[hokhay]

Hi all, I am using k-omega SST turbulence model for running car aerodynamics simulation and I have verified the computational model using Drivaer and other two car models which their drag results are accurately predicted against the experimental results, error within 3%.

However when it comes to Tesla Model S, the drag result is very inaccurate where it predicts 10% higher than the official number (0.24). I have tested with coarse and fine mesh but the predicted drag is still higher than official number by 10-20%.

I then switched the turbulence model to realizeable K-epsilon and the predict drag was much closer to 0.24, around 0.238.

I am somewhat confused about the prediction of the turbulence model. How can one do well on few cars but not on Model S? What does these phenomena suggest?

Mesh settings are the same for all models with refinement behind car model and behind mirrors. The resulting mesh number is 22 millions.

[tbwake]

I would suggest post processing the results to look at where the flow is attached and where it is separated. Your drag will be highly dependent on this. You may find that with the Sst model you have a larger region of separation compared to realisable K epsilon.

In any case, how do you know your official CD is correct anyway?

[Gerry Kan]

Dear Hokhay:

A Model S is no different than any other vehicle. I would say it is a coincidence that you obtained a drag coefficient that is more agreeable with wind tunnel measurements using one RANS model over another, nothing more. It does not mean that one model is superior than another. Besides, a 10% difference in CD is not really a deal breaker; after all, drag is one of the most difficult quantitties to match using a RANS model.

Also, much of these will also depend on how the wind tunnel measurement is conducted, and how the vehicle model is constructed. Judging from the 22 million cell count and the model of the vehicle, you are probably doing this commercially. So in this case, all I can tell you from experience is that you should know better to earn that money.

Gerry.

[FMDenaro]

Why do you think that validating your mesh setting on a test car would allow you to generalize the validity for a different car? Have you checked you guarentee the same fine resolution of the BL over the whole car? And the test car and model S are computed at the same Re number? Finally, are you sure to have considered the correct geometry tested in the wind tunnel?
Turbulence model in RANS can have real impact on the solution but you need to be sure of what you are doing.

[duri]

Originally Posted by hokhay

Hi all, I am using k-omega SST turbulence model for running car aerodynamics simulation and I have verified the computational model using Drivaer and other two car models which their drag results are accurately predicted against the experimental results, error within 3%.

However when it comes to Tesla Model S, the drag result is very inaccurate where it predicts 10% higher than the official number (0.24). I have tested with coarse and fine mesh but the predicted drag is still higher than official number by 10-20%.

I then switched the turbulence model to realizeable K-epsilon and the predict drag was much closer to 0.24, around 0.238.

I am somewhat confused about the prediction of the turbulence model. How can one do well on few cars but not on Model S? What does these phenomena suggest?

Mesh settings are the same for all models with refinement behind car model and behind mirrors. The resulting mesh number is 22 millions.

SST model has predicted 3 car models within 3% accuracy but not model-s. The turbulence model is expected behave consistent with same physics and mesh distribution. Probably investigating on mesh difference, Reynolds number, loading and streamline curvature difference, component of skin friction, wake, separation pattern etc. would give you a clue on error. This is typical CFD model or process development problem where goal is not to match (matching the test is great but not always possible) exactly the test results but to understand the uncertainty of CFD model in different flow regions.

[hokhay]

Thanks for your replies and they are useful to me. To conclude, the two turbulence models should not be source of making the difference if use probably. The problem could be due to the difference between the wind tunnel model and the CFD model or modelling mistake that did not show up in the 3 cars but on the Tesla.

The Reynolds number should fine as I am using 1:1 scale model and same air velocity as the wind tunnel test. The y+ value over most of the car is around 20, so the it should be fine too. I will try to refine the region where flow separate and high velocity gradient to see if these help.

Thank you very much for the help from you guys.

P.S. I am doing these simulations as a hobby and the results will be posted on my own website. Please feel free to check it out .
Jason

[FMDenaro]

Originally Posted by hokhay

Thanks for your replies and they are useful to me. To conclude, the two turbulence models should not be source of making the difference if use probably. The problem could be due to the difference between the wind tunnel model and the CFD model or modelling mistake that did not show up in the 3 cars but on the Tesla.

The Reynolds number should fine as I am using 1:1 scale model and same air velocity as the wind tunnel test. The y+ value over most of the car is around 20, so the it should be fine too. I will try to refine the region where flow separate and high velocity gradient to see if these help.

Thank you very much for the help from you guys.

P.S. I am doing these simulations as a hobby and the results will be posted on my own website. Please feel free to check it out .
Jason

No, the grid is not good to compute the drag!

[EmmTheof84]

Originally Posted by hokhay

Hi all, I am using k-omega SST turbulence model ...

I then switched the turbulence model to realizeable K-epsilon and the predict drag was much closer to 0.24, around 0.238.
.

Originally Posted by hokhay

The y+ value over most of the car is around 20, so the it should be fine too...

a y+ of 20 is fine for k-epsilon (you can get away with 30 or even higher). For k-w SST (or any k-w) you should have a y+ of less than 5 or ideally close or less than 1

I would suggest trying with k-epsilon, a y+ of 30 to 40 in order to save some mesh count and then trying mesh adaptation

[JBeilke]

Make sure you run your simulations transient in DES mode.

[jordan23]

Thx for post

[CFDfan]

Originally Posted by FMDenaro

No, the grid is not good to compute the drag!

Good job with the website. What software do you use to simulate the drag coefficient?

[DLuo]

How are you confident you're replicating the wind tunnel test accurately?
Do you know what the freestream turbulence intensity and length scales are?
Are you certain your BL profile on the WT floor matches the wind tunnel?
Is your CFD model and exact match in the wheels? If they aren't using simplified wheels are you modeling every tread feature? How are you handling the rotating fluid in between the wheels?
Are confident are you that you're modeling radiator porosity correctly? Do you know if the WT test removed or blocked the radiator?
If you can't answer in the affirmative to these questions I think jumping to the conclusion that the Turbulence Model may be a bit premature.

It's extremely difficult to match a WT test and even more so if you're not from the same organization that ran said WT test.

Anecdote time. I was given a STL of a Model-3 directly from Tesla's Aero dept themselves as a benchmarking exercise. Even though they provided the STL, there was some non-physical geometry in the wheel wells that made me and my colleagues say ??? After conferring with Tesla they just shrugged and said, just go with it. So even working directly with them we couldn't say with 100% confidence the geometry we were given was correct. IIRC the steady state RANS was run with kOmegaSST and we also had a high CD. But running it through DES brought the number fairly close to the published/official CD.
In general for bluff-body vehicle CFD we primarily used kOmegaSST but only relied on RANS for design iterations. If we were trying to validate/correlate to a set of WT values we would focus on a modified DES.