Correlation between CFD and wind tunnel results

[PNSD]

It's good for drag they just setup the thing too stingily for the computing power they have.

I'd also disagree, but primarily with that sentence. The internal flows are likely to be a source of error, but then again I would say the wind tunnel would struggle to get decent correlation with the real thing regarding the internal flows. CFD is getting less processor and memory limited, its storage limited these days, I guess it becomes a compromise of what you want to know, and what you consider more crucial than other aspects.

And when I say CFD is not good for drag, I mean it in every sense, but it depends what you define as acceptable parameters. In aerospace applications 0.1% error in drag is highly unacceptable when you consider that there are numerous wind tunnels with the ability to calculate Cd upto +-0.0001 accuracy with a repeatability error of 0.00005 Cd. (Consider a 0.1% error in Cd in relation to fuel burn, it gets quite costly).

Drag is a difficult subject to understand, let alone calculate. Its far more complex than people initially think. It just so happens the 4 black round things on an F1 car mean drag is the least their worries. So in that respect, maybe CFD calculation of drag is acceptable for F1. But in general terms its far from it.

Even given experimental setup data, WT turbulence, calculating a Cd on a simple aerofoil its hard enough to get within 0.5% in CFD.

[marcush.]

ah.. with 4 open rotating wheels dominating the drag numbers the qquestion is by how much? half the drag of the car ?
still any drag reduction will help overall drag numbers,right? And if you come up with a solution reducing wheel drag there is maybe even scope for removing some wheel drag methinks..

brings me to a question to the aero pros :

couldn´t a set of cleverly positioned vortex generators in front of the wheels help reducing drag of those? Maybe one could really look for drag reductions only and not look single minded for downforce and afterwards try to reduce drag..

[ESPImperium]

There have been plenty of teams been mired by faulty tunnel calibration including Williams, Toyota and Honda. So now it is Ferrari's term.

Correct, and for exact terms;

Toyota: 2006/2007
Honda: 2007/2008 (Both tunnels)
Renault: 2007
BMW Sauber: Late 2008/2009
Williams: 2006-2008
Red Bull/Toro Rosso: 2006/2007
McLaren: 2009

So with the law of averages, its Ferraris turn, last time this happened for them was early 2005.

I think its why Ferrari have been doing constant speed runs on both cars at 200Kph at Malasia arround turn 5 and 6 and on the straights at Malasia and the back straight at China to get a number so they can correct the tunnel and CFD data they are getting.

[marcush.]

To me it´s astonishing that teams do not do this on a regualar basis now.You Really need some benchmarks to validate against .So why not value your development against your starting point now and then for example...I think a lot of teams are too much preoccupied with their advances in development.

with the limited running in my view Fridays need to be extensively used to have a rolling lab with two cars solely to validate your simulations and assumptions.Who cares about laptimes.Qualy is for that.

[ringo]

If you have that level of computing power, you should model the radiators fin by fin and all that is in the side pod and forget about porous surface boundary conditions.

It's already taken them 3 days to mesh it with a porous surface and you're saying they should model the radiator fin by fin and the complicated internal flow associated with such a mess? That would be over complicated for the sake of it if you ask me.

You want accurate results, you build the model accurately.
I don't know what the other guys are talking about above.
I guess they've done it before.

There is no real mathematical reason for the drag to be off by 20% and the lift only by 5%.
Something is wrong with the model and the boundary conditions.

Unless someone can explain why.

Drag is a difficult subject to understand, let alone calculate. Its far more complex than people initially think.

a statement like this should be supported PNSD.

[shelly]

Unless someone can explain why.

RANS can not predict drag accurately, that is why.

[ringo]

Conclusion at the end of the paper:

Correlation with experimental results show that our
model still need to be improved.

[PNSD]

a statement like this should be supported PNSD.

Simple. Go find an aerodynamic consultancy/aircraft manufacturer who use CFD for their drag prediction. You wont find many. Put it this way when aircraft manufacturers pay a British aerodynamic consultancy £100'000's per day to provide drag estimates for their aircraft through use of their WT facilities and help with CFD prediction, it tells you something.

CFD can predict pressure forces very well, but not viscous forces. One of the main sources of error is predicting transition (basically, CFD cant! So its done through modelling which is not great). Another source of error, and related to this model would be the use of an unstructured grid and the increased numerical diffusion you get through that compared to a structured grid.

The short answer is, as Shelly has mentioned, RANS struggles, and I guess its drag prediction where the RANS wall boundary, and viscous assumptions fall short.

Should add - As mentioned, for a racecar this issue isnt really an issue, but it does go someway in explaining the difference in drag. Also, google work done on the Onera M6 wing. Its a popular aerofoil for validation purposes and you'd be surprised just how difficult it is to get drag right for even an aerofoil, albeit in the transonic range.

Edit - Ringo, a quick google search yielded this;

http://aaac.larc.nasa.gov/tsab/cfdlarc/ ... degger.pdf

nothing detailed but it's something. Im sure there's an AIAA article that will provide any information to support this argument . Ill get hunting. My opinion on CFD as it stands is that it is for understanding rather than data acquisition. Still a long way to go yet.

[ringo]

20% is too much IMO and RANS is just a calculation feature. I don't think modern FEA package are restricted to single simulation techniques either. The user may chose based on their needs.

I'm no CDF genius, but there isn't enough information in that first article to suggest it's the cad package that is the problem.
They are at fault with the radiator model and probably a few other things, maybe the turbulence methods, such as the RANS, but it is not suggested by the article as the problem.
The radiator thing is the elephant in the room to me.

