CFD - Computational Fluid Dynamics, Motorsport, Formula 1

[Gecko]

There is no real discrepancy. Having a finer mesh of similar quality will bring about better results than a coarser mesh, no question about it. If you can afford a finer mesh then by all means use it. The mesh sizes as used by F1 teams are therefore increasing all the time, simply because the computing facilities are getting better.

The point I was trying to make is that the advantages thus obtained just might not be as significant. I was trying to argue that, instead of doing a single run on a fine mesh, you might, in the same amount of time, be better off by doing five or ten runs on coarser meshes by varying parameters such as boundary layer cell thickness and turbulence models. But in any case, you should use meshes as fine as your computing facilities allow.

[AeroGT3]

Gecko, AeroGT3,

I am no expert on this subject, but you guys seem to be saying that small models are sufficient, but my information is that F1 models are big and getting bigger.

So there appears to be a discrepancy here, can you explain why? Is it because you are talking about components (like a front wing) and i am talking about whole car models?

G

Appears to be a discrepancy? Where? Can you point to something that says most F1 models are more than 30M cells for a single car? I'm personal friends with a guy who's 3 years out of a very senior aero position with a team. I know some others, too. And I've got my own work. All say that 30M cells for a single car just isn't necessary. Especially for design.

[connollyg]

Gecko, AeroGT3,

I am no expert on this subject, but you guys seem to be saying that small models are sufficient, but my information is that F1 models are big and getting bigger.

So there appears to be a discrepancy here, can you explain why? Is it because you are talking about components (like a front wing) and i am talking about whole car models?

G

Appears to be a discrepancy? Where? Can you point to something that says most F1 models are more than 30M cells for a single car? I'm personal friends with a guy who's 3 years out of a very senior aero position with a team. I know some others, too. And I've got my own work. All say that 30M cells for a single car just isn't necessary. Especially for design.

Robert,

What can i say, i have been reliably informed that some of the teams are routinely using models in excess of 30M cells, and if they are not, i cannot understand why they are all rushing to buy such big supercomputers? You do not need a 4K core machine to do 30M cells!

G

[AeroGT3]

Robert,

What can i say, i have been reliably informed that some of the teams are routinely using models in excess of 30M cells, and if they are not, i cannot understand why they are all rushing to buy such big supercomputers? You do not need a 4K core machine to do 30M cells!

G

No, you don't need 4k cores for 30M cells but you do need them to do many runs of 30m cells. Are these teams routinely doing multi car models? Are they routinely using higher order turbulence models? Are they routinely NOT using symmetry planes? Are they including engine or radiator models?

I maintain that for single car, external models of F1 cars with symmetry planes, more than 30M cells is not necessary. There are teams doing considerably better than Renault using this philosophy.

[kilcoo316]

That's complete garbage - software developer's talk. I've done countless grid studies that show refinement of 1.3 is not necessary.

Do the Fourier series expansion for a central difference scheme yourself then.


All of a sudden, I've alot less confidence in your results vis-a-vis BMW. You obviously cannot have 10 or more cells within your boundary layer if you use expansion ratios greater than 1.3 with y+ of around 100. Thus, it is not correctly resolved, therefore any detachment will not be right, and the wakes will not be right, nor will the impact of the wakes on any downstream components.


There is no such thing as a free lunch.

Your missing my point.

Your completely missing mine.

Wing A affects the performance of suspension B.

Get the wake of wing A wrong and you get suspension B wrong.


CFD is excellent for comparing isolated components. Bring interactions into it and things get very tricky, very quickly.

Actually, that isn't true of aircraft. The tail can change Cp at the nose by 15 or more percent. It is an equally coupled problem.

The angle of attack of the aircraft can (and will) change the nose pressure distribution.


Detail changes to the vertical fin or horizontal stabiliser profiles won't. Even shifting from a low tail to a T-tail won't do that.

[kilcoo316]

http://www.fluent.com/solutions/sports/ ... onf_05.pdf

35 million cells.... in 2005.

