Adaptation of Eddy-Viscosity Turbulence Models to Unsteady Separated Flow Behind Vehicles F.R. Menter and M. Kuntz Ansys-CFX Introduction Turbulence model development for aerodynamic applications has for many years concentrated on improving the capabilities of CFD methods for separation prediction. Validation studies of turbulence models in the ‘80th have clearly shown that most turbulence models were not capable of predicting the development of turbulent boundary layers under adverse pressure gradient conditions. Based on that observation, new models were developed to specifically meet this challenge, resulting in a series of models capable of capturing boundary layer separation in good agreement with experimental data (Johnson and King 1984, Menter 1993, Spalart and Allmaras 1994). From the experience with the SST turbulence model (Menter 1993), the present authors would argue that the capability of the model with respect to the prediction of the onset of separation is within the accuracy of the available experimental data and that no systematic deviation between the simulations and the data is observed. Based on the experimental evidence (which is admittedly limited for three-dimensional flows), there is currently little need for model improvements for that type of flows. Off course, this does not imply that aerodynamic flows can be predicted within experimental uncertainty levels, as these flows involve other effects, which pose additional challenges to the turbulence model. The main areas of concern are the behavior of the flow downstream of the separation line, including the flow recovery after reattachment (Johnson et al. 1994), the proper simulation of vortex flows and questions related to laminar-turbulent transition. Particularly the flow development downstream of separation is of major importance from an aerodynamic standpoint and can have a significant effect on the characteristics of aerodynamic bodies. This is particularly true for ground vehicles, as they generally exhibit significant regions of separated flows, even at design conditions. From a modeling standpoint, it has been observed for a long time that RANS turbulence models underpredict the level of the turbulent stresses in the detached shear layer emanating from the separation line (Johnson et al. 1994). This in turn seems to be one of the main reasons for the incorrect flow recovery predicted by the models downstream of reattachment. The issue is some- 340 F.R. Menter and M. Kuntz times masked by the tendency of models of under-predicting the onset and therefore the strength of the separation zone, which in turn results in an acceptable agreement in the recovery region. However, the improvement is only the result of a cancellation of errors, as one cannot trade separation prediction capabilities against improved velocity profiles in the recovery region. The delayed recovery of the boundary layer downstream of the reattachment line can lead to a premature separation under a second adverse pressure gradient. A second, and even more disturbing uncertainty resulting from the incorrect prediction of the detached shear-layer, concerns the flow topology downstream of the separation line. Current turbulence models cannot reliably answer the question, whether the flow is forming a closed separation bubble, or a fully stalled flow regime. This question is of major importance for the prediction of the aerodynamic characteristics of automobiles, which almost always exhibit regions of separated flow. The topology of these regions has a strong influence on the drag and more pronouncedly on the lift of the car. In an attempt of improving the predictive capabilities of turbulence models in highly separated regions, Spalart (1997) proposed a hybrid approach, which combines features of classical RANS formulations with elements of Large Eddy Simulations (LES) methods. The concept has been termed Detached Eddy Simulation (DES) and is based on the idea of covering the boundary layer by a RANS model and of switching the model to a LES mode in detached regions. Ideally, DES would predict the separation line from the underlying RANS model, but capture the unsteady dynamics of the separated shear layer by resolution of the developing turbulent structures. Compared to classical LES methods, DES saves orders of magnitude of computing power for high Reynolds number flows, due to the moderate costs of the RANS model in the boundary layer region, but still offers some of the advantages of an LES method in separated regions. There are two main concerns with the current DES formulation. The first is how quickly the unsteady turbulent structures develop after the model has switched from the RANS to the LES