Radiator fins may have a spacing of maybe 5mm or so, and when you are meshing to the point you can fit 50 cells in 5mm for argument sake, i don't see why they didn't just create the radiator to begin with.

I don't buy this 20% error thing as the norm for CFD users in racing, i think it's a problem this specific group have and i think that as they themselves say in the conclusion that the model needs to be improved. And who knows they could have corrected themselves right now.

The RANS, LES and so on are just meshing and calculation methods. Having a bazillion cells i don't suppose one is going to be off by 20% from wind tunnel results; assuming they have modeled the conditions perfectly and the other will be within 5%.
In fact there is no reference to using LES or other simulation method to improve the accuracy.

The difference between drag and lift, is simply an axis and wind direction in a CFD environment. There is no discrimination to direction in the calculation. It just so happens that in the z direction something was off with this model.

The fact that they use a full scale model in the cfd and a scale model in the WT is enough to cause a whole new set of variables to reduce accuracy.

[PNSD]

They are using K-Omega SST which I would say is adequate for the problem...

The reason RANS is being used is because they are interested in the time averaged forces I would guess. LES might be successful for modelling the exposed wheel's but when wanting to resolve forces I would say RANS would be more appropriate in this case.

I wouldn't say 20% is the "norm" but I would say to expect large inaccuracies in drag calculation. As you say, the presentation is rather brief, but thats because it is just a presentation rather than report.

I suppose the reason they have gone for 60M cells would be to do with resources and/or some sort of grid convergence based on appropriate criteria.

The 20% variance will come down to alot of things, but generally if you have "bang on" cd values, its often by accident :p

Just spotted this thread, a nice read.

viewtopic.php?f=4&t=2405

edit -

Ringo, supports what has been mentioned about LES but states the time scale issues. Another more than interesting read

http://www.stanford.edu/group/uq/pdfs/c ... 9-0082.pdf

[ringo]

Interesting paper, but it does emphasize how important the geometry is to give accurate results. A slight discrepancy in tyre shape was enough to poke some holes in the relatively close stimulation observations to the real world.
If this happens with a tyre sidewall, imagine what happens with a radiator?
The big difference with LES and RANS does seems to be in small regions of separation in the near wake. They basically behave very similar just 1 wheel diameter away. The difference between calculations doesn't over shadow the much bigger difference if the simulation is fundamentally modeled incorrectly.

If it's quicker to do a sim with square wheels than with perfectly round wheels, i wouldn't blame the CAD package for being off reality by 80%, when i have a sponge as a radiator and polygon wheels.

[marekk]

All Navier-Stokes based solvers compute velocity field of flow.
That means we have some data about how quick the flow at given position is.

Using this information is relatively easy to calculate pressure and stress maps.
Using speed and stress info one can estimate drag forces resulting from dynamic pressure and shear forces (viscosity), but doesn't help much with assesing induced and vortex related drag forces. Under heavy research now, but still far from finding correlations between velocity fields and induced/vortex drag.
And for a wing induced/vortex component is about 90% of total drag.

Using LES models one can achieve more accuracy, but to have some decent results, it needs far more detailed meshes and timely resolution - and this means orders of magnitude more computation power needed.

I can imagine datasets usable to do LES will be HUGHE - few terabytes at least. With openwheelers you can't make any shortcuts really.

To accomodate this big dataset one have to use many separate computing nodes (typically 1 current node = 12 cores at 2,6GHz and 48-96GB of local memory), and during the simulation those nodes need to exchange lot of boundary data.

Computing clusters, even those with sufficient theoretical computing power, don't scale well when using big number of nodes for this type of tasks, due to limited interconnects bandwith (currently state of the art at 40Gbit/s using QDR Infiniband). At some point they achieve their peak and increasing number of nodes doesn't help anymore.

Under current budget cap teams are restricted in using their compute clusters for CFD, and can exchange some hours of CFD to wind tunnel time and vice versa (at a rate of 1 Wind tunnel hour = 1 compute hour at 40Tflop AFAIK).

[Lurk]

Toyota: 2006/2007
Honda: 2007/2008 (Both tunnels)
Renault: 2007
BMW Sauber: Late 2008/2009
Williams: 2006-2008
Red Bull/Toro Rosso: 2006/2007
McLaren: 2009

So with the law of averages, its Ferraris turn, last time this happened for them was early 2005.

Domenicalli just says to Autosprint they have a structural problem with their wind tunnel

=D>

[hardingfv32]

PNSD - Wind tunnel design

How do they get .0001% accuracy in a wind tunnel? I can barely fined a .001% force measuring device. Would this also mean the the airflow in the tunnel must be maintained at .0001% or at least measured to that accuracy? With total accuracy at .0001%, that would mean all the individual tolerances of the system must be lower. How is this possible?

Brian

[DaveW]

I have long been away from a close interest in steady state aerodynamics, so this post is intended to elicit information, rather than to supply it.

I suspect PNSD may have been referring to W/T measurement accuracy & repeatability, rather than "representational" accuracy (for want of a better word). I recall dimly that moving engines to the tail area (DC9, 727, VC10, BAC111, Trident, etc.) had a major impact on aircraft drag estimation. If I recall correctly, the closest estimate was around 13 percent in error. I understand that all tunnels require experimental "correction factors" to account for wall constraints, and "scale" errors, & quite small deviations from the norm can require large changes to correction factors.

I also understand that most (if not all) CFD codes work with inviscid flow equations & employ relatively crude approximations to account for viscous flow effects. If that is the case, then it is not surprising (perhaps) that CFD can sometimes yield quite misleading absolute results.

I would be happy for any of the above to be shot down in flames......