any suggestions for using tetra meshes in star. I have noted that the convergence is much slower than in fluent. Any suggestions to speed it up.

Hi, One thing you can try is tetalign in prostar. This will reorder the vertex numbering, reducing memory and cputime requirements. However, we had big problems with the predicted stagnation pressure using tet meshes in Star. We did a series of runs evaluating tet meshes in Star and Fluent on the Asmo car. It was exactly the same mesh in both cases and StarCD overpredicted the stagnation pressure with a horrible 16 % while Fluent overpredicted with 4 %. This case was sent to StarCd but no explenation was given to us. So our conclusion was that StarCD did not perform well on this exteriour form while Fluent did really good.

Regards Anders

what convection scheme did you use in STAR. Also what kind of convergence did you get for mass residual. Please tell me the version of STAR you used and switches and real constants you turned on

Hi, We used Mars for convection and converged below 1e-3 in scaled residuals for mass and momentum. This was version 3.05, standard k-epsilon turbulence model.

Regards Anders

Hi Anders! I'm curious - what do you think is the main reason for the overprediction of the stagnation region pressure? Is it an accuracy problem or a model problem? Does improving the grid-resolution or switching to a higher-order scheme give you better results?

Without any experience from your application my guess is that it is mainly a turbulence modeling problem and that the numerical schemes used are of secondary importance. Do you agree?

With the standard k-epsilon model you'll certainly over-predict the turbulent viscosity in stagnation regions and this will also lead to an over-prediction of stagnation-pressure, as you describe. Did you use the same standard k-epsilon model in both Star-CD and Fluent? Details in the implementation of this model in the codes might explain the difference. I'm sure that you have tried other models that don't have this "stagnation" problem (Realizable k-eps, Kato-Launder ...) - does it help?

Did you compare the turbulence levels and eddy-viscosity levels that Star-CD and Fluent gave in the front region?

Hi Jonas, The work was done as a Master Thesis, hence it was a bit limited in time but a few things where tested.

Mesh resolution: We had the same mesh, and I mean identical mesh between Fluent and starruns. Comparing first order UD and Second order Marsscheme gave very small changes. The same behaviour for Fluent! Case Cp,stagnation Fluent UD 1.0491 Fluent 2:nd order 1.0462 Star UD 1.1615 Star MARS 1.1601

Looking at pressure distribution on the car symmetry line shows very little difference between the schemes in the front. Star UD and Mars scheme behaves very similar to each other and Fluent UD and their second order scheme also behaves very similar to each other. This, to me, indicates that it is not a grid refinement effect, do you agree?

Of course, since we used std k-epsilon, we expected overprediction of the stagnation pressure. But 16% is much higher than expected, fluents 4 % is more what we had expected. What intrigued us is what produces the difference between the codes.

It is more or less the same turbulence model, i.e. the coefficients are the same, execpt \sigma_epsilon which was 1.3 for Fluent and 1.219 for Star. Do you think that this is a significant difference?

We tried to run the cubic model and RNG in Star, which could improve stagnation region prediction, but none of those converged.

Regards Anders

Interesting. Your obeservations confirm that the problem is model related and not primarily related to the numerics/grid.

It is not that surpricing that you get different results with the same model in different codes. In this type of stagnation flows the details of how k-epsilon is implemented is often important for how much it "explodes". I'd really recommend you to check your turbulence levels and eddy-viscosities in the front region - if you see a clear difference between the codes there then that is your explanation. I don't think that the small difference in sigma_epsilon can explain this.

The fact that StarCD gives worse results than Fluent in this particular case does not necessarily mean that Fluent in general is better. My guess is that Fluent uses some internal limiter or some other trick on this standard k-epsilon model which reduces the overproduction of k. In fact, StarCD might well have the most correct implementation of this model!

Hasn't StarCD got the Kato-Launder fix? That one should remove all over-production in the front region - but it might also make your simulation more unstable since it will lower your turbulence-levels quite significantly. I would guess that if you could get a converged solution with Kato-Launder, RNG or a cubic model then it would be much better.

Btw, do you see the same problems on structured grids? Do you use the standard k-epsilon model there also?

I am aware of that different implementations of the same model might produce different results, but I did not expect them to be this large. It is a very good point that Fluent might have a limiter that fixes the turbulence levels.

We did try other turbulence models, or variants, RNG and cubic but unfortunately they did not converge.

It is a good point to look at the turbulence levels, I think I will dig up the data and have a look at it soon.

We tend to use the standard k-epsilon with structured meshes and the overprediction is much lower, say 2-5 percent depending on front mesh density etc.

Anders

It's strange that the structured mesh in StarCD gives you so much better results when the change in scheme on an unstructured mesh didn't make much difference. That would indicate that the problem is not only model related. Perhaps Fluent has a better way of implementing k-epsilon on tets, I don't know, haven't used StarCD myself yet. Thanks for sharing your experience btw!

I followed this with great interest - the picture sometimes can be confusing. We have some clients reporting very good results when they compared different codes, the best results are achieved with STAR-CD's k-e cubic model.

It is noted that when better turbulence model is used, the flow, especially the region behind the wheel, becomes unsteady as in the reality. So maybe it's not surprising that the steady approach may hit the convergence problem. But it may not affect your analysis even with steady method.

Hi Jonas,

could please explain the connection between increased eddy viscosity and increased stagnation pressure.

Thanks, Mark

If you have a very viscous region around the front of the car then the stagnation pressure (total pressure) along the stagnating streamline will increase due to viscous effects from the surrounding free-stream passing by. This effect is only visible if your turbulence models behaves very badly and creates "honey".

The k-e model is known for generating to high eddy viscosity in stagnation regions. Does this mean that I this will result in a too high stagnation pressure ?

Mark

Anders,

Were both codes using the same energy equation? (e.g. static enthalpy, total enthalpy, etc..)

SC

Hi, The energy equation was not solved at all since this was an isothermal simulation.

Regards Anders