ENGINEERING AND CFD

Details

Turbulent supersonic wall-bounded flows impose several challenges. A huge difference in length scales due to the turbulence structures occurs generally, requiring a very fine grid resolution to resolve the smallest eddies appropriately. The wall cooling required to generate supersonic channel flow leads to huge gradients in temperature, density and consequently viscosity and heat conductivity very close to the wall, leading to large variations in flow parameters like Reynolds and Mach numbers. Additionally, very large coherent structures are generated, responsible for momentum and energy transfer, requiring large domain sizes to capture them accurately, which required the present project extension.

With higher Reynolds numbers, very large-scale anisotropic structures (VLAS) become additionally important, usually being found closer to the channel center. Due to the control, peaks in the corresponding energy spectrum at large wavelengths moved closer the wall and showed less energy decay compared to intermediate and small vortices in the wall vicinity. Therefore, the VLAS turned out to be much more dominant in the controlled flow, consequently influencing the energy transfer in controlled wall-bounded flows significantly. This increased importance on the other hand required simulations of larger domains and therefore high performance computing, to check the appropriateness of smaller domains, necessary to perform the multitude of simulations required when varying flow parameters during the overall project.

Fig. 1 depicts the velocity fluctuations in different wall parallel planes, to illustrate the difference and size of the coherent structures observed as streaks. After the control oscillations are applied, the turbulent stresses break down, the wall friction is decreased. Smaller vortices are grouping together to form larger clusters close to the wall – a direct connection to the positive and negative velocity fluctuations at the channel center is visible (Fig. 2).

As mentioned above, the strong gradients and small vortices require a fine computational grid in the relevant regions, whereas the increased importance of the VLAS, after the control is applied, leads to larger domain sizes to accurately capture them. The fine grid in addition influences the time step required to stably simulate the flow, significantly reducing it. Together this results in huge numerical grids and long computation times, requiring supercomputers to be able to appropriately simulate the flows. Here, a more than an order of magnitude larger computation time was required for the largest Reynolds and Mach numbers, compared to uncontrolled flows. For further details please refer to [3], [4] showing results of the first and this second project phase, respectively.

PyFR [1] is especially suitable for these types of simulations. The code generates platform specific code directly at runtime, being able to run on CPUs or GPUs alike. Due to its CUDA backend, it can make full use of the JUWELS booster. Selected simulations used for the characterization of the very large scales and one high Mach number case are exemplarily given in Table 1. 128 nodes (each including four A100 GPUs) were used for the largest simulation.