mode. This is of significance for the prediction of separated shear layers, as a delayed onset of resolved turbulent structures would aggravate the underprediction of the turbulent stresses due to a reduction of the unresolved turbulence level by the DES formulation. The second concern is with the switching mechanism employed by the current DES methods. In order to prevent the activation of the DES limiter in attached boundary layers, it is typically required to ensure a lower limit on the local surface grid resolution. If this condition is violated, the integrity of the RANS model is severely compromised resulting in most cases in grid induced separation. This issue will be addressed in the section on turbulence model formulation. For an alternative of a hybrid turbulence model without an explicit grid dependency see Menter et al. (2003). The capabilities and limitations of advanced aerodynamic RANS and DES turbulence models for automotive applications will be evaluated and discussed. Alternatives to the current RANS/DES switch will be discussed. The models will be applied to the flow around a simplified generic car shape, known as the Ahmed car body (Ahmed et al. 1984, Lienhart and Becker 2003). The RANS Adaptation of Eddy-Viscosity Turbulence Models to Unsteady Separated Flow 341 simulations have been presented at the ERCOFTAC workshop on Refined Turbulence Modelling (Durand et al. 2002) in a comparison study of different CFD methods for specific testcases. A detailed report is available from the authors upon request. All simulations have been computed with the commercial CFD method CFX-5 of ANSYS. Flow Physics The geometry of the Ahmed car body is shown in Figure 1. It consists of a box with rounded edges and a slanted back. The angle of the slant is adjustable and is the main variable model-parameter in the experimental investigations. Two sets of experiments have been carried out for this geometry (Ahmed et al. 1984, Lienhart and Becker 2003).The present comparison is mainly based on the data of Lienhart and Becker (2003). In this experiment, the Reynolds number with respect to the length of the car was ReL=2.6x106. Slant angle of 25° and 35° were investigated. The emphasis of the experiments was on the flow structure in the slant region and downstream of the body. Figure 1: Ahmed car geometry and drag forces on body It was observed by Ahmed et al. (1984) that a change in the slant angle gave a significant change in the drag coefficient, as shown in Figure 1 taken from Ahmed et al. (1984). An increase in the slant angle from zero to around 30° results in a gradual increase in the drag coefficient with a significant shift in 342 F.R. Menter and M. Kuntz the contributions from the different surfaces. While the viscous drag forces show little sensitivity to the slant angle, there is a distinct increase in the drag portion coming from the slant surface combined with a decrease in the contribution from the back surface. At a critical slant angle of around 30°, the drag coefficient drops significantly, mainly due to a reduction in the contribution coming from the slant. At higher slant angles, the drag coefficient stays virtually constant. The change in drag coefficient is due to a change in the flow topology in the aft region of the car body (Lienhart and Becker 2003). Figure 2: Flow topology in the aft region of Ahmed car body (left: 25° and right: 35°). Courtesy Lienhart, LSTM Erlangen. Figure 2 shows a sketch of the flow topologies identified by Lienhart and Becker (2003) for 25° and 35° of slant angle. One of the main differences is that the flow at 25° shows a separation bubble at the onset of the slant, whereas the 35° case is fully stalled over the entire slant. A possible interpretation of the flow dependence on the slant angle is as follows: the initial increase in the slant angle results in a change of the direction of the free flow (similar to the effect of a flap on an airfoil). The change in flow direction requires an opposing force on the body, resulting in an increase of lift. An increase of lift in turn results in an increase of the induced drag and a strengthening of the trailing vortices (tip vortices). The effect of the induced drag results in an increased drag on the slant, as shown in Figure 1. At a critical slant angle, the flow over the slant can no longer follow the contour, resulting in a fully stalled flow with no reattachment on the slant. At this point, the induced lift breaks down and the drag comes mainly from the base drag of both the slant and the back surface. Note that the flow topology has to be identical for 0° and 90°. The present interpretation is consistent with the significantly reduced vortex strength