In this project, the goal would be to learn the evolution of chemical species within a thermochemical manifold using neural ODEs. The machine learning-driven framework will be developed to approximate the thermochemical state of a system at a future time, t+1, given the state at the current time, t. There are some unique advantages to this approach. First, this can be used for trimming of the dimensionality of the manifold, where neural networks are used to learn appropriate formulae for updating the species concentrations in a closed form after minor species with negligible concentrations have been eliminated. This has the potential to significantly speed up simulations since the number of species to be transported would be lower. Secondly, current stiff ODE solvers are not amenable to GPU-acceleration. Thus, the framework developed in this work, which involves explicit prediction of future states can be beneficial, especially for computing platforms that contain hybrid CPU-GPU architectures. In this scenario, hydrodynamics and species diffusion-convection can be solved on CPUs, while the chemical kinetics can benefit hugely from parallelization on GPUs.

The research work will include training/testing data-driven models and coupling them with reacting flow CFD simulations to demonstrate acceleration compared to traditional numerical solvers.