Fully developed pipe flow - different turbulence models
 

Dear all,

I try to compare measurements of fully developed pipe flow with different low-Re turbulence models (from Fluent).
So I found some nice data here:
http://www.princeton.edu/mae/people/...e-data/mckeon/
and used the Re=74343 data set.

In my simulation I just took a short piece of pipe with periodic boundary conditions. I normalized all data, that it becomes U=1 in the middle of the pipe.
This is what I get:
Attachment 37312

To be honest: I thought that these models work much better for such a simple "standard" problem. What do you think? Is this the best I can get, or am I probably doing something wrong?

I suggest using the (U+,r+) representation in log scale so that the difference will be better highlighted.
Anyway, the geometry of the test-case is simple but the generated turbulent flow is quite complex and RANS and similar models can fail, too.
How about the grid? I think your Re_tau is O(10^3), rhight?

Filippo,

Thanks for the quick answer. The reason I did not use u+ for plotting is that I thought, if the wall shear stress is different between the cases i will get different plots for u+. Even if the curves look the same in real values. For me (in this case), the shear stress is not important, but just the shape of the velocity curve. That's why I plotted everything from 0 to 1. What do you think, does that make sense?

I just made some fast o-grid, y+=0.5, there should be way enough points for RANS...
Attachment 37367

Subsequently, I want to compare laser-measurements of velocity profiles in pipes with different barriers and curves in the pipe to simulations. That's why I wanted to know which of the turbulence models represents the fully developed flow the best.

try a better grid, put 3-4 nodes within y+<1

Hi,
I made a ridiculously fine mesh with y+=0.05 and the coarse mesh profile perfectly matches the fine mesh profile.
I am not sure whether I do something wrong by i) normalization, ii) numerics, iii) expecting too much from RANS models...

first, if you want to assess what a turbulence model really do, nothing better than a simulation on the same grid without any model!
This is very common in LES studies.
Try if your case remains numerically stable without RANS model, perhaps you could have problema in getting a steady case without modelling therefore you have to do a time average of the oscillating solution

Good news. After thinking a little longer :cool: I realized that normalizing the velocity to the maximum value (thus to a maximum of "1") is not a good idea.
It distorts of course the whole profile, if the maximum value isn't accurate. Also, if two curves match exactly except that one point, the two curves would look pretty different. I think this was the problem here, too.

Now, I normalize to the mean velocity (from calculation of pipe Reynolds number) and it looks much better. Only the center of the pipe is not reproduced by the models.

Take a look. This is a quite large comparison in x-log scale:
Attachment 37419
and without log-scale:
Attachment 37420

Then, I finally found out that Fluent offers different low-Re k-epsilon models, which I didn't know:
Attachment 37422

And the winner is Yang-Shih k-epsilon model:
Attachment 37421

What do you think?

some observations:

1) too many curves in the same log-plot... I was unable to see some details

2) I see some solutions having no nodes in the viscous sub-layer, I suppose they use wall-model and should be considered in a different comparison

3) Even if the log plot contains too curves, I see some solution that do not obey to the log law but show a different slope

4) The "winner" is a wall-model based, right?

5) usually the velocity profile is just a zero-order statistics and says very few about the quality of the solution. On the other hand, RANS can only provide zero-order statistics and I dont know if you can extract other useful curves

1) Ok, I just wanted to show that they look quite better than in the original post.

2) They all have the same grid, with low-Re boundary conditions. The openFoam plots are just sampled differently than in Fluent, so they look differently. All of them have an y+ of about 0.5. I tryed your advice to use more gridpoints here, but didn't see any difference in the solution so I sticked to the coarser grid.

4) see 2)
5) Yes that's a good hint, but we are only interested in getting accurate velocity profiles, so "U" is the first thing a have a look at. If I can get the profile correctly, I actually don't care if "k" and "epsilon" are completly unphysical ;)

I am very interested for flow in pipe and more precisely the bubble flow and turbulence walls are you simulated a bubbly two phase flow in pipe and are you find a wall peak of void fraction gaz near walls