between the 25° and the 35° case measured in the experiments. This flow poses a severe challenge to RANS turbulence models, as it requires models which are capable to correctly predict separation and reattachment at the lower slant angles in order to capture the correct flow topology. Otherwise, the overall flow characteristics will be missed resulting in an incorrect prediction of drag and more severely of lift. Adaptation of Eddy-Viscosity Turbulence Models to Unsteady Separated Flow 343 Turbulence Model RANS simulations have been carried out with the standard k-ε model (Launder and Spalding, 1974) with wall functions and the SST model (Menter 1993) with automatic wall treatment (Esch et al., 2003). In addition, the SSTDES model proposed by Strelets (2001) has been modified and applied to overcome some of its deficiencies in the RANS regions. SST-DES Formulation Strelets et al. (2001) The idea behind the DES model of Strelets (2001) is to switch from the standard SST-RANS model to an LES model in those regions where the turbulent length, L t, predicted by the RANS model is larger than the local grid spacing. In this case, the length scale used in the computation of the dissipation rate in the equation for the turbulent kinetic energy is replaced by the local grid spacing, ∆. ε = β *kω = k 3/ 2 k 3/ 2 → Lt CDES ∆ for CDES ∆ < L t; ∆ = max(∆ i ); Lt = k βω * The practical reason for choosing the maximum edge length in the DES formulation is that the model should return the RANS formulation in attached boundary layers. The maximum edge length is therefore the safest estimate to ensure that requirement. The DES modification of Strelets can be formulated as a multiplier to the destruction term in the k-equation: Lt ,1 ε = β *kω → β *kω ⋅ FDES with FDES = max CDES ∆ with CDES equal to 0.61, as the limiter should only be active in the k-ε model region. The numerical formulation is also switched between an upwind biased and a central difference scheme in the RANS and DES regions respectively. SST-DES Formulation CFX The main practical problem with the DES formulation (both for the Spalart Allmaras and the SST-DES model) is that there is no mechanism of preventing the limiter of becoming active in the attached portion of the boundary layer. This will happen when the local surface grid spacing ∆s is less to the boundary layer thickness ∆ S < cδ with c of the order of one. This is not a situation unlikely to occur, especially when unstructured grids are used in the simulation. 344 F:R: Menter and M. Kuntz In case that the limiter is activated in the boundary layer, the result will in most cases be grid induced separation. In other words, the separation onset and therefore the flow topology can be altered by the grid provided by the user. Figure 3 shows an example of a grid-induced separation based on the grid spacing shown by the grid lines. It can be argued that the grid induced separation could be avoided by a larger grid spacing in the lateral direction, but that implies that the flow direction is known at the grid generation stage, which is not the case in most complex three-dimensional simulations. Furthermore, unstructured prism/tetraeder based grids are typically isotropic on the surface, eliminating this option. It should also be noted that a grid-spacing in the lateral direction (spanwise for a wing) which is solidly on the “safe” side of the DES limiter, would prevent the DES mode in the region downstream of separation and thereby limit the effectiveness of the DES model to produce unsteady structures in the separating shear layer. Figure 3: Regions of negative U-velocity for NACA4412 simulation for SST model (left) and SST-DES model by Strelets (right) – separation point indicated by arrow One way of alleviating the grid induced separation problem is to “shield” the boundary layer from the DES limiter, thereby avoiding/reducing the problem. As the SST model is based on a zonal formulation, differentiating between the boundary layer and the rest of the flowfield, the blending functions of the model can also be used to formulate a zonal DES limiter. The following modification is therefore proposed for the SST-DES model: Lt (1 − FSST ), 1; with FSST = 0, F1 , F2 FDES −CFX = max C DES ∆ FSST=0 recovers the Strelets et al. model. F 1 and F 2 are the two blending functions of the SST model. F2 shields more of the boundary layer and is therefore the prefered default. It should however be noted that even F2 does not completely eliminate the problem, but reduces it by an order of magnitude, ∆ S < cδ where c is now of the order of 0.1. Figure 4 shows the same simulation as for the standard SST-DES formulation computed with the SST-CFX-F2 model. The influence