[CSMDakota]

Hi All,

I could use some feedback on a LES case I'm trying to run in SU2 v7.3.1.

I've been bencharking SU2 against the Sajeben Transonic Diffuser case: https:https://www.grc.nasa.gov/WWW/wind/va.../transdif.html.
I ran it as a 3D problem with a finite channel width as I had a vision for eventually running LES.

• For the weak shock case, the shock position is stable, and the RANS results are good.
• For the strong shock case, the shock position oscillates at 300 Hz according to Robinet & Casalis https://arc.aiaa.org/doi/10.2514/2.755 , and running as a uRANS case in SU2 I was able to achieve 300 Hz as well.

I decided to extend my study of the strong-shock case by trying LES; not so much to gain more insight into the problem but rather to use as a learning case as I've not ran LES before.

There are two issues I see:

1.) Through several attempts my solutions diverge once the eddy structures reach the outlet; a small zone of high Mach number flow develops, and the otherwise well-behaved setup soon diverges. I tried an extended diffuser length to give the eddies more opportunity to dissipate but ran into the same behavior at the outlet.
2.) The shock structures just downstream of the throat, which are resolved from the RANS initialization, seem to break down and migrate upstream as the LES solution progresses; I suspect this is non-physical since the pressure and velocity profiles from the RANS cases were pretty accurate.

The *.cfg is attached, but a few points about my setup:

• The mesh is structured hex and resolved to the wall. Nominal cell size is 0.6 mm; standard diffuser has ≈ 6e6 cell count, 10e6 for the extended diffuser, and both were 20 cells in spanwise direction. I’m quasi-confident the mesh is fine enough to resolve a decent portion of the turbulent energy cascade but may not be optimal, yet, e.g. perhaps there is still a significant portion is still reliant on the sub-grid scale. I suspect this may play a role with issue 2) moreso than issue 1).
• The channel side boundaries are periodic for LES instead of symmetry for RANS.
• The time marching is DUAL_TIME_STEPPING with dt = 1e-6 seconds, up to 30 inner iterations unless RMS density reaches 1e-8 (which is does for most of the time steps, some expire at 30, but residual behavior is quite good regardless), and CFL locked at 1 for the inner steady iterations.
• Using the WALE sub-grid model

There is also an image of the of the solution at crash for the extended diffuser case, showing the flow field and a zoomed-in section where the divergence occurs near the upper region near at the outlet.
The solution animation can be viewed here: https://www.youtube.com/watch?v=oKbh7cmLm4A

My questions are:

a.) is there something wrong my setup strategy for this case?
b.) is the outlet boundary known to be robust for LES cases with eddies driving local recirculation regions?
c.) should I add more numerical dissipation by switching to JST or a different upwind scheme (vs. SLAU2)?
d.) is this an unreasonably high Reynolds number case upon which to start learning LES? 😊

Otherwise, SU2 continues to demonstrate solid results for RANS cases I've played with and is easy to use; Open Source CFD for the masses makes the world a better place <3

Thanks,
Brandon

[SahaSK]

Hello Brandon,

I have been trying to implement LES for a transonic wing in a FSI case, but the solution is diverging. I am wondering if you have been able to solve your issues or not.

Thanks, Saha

[CSMDakota]

Hello, Saha,

No, I never got that simulation to solve satifactorily as a scale-resolved transient case. Are you having similar issues at an outlet/farfield boundary, or can you tell where in your domain the solution is diverging?

Regards,
Brandon

[SahaSK]

My LES case is diverging at the second time iteration with an error of FGMRES orthogonalization failed (no matter how fined my mesh is).

[CSMDakota]

Sounds like you've experimented with the mesh density, but how about the mesh quality? A few bad cells can cause issues. Does it run OK as a RANS case, steady state and/or unsteady? You could try starting as a first-order run, smaller time step size, or both?

[SahaSK]

Thank you very much! There might be very few bad cells in the mesh, I will check that again. I was trying the unsteady case with the NS solver, 2nd order though.