Sometimes even 100 million back then:

Larsson recalled a study in 2003 when the team’s cars were exhibiting an aerodynamic problem that was proving diffi cult to identify. They were losing a lot of rear downforce in yaw but the wind tunnel could only simulate up to 4.5 degrees of side slip. However, CFD was able to conduct a study on a fi le of 100 million cells that could simulate six degrees of yaw

More evidence:


http://www.fluent.com/solutions/sports/tn272.pdf

Computational Fluid Dynamics has become an integral part of modern development in motorsports, the product FLUENT has become the CFD-code of choice for a clear majority of racing teams world-wide. The availability of massive parallel computing systems and LINUX clusters in combination with efficient and reliable solution methods have pushed up the limits of applicability substantially. Total cell counts beyond 100 Million are almost a standard today, the solution of transient tasks with complex geometry and demanding physics is tackled more often now, and even the simulation of two racing cars during overtaking seems to be feasible in near term.

http://www.f1complete.com/content/view/2759/900/

Albert2's enormous technological potential is used for analyses in the area of aerodynamics. By means of CFD aerodynamics components for Formula 1 cars are calculated on the computer, using numerical grid models that often consist of more than 100 million cells.

I even read pieces about BMW considering 1 BILLION cells... probably not feasible yet, but if its being considered, it means they feel there is something to be chased.

[connollyg]

Robert,

What can i say, i have been reliably informed that some of the teams are routinely using models in excess of 30M cells, and if they are not, i cannot understand why they are all rushing to buy such big supercomputers? You do not need a 4K core machine to do 30M cells!

G

No, you don't need 4k cores for 30M cells but you do need them to do many runs of 30m cells. Are these teams routinely doing multi car models? Are they routinely using higher order turbulence models? Are they routinely NOT using symmetry planes? Are they including engine or radiator models?

I maintain that for single car, external models of F1 cars with symmetry planes, more than 30M cells is not necessary. There are teams doing considerably better than Renault using this philosophy.

Renault aren't the only team with a computer that size (or planning one of that size). All i can say is, that i was told WHOLE car models, and that model sizes were substantially bigger than 30M, and they are running these models on machines with 1K cores or less!

Yes i realise that having a machine that big doesnt always mean that the jobs will be that big, especially as some of the codes don't scale well, but i know of one team that is working on superfast job turn-around and they are looking at a big machine to realise that!

Its interesting to note that most of the teams are now moving away from their traditional SMP machines towards a cluster, but these things are cyclic.

Intel Xeon has recently been in favour, but now AMD's Barcelona chip is finally shipping and seeing how its faster i expect some drift that way, although Intel have their Nehalem chip shipping this year and with its replacement of the FSB for QPI there is a good chance they will have leapfrogged back, maybe at the right time for next seasons upgrades! All in all the machines will be getting bigger and faster!

G

[connollyg]

Hey Kilcoo, thanks for the backup

[AeroGT3]

Do the Fourier series expansion for a central difference scheme yourself then.

How about a grid independence study instead? Hypothesizing what error might be with math is great, but actually doing a case to find it better. What's the order of accuracy on this central differencing your doing?

Thus, it is not correctly resolved, therefore any detachment will not be right, and the wakes will not be right, nor will the impact of the wakes on any downstream components.

Dead wrong. Take a look at the image below. The middle wake is for what would be about 25M cells on a car. The lowest image about 120M cells. Care to point out how "wrong" the middle mesh (which later I used on a full car) is? What am I losing vs the bottom mesh, which is 4+ times the cell count?



You are beyond completely missing mine. I know components are coupled. That's bloody obvious. My point is that you don't need an excessively find mesh to capture that coupling.

Detail changes to the vertical fin or horizontal stabiliser profiles won't. Even shifting from a low tail to a T-tail won't do that.

Wrong again. Check out two papers presented at AIAA Aero conference in January of this year. On that exact topic . . . showed significant changes by changing what you claim makes no difference.

Total cell counts beyond 100 Million are almost a standard today,

That case wasn't using symmetry planes, so cut off half the cells there. How detailedly are they treating the brakes and heat transfer there? That could be 5+ million cells per axle a long if you're getting the entire caliper plus the innards of the discs all at y+ of 1.

I maintain that for external flow in a pure headwind 30M cells is adequate.

[dnomdec]

Glad to see some active discussion going on this thread again.

I have to say I am mostly with Robert on this one as well.

Larsson recalled a study in 2003 when the team’s cars were exhibiting an aerodynamic problem that was proving diffi cult to identify. They were losing a lot of rear downforce in yaw but the wind tunnel could only simulate up to 4.5 degrees of side slip. However, CFD was able to conduct a study on a fi le of 100 million cells that could simulate six degrees of yaw

In this case, they were dealing with tyre wake with yaw. The tyre wake alone is the 2nd biggest (if not the biggest) beast, just because of the huge amount of separation. In yaw, it gets even more nasty. They probably piled on lots of cells behind the front tyre to try to resolve the wake as much as possible to get the wake/sidepod/indy fin interaction right, and hence the 100M+ cells.

