High Performance CFD Computations for Ground Vehicle Aerodynamics 2011-26-0107
In this paper, benefits of high performance CFD computations for ground vehicle aerodynamics are discussed with the examples of Ahmed body and Formula SAE external aerodynamics. Both problems are highly compute intensive in terms of physics and their grid resolutions. Grid resolution is usually kept high to resolve flow separation, wake kind of complex physical phenomenon. For this, computations have been performed for more than 20 million cells grid. Firstly, Reynolds Averaged Navier-Stokes (RANS)-based turbulence study of Ahmed model is done for three slant angle variations, 25°, 30°, and 35°. Sensitivity of solver on solution is analyzed by a grid refinement study. The 2 million and 22 million cells mesh are used for the detailing of solution accuracies and solver speed up. Commercial solver, CFD++, and open source solver, OpenFOAM, are used for RANS computations. Their speed up test is done on supercomputer Eka. Grid refinement study shows moderate sensitivity of grid towards turbulence model, i.e., Realizable k-epsilon. A dense grid produces closer drag numbers towards experimental values. Drag coefficients, CD, obtained from two solvers, are in good agreement with the experimental data except for the configuration with slant angle 30°, which was reported earlier in the literature as a hard case to predict numerically. Also in contrast of attached flow observed in experiment for 25°, separation and reattachment is observed in simulations. This is validated well with other numerical studies. Speed up study is performed for 2 million and 22 million grids for two solvers on massively parallel supercomputer, Eka, using maximum 1024 CPUs. Scalability of the two solvers is compared for two grids and also for two solvers. Speed up and efficiency wise OpenFOAM seems better choice over CFD++ but it lacks a better time per iteration when compared to CFD++. OpenFOAM exhibits super-linearity with 300% efficiency up till 256 CPUs, while CFD++ is less than 90% efficient when CPUs exceeds 128 for coarse grid. Furthermore, another high end CFD computation has been used to analyze the flow behavior over generic Formula SAE car. CFD++, Fluent and OpenFOAM are used with RANS for baseline and modified configuration. A comparable prediction, with very dense grid for overall drag, is done for baseline shape, and drag numbers are reduced by 31% in modified shape. In the absence of any test data for race car model, the results are compared among CFD++, Fluent and OpenFOAM. A good agreement is found in obtained drag values. Overall CFD cycle time for these computations have significantly been reduced while using Parallel computations.