A Data-Driven Extension to the Evolution-Variable Manifold Model in High-Speed Hydrogen Combustion

Simulating high-speed combustion is computationally costly, motivating the development of tabulated chemistry models. One such model is the Evolution-Variable Manifold (EVM) approach, which precomputes and tabulates a representative chemical source term for use in a simulation at runtime. A variant of the EVM approach is proposed and discussed here which replaces the majority of these tabulated states with two pre-trained Artificial Neural Networks (ANNs). These ANNs use identical inputs and outputs to the original EVM model, but have been trained using data from a Reynolds-Averaged Navier-Stokes (RANS) flowfield with full finite-rate chemistry. This extends the EVM framework to all thermodynamic states within this flowfield, rendering the representation more robust. A proof-of-concept is shown, where a RANS simulation of a hydrogen jet in supersonic crossflow is conducted to verify the model against its training data. A Large-Eddy Simulation (LES) is conducted on the same configuration and compared to an LES with full finite-rate chemistry.