Menu

Discussion Forum

New Discussion
Filtra

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.

How to know if your model is going to fail during study?

Posted 30 set 2020, 19:26 GMT-4 General, Studies & Solvers, Equation-Based Modeling Version 5.5 2 Replies

Mohamad A
Send Private Message Flag post as spam

Please login with a confirmed email address before reporting spam

Realistically, how long should you wait before throwing in the towel?

I'm currently trying to model a 3D parallel plane plasma discharge, my dof is around ~130k after playing with the meshing parameters. Sometimes the model takes up to 4-6 hours on a 32 gb ram rig just to give me an error on convergence. Is there a sort of intuition behind knowing when to give up and cancel?

Can you tell if your model is going to fail by looking at the convergence plot? (I've attached mine below). The reciprocal step size oscillates downwards, but then shoots up after ~500 timesteps, from your experience is that a precursor to the model failing to converge or is that normal?

Attachments:

2 Replies Last Post 1 ott 2020, 01:10 GMT-4
Robert Koslover Certified Consultant
Send Private Message Flag post as spam

Please login with a confirmed email address before reporting spam

Posted: 4 years ago 30 set 2020, 23:08 GMT-4

Time domain models can be tricky in regard to convergence. You need to work on your solver settings. (There are a lot of settings and you may have to do quite a bit of experimenting.) You may also have to work on your meshes. Higher quality meshes are better. You should follow along in the "log" window and see if errors are piling up at each or any step. If you have enough memory, consider using direct instead of indirect solvers. You may also have trouble-making boundary conditions or a time-domain excitation that is too sudden somewhere and is not properly being time- or space- resolved, so consider changing that. Sometimes enforcing "strict" time steps is helpful in convergence, but can still cause a long time for execution to occur. And if it would make sense at all, try a 2D model first. There are many many possible causes of your problems. Consider posting your model to the forum for others to look at.

-------------------
Scientific Applications & Research Associates (SARA) Inc.
www.comsol.com/partners-consultants/certified-consultants/sara
Time domain models can be tricky in regard to convergence. You need to work on your solver settings. (There are a lot of settings and you may have to do quite a bit of experimenting.) You may also have to work on your meshes. Higher quality meshes are better. You should follow along in the "log" window and see if errors are piling up at each or any step. If you have enough memory, consider using direct instead of indirect solvers. You may also have trouble-making boundary conditions or a time-domain excitation that is too sudden somewhere and is not properly being time- or space- resolved, so consider changing that. Sometimes enforcing "strict" time steps is helpful in convergence, but can still cause a long time for execution to occur. And if it would make sense at all, try a 2D model first. There are many many possible causes of your problems. Consider posting your model to the forum for others to look at.
Mohamad A
Send Private Message Flag post as spam

Please login with a confirmed email address before reporting spam

Posted: 4 years ago 1 ott 2020, 01:10 GMT-4

Hi Robert,

My model is heavily influenced by the 2D positive column application provided by COMSOL (You'll see a lot of the same reactions and data used here as well). What I'm trying to do is transform that 2D positive column analysis to a 3D parallel-plate analysis. The 2D axisymmetric model that was in the application gallery had ~14k dof and took 20 seconds to compute while my model has 130k dof.

I totally agree with what you're saying, which is why I asked whether there was a faster way to go about this (Can't spend 3-5 hours testing out a single change in solver configurations). I suspect that my model is still flooded with minor errors that could contribute to this, but I don't want to spend a lot of time computing it just to find out. (Maybe starting off with a 2D model would be better here). My model is also at the preliminary stages, so for the most part, I'm fine with reasonable error margins, it's just a matter of confirming that the physics is valid.

For now I'll take your and Apotik's advice of starting off with 2D and see where I go from there.

Please find attached my model, Thanks

Hi Robert, My model is heavily influenced by the [2D positive column](https://www.comsol.ch/model/dc-glow-discharge-8640) application provided by COMSOL (You'll see a lot of the same reactions and data used here as well). What I'm trying to do is transform that 2D positive column analysis to a 3D parallel-plate analysis. The 2D axisymmetric model that was in the application gallery had ~14k dof and took 20 seconds to compute while my model has 130k dof. I totally agree with what you're saying, which is why I asked whether there was a faster way to go about this (Can't spend 3-5 hours testing out a single change in solver configurations). I suspect that my model is still flooded with minor errors that could contribute to this, but I don't want to spend a lot of time computing it just to find out. (Maybe starting off with a 2D model would be better here). My model is also at the preliminary stages, so for the most part, I'm fine with reasonable error margins, it's just a matter of confirming that the physics is valid. For now I'll take your and Apotik's advice of starting off with 2D and see where I go from there. Please find attached my model, Thanks

Attachments:

Note that while COMSOL employees may participate in the discussion forum, COMSOL® software users who are on-subscription should submit their questions via the Support Center for a more comprehensive response from the Technical Support team.

Suggested Content