Vous êtes sur la page 1sur 12

WEST-EAST HIGH SPEED FLOW FIELD CONFERENCE 19-22, November 2007 Moscow, Russia

USE OF HIGH PERFORMANCE COMPUTING FOR SIMULATIONS OF AERODYNAMICALLY GENERATED NOISE IN TURBULENT SEPARATED FLOWS
A.Yu. Snegirev*, S.V. Lupulyak*, Yu.K. Shinder*, B.S. Grigoriev*, K.Yu. Zamotin* and B. Khalighi**
* Laboratory for Applied Mathematics and Mechanics Saint-Petersburg State Polytechnic University Polytechnicheskaya, 29, Saint-Petersburg, 195251, Russia Email: a.snegirev@phmf.spbstu.ru, web page: http://lamm.spbstu.ru ** General Motors R & D Center MC 480-106-256 30500 Mound Rd., Warren, MI 48090 USA Email: bahram.khalighi@gm.com

Key words: Computational aeroacoustics, turbulent separated flow, high-performance computing. Abstract. Computational studies are presented of sound generation in turbulent separated flows behind the bluff bodies replicating vehicle side mirrors. The decoupled two-stage methodology is applied that requires massive CFD simulations to collect pressure variations at solid surfaces followed by acoustic post-processing of the data recorded. In flow simulations, URANS, DES and LES turbulence modeling approaches are applied, and the flows around two mirror models are studied for different flow velocities. Curles formulation in low Mach number limit is used to calculate acoustic pressure produced by pressure fluctuations at solid walls. Reasonable agreement between measurements and predictions has been achieved both for flow and sound characteristics. It has been found that the URANS and DES approaches heavily underestimate surface pressure fluctuations and therefore result in underestimated sound pressure levels. Alternatively, LES approach (particularly that with the dynamic subgrid viscosity model) provides better agreement for time-averaged surface pressure distributions and for sound pressure levels.

1.

INTRODUCTION

Quantitative predictions of aerodynamically generated noise require simulations of separated turbulent flows coupled with modeling of sound production and spread. In this work, a computational study is undertaken of turbulent separated flows around two vehicle side mirror models to assess its noise generation potential. The methodology used is applicable to low Mach number unsteady flows where the radiated noise is a small byproduct of the flow that is not altered by it. This assumption is central to Lighthills formulation of aerodynamic sound generation which implies that sound characteristics are obtained in two stages. Turbulent flowfields are predicted at CFD stage followed by sound post-processing stage. One of the objectives of this study is to select turbulence modeling technique that is computationally affordable and yet capable in producing appropriate data for further acoustic post-processing. To attain the

A.Yu. SNEGIREV et al./Aerodynamically generated noise

objective, we explored a range of turbulence modeling approaches including RANS, DES and LES. To calculate the acoustic pressures at the sound post-processing stage, we use the Curles formulation. The latter uses a Greens function-based solution for the sound propagation equation with the Lighthills form of noise source-terms in the special case of immovable and impermeable boundary submerged in an inviscid fluid. Note, that such a decoupled methodology is normally referred as a hybrid approach [1, 2] to noise prediction and is widely used to avoid simultaneous resolution of both flowfield and acoustic scales. Thus, another objective of this work is to develop acoustic postprocessing methodology and associated software. Flow simulations presented in this paper were conducted using 48-core AMD Opteron cluster (161.3 GFLOPs)1 at the Laboratory for Applied Mathematics and Mechanics (Faculty of Physics and Mechanics, Saint-Petersburg State Polytechnic University).

2.

MODEL DESCRIPTION

2.1. Statement of the problem and computational grids Flows around two different full-scale mirror models were considered. Computational domain (Figure 1 a) replicates the wind tunnel conditions. In the nozzle outlet, a uniform mean velocity profile was assumed with turbulence intensity of 0.01, and integral length scale of 0.7 m. The inlet flow velocities, 20, 30 and 40 m/s, have been considered in this work.

a)

b)

Figure 1. The test section of the wind tunnel computational domain (a) and a computational grid (b) used in flow simulations

Multi-block structured computational grids (Figure 1 b) used in the simulations were constructed with Ansys ICEM CFD 10.0. The computational domain was divided onto several sub-domains, each covered by a separate curvilinear structured grid. The grid blocks were constructed according to the following principles. Near-surface grid resolution in boundary layers (in the vicinity of the mirror model and the ground surface in near wake downstream the model) should provide

At the moment of writing this paper, 256-core cluster (64 nodes, AMD Opteron 280, 1035 GFLOPs) is also used.

A.Yu. SNEGIREV et al./Aerodynamically generated noise

sufficiently small y + value (from a fraction of one in the recirculating zone to about 20 in the wake). Grid resolution in a detached shear flow away from the surfaces should be fine enough to resolve large scale flow vortical structures. The recirculating zone and reattachment zone need a highest grid resolution whilst the remaining sub-domains may be covered by coarser grids. The ratio of neighboring grid element sizes inside every sub-domains should not exceed the value of 1.3. Total number of grid elements was about 2 millions. 2.2. Turbulence modeling In flow modeling and simulations, large scale flow vortical structures, dominant frequency, overall flow topology were investigated. In the unsteady RANS approach, the shear stress transport (SST) turbulence model by Menter [4, 5] was used. Finer flow structures and wider frequency range were predicted using a hybrid URANS-LES methodology (DES [6, 7]) as implemented in Ansys CFX and pure LES with a range of subgrid models. The conventional Smagorinsky model, the dynamic Smagorinsky model, and the kinetic energy equation model were exploited in this work. Two latter subgrid models were used with Ansys Fluent [8], since the current (at the time when the simulations were carried out) version of Ansys CFX [9] offers only conventional Smagorinsky model.

3.

FLOW SIMULATION RESULTS

3.1. Flow topology In flow structure analysis, URANS simulations provided valuable information that was then further interpreted in terms of DES and LES data. Several distinctive flow regions have been identified when analyzing the URANS results: Stagnation region with high pressure at the windward mirror side; Rarefaction region at mirror side and top surfaces. Adverse pressure gradient may inspire flow separation here. If such a separation does not occur, it happens behind the sharpened edges; Massive separation zone with recirculating flow inside (recirculating zone). This zone is characterized by a considerable pressure drop below the reference pressure; Zone of flow reattachment to the ground surface (later referred as reattachment zone) behind the recirculating zone (pressure recovers here producing values above the reference one) followed by the wake.

A.Yu. SNEGIREV et al./Aerodynamically generated noise

4 5

Model A

Model B

Figure 2. Large-scale flow structures identified in URANS simulations around two mirror models (wind speed 30 m/s, Ansys CFX). 1 transient vortices shed by the mirror side surface, 2 streamwise vortex tube, 3 and 4 steady zones of horizontal circulation, 5 and 6 steady zones of vertical circulation.

Upper Figure 2 (Model A) clearly shows formation of pressure spots at the ground surface downstream the recirculating zone of the mirror model. The spots indicate local minima and maxima of the pressure which in its turn are associated with the vortexes shed by the left mirror side. Local pressure maxima and minima correspond to alternating direction of fluid rotation in the vortexes. The vortexes reach the ground surface and corresponding pressure spots appear in the flow reattachment zone. One of the vortexes of this type is highlighted and designated as 1 in Figure 2 (Model A). Such a flow behavior is similar to formation of von Karman vortex street behind a bluffbody. Alternatively, for the mirror model B shown in upper Figure 2 b flow reattachment zone is represented by a single pressure maximum just weakly fluctuating in time. No large scale vortex shedding has been observed in the simulations. In the wake, streamlines concentrate in the vortex tube aligned with the bulk flow and designated as 2 in Figure 2. In the recirculating zones produced by both mirror models fluid circulation has been observed not only in horizontal (x-y) but also in vertical (x-z) direction. Vertical

A.Yu. SNEGIREV et al./Aerodynamically generated noise

circulation drives fluid particles from the region near ground surface towards the mirror top and back. For the mirror model A, the flow in this zone is heavily fluctuating and more homogenous than that for the mirror B. For the latter, both vertical and horizontal stratification can be concluded from flow visualizations. Indeed, top view and side view shown in Figure 2 (Model B) demonstrate existence of separate recirculating sub-zones with vertical (designated as 3 and 4) and horizontal (designated as 5 and 6) axis of rotation. Thus, in the flows studied, several distinctive flow structures have been observed (Figures 2 and 3). Transient vortices shed by the mirror side surface (1). Streamwise vortex tube (2). Steady zones of horizontal (3 and 4) and vertical (5 and 6) circulation. Combination of the zones 3 to 6 form an arch-type structure. Flow topology is more convenient to be visualized and investigated by plotting the isosurfaces of the strain-vorticity invariant = S 2 2 , where S and are the magnitudes of strain rate and vorticity tensors with the components
S ij = 1 u i u j + 2 x j xi u , ij = 1 u i j 2 x j xi .

(1)

Figure 3 demonstrates the I = const iso-surfaces for the two mirror models and the wind speed of 30 m/s. The above mentioned flow structures are shown by iso-surfaces corresponding to a negative (blue surface) and positive (red surface) values of . Quasiperiodic vortex shedding (1) from the left side of mirror model A and formation of the streamwise vortex tube (2) is clearly evident. Our simulations have also shown that velocity increase from 30 to 40 m/s does not cause qualitative transformation of the flow.

(a)

(b)

(c)

(d)

Figure 3. Flow structures resolved by URANS (a, c) and DES (b, d) approach in 30 m/s flow (Ansys CFX). (a, b) - mirror model A; (c, d) - mirror model B

A.Yu. SNEGIREV et al./Aerodynamically generated noise

Use of LES (and DES) approach allows for fine turbulent structures to be resolved. In Figure 3 a and b, simulation results obtained by URANS SST and DES SST (as implemented in Ansys CFX) are compared for the mirror model A. First, it can be concluded that the overall large-scale flow structure predicted by URANS can also be recognized in DES results. Indeed, quasi-regular shedding of vortical structures is observed from the left mirror side. However, those structures are further composed of several sub-structures with opposite signs of (Figure 3 b). Pressure spots moving over the ground surface are rather disordered compared to those predicted by URANS (Figure 3 a). Worth noting that finer details are now predicted of the surface pressure, which is very important for further use of this information in simulations of noise produced by dipole sources. Dissimilar to the Model A, remarkable difference can be concluded between the predictions by URANS and DES methods for the flow around another mirror model considered (Figure 3 c and d, Model B). The DES methodology predicts a long and unstable wake, very distinctive from that predicted by URANS. However, the conclusion on formation of an arch-type trailing vortical structure that surrounds the recirculating zone behind the mirror is still applicable to DES predictions as well. Ground surface pressure distributions Adequate prediction of surface pressure distributions is necessary for accurate simulations of noise originated from the dipole sources. Considerable attention has therefore been paid in this work to comparisons of the predicted surface pressure distributions with those experimentally measured in General Motors Research and Development Centre. In the experiments, static pressure was measured along the centerline and in three points at the reflecting surface of the mirror models shown in Figure 4 a and b. Not surprisingly, choice of turbulence modeling methodology significantly affects the pressure field resolution. Figure 4 shows the result obtained with the dynamic Smagorinsky subgrid model as implemented in Ansys Fluent. Note that the instantaneous flow differs remarkably from the average one. Indeed, Figure 4 a demonstrate formation, transport and decay of surface pressure spots. Furthermore, it was observed that vortex break-up also occurs producing doubled street of pressure spots moving streamwise along parallel directions. The averaged flow field (Figure 4 b) does not show the above features, and it appears to be quite similar to those predicted by the steady RANS model.

a)

b)

c)

Figure 4. Surface pressure distributions for 30 m/s flow around the mirror model by LES (dynamic Smagorinsky subgrid model, Ansys Fluent). a) Instantaneous resolved field, b) Time averaged field, c) Centerline surface pressure distributions. Vertical bars correspond to root mean square fluctuation magnitude. The streamwise locations of the reflecting surface points are not representative

