open TELEMAC-MASCARET The mathematically superior suite of solvers

Hi,

Recently I got the access to a cluster so I'm able to test the stability of our 3D model. I use the k-epsilon model for turbulence and the numerical schemes I use is as follows:

SCHEME FOR ADVECTION OF K-EPSILON = 2
SOLVER FOR DIFFUSION OF K-EPSILON = 7
OPTION OF SOLVER FOR DIFFUSION OF K-EPSILON = 6
PRECONDITIONING FOR DIFFUSION OF K-EPSILON = 2

However, the computation can continue without interruption but the RELATIVE PRECISION for epsilon is NaN at some timestep. Is this a error that means the solver cannot solve the equation or just means the true value is zero?

Thanks in advance!
Qilong

Hello,

No, it is really a problem, but beware that GMRES does not like easy problems, it is unable to invert a diagonal, so it would be safer to check with solvers 3 or 4 instead of 7. Another possibility is to use characteristics on k and epsilon instead of SUPG, and then you can have the conjugate gradient as solver.

Another thing to look at is the minimum epsilon accepted in the computation, set in CSTKEP, that you can change.

Regards,

Jean-Michel Hervouet

Thanks for your suggestions!

I tried to use solvers 3 and 4 with the SCHEME FOR ADVECTION OF K-EPSILON=2, but none of them worked, giving a error read
EXCEEDING MAXIMUM ITERATIONS: 107 RELATIVE PRECISION: NaN
(107 is the maximum iteration number I give to diffusion of k-epsilon).

I also tried different solvers with the characteristics scheme on k and epsilon but none of them worked. The error is similar to the above one or the error I mentioned in the previous post.

I changed the minimum epsilon too and the problem remains.

I have k-epsilon working in a simple model (a rectangular tank with flat bottom) so I guess it may be caused by the complex bathymetry in the estuary or tidal flats or model initialization. Is there anything I can try in my model?

Regards,
Qilong

Hello,

Well you can play with the many k-epsilon options in cstkep.f but probably it occurs on tidal flats and there is a division by 0 somewhere but we would need your files to see what happens in a debug mode that stops at the first NaN.

Regards,

Jean-Michel Hervouet

Thanks! I'll prepare the model because the boundary data is too large so I have to extract part of it. The error occurs at the beginning so I will make a smaller version for uploading to the forum.

Regards,
Qilong

Hello,

I make a smaller version of my model and it can be downloaded at
Model Download

I also checked the info in the debug model and the text file is attached.

It seems that the subroutine CVDF3D has difficulties dealing with epsilon.

Regards,
Qilong

Hello,

Got it! Please expect some delay before an answer.

Regards,

JMH

No problem!Please take your time and many thanks!

Regards,
Qilong

Hello,

The diagnostic is that in kepcl3.f there is a division by DIST=0 at line 171, so a provisional bypass would consist of adding after line 168 :

DIST=MAX(1.D-5,DIST)

however this DIST=0 should not happen here and the real reason is certainly that array IPBOT is not consistent with array ZPROP which is used in KEPCL3, so this would be rather a mistake in telemac3d.f, and I still have to find where... tomorrow.

Regards,

Jean-Michel Hervouet

Thanks a lot!
Now we have a way to bypass this problem. Looking forward to your final solution!

Regards,
Qilong