Discussion Closed This discussion was created more than 6 months ago and has been closed. To start a new discussion with a link back to this one, click here.
optimising solution methods in parametric sweeps. newton step suddenly going wrong
Posted 20 août 2014, 06:19 UTC−4 Computational Fluid Dynamics (CFD), Mesh, Studies & Solvers 1 Reply
Please login with a confirmed email address before reporting spam
I sometimes have my computations converging reasonably but then suddenly one step completely changes the picture and the error drastically increases, almost as if a calculation error had been made, probably something non-linear or a division by near-zero. See the images for an example of the convergence and how the solution step changes with a sudden peak in a small area of the domain.
Of course, I can solve this problem by changing the damping factor and the size of the mesh elements. However, when I am performing a sweep varying many parameters, these changes are an overkill. The problem with convergence only occurs in a few cases, for instance with very high velocities.
- Are there solutions to have the solution method adapt during a parametric sweep? I don't see these options but maybe some people have ways to tweak it.
- Also are there ways to have the solver realise earlier that it is going completely wrong? Like in the attached image I see clearly that it is wrong but the solver doesn't know yet, are there ways to forumulate this and let the solver work with it? I would be happy with all the rest of the parametric sweep and just have these problematic cases discarded. In my current case the solver gets stuck in these cases, taking much time to continue the calculation. With the result that after a weekend of computation I see my model is not finished because it got stuck in one particular case (I know of the option to continue if an error occurs but sometimes the solver takes a lot time before it realizes the error).
Of course, I can solve this problem by changing the damping factor and the size of the mesh elements. However, when I am performing a sweep varying many parameters, these changes are an overkill. The problem with convergence only occurs in a few cases, for instance with very high velocities.
- Are there solutions to have the solution method adapt during a parametric sweep? I don't see these options but maybe some people have ways to tweak it.
- Also are there ways to have the solver realise earlier that it is going completely wrong? Like in the attached image I see clearly that it is wrong but the solver doesn't know yet, are there ways to forumulate this and let the solver work with it? I would be happy with all the rest of the parametric sweep and just have these problematic cases discarded. In my current case the solver gets stuck in these cases, taking much time to continue the calculation. With the result that after a weekend of computation I see my model is not finished because it got stuck in one particular case (I know of the option to continue if an error occurs but sometimes the solver takes a lot time before it realizes the error).
Attachments:
1 Reply Last Post 20 août 2014, 08:28 UTC−4
Please login with a confirmed email address before reporting spam
Posted:
1 decade ago
These are the solution and convergence with a finer mesh. I believe the error with convergence in a larger mesh occurs because some of the small scale features are not well resolved. But it takes almost twice longer to calculate (about 10 minutes). Say I whish to do 12 x 4 x 4 x 4 x 3 of these calculations in a parametric sweep (16 days), then it is very usefull if I could use a coarser mesh in some of them.
Attachments: