Discussion Forum

Jiuan,

What precisely do you mean by "quite different"? For both of the solvers you mentioned the accuracy is determined by the convergence criteria/tolerances you specify. If the solvers each determined a "converged" solution that are very different perhaps the convergence criteria are set differently?

Usually folks would consider trying a different solver when one was not able to arrive at converged solution. It's curious that you would "converge" to different solutions to the same FE problem.

Would you be able to upload your model for the community to see?

Best regards,
Josh Thomas
AltaSim Technologies

Hi Josh,

Thanks very much for your reply.

For my heat transfer case with surface-to-surface radiation, when I usee segregated or fully coupled solver, I used the default settings for convergence criteria, and converged resutls are obtaiend from both solvers.

Then I compare the temperature distribuiton in the same surface, I found that temperature difference in the same points can be 50K ( the value of temperature is around 1500K). And pattern of temperature contours are also very different.

Due to confidential reason, I am not able to upload the case. Sorry for that.

Jiuan

Jiuan,

Very interesting behavior. Maybe somehow the problem posed does not have a unique solution and both returned solutions satisfy the equations (although the solutions are different). One solver converged to one local solution and the other solver converged to a second solution.

If confidentiality is important for your project, the discussion forum probably is not your best avenue for support. I would recommend contacting support directly, or hiring a consultant with an NDA.

If you do discover why this behavior is happening, I would be interested in hearing back from you on this thread.

Best regards,
Josh Thomas
AltaSim Technologies

Hi All,

I still have doubt on the difference between segregated and fully coupled solver.

From my previous experience, normally fully coupled solver requires a large memory, and due to this reason, I have to select segregated solver.

But I have compared these two solver in some testing case, and got different converged result. Then I am not clear which solver is better for a certain case, for my case, it is a radiation heat transfer case. Is there any guidline to select the solvers?

Thanks.

Jiuan,

This is some very interesting behavior that you are converging to two different solutions. Would you be able to upload your "testing" case or would this also violate confidentiality?

You are correct that the segregated solver in most cases will require less memory. The fully coupled solver in general requires more memory, but don't assume that this means that the fully coupled solver is always more robust. For some combinations of physics/variables it is better to use a segregated solver (even if you have the memory to solve fully coupled). One nice thing about the segregated approach is that each segregated sub-step can have its own linear system solver that is tuned to that particular set of variables. This approach can lead to better convergence for some physics.

For your radiation case, I would go with the COMSOL default solver settings unless you have a good reason to change. The default solver selections are tuned to be the best they can be in a general sense.

I suppose it is good news and bad news that both solvers are converging for you. Good news because at least you are getting converged solutions. Bad news because they aren't coming out the same! If your convergence tolerances are set the same for both solution techniques, it's strange to me that this behavior is happening.

Best regards,
Josh Thomas
AltaSim Technologies

Dear Jiuan,

The default convergence criteria are sometimes too loose. I would really tighten the convergence tolerances with both solvers to make sure that they reach the same solution. They should reach the same solution expect. as Josh described above, when your model has no unique solution.

Nagi Elabbasi
Veryst Engineering