of the DES limiter Adaptation of Eddy-Viscosity Turbulence Models to Unsteady Separated Flow 345 is avoided and the DES model does not affect the separation point. It can be seen that even a more severe grid refinement does not lead to separation (right picture). However, refinement of the surface grid below ∆ S = 0.1δ should be avoided. Figure 4: Regions of negative U-velocity for NACA4412 simulation for SST-DES-CFX-F2 model– separation point indicated by arrow. Right – locally refined Numerical Results Simulations for the two experimental cases with 25° and 35° slant angle have been computed using three different grids with 0.65, 1.3 and 2.6 million nodes for the half-model. The results were grid converged to plotting accuracy on the medium grid. For the 35° case the k-ε and the SST model predict a fully separated flow at the slant region, which is in good agreement with the experimental data. In the following, the flow for the 25° slant angle is analyzed, as it is the more challenging configuration. A detailed report for the 25° and 35° slant angle, including a grid refinement procedure, can be obtained upon request from the authors. Note that both cases have been computed by several authors for a recent ERCOFTAC workshop (http://labo.univpoitiers.fr/informations-lea/Workshop-Ercoftac-2002/Index.html). RANS Simulations 25°° Slant Angle For the 25° case, the agreement between the numerical simulations and the experimental data was much less satisfying than for the higher slant angle and severe differences between the solutions have been observed. It was also found that this case was very sensitive to numerical details – a first order upwind simulation with the SST model resulted in an attached flow, whereas a second order solution produced a separated flow. 346 F.R. Menter and M. Kuntz Figure 5 shows a comparison of the velocity profiles in the symmetry plane. For this case, the models produce very different results. The k-ε gives a fully attached flow, whereas the SST model results in a stalled flow. Both solutions are not in agreement with the data, which show a separation and subsequent reattachment along the slant. The SST-DES results are discussed in the following chapter. For this case, the differences in the turbulent stresses are even more pronounced, as can be seen in the turbulent kinetic energy shown in Figure 6, which is underpredicted by an order of magnitude by the SST model. The k-ε model does not predict an increased level of turbulence at all, as it fails in capturing the flow separation. Shear layer experimental data suggest however, that there is an excessive level of turbulence in the experimental data, which seems to be a result of a large-scale unsteady behavior of the flow. Spalart (2003) argues that the observed level of turbulence could be obtained by a repeated switch in the flow topology from attached to separated and back. The large level of the experimental turbulence intensity in the spanwise direction suggests that this motion should be accompanied by a strong lateral movement. None of these effects is apparently predicted by any of the RANS models. Figure 5: Velocity profiles computed with different turbulence models for the 25° case Adaptation of Eddy-Viscosity Turbulence Models to Unsteady Separated Flow 347 Figure 6: Profiles for the turbulent kinetic energy for the 25° case From the standpoint of the steady state flow topology, it appears that the k-ε model is in closer agreement with the experimental data than the SST model. For the 25°, the experiments show a topology associated with high lift and strong trailing vortices. The confined separation bubble does not alter the overall flow topology. RANS models, which predict no separation, are therefore in closer agreement with the experiments. This does however not imply that the k-ε model is as superior turbulence model for this type of flow. The agreement in flow topology is a result of the failure of the model to predict the separation and not of its ability to accurately predict reattachment. The ability of a turbulence model of predicting the onset of separation is still the first priority for aerodynamic flow simulations and cannot be traded against other model characteristics. DES Simulations 25°° Slant Angle The 25° testcase is a good candidate for the evaluation of the improvement achieved by the DES formulation in the separated flow region. The goal is the resolution of the unsteady features of the shear layer separating from the edge of the slant. The vortex-shedding and break-up is expected to be the main mechanism to increase the turbulent energy in the separated shear layer, thereby forcing a reattachment of the flow. However, as the high levels of turbulence in the measurements cannot be explained by classical shear layer physics, the next question is, if the large-scale unsteadiness observed in the experiments can be triggered by the turbulent structures emanating from the slant. 