Dear Friends

I am doing validation for a simple set up in terms of temperature. What is the usual accuracy of CFD models when compared with experiments.

Best regards

The accuracy of a CFD simulation is the difference between the results it gives and the actual values!! I'm guessing this doesn't help you much! The accuracy can vary depending on a large number of factors - mesh, accuracy of boundary conditions, advection scheme etc. First thing to make sure of when comparing to experimental data is that your boundary conditions match those of the real case - any approximations will affect accuracy. This might involve setting up profile inlet boundaries, setting the correct turbulence levels at inlets etc. Next is to make sure your mesh is OK - you should do a mesh dependancy study, which basically means refining your mesh until your solution doesn't change between successive refinements. Try to read "Quantification of Uncertainty in CFD", Annual Review of Fluid Mechanics, P.J.Roache, 1997, for a discussion on this issue. A higher order advection scheme should reduce the error IF your mesh is fine enough (use 2nd Order High Resolution in CFX-5). The turbulence model choice will also affect things. So, basically there's no way to say how accurate CFD is in general. If your boundary conditions are perfect, and everything else is set up correctly, then you may expect an accuracy within a few percent. However, this isn't always the case.

How accuarate are your experimentental results ! always remember there are also many errors in experimental work.

freely quoted- "Nobody trust the results from a CFD calculation except the person that has done them. Everybody believe the results from an experiment, except the person that have carried them out"

"A higher order advection scheme should reduce the error IF your mesh is fine enough"

A higher order advection scheme will always provide more accurate results than 1st order upwind. Use 2nd Order High Resolution or a Specified Blend Factor of .75 up to 1 (1 is fully second order).

Robin

I don't agree the a higher order scheme will ALWAYS give you more accurate results than first order. For example, a Specify Blend Factor of 1 may well produce over/undershoots in areas of high gradients. I wouldn't consider this more accurate than first order in this case. Also second order accuracy means that the truncation error is proportional to the square of the mesh spacing - so this means if you half your mesh spacing the truncation error would be reduced by a factor of 4. However, it also means that if you double your mesh spacing the truncation error increases by a factor of 4. On a very course mesh (when the mesh spacing is greater than unity) a higher order scheme may be less acurate. Having said all that, in the vast majority of cases using the 2nd Order High Resolution scheme will be more accurate - this scheme also prevents the over/undershoots.

Sorry Mike, but I cannot disagree more.

A second order scheme may produce overshoots or undershoots, but a 1st order UDS scheme will completely smere the gradient. The global solution will still be better with 2nd order.

As for error, you are right to say that the error increases by a factor of 4 when increasing the mesh size. I suppose if one were to produce a horendously inappropriate mesh, there is the possibility that it could result in magnifying the error. However, the solution would be so far off in that case that it wouldn't really matter what advection scheme they used. For practical meshes, a first order scheme would require significantly smaller mesh spacing to acheive the accuracy of the second order scheme.

Now that unstructured mixed element meshes are common, it is even more important to run 2nd order schemes. The flow will never be aligned to the grid in a tet mesh, thus the level of truncation will be very high for a 1st order scheme, resulting in significant numerical diffusion. For these meshes it is absolutely necessary to have the higher order terms.

To quote Ferziger and Peric from Computational Methods for Fluid Dynamics: "First order UDS is inaccurate and should not be used."

Out of curiosity, how many of you reading this are still running 1st Order Upwind?

Robin

Robin, First I must say that I agree with you, but I also have the experience that it can be more difficult to obtain a converged solution with 2nd order schemes. So I am wondering what is the most accurate solution, 1) a well converged solution with a 1st order scheme 2) a poor converged solution with a 2nd order scheme On a fine mesh I would prefer 1 I must admit that I still some time use 1st order scheme, but I do not use unstructured mixed element meshes and all my cells are alligned with flow so it is not so bad. And do remember the largest error in any cfd solution is placed ca. 30cm from the screen.

I agree with all your points but I don't see what you disagree with? UDS is less accurate and everybody should be using higher resolution schemes in general - but that doesn't mean they are unconditionally more accurate. Mike

Hi Jan,

I would suggest being very critical of any UDS solution or any unconverged solution. If you cannot converge with a fully second order scheme, try specifying a blend factor or using the 2nd Order High Res scheme.

As for which to choose from, well there are a lot of reasons why a solution will not converge, some of which may be tolerable, some not. When it comes to UDS, however, there are very clear and documented reasons why UDS is bad. If you are in doubt, contact technical support.

Ultimately I must agree with you on the last point; regardless of convergence, discretization and all other factors, no level of technical support or product development can make up for problems in the wetware, but that is precisely what discussions like these are aimed to fix.

Robin

That's a good point. Sometimes it is indeed quite wrong not to question the experimental data. In my opinion the CFD engineer should in fact be quite involved in the data collection, or at least the data collection procedure, and be able to assess how accurate they are.