A.Yu. SNEGIREV et al./Aerodynamically generated noise

To validate simulation results the latter were compared with the static pressure measurements. Surface pressure distributions along the centerline y = 0, z = 0 were normalized to obtain mean-pressure coefficient,

CP =

p(x ) p0 , q

(2)

where q = V2 2 is the dynamic pressure and p 0 is the reference pressure. Comparison of predicted and measured pressure distributions along the centerline and in three points at the reflecting surface is shown in Figure 4 c, which shows, first, good agreement for the time-averaged data and, second, quite a significant variance of the numerically resolved field. It was observed in the simulations that in RANS and DES simulations (Ansys CFX) pressure drop behind the mirror revealed underestimated pressure drop behind the mirror. In the pure LES simulations, the latter pressure drop increased giving observably better agreement with pressure measurements, particularly at the reflecting surface. Use of dynamic subgrid-scale modeling provides further improvements in reproducing the measured flow. That indicates the importance of appropriate subgrid model in LES.

4.

ACOUSTIC MODEL AND SOUND SIMULATION RESULTS

4.1. Acoustic model description The decoupled (hybrid) methodology used in this work includes two consecutive stages. At the first one, the CFD simulations are performed including transient simulations required to establish statistically steady flow followed by simulations in statistically steady flow to record transient surface pressure distributions. At the second stage, acoustic post-processing is carried out of the recorded surface pressure distributions. Acoustic post-processing of the recorded surface pressure distributions includes: p 1. Calculating transient acoustic pressure ~(x, t ) at a given location(s) x using the equation:

~(x, t ) = 1 p 4c0

(xi yi )ni [ p ] dy +
r
2

1 ( xi yi )ni [ p]dy , 4 r3 S

(3)

where r = x y is the distance between the source point y at the sound radiating solid surface S and the observer position x ; ni corresponds to a component of the unit normal vector at the wall surface; c0 is the speed of sound. Square brackets in Eq. (3) indicate the retarded time, t r c0 . This is the Curles formulation in low Mach number limit [3] that takes into account sound radiation by solid walls at rest whilst neglecting volumetric (quadrupole) sound emission. p 2. Calculating the power spectral density (PSD) S (x, f ) of sound pressure ~(x, t ) at a given location(s) x using the fast Fourier transform technique (FFT). About 16 thousands of time steps recorded during flowfield simulations were used in acoustic post-processing. It was found in sensitivity studies that further increase of the

A.Yu. SNEGIREV et al./Aerodynamically generated noise

observation period (the time step is constant and equal to 510-5 s) does not alter the results. 3. Calculating sound intensity
I (x ) = 1 1 var( ~ (x )) = p 0 c0 0 c0

S (x, f )df ,
0

(4)

and sound pressure level SPL(x ) = 10 log 1 2 p ref

S (x, f )df
0

(5)

p at a given location(s) x , where var( ~(x )) is the sound pressure variance, 0 is the 5 undisturbed air density, p ref = 2.0410 Pa is the reference threshold pressure.

a) Mirror only

b) Mirror and ground surface

Figure 5. Predicted sound pressure spatial distributions in the horizontal plane located at elevation of 18 in above ground surface (Model B, 30 m/s wind speed, LES, Smagorinsky subgrid viscosity model, Ansys CFX). Wind direction is given by the arrow. a) contribution of the mirror model surface only, b) total contribution of the mirror surface and the ground surface

A.Yu. SNEGIREV et al./Aerodynamically generated noise

4.3. Acoustic pressure and sound pressure level Instantaneous spatial distributions of acoustic pressure over the horizontal plane at an elevation of 18 in above the ground surface are shown in Figure 5 for one of the mirror models considered. Note that both the surface of mirror model and the ground surface contribute into the sound emission. The role of mirror model is to disturb the flow downstream thereby affecting the power of sound generated by the ground surface. The greater is the intensity of pressure fluctuations in the recirculating zone behind the model, the higher is the noise generation potential of the given model geometry. It can be observed that the mirror-induced acoustic pressure is fairly symmetric (Figure 5 a) while the spatial distribution of total acoustic pressure simultaneously produced by mirror and ground surface is shifted downstream (Figure 5 b). Sound pressure level (SPL) simulations have been performed in a number of points at an elevation of 18 in above the ground surface. The results for wind speed of 30 m/s are shown in Figures 6 and 7 for the two turbulence modeling approaches used (only two positions located above the centerline, y = 0, are shown conventionally designated as 2 and 5). The following conclusions can be made.

