Welcome to the forums. Please post in English or French.

You are not logged in.

#1 2019-04-18 08:56:36

m_golbs
Member
From: Germany
Registered: 2009-11-09
Posts: 116

STAT_NON_LINE basic convergence question

Hello,
FEM convergence in none linear static is a big "work of art". Over years I have in CA in STAT_NON_LINE often this effect:

The convergence by RESI_GLOB_RELA goes not to zero but to a little offset value.

In this fist time step I have :
- linear material behavior, no plasticizing
- short stress and strain
- very short displacement
- static determinacy
- no activ contact
- no degenerate elements (jacobian test must have?)
- but very much elements, big modell
- effect are in different modells with different element types (1D, 2D or 3D Elements)

A model test with MECA_STATIQUE and check the depl results by Dx, DY, DZ, DRX, DRY or DRZ brings no indications.

Whichever effect can generate this offset by RESI_GLOB_RELA? About informations I'm very happy.

Thanks Markus

Last edited by m_golbs (2019-04-18 16:18:35)


Attachments:
2019-04-17-Konvergenzprobleme-Code-Aster.jpg, Size: 18.22 KiB, Downloads: 50

The superfluous chase, is to miss the essential.
Jules Saliège

Offline

#2 2019-05-09 14:43:39

m_golbs
Member
From: Germany
Registered: 2009-11-09
Posts: 116

Re: STAT_NON_LINE basic convergence question

Hello,

here the next project with the same effect.

I5GF23B=AFFE_MODELE(MAILLAGE=NETZ,
                    AFFE=(_F(TOUT='OUI',
                             PHENOMENE='MECANIQUE',
                             MODELISATION='3D',),
                          _F(GROUP_MA=('EL2DPR01','EL2DPR02','EL2DPR03','EL2DPR04','EL2DPR05','EL2DPR06','EL2DPR07','EL2DPR08',
                                        'EL2DPR09','EL2DPR10','EL2DPR11','EL2DPR12','EL2DPR13','EL2DPR14','EL2DPR15',
                                        'EL2DPR16','EL2DPR18','EL2DPR19','EL2DPR30','EL2DPR50',),
                             PHENOMENE='MECANIQUE',
                             MODELISATION='DKTG',),),);

...
...

ERGEBNIS=STAT_NON_LINE(MODELE=I5GF23B,
                      CHAM_MATER=MATERIAL,
#       	      CONTACT = KONTAKT,
                      CARA_ELEM=CARA,
                      EXCIT=(_F(CHARGE=LAGER01,
                                   TYPE_CHARGE='FIXE_CSTE',),
                                   _F(CHARGE=Lastseit,
                                  FONC_MULT=Lastfunk,
                                  TYPE_CHARGE='FIXE_CSTE',),),
                       COMPORTEMENT=(_F(DEFORMATION='GROT_GDEP',
                                   RELATION='ELAS_VMIS_TRAC',
                                    TOUT='OUI',),
                                    _F(DEFORMATION='GROT_GDEP',
                                        GROUP_MA=('ALLE2DEL', ),
#                                      RELATION='VMIS_ISOT_TRAC',),),
                                        RELATION='ELAS_VMIS_TRAC',),),
                      INCREMENT=_F(LIST_INST=SolSchrX,
                                    PRECISION=1e-06,),
                       METHODE='NEWTON',
                       CONVERGENCE=_F(RESI_GLOB_MAXI=10,
                                      ARRET='NON',
                                      ITER_GLOB_MAXI=25,),
                       SOLVEUR=_F(METHODE='MUMPS',
                                  RENUM='METIS',
                                  NPREC=12,
                                  ELIM_LAGR='NON',
                                  STOP_SINGULIER='NON',),
                       ARCHIVAGE=_F(LIST_INST=ArcSchri,
                                    CRITERE='RELATIF',
                                    PRECISION=1e-06,),);

FIN();

How can I accelerate the solution/convergence? Is the modell without COMPORTEMENT(...) PETIT and ELAS?
Why goes the RESI_GLOB_RELA not in typical power of ten down ("primary at low force time with static determination")? Is it a general effect of large meshes, stiffness jumps between mesh regions, bad element quality,...?

