[Yonadav]

I am trying to use the FULLMG_CYCLE option in order to initialize the flow field to values close to the solution even if not as accurate as fine grid.
The application always stops with inf from the first iteration.
No matter how many mg levels (even 1) or cfl value
With the same parameters but no multi grid or other multi grid algorithms are working but very slowly since large parts of the filed have to be updated.

Ideas?
Working example?

Thanks

[giovanni.medici]

Difficult to provide an answer without having more details on the problem you are solving, grid quality, and geometric features, Mach regime ...

Nevertheless some hints were provided in this post:

Multi-grid-initialisation-flowfield

[Yonadav]

Thanks for the reply.
I checked the thread you mentioned several weeks ago, he did not get an answer to his question as well.
He was answered about v cycle, v cycle or w cycle are totally different than FULLMG Cycle.
As I mentioned V cycle works for me but in my case I do not find it worth the extra computational cost and anyhow I need to initialize the flow field to a values somewhat closer to the solution than a uniform value.
FULLMG starts at the coarse grid and solves until convergence criteria achieved or maximum user defined number of iteration is reached, then it interpolates to the next finer grid and solves and vice versa up to the actual mesh which is the finessed grid.
Unlike v or w MG cycles FULLMG is expected to be more stable, the coarser the grid the solution has better convergence features with the price of losing high resolution features of the solution and not solving correctly the boundary layer.
However, we eventually reach a fine grid and solve all these features.

My problem is that FULLMG gives inf at the first iteration (when using FGMRES you gen non convergence as the mention thread, I am using BCGSTAB)

As an example use TestCases/rans/rae2822/turb_SA_RAE2822.cfg
modify the parameter MGCYCLE= FULLMG_CYCLE

It would diverge in the first iteration
I think it is a bug, but addressed the forum in case someone used FULLMG successfully.
I am using version 7.04 it happened in 6.X as well.
In another software which I use, a structured grid software, so generating meshes for complex cases is a problem, FULLMG works very well even in low quality grids.

The actual problem I am trying to solve is a motor plume so unlike an airfoil or a wing, in that case large parts of the solution field are updated to values significantly different then the initial value, which takes a lot of time and requires low CFL for the finest grig, therefore, FULLMG might be an excellent option for me.

Thanks

[pcg]

Hi,

I think the full cycle does have some issues, it is not used in any of our regressions tests and it is a relatively old feature (which makes it a good candidate for bugs).

You can try to emulate the behavior by playing with the number of PRE or POST smoothing iterations on the coarsest level.
Note however that MG is only used for the flow equations, turbulence and other scalars are never solved on the coarse grids, they are recursively interpolated from the finest grid. So if you have turbulence I doubt this initialization strategy is going to help you.

Unfortunately SU2 is not great for bootstraping the solution of complex cases from coarser grids, you could try using a lot of numerical dissipation to be able to run at high CFL and then reduce the dissipation to normal levels.

[giovanni.medici]

Yesterday I tried to run the rae2822 case with the original and a revised grid, with FULLMG_CYCLE enabled, without success.
The "closest" I got to have something running was iterating but always with the residuals to -Inf ( I added PRE and POST smoothing and correction), and the solution was (not surprisingly) not right.

Thanks Yonadav and pcg for spotting it out and explaining a little bit more what is the philosophy / under the hood.

[Yonadav]

Pedro many thanks for your quick reply
I thought it's a bug (inf on the first coarse grid iteration), but tried to address the forum, maybe I missed something.
As well as I know in V, W cycles no one solves the turbulence in the coarse grids, in the boundary layer the turbulence cannot be solved correctly in a coarse grid so it would only disrupt the solution.
In FULLMG there is no reason not to solve turbulence in the coarse grids even if the solution is not correct in the boundary layer it would be polished in the fine grid.
I actually converged to the solution of using high dissipation as an initializer before addressing the forum. FULLMG might be faster.
Shuld I open an issue in the SU2 github?

Thank you Giovanni for trying the problem as well.
RAE2822 is just an example for the bug, it is probably not the best candidate to gain much from FULLMG (it does gain from V or W cycles), but it should run without inf, NaN etc.

[pcg]

You can open an issue but I don't know anyone actively working on improving/fixing the MG. So it may take a while for someone to pick up the issue.