Direct Numerical Simulation of a Low Pressure Turbine Cascade. Christoph Müller

Transcription

1 Low Pressure NOFUN 2015, Braunschweig,

2 Overview PostProcessing Experimental test facility Grid generation Inflow turbulence Conclusion and slide 2 / 16

3 Project Scale resolving Simulations give insight into physical flow parameters Reproduction of RANS transition model quantities Comparison between modeled results and physical results Improvement of RANS models NOFUN 2015 ASME Turbo Expo 2016 Work in progress Final results slide 3 / 16

4 Simulations Method: Direct solution of compressible Navier-Stokes equations Demands: No turbulence model: High grid density in order to resolve all spatial turbulent scales (Δ η Kolmogorov ) High temporal resolution to resolve all temporal turbulent scales (CFL << 1.0) Low dissipative numerical schemes Low computational stability No RANS averaging / unsteady simulation: Long simulation time to achieve a steady state solution Long sampling time for statistical evaluation of results Turbulent inlet boundaries must describe all fluctuations Lots of computational power : Full insight into highly detailed unsteady data slide 4 / 16 High accuracy - comparable to experimental results - no interactions with probe heads High flexibility (operating point, spectrum of inflow turbulence)

5 Cascade wind tunnel Facility Stadtmüller (2001) Test rig at Universität der Bundeswehr (München) Linear cascade in pressure tank Pressure can be lowered down to 4000 Pa Reynolds and Mach number can be adjusted independently Realistic flight conditions slide 5 / 16 Ma 2,th 0.2 to 1.05 Re 2,th 20k to 1.6M (0.1m profile chord) Tu 1 0.4% to 7.5%

6 Cascade wind tunnel Cascade box Flow parameters: Re 2,th = 50k Ma 2,th = 0,6 (similar cascade)t106d-eiz cascade (Stadtmüller 2001) slide 6 / 16 Casade wind tunnel: Evaluation of individual profiles 5 to 7 linear arranged blades Simple setup Good accessibility with measurement techniques

7 Numerical domain specifications Inviscid walls due to changing channel height Cyclic * (Δ/η Kolmogorov ), max Outlet: U wavetransmissive p wavetransmissive T wavetransmissive Inlet: U loopvaryingmappedfixedvalue p wavetransmissive T wavetransmissive Solver: rhopisofoam : (rhopimplefoam in piso mode) temporal scheme : backward spatial schemes : linear / limitedlinear *profile dummy slide 7 / 16 Hexa grid: Nodes y + y + (layer 10) (Δ/ηKolmogorov), max v1 Baseline 26.8 Mio. <0.5 > v2 Baseline (running) 37 Mio. <0.5 <0.5 - v2 Refined (planed) 300 Mio. <0.25 <0.25 -

8 Grid generation Low diffusive grids: structured Hexa grids I ) HCO-topology Grid generation with Ansys-ICEM (3.3 Mio. Nodes): Flexible grid generator (non automated) High influence on the grid quality (3.6% below 3x3x3 determinant > 0.95) Smoothing routines II ) Grid refinement : III ) refine3dhexamesh (OpenFOAM tool) Factor eight refinement 26.8 Mio. Nodes Linear splitting mesh cells Improving grid quality: Projection of the blade surface on a highly accurate profile contour (individual python script) Adjustment of the boundary layer resolution is possible slide 8 / 16

9 Generation of inflow turbulence Multiples of box length Damped frequencies k=1 k=2 k=3 k=n Minimum frequency defined by box size Method: A turbulent box is used to generate the turbulent spectrum: Matching turbulence field from measurements in the cascade box Continuous amplification of frequencies that are insufficiently strong Volumetric wavelets allow to amplify wavelengths between multiples of box lengths slide 9 / 16 Implications: Box size is defined by periodic boundaries in the cascade: Insufficient energy in large scales Resulting spectrum is characterized by sharp peaks

10 Scalability of the OpenFOAM domain Simple decomposing Hierarchical decomposing Optimum slide 10 / 16 Expensive DNS calculations demand a proper parallelization Study was conducted to find an optimal decomposition technique and CPU count Result: Highest efficiency in terms of CPU h / averaging period achieved with 1920 Cores Corresponds to cells/cpu (similar value observed on different testcases)

11 PostProcessing Process Chain HPC Calculations (North-German Supercomputing Alliance) Parallelized OpenFOAM calculations up to 1960 MPP cores 3D data (10 Gb/ timestep) Transfer node Reconstruction of the domain Extraction of slices Output: U probes 200kHz 3D data sampling (decomposed) Output: (ParaFOAM / batch mode) up to 16(+) x 1 SMP core 2D raw data slices (U,grad(U),p,T,rho) Initiating data transfer Starting p.p jobs Available timesteps slide 11 / 16 Desktop PC Job control Calculation of derived quantities TFD - Cluster Controlling post processing Data storage 1 cluster core 2D raw data (80 Mb / timestep) TFD storage system (100 Gb)

12 Numerical stability Effect of the geometric accuracy Very low damping due to the absence of turbulence models and diffusive schemes. Small gradients in the surface curvature cause significant oszillations and propagating pressure waves. slide 12 / 16 The profile geometry has to be extremely accurate.

13 Prediction of experimental data Very accurate prediction of the pressure distribution Wake differs from experimental data (adjustment of boundary condition still in progress) slide 13 / 16 Longer averaging time demanded Presented: 100 samples ( 0.32 pass troughs) Needed: 5 pass troughs

14 Flow animation PostProcessing Profile is masked with dummy slide 14 / 16 Shear stresses are visualized by the magnitude of the strain rate S Level of inflow turbulence is constantly increased to maximum

15 Conclusions: DNS with FiniteVolume solvers in OpenFOAM are possible DNS are highly expensive and time consuming / OpenFoam offers a high degree of parallelization Complex post processing routines are necessary to handle the data volume Experimental data can be reproduced very well : Implementation of high order temporal discretization methods Conducting a grid refinement study Reproducing RANS model quantities with DNS data slide 15 / 16

16 Thank you for your attention! Any questions? slide 16 / 16