experiences with RSTMs?
I calculated a u-duct geometry using k-epsilon- and Launder-Gibson-Turbulence-Models. A little bit surprising was that the results for the k-epsilon model are often much closer to the experimental data than the RSTM simulation.
I thought that the RSTMs perform ways better in simulating flow variables of strong curvature than the EVMs and that therefore u-ducts are better dicribed by Reynolds-Stress-Models?
I really would be happy to hear abput your experiences concerning OpenFOAMs LaunderGibsonRSTM - what are your findings?
Thanks pretty much
Perhaps it's also just a stupid mistake in my initial and boundary conditions:
- converged k-epsilon simulation
- R generated with OpenFOAM post-processing tool 'R' and rearranged boundary field
- simpleFoam solver
- under-relaxation: p=0.2, other parameters=0.5
- standard fvSchemes like in simpleFoam tutorial cases
I have also studied the Uduct with linear and RSM models. In my case the RSM's were clearly superior to the linear models like LS-k-eps.
I should mention that I have done this study in Fluent and not OF.
My boundary conditions are a little bit different. I run a precursor channel flow with the RSM and LS model. I then specified the solution of the channel flow, which has been matched to the experimental data, data as inflow boundary conditions for the Uduct.
Thanks for the reply Markus
I began to wonder what kind of dicretication scheme should be used for a problem like this...
You using fluent - anyway it would be nice to hear how you had choosen the discretication schemes?
Thanks pretty much
A couple of days later I have collected a lot of experiences but I'm still searching for the optimal settings for my u-bend calculations.
It turned out that using a 2nd order discretication scheme for the velocity brings the numerics much closer to the experimental data.
What me really surprised was the influence of the 1st cell size at the walls. A variation of it has a huge influence for the WHOLE cross section of the channel! I thought - the RSTM are running with wall-functions therefore I have to be out of the viscous sublayer and that's it... No - the setting for these 1st cell has a huge impact for the whole calculation. I ran a little study of this influence with the hope to figure out the best near wall cell size. Sadly the outcome is really ambivalent.
I simulated settings for an u-duct geometry like it is published in 'Case06' in the ERCOFTAC Classic Database.
Perhaps you have collected some experiences with this topic? I would be really glad to hear them.
Thanks - Sven
> Perhaps you have collected some experiences with this topic?
What is the objective of your exercise?
In boundary layers pretty much everything scales off the wall shear stress and so changes to it would be expected to change the main flow.
In the cell next to the wall the gradients in a number of quantities are large and unresolved. Wall functions provide models for some of these gradients as a function of the values of solution variables around the wall. Wall functions can be implemented in a number of ways which give slightly different answers.
The use of wall functions normally precludes getting grid independent solutions. One way to get a feel for your sensitivity to wall functions is to calculate a fully developed flow (i.e. 1D variation) by:
- normal grid refinement everywhere with first cell in valid range (only a few valid grid sizes but enough to see the trend)
- keeping the size of the first cell fixed and progressively refining the rest of the grid until numerical error negligible.
- changing the size of the first cell with a very fine grid everywhere else
RST models represent the convection/production term exactly and this is the dominant part of the physics for the large suppression/generation of turbulence due to curvature. The k-e model gets it wrong by rotating the modelled stress in sync with the rate of strain and so effectively ignoring the effect.
The rest of the terms in the RST models are pretty much curve fits to experimental data with only a little bit of recognisable modelled physics. In regions where these become relatively large but the flow is not similar to that used for the curve fits (in my day an example of this was impinging flows but people later reworked the form of the terms and curve fits to improve matters) then the predicted flow could become unrealizable.
Your flow is one for which an RST model would be expected to perform well so long as the coefficients and the terms were appropriate for this type of flow. I am guessing openfoam has implemented what I would have called the Gibson Launder model in which case it is effectively the simplest RST model but it will need a "wall surface integral" term in order to predict boundary layers in any semi-sensible way. (It will make the v' and w' components different to help drive your corner secondary flows). This needs a distance from the wall not only for the near wall cell but all the cells and some way of determining which is the closest wall or a way of summing the wall contributions. It may be a good place to start looking for implementation problems. However, there were later models that reworked the term using a turbulent Reynolds number to avoid this issue and so this may be a red herring. Perhaps that is why the authors have been reversed? Whatever, I do not know what has implemented in openfoam and so I guess you may have to look in the code to find out.
|All times are GMT -4. The time now is 16:00.|