Does under relaxation factor influence the accuracy of solution？
 

Hey everyone,
I'm confused whether the under relaxation factor influences the accuracy of solution.
I'm trying to solve a transient problem with Reynolds number=10000 using LES.
When I use SIMPLEC with default under relaxation factor, it needs more than 20 inner iterations per time step to converge with time step=0.0002s, and about 10 inner iterations with time step=0.0001s.
If I increase the under relaxation factor to pressure of 1 and momentum of 0.5, it needs only 5 inner iterations with time step=0.0002s and 1 inner iteration with time step=0.0001s.
There is no oscillation or divergence problem if I increase the URF. But the results with large URF and time step=0.0001s is not reasonable.
Does it means URF influences the accuracy of solution even no stable problem happens?
Is it reasonable to adjust the URF and time step to make every time step need 5-15 inner iterations to converge?

URFs don't influence the accuracy of your solution. But they have influence of the time until your solution is converged. The URFs are something like a damping factor. If your simulation for example gets unstable and starts to oscillate you could lower the URFs and so cut out the steep oscillations. But you should only change the URFs from default of there are convergence problems, for example in difficult cases with combustion.

Hope it helps. Somebody could confirm if I'm right? :D

Stable doesn't necessarily mean accurate.


You should tinker a lot with urf's and time-step size for any particular problem if you're trying to min/max. But you need to make sure it has appropriate accuracy. How are you judging convergence? You need more than just monitoring residuals.

Thanks for your reply.
For the problem of oscillation or divergence, URF should be reduced as you say.
But if the calcultion is always stable, URF can be improved to save calculation time in my opinion.
My question is about the strange change of simulated results with enlarged URF.

Thanks for your suggestion.
I use the default convergence criteria in FLUENT, the absolute criteria of 0.001 for residual.
Is there any better criteria to judge convergence?

You need to actually monitor the solution. Like make a point probe and look at the velocity and make sure that it converges.

Also a residual of 0.001 is way too high for time-resolved simulations like LES. You should be able to achieve 1e-05 or 1e-06 or so. You'll see this once you start looking at the actual solution.

I understand. I will try the method as you say. Thank you so much!