MgDynPost Convergence History

Numerical methods and mathematical models of Elmer
Post Reply
modelman75
Posts: 11
Joined: 03 Oct 2020, 23:33
Antispam: Yes

MgDynPost Convergence History

Post by modelman75 »

The convergence behavior of the MgDynPost solver is a little puzzling. MgHarm converges nicely for the problem run; Then MgDynPost calculates the variables of interest.

The issue here is that the attached "convergence history" plot shows that MgDynPost hasn't converged to the requested 1e-10 tolerance. However, the "solver log" (attached) indicates that the 1e-10 tolerance is satisfied.

Can anyone please comment on this apparent discrepancy?
Attachments
Solver-Log_NPROC-3_WIN10.txt
(14.25 KiB) Downloaded 142 times
Convergence-History.PNG
(30.63 KiB) Not downloaded yet
raback
Site Admin
Posts: 4812
Joined: 22 Aug 2009, 11:57
Antispam: Yes
Location: Espoo, Finland
Contact:

Re: MgDynPost Convergence History

Post by raback »

Hi

Yes, that may be confusing. MgDynPost solver computes nodal fields by the Galerkin method which involves solving a simple mass matrix problem with varying r.h.s. vector for different solution vectors. Now this looks on the output as if there where multiple nonlinear iterations while there actually is not. Unfortunately the ElmerGUI analyzes the output and picks the convergence behavior from there which is completely bollocks.

I could try to supress the computation of change so this would not be seen.

Also, I would usually recommend using elemental fields for postprocessing. So for MgDynPost have

Code: Select all

Calculate Nodal Fields = False
Calculate Elemental Fields = True
This will in effect create Discontinuous Galerkin type of fields that do not involve large linear systems. Only local small linear system is solved multiple times, and hence there is no convergence behavior.

Then, in the VTU output you can request

Code: Select all

Discontinuous Bodies = Logical True
which will average the DG fields within bodies but maintain the discontinuous fields only between the different bodies.

You can find example of these settings in test case "mgdyn_steady_wire", for exampe.

-Peter
modelman75
Posts: 11
Joined: 03 Oct 2020, 23:33
Antispam: Yes

Re: MgDynPost Convergence History

Post by modelman75 »

Peter, I checked the sif to see if elemental fields were being used for the postprocessing. Although I explicitly deselected "skip elemental fields", and selected "skip nodal fields" in the GUI, the sif file did not reflect those requests.

I have since manually added your suggestions to the sif, and re-ran successfully. Many of the issues that I have encountered involve the GUI lacking options, or selected options not entered into the sif. I have relied too heavily on the GUI. As my familiarity with options improves, I will try to rely more on manual editing.

Once again, thank you for your prompt attention to my questions and your useful suggestions.
raback
Site Admin
Posts: 4812
Joined: 22 Aug 2009, 11:57
Antispam: Yes
Location: Espoo, Finland
Contact:

Re: MgDynPost Convergence History

Post by raback »

Hi

The problem with ElmerGUI is that very few people use it for serious business. It is a nice educational tool but when you go beyond that people tend to leave it behind and build their own workflow using various scripts etc. The menus would basically be easy to extend as they are described by XML files but unfortunately they may indeed be lacking some features.

I suggest using the "free text input" field in ElmerGUI for missing keywords. This way the project can be loaded back.

-Peter
Post Reply