I am using machine learning to obtain improved approximate solutions from models of gas dynamics such as the BGK equation. We aim to find models for the gas collision frequency which are able to accurately reproduce results from higher fidelity methods such as DSMC at a reduced cost.
Collaborators: Justin Sirignano, Jonathan MacArt, Marco Panesi, Narendra Singh
I am also using machine learning to enhance and optimise over particle methods for rarefied gas dynamics. High fidelity methods such as Direct Molecular Simulation can be computationally intensive, requiring expensive classical trajectory calculations. We have shown that replacing this part of the method with an appropriately trained neural network allows for significant faster simulations with no loss of accuracy, at least for monatomic gases. I am currently interested in extending this method to gases with internal degrees of freedom, for which the relevant models must be significantly more complicated.