They also didnt' have a wind tunnel for that case. My bet is 100M cells with some k-eps RANS may still be a couple % off, but when your wind tunnel doesn't work well enough, which was the case, piling on cells become the only pre-manufacturing thing you can do to get the best accuracy. To capture design trends, 3-5% error is generally good enough. Every now and then you may run into a case that CFD solns are not good (more off rel. to the wind tunnel) at a certain attitude, but it's exactly why the tunnel is there.

While most teams certainly have some limited quantities of models/cases with over 100M cells, I highly doubt that the majority of the models have that many. For most studies, which is SS, straight-line stuff, 20-30M is quite adequate.

Total cell counts beyond 100 Million are almost a standard today

Well I came across that piece a while back too. As mentioned above, 100M+ cells aren't typical. And, Fluent wrote that article. More cells => more nodes for the same turnaround => more parallel licenses => more $$. There is that possible incentive we need to beware of.

Hey Robert, haven't talked to you in a while. Still in SLO?

[connollyg]

More cells => more nodes for the same turnaround => more parallel licenses => more $$. There is that possible incentive we need to beware of.

Well that is point i was trying to make earlier, if 30M cells is all that is required, why are all the teams rushing to buy bigger and bigger machines, in terms of $$$ todays 2K core machine is probably the same cost as the 600 Core machine that teams bought 2-3 years ago and that is including the licences.

G

[slimjim8201]

More cells => more nodes for the same turnaround => more parallel licenses => more $$. There is that possible incentive we need to beware of.

Well that is point i was trying to make earlier, if 30M cells is all that is required, why are all the teams rushing to buy bigger and bigger machines, in terms of $$$ todays 2K core machine is probably the same cost as the 600 Core machine that teams bought 2-3 years ago and that is including the licences.

G

Easy, speed. Bigger machines turn around simulations faster.

[kilcoo316]

How about a grid independence study instead? Hypothesizing what error might be with math is great, but actually doing a case to find it better. What's the order of accuracy on this central differencing your doing?

CFD is mathematics.

Getting one or two results on an individual study does not circumvent the maths behind it.


Its pish like that which leads to crap results resulting in CFD getting such unwarranted criticism.



Virtually all simulations done using 2 eqn RANS will be 2nd order, although some may have a hybrid of 1st upwind with 2nd order for improved stability.


Oh, and yes. I've seen massive differences in simulation quality going from wall function to low-Re modelling.

And even from going from a disc of 1.25 to 1.1.

The difference is mainly coming from the number of elements in the BL.

Dead wrong. Take a look at the image below. The middle wake is for what would be about 25M cells on a car. The lowest image about 120M cells. Care to point out how "wrong" the middle mesh (which later I used on a full car) is? What am I losing vs the bottom mesh, which is 4+ times the cell count?

Even with that most rudimentary method of comparison (the pretty picture that tells very little) there are differences in your pressure zones.

You produce some graphs of pressure along various Y locations (plotting pressure against spanwise location) and you'll see proper differences.

You are beyond completely missing mine. I know components are coupled. That's bloody obvious. My point is that you don't need an excessively find mesh to capture that coupling.

I believe that is incorrect.

And its is not an excessively fine mesh.

30mil is coarse.

Wrong again. Check out two papers presented at AIAA Aero conference in January of this year. On that exact topic . . . showed significant changes by changing what you claim makes no difference.

Can you name the papers?

If not, just give me the conference name and location.

That case wasn't using symmetry planes, so cut off half the cells there. How detailedly are they treating the brakes and heat transfer there? That could be 5+ million cells per axle a long if you're getting the entire caliper plus the innards of the discs all at y+ of 1.

I maintain that for external flow in a pure headwind 30M cells is adequate.

Oh right.

SO 30 million is ok as long as you only model the most rudimentary aspects of the car and do it in conditions which are of little actual use?


Very little F1 testing is pure headwind now - when do you need downforce in a straight line? Its all yaw, hence all needing the full car.

[kilcoo316]

For most studies, which is SS, straight-line stuff, 20-30M is quite adequate.

Most studies are in yaw now.

As I said above - when do you need downforce in a straight line?

Well I came across that piece a while back too. As mentioned above, 100M+ cells aren't typical. And, Fluent wrote that article.

In 2004

[kilcoo316]

Easy, speed. Bigger machines turn around simulations faster.

Slim, maybe you missed my post earlier, are you looking at porting CFdesign onto GP-GPUs?