Model reduction for hypersonic flows in thermochemical non-equilibrium.

Abstract

Projection-based Reduced Order Models (ROMs) are proven physics-based alternatives to high-fidelity simulations that usually provide answers at a significantly cheaper computational cost with a tolerable loss in accuracy. In the case of hypersonic flows with strong nonlinearities, the strategy is to use nonlinear manifold projections (ISOMAP) near shocks and linear maps (POD) in quieter regions to strike a balance between computational cost and accuracy. Along with ISOMAP, Convolutional Neural Networks (CNNs) have been proven successful in model reduction for flows with non-linearities, albeit at lower Mach numbers and where thermochemical non-equilibrium is absent. In this paper, we initiate a framework to test the applicability of a U-Net based CNN and ISOMAP using a decomposed domain strategy for steady, inviscid hypersonic flows.