Numerically Inspired Deep Neural Nets for Chemically Reacting Flows

The project will develop numerically inspired deep neural nets (NINNs) in order to replace the stiff ordinary differential equation (SODE) solvers currently being used to integrate chemical species in high-fidelity computational fluid dynamics simulations. Unlike traditional deep neural nets, the architectures and optimization strategies used to learn the physics of a problem will be based on the numerical schemes used to integrate SODEs, thus allowing higher fidelity and robustness. Tools will be developed to rigorously quantify training data sets in higher dimensions, allowing to identify regions of sparse or overabundant data, as well as noisy or multivalued data. A large and complete training database will be generated for two example cases found in the literature: 9-species hydrogen and 30-species methane combustion. This will allow an impartial assessment of the NINNs developed. The expected speedups of this replacement are of two orders of magnitude, opening the way to whole classes of combustion and chemical manufacturing problems that are currently out of reach – even on DoD HPCMP class machines. The techniques developed will be commercialized with a paradigm shift: instead of selling `the code’, the product sold will be ‘the NINN’ for the particular chemical species and reactions of a field.