348 F.R. Menter and M. Kuntz The DES simulations require a significantly refined grid in the slant region, particularly in the spanwise direction, to activate the DES limiter. In order to keep the total number of nodes at an acceptable level, two steps of local grid refinements have been introduced in the slant region. The total number of nodes is 1.783.000, and the number of spanwise cells is 70 on the slant. Off course, the entire body is now computed without a symmetry assumption. The time step for the simulation is ∆t=10-4[s], which corresponds to 250 time steps for one passage of the car at freestream velocity. A large number of steps were required to ensure proper statistical behavior of the solution. Due to the high computing costs of DES, only one grid could be analyzed. Care was taken, that the resolution in the RANS region was sufficient to capture the attached boundary layers. In the separated region, local grid refinement was applied to ensure an optimal distribution of nodes. Figure 7 shows the time history of the forces acting on the body in the three space dimensions. It can be seen that the forces are non-periodic and even after ~7000 time steps do not follow a regular pattern. Nevertheless, the unsteady structures in the separating shear layer have changed the flow topology as can be seen in Figure 8, showing the wall-shear stress vectors on the slant for the SST and the SST-DES-CFX-F1 formulation. Instead of the fully separated flow topology, with low lift and weak trailing vortices, the DES topology is closer to the experiments with a confined separation zone, increased lift and strong trailing vortices. Note that the level of fluctuations in the lift forces is more than 50% of the mean lift. Figure 7: Time history of forces on Ahmed car body Adaptation of Eddy-Viscosity Turbulence Models to Unsteady Separated Flow 349 Figure 8: Flow structure on the slant for SST (left) and SST-DES-CFX-F1 (right) More details can be seen in Figure 9 showing the velocity profiles at a plane at 180 [mm] (the half width of the body is 194 [mm]). In the symmetry plane (see Figure 5 and Figure 6), the advantage of the DES simulation is not very pronounced, but in the off-symmetry plane, the change in flow topology and the improvement in the predicted results is apparent. The unsteady simulations give a first indication of the underlying mechanism of the large-scale unsteadiness of the flowfield. It is observed that the separation zone has a strong lateral movement, which at some instances interacts with the side vortices, leading to a vortex-breakdown. This is shown for a certain instance in time in Figure 10. The right side vortex is unaffected by the separation, whereas on the left side of the body the separation interacts with the vortex leading to a vortex-breakdown. The time value correlates with a maximum of the side force (compare Figure 7). While it is likely that this is the main mechanism for the experimentally observed global unsteadiness, it is not (yet) of sufficient strength to reproduce the experimental fluctuation level. It is not clear if this is a question of an insufficient length of the time integration, or a shortcoming of the DES model. 350 F.R. Menter and M. Kuntz Figure 9: Comparison of velocity profiles at Y=180[mm] for SST and SST-DES-CFX-F1 model for the 25° case Figure 10: Vortex structures on slant of Ahmed car body (25°) Grid Induced Separation - Revisited After the simulation of the Ahmed car body using DES, it is worthwhile revisiting the issue of grid induced separation. The present application required a significant increase in the lateral grid spacing (compared to RANS) in order to activate the DES limiter in the separating shear layer emanating from the car roof. As the 2D vortices originating from the separation are smaller than the boundary layer thickness and have to be resolved with several grid nodes, it is also required to have a streamwise grid spacing lower than the boundary layer thickness. If these grid resolution requirements would not be satisfied in the region close to the separation line, the onset of the DES limiter would be delayed to a significantly further downstream location, thereby reducing the Adaptation of Eddy-Viscosity Turbulence Models to Unsteady Separated Flow 351 chances of capturing the essential flow features. For the present application, the flow separation takes place at a corner - it is therefore not of major consequences if the DES limiter