Figure 6. Acoustic post-processing of DES results at two positions at elevation of 18 in (Model A, 30 m/s wind speed, DES SST, Ansys CFX): a) acoustic pressure variation in time. b) SPL frequency distributions

Figure 7. Acoustic post-processing of LES results at two positions at elevation of 18 in (Model A, 30 m/s wind speed, Smagorinsky subgrid viscosity model, Ansys CFX): a) acoustic pressure variation in time. b) SPL frequency distributions

A.Yu. SNEGIREV et al./Aerodynamically generated noise

1. Predicted SPL frequency distributions are qualitatively similar to the measured ones2. They exhibit low-frequency spike followed by weaker frequency dependence (at higher frequencies). Note that frequency range beyond 2 kHz is strongly affected by numerical errors rather than controlled by fluctuations resolved in the simulations. Indeed, the range of acoustic frequencies that can be captured in LES depends primarily on the value of lt in the dominant noise-generating region ( lt is the integral length scale of turbulent motions, and is the local grid spacing). Maximum eddy-frequency f max resolved in LES is estimated in [1] using the model turbulence spectrum:

f max 1 lt 2 ft

2/3

(6)

where f t = t1 = u lt , u being the rms fluctuation velocity. For the flows of interest, the turbulence integral scale can be estimated as lt = 0.07 m (half of the characteristic size of the mirror models), fluctuation velocity is of order of u = 0.5 V = 15 m/s and the characteristic value of = 310-3 m can be used for the grid spacing. A numerical estimate of the maximum LES-resolved eddy-frequency f max is therefore of order of 1 kHz. For the frequency estimated, the corresponding sound wavelength in air is of order of 0.3 m. Note, the wavelength that is equal to the characteristic size of the mirror models (0.14 m) would correspond to the frequency of order of 2.4 kHz. It can therefore be concluded that for the frequency range resolved in LES, the mirror models can be approximately regarded as acoustically compact bodies (which is not the case for higher frequencies). 2. SPLs computed using the flowfield predictions by DES are observably lower than measured values (see Figure 6 b). This indicates that intensity of near-surface pressure fluctuations is underestimated by this turbulence modeling approach. Although these are certainly useful to determine important characteristics of the flowfield (drag, dominating frequency, shape and size of separated zone), higher numerical resolution is required for subsequent acoustic post-processing. 3. Predicted SPLs are quite close to the measured ones if LES results are used for the acoustic post-processing (see Figures 6 c). 4. The effect of wind speed on the SPL frequency distributions has also been investigated. As expected and in accordance with the experiments, sound pressure level increases as the wind speed increases. 4.4. Sound intensity Recall that both the mirror model surface and the ground surface contribute into the sound emission. Furthermore, as it was suggested in Figure 5, the ground surface contribution appears to be greater. Distributions of sound intensity produced by the mirror surface and ground surface together are shown in Figure 8 for 1 m radius sphere centered in the coordinate origin inside the mirror model. It can be seen that there is the single maximum of the intensity angular distribution, and most of acoustic energy propagates in a downstream direction.
Sound measurements have been performed in the experiments with the mirror models mounted on the table rather than on the ground surface as assumed in this work. Investigation to assess the effect of this difference is currently in progress.
2

A.Yu. SNEGIREV et al./Aerodynamically generated noise

Distributions of sound intensity produced by the mirror surface itself (shown in Figure 9) is just a fraction of total intensity. Dissimilar to the total intensity, its angular distribution reveals two propagation directions. In addition to the primary downstream direction, a pronounced local maximum is visible in opposite (upstream) direction (see Figure 9 b). It was concluded that such a distribution is coupled to the pressure variance distribution over the surface.

