I tried the AMG linear solver on a 1D nonlinear diffusion problem and it crashes the FEATool app and Matlab - no error messages, just closes everything. The same problem runs no problem on all the other solvers. Anyone else seeing this? I'm running the latest FEATool 1.12.4 build 20.08.221.
The algebraic multigrid (AMG) solver and interface it not yet robust enought to be able to automatically handle model problems of any type. Generally, problems involving positive semi definite matrices should be fine, but ill-conditioned matrices, matrices with zeros on the diagonal, and potentially matrices that for some reason can not be split in to some multigrid hirearchy can be problematic.
Multigrid in general might be of limited benefit for small 1D problems (with low ammount of degrees of freedom). Multigrid is designed to combine memory efficiency of iterative solvers with constant number of iterations with respect to increasing problem sizes.
Thanks for the reply. That explains why the other solutions worked fine. Now if only Win 10 feature update 2004 hadn't broken things......there was also a .NET Framework 3.5 and 4.8 update. Does FEATool use any of those?
Now if only Win 10 feature update 2004 hadn't broken things......there was also a .NET Framework 3.5 and 4.8 update. Does FEATool use any of those?
No, FEATool does not use any OS dependent features. The binary dependencies (geometry and mesh engines) have been specifically statically compiled to avoid these types of issues. It could be that Matlab uses some feature that might now be incompatible, it might be worthwhile to also look in the Mathworks forums if anyone has issues with the new update and Matlab as they have a significantly larger user base to catch and report such errors.
I did a cursory check of Matlab forums and I didn't find anything yet. The issue occurs, it seems, when the nonlinear solver is called, and it occurs both if the problem is run within the FEATool app as a fea file, or if the file is run as an m file in Matlab using the FEATool functions. Like I said, the Space transform sample problem failed to run, when it ran fine before the update. I'll keep checking.
The vagaries of Matlab have struck again.....I uninstalled FEATools from the Matlab version I was running (2020a v4) and installed it in 2019b v6, and everything runs again, including the AMG solver, without error. I have had perfect success running all kinds of problems in 2019b now since the beginning - just thought I'd try 2020a - unfortunately, no go.
Thank you for reporting the issue and "fix". It does seem the Mathworks changed Matlab mex (shared library) file compatibility very significantly in 2020 (which are used by the linear solvers mumps, amg, fsi). I had recompiled them so that they should work, but maybe a Windows update introduced some incompatibility again. At the moment, it does seem more significantly stable to use Matlab versions up to 2019b, but I have also typically found no issues having several versions of Matlab installed (and actually running simultaneously for testing) at the same time.
I wanted to report back that Matlab informed me that since FEATool is a third party app, there is nothing they can or will do, even though it's their own update that broke it. 2019b v6 works fine with the windows 2004 update, 2020a v4 does not, and they're evidently fine with that.