is activated already in the region upstream of the separation line, as the boundary layer there is not exposed to an adverse pressure gradient. Grid induced separation is therefore not a problem for the present geometry. However, for a general car geometry, where the separation can be induced by an adverse pressure gradient from a smooth surface, standard DES would face severe difficulties. One choice is generating a fine grid, which would allow the activation of the DES formulation in the separation zone, at the danger that the DES limiter will change the RANS part and produce grid-induced separation. The second choice is the use of a coarser grid and thereby delay the DES impact far downstream of the separation line and miss the physics of the flow. As the separation line is not known during grid generation, it is difficult to imagine how a suitable DES-grid could be generated for such a flow. The use of the proposed zonal DES formulation, based on the SST model blending functions, will at least reduce the risk of a grid induced separation occurrence, however, it still has to be tested if the blending function will switch quickly enough from RANS to DES to activate the DES mode for a pressure induced separation. Nevertheless, this approach is preferred, as it reduces the influence of the user (grid) on the solution. Conclusions CFD simulations have been carried out for the generic Ahmed car body at 25°. Strong turbulence model differences where observed for this case. The k-ε model produced an attached flow over the entire slant, whereas the SST model predicted a fully stalled flowfield. Both solutions are in disagreement with the experiment, where the flow separates and reattaches at about 50% of the slant. DES simulations based on a modified version of the SST-DES formulation, resulted in a significant improvement of the solution compared to the SSTRANS model. Instead of a fully stalled flow, the time-averaged DES solution shows a confined separation zone and strong trailing vortices, associated with the experimentally observed flow topology. The DES solution also gives first insight into the mechanism driving the strong unsteadiness of the flow, as observed in the experiments. The most likely explanation is a strong lateral movement of the separation zone, which interacts with the side vortices, leading to periodic vortex-breakdown. It could not be determined if the full unsteady effect could be obtained by further continuation of the simulation, due to constraints in computing power. 352 F.R. Menter and M. Kuntz Acknowledgment This work was supported by research grants from the European Union under contracts GRD1-2001-40199 (Flomania) and EVG1-2001-00026 (EXPRO). The authors want to thank H. Lienhart from the University of Erlangen for the provision of some of the figures concerning the physical interpretation of the flow. Literature Ahmed S.R., Ramm G. and Faltin, G., 1984, “Some salient features of the time-averaged ground vehicle wake” SAE technical paper 840300. Durand, L., Kuntz, M. and Menter, F., 2002, “Validation of CFX-5 for the Ahmed Car Body”, CFX Validation Report, CFX-Val 13/1002. Esch, T., Menter, F.R. and Vieser W., 2003“ Heat transfer predictions based on two-equation turbulence models”, 6th ASME-JSME Thermal Engineering Joint Conference, March 16-20, 2003, Hawaii. Johnson, D.A. and King, L.S., 1984, “A new turbulence closure model for boundary layer flows with strong adverse pressure gradients and separation”, AIAA Paper 1984-0175. Johnson, D.A., Menter, F.R., and Rumsey C.L., 1994, “The status of turbulence modeling for aerodynamics”, AIAA Paper 1994-2226. Launder B.E. and Spalding D.B., 1974, “The numerical simulation of turbulent flow”, Comp. Meth., in Appl. Mech. and Engng., 3, pp 269-289. Lienhart H. and Becker S., 2003, “Flow and turbulence structures in the wake of a simplified car model”, SAE Technical Paper 2003-01-0656, Michigan. Menter, F.R., 1993, “Zonal two-equation k-ω turbulence model for aerodynamic flows”, AIAA Paper 1993-2906. Menter, F. R. Kuntz, M. and Bender R., 2003, “A scale-adaptive simulation model for turbulent flow predictions”, AIAA Paper 2003-0767. Spalart, P.R. and Allmaras, S.R., 1994, “A one-equation turbulence model for aerodynamic flows”, La Rech. Aerospatiale, V.1, pp.5-21. Spalart, P.R, Jou, W.-H., Strelets, M. and Allmaras, S.R., 1997, “Comments on the feasibility of LES for wings, and on a hybrid RANS/LES approach”, 1st AFOSR Int. Conf. On DNS/LES, Aug.4-8, 1997, Ruston, LA. In Advances in DNS/LES, C. Liu & Z. Liu Eds., Greyden Press, Colombus, OH. Spalart, P.R., 2003, Private communication. Strelets, M., 2001, “Detached eddy simulation of massively separated flows”, AIAA Paper 2001-0879.