Convergence, solver selection questions


#1

After some tests I would like your opinion on my preliminary conclusion ideas.
I did several simulations with different flows, from initial speed v0, to v0 x10.

  • In several cases simulation seems to converge and stabilizes with residuals 1E-4 or less, but blows up in just some steps., can the solution be used before the blow up? If the residuals are low I dont´t see why not…

  • I found more difficult to achieve convergence, or just low residuals, with increasing flow, why could be the reason?

  • I found from some suggestions, googling, and trial/error this very basic setup to get stabilized convergence, or at least to get to low residuals (1E-4 or less), although with enough steps, (1000 or so) a lot of times seems to suddenly (in just 4-5 steps) blow up
    1- Reduce relaxing factors some 20-30%
    2- Change default solvers to pressure GAMG, rest (Smoothsolver)
    3- Reduce relaxing factors some additional 20-30%
    4- Change default solvers to all (Smoothsolver), slows very much the convergence

I read regarding the inner working of the SIMPLE algorithm but, Is there any additional info regarding the solvers and their detailed parameter setup, in order to improve convergence or selecting the most suitable for any kind of problem?

Thanks for your comments


#2

Hi @rarrese,

it would be great to know more about the application you are simulating and the mesh you are using to provide a more tailored answer. But maybe some first hints:

  • What turbulence model are you using? Did you check on the y+ values of your mesh? Max and average y+ can be calculated quite simple via a result control item which gives you more info if your mesh is well suited for the simulation. What are your y+ values right now (avg. and max)?
  • Did you change the tolerance values of the linear system solvers? You could try to lower them, especially for the turbulence quantities and the velocity. I tend to use GAMG for pressure but stick to PBiCG for the other quantities - from my experience the smoothSolver is more robust but slower.
  • Increasing the velocity gives you a higher Reynolds number flow which can have different implications. Also there might appear transient flow characteristics that are not well resolved by a steady-state solver, but it’s hard to say without knowing more about the sim

For these blow ups, any chance you could provide a screenshot? I think @jprobst lately looked at something similar, maybe he has a good idea on that.

Hope that provides some initial help!

Happy simulating,

David


#3

Hi,
Just general questions,

  • what are the consequences of having a too high(>300) of too low (<30) value of y+ in k-e model? Regarding convergence and accuracy.
  • when you suggest increase the tolerance of solvers, do you mean to tight it or to loose it?
    Thanks