Hello ! I finally talked myself into making an account to contribute (and more likely, add to the pile of questions)
Quick background : I am a fairly new Linux user, which is less than a year. I find myself struggling when I see how experienced folks handle their problems on the distro and softs. Add this line there, recompile, huh ? recompile what and how ? That is how bad it is and hopefully it cannot get any worse.
As for the EDF FEM/CFD software ecosystem, I am 6 months in. Prior to that my experience was only on commercial softs + windows. They do handle quite differently. Big thanks to the EDF teams for their amazing work at putting together an open source simulation ecosystem of professional quality.
Back to the problems. I use Salome CFD and mostly work with mixes of radiative/conductive heat transfers on SYRTHES. Runs smooth when the geometry is simple but rapidly go south when the number of interacting faces grows.
What puzzles me is that I have no hint on why it diverges. It can start smoothly, easily converging the first steps and then throws NaN all over the place. I will attach an example of one of my listing files.
I tried to play with timestepping, reducing the dt to get more steps converging but overall it seems useless. About that, is there no way to automatically rescale the dt and restart a step when divergence appears ?
Most likely these problems are mesh related so I tried remeshing a lot. But then I don't know how to get information about where things are failing, I also don't fully understand SYRTHES behavior. Instead of a refined radiation mesh, SYRTHES likes it as coarse as it can get.
So as a conclusion, I would love to know if you have some pointers to address the above mentioned issues when models get larger. Thank you in advance if I eventually get some answers. You can find some of my files attached to this post.
Radiation convergence
Radiation convergence
 Attachments

 syrthes_rad.tar.gz
 (397.04 KiB) Downloaded 7 times

 Posts: 2638
 Joined: Mon Feb 20, 2012 3:25 pm
Re: Radiation convergence
Hello,
Note that regarding Syrthes installation, an up to date/cleanedup install version (Python3/PyQT5) is available on this site, in the download/prerequisites version. It basically matches the version in Salome_CFD 9.3, so numerics have not changed, but there might be a bug fix or 2, and installing the GUI should be less of a pain.
As for the radiation module from Syrthes, I have not used it since Syrthes 3.4 (before parallelization), so do not know too much about its behavior, but I know it uses a visibility factor logic, leading to dense matrices, which is why meshes need to be coarse, as data size grows very fast (as nbf^2 where nbf is the number of boundary faces; for a cube mesh with n elements on each axis, as the nbf evolves as n^2, the general behavior is n^4).
The semitransparent radiative model in Code_Saturne may be less precise (it uses a discrete ordinates model with n directions), but possibly better adapted to large cases (grows as n^3 in 3D, as it uses the regular volume mesh). Relevance of one of the other depends on what you need to do.
I'm not sure this is useful information to you, but it is the best I have so far. I know the Syrthes team is working on a new version, but have no idea when it will be released.
Best regards,
Yvan
Note that regarding Syrthes installation, an up to date/cleanedup install version (Python3/PyQT5) is available on this site, in the download/prerequisites version. It basically matches the version in Salome_CFD 9.3, so numerics have not changed, but there might be a bug fix or 2, and installing the GUI should be less of a pain.
As for the radiation module from Syrthes, I have not used it since Syrthes 3.4 (before parallelization), so do not know too much about its behavior, but I know it uses a visibility factor logic, leading to dense matrices, which is why meshes need to be coarse, as data size grows very fast (as nbf^2 where nbf is the number of boundary faces; for a cube mesh with n elements on each axis, as the nbf evolves as n^2, the general behavior is n^4).
The semitransparent radiative model in Code_Saturne may be less precise (it uses a discrete ordinates model with n directions), but possibly better adapted to large cases (grows as n^3 in 3D, as it uses the regular volume mesh). Relevance of one of the other depends on what you need to do.
I'm not sure this is useful information to you, but it is the best I have so far. I know the Syrthes team is working on a new version, but have no idea when it will be released.
Best regards,
Yvan
Re: Radiation convergence
Thank you for your prompt answer
I will try to see what I can get from the up to date version of SYRTHES and the Code_Saturne radiative model.
I did understand that the matrices grew up pretty fast with the number of boundary faces and elements by observing the view factors calculation times, but I still don't understand why it leads to divergence. Statistically it increases the chances of having one of your discretized subdomain to go bad because the number of parameters skyrockets, but it should also make the energy minimization process smoother thanks to more interpolation points.
Or, radiation models works in a completely different way and I should get back to dig in the theory haha
Anyway, I did not know the growth was n^4 correlated, that certainly put strain on the solver and is probably the main issue here. Thanks for pointing it out !
Have a nice day
I will try to see what I can get from the up to date version of SYRTHES and the Code_Saturne radiative model.
I did understand that the matrices grew up pretty fast with the number of boundary faces and elements by observing the view factors calculation times, but I still don't understand why it leads to divergence. Statistically it increases the chances of having one of your discretized subdomain to go bad because the number of parameters skyrockets, but it should also make the energy minimization process smoother thanks to more interpolation points.
Or, radiation models works in a completely different way and I should get back to dig in the theory haha
Anyway, I did not know the growth was n^4 correlated, that certainly put strain on the solver and is probably the main issue here. Thanks for pointing it out !
Have a nice day