Robust AI Unrolling and GPU-Accelerated High-Speed Flow Simulations in Passive Control Applications

<p dir="ltr">The challenge of improving simulation quality and speed for high-speed flows is addressed in this work through three distinct approaches. The first approach focuses on using machine learning to improve or replace existing Computational Fluid Dynamics (CFD) solvers. First, a physics-informed neural network was integrated into a finite volume solver to compute the solution of the Riemann problem at the interface between elements, leading to a three order of magnitude speed-up compared to the exact solver. Second, a new approach was developed to unroll accurately long trajectories using existing ML solvers for CFD, leading to up to a 10x improvement in solution accuracy.</p><p dir="ltr">The second approach to increasing simulation speed focused on improving the parallelization of a pseudo-spectral numerical solver through the integration of a package that enables multi-platform execution. With this package, the code is now capable of running on multiple device architectures. Compared to the previous version, the new code is twice as fast and scales efficiently across three GPU architectures on High Performance Computers (HPCs).</p><p dir="ltr">The third and final approach focused on developing a new combined sub-filter scale (SFS) turbulence and shock-capturing model for high-order finite volume schemes, called the Block Spectral Stresses (BSS) method. The model was validated on the Sod shock tube and shock-vortex interaction cases for its shock-capturing capability, and on subsonic Taylor–Green Vortex (TGV) and channel flow cases for SFS modeling. Compared to existing LES models, BSS showed better performance at low numerical orders in the TGV case, and the opposite trend in channel flow cases. In a supersonic TGV case, the model successfully handled both turbulence modeling and shock capturing, showing similar behavior to the subsonic case.</p><p dir="ltr">Direct Numerical Simulations (DNS) of a flat plate with wavy inserts showed that changes in wavy wall amplitudes led to a shift toward lower frequencies in the power spectra of the second-mode waves. The DNS also showed frequency locking over the wavy region for wall amplitude of 1 mm, which was validated by experiments. A reduction in the skin friction coefficient and Stanton number was observed downstream of the region for certain geometries, likely because the wavy surface affects only the viscous sublayer of the flow.</p>