Session: 34-08 LES Solvers and applications 2

Paper Number: 128238

128238 - GPU-Accelerated Full-Wheel Large-Eddy Simulations of a Transonic Fan Stage 

The application of traditional high-fidelity numerical approaches, such as direct numerical simulation (DNS) and large-eddy simulation (LES), to industrial turbomachinery analysis and design is limited due to their computational expense and relatively high requirement on computing resources. On the other hand, the three-dimensional unsteady flow features in turbomachinery, the rotor and stator blade counts in real geometries, and the importance of the relative motion between rotor and stator and the associated flow characteristics render the need of unsteady and full-wheel multi-row simulations. Leveraging the recent development in numerical algorithms and computer hardware, especially the graphic processing unit (GPU), Cascade Technologies, now part of Cadence Design Systems, has developed computational technologies to the level that an efficient and affordable solution can be provided. These technologies include: 1) advances in wall modeled large-eddy simulation (WMLES) techniques, 2) a highly efficient moving mesh solver based on Voronoi diagrams that is well-suited for turbomachinery flow simulations, and 3) high throughput computing using GPUs.

The simulations in the present work are performed with the GPU-accelerated moving-mesh compressible flow solver Fidelity CharLES. The solver employs a low-dissipation, nonlinearly stable numerical scheme. An extended gradient operator is constructed for all interior cells, which is formally second order accurate (in an L_ inf  sense) on arbitrary meshes with small dispersive errors. An efficient one-sided compact gradient operator is used to treat the stationary-moving part interfaces. Away from the interface, the spatial operators in the stationary region are left unchanged while those of the moving region are linearly transformed. Due to its flexibility in building part interfaces and its numerical accuracy, the solver is particularly suitable for turbomachinery applications and aeroacoustic simulations.

A single-stage transonic axial-flow fan NASA 67 is simulated. The stage consists of a rotor with 22 blades and a stator with 34 blades. The rotor has a tip chord of 0.095 m and a tip clearance of 0.5 mm. When it operates at 100% rotation speed of 16043 rpm, the relative tip Mach number is 1.38 and the tip Reynolds number based on the blade chord is approximately 3.0x10⁶. The length of the computational domain is 0.735m with the inlet located at 0.16m upstream of the spinner. The full annulus of the rotor and stator is included in the domain. At the inlet, a total pressure of 101,325 Pa and a total temperature of 288.15 K are specified. A Navier-Stokes characteristic boundary condition (NSCBC) with specified back pressure is applied at the outlet. The back pressure is adjusted to move the operating condition along the performance curve. An algebraic wall model with equilibrium boundary layer assumption is employed on all walls including rotor and stator blade, spinner, hub and shroud. The Vreman sub-grid model is used to account for the effects of the unresolved turbulence structures. Two meshes are employed for the simulations. The first mesh uses purely isotropic hexagonal close-packed (HCP) seeding with a background mesh spacing of 4 mm. Refinement with the smallest mesh spacing of 0.5 mm is applied to the rotor and stator blades and results in a total number of 58 million control volumes (Mcvs). The second mesh employs anisotropic boundary-layer type of mesh around the blade surfaces with a minimum grid spacing of 0.125 mm in the wall normal direction and the total mesh size is 84 Mcvs. The operating conditions at 100% rotation speed are simulated and a sweep of the performance curve is carried out. Good agreement against experimental measurement is observed and the performance curves will be presented in the full paper.

The simulations are performed on both Nvidia V100 GPUs and AMD MI210 GPUs. Compared to the performance on a CPU-only cluster, a core equivalence of 500~600 CPU cores is observed. Detailed solver performance will be included in the final paper.

Presenting Author: Kan Wang Cadence Design Systems Inc

Presenting Author Biography: Kan Wang received a PhD in Aerospace and Mechanical Engineering from University of Notre Dame, and his BS and MS degrees in Engineering Mechanics from Tsinghua University. He joined Cascade Technologies in 2018 as a Research Scientist. At Cascade, he focuses on the moving solver with applications in turbomachinery. Before joining Cascade, he was a postdoc and then a Research Assistant Professor at University of Notre Dame performing research on computational aeroacoustics and aero-optics. He joined Cadence Design Systems with the Cascade team in 2022. Now he is a Principal Software Engineer in Cadence.

Authors:

Kan Wang Cadence Design Systems Inc
Sanjeeb Bose Cadence Design Systems Inc
Christopher Ivey Cadence Design Systems Inc