Application of Machine-Learned Turbulence Models for Improved Predictions in Hypersonic Flows

The current work describes machine learned modifications to the Menter SST-V turbulence model for improved Reynolds-Averaged Navier-Stokes (RANS) model predictions. A neural network learned a physical mapping between the scalar invariants of a RANS simulation flow field and turbulence closure variables based on Pope’s more general effective-viscosity hypothesis. Model training was performed on an axisymmetric large cone-flare geometry at Mach 6.17. The trained Menter SST-V model demonstrated improved predictions in surface pressure and heat transfer for the same large cone-flare at Mach 5.9 and for a Mach 7 ogive cylinder-flare.