Open MFraters opened 1 month ago
The continental extension experiment fails for me too.
I also tested the Newton solver and it fails for both AMG solver and GMG solver :-( It seems that there is something wrong in the Newton assemblers when elasticity is not taken into account.
Supplement to the previous comment: the Newton solver converges for both AMG and GMG methods when Use Newton residual scaling method
is set to true. So maybe at the first time step the initial guess (0) is too far away from the yield surface, and we have to scale the Newton step to make the convergence stabler (the following time steps don't have the problem, because the initial guess current_linearization_point
is not too far from the yield surface)?
On my machine, the Newton solver with AMG or GMG method is slower than the defect correction solver. So defect correction solver with AMG seems to be the best choice for this example.
I've seen the same behaviour, in that Picard is much faster than Newton for this example.
@MFraters - Thanks for posting this and yesterday a colleague and I found highly different behavior between AMG and GMG for a 3D continental extension problem with no elasticity.
@YiminJin @anne-glerum - Likewise, thanks for doing all of that testing.
Issue #4563 is still open and I wonder if we should just disable use of the GMG + Newton/defect correction + Visco Plastic material model for now? Are we confident from the existing test suite that all the other combination of solvers (AMG + Newton/defect correction, GMG + regular picard) are working?
The continental extension cookbook currently fails for me with the output as shown below.
When I change the solver to AMG, the model runs without a problem.
I would suggest to switch this cookbook to AMG for now, untill we have found a fix. I will make a pull request for that.
Likely related #5736