Development of hypersonic aircraft and weapon systems has become a critical focus for the Department of Defense to maintain global strike and projection of force capabilities. Despite decades of research, traditional computational fluid dynamics (CFD) methods are either incapable of adequately predicting complex features in hypersonic flows or too expensive to be of practical use for vehicle design in this regime. Therefore, a new modeling methodology is required that approaches the accuracy of scale-resolved CFD simulations at a cost similar to Reynolds-averaged Navier-Stokes (RANS). ATA Engineering, in partnership with the University of Arkansas, has developed a data-driven framework, known as the HYpersonic Physics-informed Energy-tracing RANS (HYPER) Tuner, to tune RANS turbulence closures using machine learning (ML) to modify several terms in a standard RANS turbulence model to improve its accuracy in hypersonic flows. The term modifications will use genetically programmed symbolic regression and constant tuning algorithms to derive the functional form of each term from scale-resolved CFD and experimental data from representative flow configurations. In Phase I, ATA demonstrated the improved accuracy of a tuned RANS turbulence closure on simple, high-speed attached boundary layer cases. In Phase II, ATA will expand and refine the ML algorithms in HYPER Tuner and validate the HYPER Tuner framework against several training cases, including both CFD and experimental data sets and running new high-fidelity CFD simulations on some of those experimental configurations. The Phase II project will also include development and validation of a laminar-to-turbulent transition model for hypersonic flows to further improve RANS predictions of wall shear stress and wall heat flux in the early portion of the developing boundary layer. At the completion of Phase II, HYPER Tuner will be an adaptable, general framework for improving the accuracy of RANS turbulence closures in the hypersonic regime, including various ML algorithms and automated features to enhance its efficacy and ease-of-use for general CFD practitioners.
Benefit: Accurate prediction of the performance of hypersonic vehicles hinges on resolving the boundary layer turbulence which directly impacts shockboundary-layer interaction (SBLI), flow separation, and wall heat flux predictions. Typically, resolving the turbulence to sufficient accuracy at the Reynolds numbers and Mach numbers associated with hypersonic flows requires costly scale-resolving simulations. The HYPER Tuner framework for modifying RANS turbulence models uses scale-resolved CFD data to learn trends in the salient mean quantities for these flows, as functions of local dimensionless parameters. It can also fine-tune the identified trends to match key experimental measurements to maximize the model accuracy while maintaining RANS-like computational cost. The Phase II prototype product will include a framework for modifying a single RANS turbulence closure, including a turbulent transition model, by training machine learning algorithms on high-fidelity CFD and experimental data from multiple flow configurations. The Phase II prototype will also include the capability to modify a second RANS turbulence closure without transition modeling, with the intention of incorporating transition modeling and other RANS turbulence closures in post-Phase II development. The prototype HYPER Tuner software tool will automate parameter selection, turbulence closure modification verification, and model export to several leading CFD solvers to improve usability for general CFD practitioners and streamline its adoption into the DoDs envisioned digital engineering environment. The final iteration of the HYPER Tuner framework will also include uncertainty quantification tools for understanding the uncertainty in the predictions of tuned RANS models. HYPER Tuner and the modified RANS turbulence closures it generates will provide a unique capability for accurately capturing the time-averaged effects of turbulence in hypersonic flows without needing to resolve the broad range of spatiotemporal scales.
Keywords: shock-boundary layer interaction, CFD, Digital Engineering, Machine Learning, Hypersonics, Turbulence Modeling, laminar-to-turbulent transition, heat transfer