Figure 8. Predicted sound intensity spatial distributions over 1 m radius sphere (Model A, 30 m/s wind speed, Smagorinsky subgrid viscosity model, Ansys CFX). Wind direction is given by the black arrow. a) side view, b) top view. Shown is the total sound emission by the mirror model and ground surface.

Figure 9. Predicted sound intensity spatial distributions over 1 m radius sphere (Model A, 30 m/s wind speed, Smagorinsky subgrid viscosity model, Ansys CFX). Wind direction is given by the black arrow. a) side view, b) top view. Shown is the contribution of the mirror models only, sound emission by the ground surface is not accounted for. Note the intensity scale is 10 times less than that in Figure 8

Note that integration of the intensity over the solid angle results in the total value of power of sound emission; the latter can be readily used as an estimate of noise generation potential of a given mirror model.

9.

CONCLUSIONS

In the paper, computational studies are outlined of sound generation in turbulent separated flows behind the bluff bodies replicating vehicle side mirrors. The decoupled two-stage methodology is applied that requires massive CFD simulations to collect

A.Yu. SNEGIREV et al./Aerodynamically generated noise

pressure variations at solid surfaces followed by acoustic post-processing of the data recorded. Curles formulation in low Mach number limit is used to calculate acoustic pressure produced by pressure fluctuations at solid walls, with volumetric (quadrupole) sound emission assumed to be negligible. A range of turbulence modeling approaches including URANS, DES and LES is applied, and the flows around two mirror models are studied for flow velocities from 20 to 40 m/s. Reasonable agreement between measurements and predictions has been achieved both for flow and sound characteristics. In the latter case, it has been found that the DES (not to mention URANS) approach results in heavily underestimated surface pressure fluctuations and therefore in underestimated sound pressure levels. Alternatively, LES approach (particularly that with the dynamic subgrid viscosity model) provides better agreement both for time-averaged surface pressure distributions and for sound pressure levels.

9.

ACKNOWLEDGEMENTS

This work was funded by General Motors Corporation (GM Research and Development Center). Contribution by Dr Alexander Smirnov and Mr Eugeny Shinder in software development for acoustic post-processing and support by Dr Alexey Ushakov (Russia and CIS R&D Science Office) are gratefully acknowledged.

10. [1] [2] [3] [4] [5]

REFERENCES T. Colonius and S.K. Lele, Computational aeroacoustics: progress on nonlinear problems of sound generation, Progr. Aerospace Sci., 40, 345 416 (2004). B.A. Singer, D.P. Lockard and G.M. Lilley. Hybrid Acoustic Predictions. Comp. Math. App., 46, 647 669 (2003). N. Curle. The influence of solid boundaries upon aerodynamic sound. Proc. Roy. Soc. (London), A 231, 505 514 (1955). F.R. Menter. Two-equation eddy-viscosity turbulence models for engineering applications, AIAA-Journal, 32, 1598-1605 (1994). F.R. Menter and M. Kuntz. Adaptation of Eddy-Viscosity Turbulence Models to Unsteady Separated Flow Behind Vehicles. Proc. Conf. The Aerodynamics of Heavy Vehicles: Trucks, Busses and Trains, Asilomar, Ca, 2002. M. Strelets. Detached eddy simulation of massively separated flows, AIAA Paper, 2001-0879 (2001). A. Travin, M. Shur, M. Strelets and P.R. Spalart. Physical and numerical upgrades in the Detached-Eddy Simulation of complex turbulent flows, Fluid Mechanics and its Applications, 65, 2002, Advances in LES of Complex Flows, Proc. of EUROMECH Colloquium 412, pp. 239-254, Kluwer Academic Publishers, Dordrecht/ Boston/ London. FLUENT 6.2 Documentation, Fluent Inc, 2005. ANSYS CFX Solver, Release 10.0, 2005.

[6] [7]

[8] [9]

Vous aimerez peut-être aussi