Physics-Informed Machine Learning for Hypersonic Turbulence

Turbulence modeling for compressible flow is essential for accurate modeling of heat transfer, skin friction, and pressure surrounding hypersonic vehicles to correctly predict vehicle survivability. High fidelity computational fluid dynamics (CFD) approaches are too computationally demanding and require faster, workhorse methods such as Reynolds Averaged Navier-Stokes (RANS) simulations for computation of flight trajectories and dynamics. However, current RANS turbulence closure models cannot accurately characterize hypersonic flowfield features. Physics-informed machine learning (PIML) methods provide an avenue for creation of turbulence models. PIML methods can use experimental and high-fidelity simulation data to learn models constrained to the physical governing equations, resulting in a more general and extendible turbulence model. In this STTR effort, Spectral Sciences, Inc. and Virginia Tech will develop an innovative PIML turbulence closure model leveraging recent advances in ensemble learning to compute accurate flowfield gradients for efficient training of the neural network-based approach. Phase I will develop a protype software tool and demonstrate the approach on a system of interest, with the aim of showing significant improvement for important flow quantities. Phase II will extend the approach to a wider range of flow conditions and integrate it into a software toolkit for hypersonic flow.