[ghorrocks]

If the Smagorinsky SGS model is used in LES with a mesh which is grossly too large then the SGS viscosity goes very high (the SGS viscosity is proportional to mesh length scale squared see eqn 2-161 in the theory manual) and is therefore highly dissipative.

This means fast, easy convergence to a solution which is completely wrong.

So Sakura's comment that RSM is harder to converge and slower than LES is further evidence that the mesh size he is using is far too large for LES.

[sakurabogoda]

Dear Glenn and OJ,
Really really thankful for helping.
I got what you meant. Yes, I can understand there is a problem with my mesh.
If RSM takes more time than LES is impossible, because it is two times slower than LES now.
I'll check mesh again.

[sakurabogoda]

Hi Glenn,
Of course I am using Smagorinsky LES.

If that so, the problem definitely with the mesh. Thank you, thank you very much for help me to find the wrong point, as I have very limited time to complete this.

[sakurabogoda]

Hi All,
I have done mesh sensitivity analysis for 5 mesh types and considered a significant point rms velocity.
M1(elements:2500000): 6.474m/s
M2(elements:3250000): 6.501m/s
M3(elements:5300000): 6.511m/s
M4(elements:6450000): 6.443m/s
M5(elements:9700000): 6.489m/s
Although there is no much variation, I used M5 for RSM as M1 is the mesh size that gave the wrong results.

Meanwhile I ran LES with mesh size M5, it runs with smaller time step than RSM, but still I can see it runs faster than RMS. (no. of time steps per a given time). Can you please tell me is this ok?

I am using adaptive time steps now and can also see LES reaches convergence quicker than RSM.

[oj.bulmer]

As Glenn suggests, the faster convergence is probably attributed to the numerical dissipation. LES can't be trusted unless you use adequate grid size and timescales, and the "mesh independence study" you are doing doesn't seem to be an appropriate way to conclude, without taking into account if adequate lengthscales are resolved.

You should resolve the scales only to the extent that below it, the turbulence is isotropic, which is handled well by the SGS models. Often this lengthscale is considered to be close to Taylor's lengthscale. The accuracy lost in modelling the anisotropic turbulence by SGS models, when you use larger than appropriate grid size, will be decided by your physics.

Situation is aggravated when you have wall bounded flows, where the grid size requirements are still critical!

OJ

[sakurabogoda]

Dear OJ,
Thanks a lot.
I can see the mesh is very fine, but as you said, still it is not appropriate for LES. Then, how can I check if adequate lengthscales are resolved?
The other problem is, by this way, are results from RSM also incorrect? It takes more computational cost, like to run 0.5S, uses 5days.

[ghorrocks]

The key check for an LES model that your mesh and timescales are adequately resolved is to plot the turbulence power spectrum (intensity versus frequency, eg http://theeternaluniverse.blogspot.c...ber-space.html). If you get the -5/3 decay in the spectrum down to your filtering size then you know you are on the right track.

[sakurabogoda]

Dear Glenn,
Thanks a lot.
But again I have a problem. Is there anyway to monitor Turbulence intensity(I) in CFX Pre?
I=u'/U
u'=(2k/3)^0.5 ----- k-kinetic energy relates to rms velocities
So how can I monitor rms velocities? I have already used them in trn statistics.
I can calculate U(mean velocity) by monitoring u,v,w velocity components.

Thanks.

[ghorrocks]

You have to define all these variables as CEL expressions. The tricky one is the u' terms - are you doing spatial or temporal averaging to obtain them? Neither of these are easy to implement in CEL.

[sakurabogoda]

Dear Glenn,
I am thinking about temporal averaging velocities.

Another issue is when use expressions in monitor tab, I cannot specify a location.

I heard TI can be calculated by FFT.

[ghorrocks]

CEL cannot calculate anything in time, so you are either going to have to do this in post processing or write your own user fortran to do it (and that will not be easy).

When I did this previously in CFX I set monitor points at the locations I wanted the turbulence intensity at and wrote the UVW velocity components to monitor points. Then you can externally process the monitor point data to get your u'v'w' parameters. Then you do FFT on the u'v'w' to get the turbulence intensity spectrum.

[oj.bulmer]

The other problem is, by this way, are results from RSM also incorrect?

I think RSM is tuned to handle anisotropic turbulence far better than SGS models. So if your grid is not small enough to capture the anisotropy-range lengthscales, they are passed on to SGS models and I am not sure if these are robust enough to deal with it, as they are tuned for isotropic turbulence.

OJ

[ghorrocks]

The SGS models in LES are generally based on the assumption (which is generally a good assumption) that all the anisotropy in turbulence resides in the largest turbulence scales and the small scale turbuelnce is isotropic. In LES the large scale turbulence is modelled by the solver and the small scale stuff is modelled by the SGS, so an isotropic SGS is adequate for LES.

But for this assumption to be correct you need to show that your mesh and filtering is in fact correctly set up for this distinction.

[sakurabogoda]

I have data of u,v,w and can do the fft. but my problem is how to get the frequency?

in matlab we can write write code,
N=--; %no. of data points (2^n value)
T=--; %sample length
fc=(0:N)/T; %sample frequency

but my problem is I have omitted first half of data, so sampling time does not start from 0S, it has a value (t).

then, fc should be,
fc=(t:N)/T; ????? Am I correct?

Thanks.