping to a reduced-order space and a data-driven neural ordinary differential equation (ODE) model allows efficient training and inference upon complex and highly stiff systems (e.g., combustion chemical kinetics). By regularizing the reduced-order mapping with an estimate of local stiffness, the nonlinear autoencoder neural network model is optimized to learn the highest reconstruction accuracy stiffness-reduced compression of the full-order system. This combination of stiffness-reduced autoencoder and neural ODE models constitutes the proposed reduced-order neural ODE (RONODE) methodology. This method was then applied to both a plug flow reactor (PFR) and a continuous stirred tank reactor (CSTR) with a 53 species and 325 reaction propane-air combustion chemical mechanism (GRI 3.0 mechanism). The RONODE model attained a 1.9 × 10⁻⁴ root-mean-square (RMS) error with a computational speed-up factor of 3.1 on the PFR system, while on the less stiff CSTR system the method attained a 2.1 × 10⁻⁴ RMS error with a computational speed-up factor of 1.3.