Thanks and greeting Markus

Last edited by m_golbs (2019-05-09 17:12:57)


Attachments:
20190509konvergenz.png, Size: 28.89 KiB, Downloads: 26

The superfluous chase, is to miss the essential.
Jules Saliège

Offline

#3 2019-05-13 11:49:01

RichardS
Member
From: Munich, Germany
Registered: 2010-09-28
Posts: 548
Website

Re: STAT_NON_LINE basic convergence question

Hello Markus,
your settings look like you have severe issues with your model:

                       CONVERGENCE=_F(RESI_GLOB_MAXI=10,
                                      ARRET='NON',
                                      ITER_GLOB_MAXI=25,),
                       SOLVEUR=_F(METHODE='MUMPS',
                                  RENUM='METIS',
                                  NPREC=12,
                                  ELIM_LAGR='NON',
                                  STOP_SINGULIER='NON',),

I would say, with such loose convergence settings (RESI_GLOB_MAXI=10,  ARRET='NON', NPREC=12, STOP_SINGULIER='NON') that your solution is probably already quite unphysical once you observe the convergence not advancing any more.

In any case it is very hard and usually impossible to argue over reasons of such behavior without having the FULL case setup available (mesh + command file).

Best,
Richard


Richard Szoeke-Schuller
Product Management
www.simscale.com
We are hiring! https://simscale-jobs.personio.de/?language=en#all

Offline

#4 2019-05-16 15:07:55

m_golbs
Member
From: Germany
Registered: 2009-11-09
Posts: 116

Re: STAT_NON_LINE basic convergence question

Hello,

many thanks! How can I calculate the optimal NPREC for a specific Modell? Is the NPREC low (6) than I have singular problems. If the NPREC high (12) then no convergence. Also the default not work. Always by the same Mesh, same *.comm file. About information I'm happy.

Thanks and greeting Markus


The superfluous chase, is to miss the essential.
Jules Saliège

Offline

#5 2019-05-16 15:24:31

RichardS
Member
From: Munich, Germany
Registered: 2010-09-28
Posts: 548
Website

Re: STAT_NON_LINE basic convergence question

Hello Markus,
there is no exact science behind choosing an appropriate NPREC value.
Usually the default (8) should be fine, if not, there is likely an issue that causes this.

You need to understand what this NPREC actually means. It is a value which is used to control at which point the linear solver should stop the solution as the loss of precision is too high. Simply speaking, a value of NPREC=8 means that we tolerate a loss of precision of up to 8 digits - (with usual double precision we still got 8 valid digits left). On the other hand the "loss of precision" is just an estimate and sometimes MUMPS over-estimates the lost precision, especially if you run the simulation on multiple parallel MPI threads.
So it might be acceptable to go up to NPREC=11 and still have a reasonable accuracy. That's also only a rule of thumb.

If your analysis still does not even converge with a high value, then maybe you should check why your simulation is that badly conditioned - maybe your input is inconsistent, the setup is just physically unstable, you have a bad mesh quality + a million of different other reasons why a nonlinear analysis might fail.

Again, without the complete case its all just speculation.

Best,
Richard


Richard Szoeke-Schuller
Product Management
www.simscale.com
We are hiring! https://simscale-jobs.personio.de/?language=en#all

Offline

#6 2019-05-16 16:18:11

m_golbs
Member
From: Germany
Registered: 2009-11-09
Posts: 116

Re: STAT_NON_LINE basic convergence question

Hello Richard,

many thanks!  "..solver should stop the solution as the loss of precision is too high.." How are calculate this precision in STAT_NON_LINE? Means this, that internal and external forces get no convergence or the different are to high to start newton method?

Why comes with NPREC=6 a singular matrix? How can precision generate singular matrix?

"..If your analysis still does not even converge .."
- I have in first step very low force, material only in elastic area
- no contact in modell (this time)
- displacements are very low, test or comparison with meca_statique
- no kinematic effect
- no RBM
- test group assignment
- test an modify material definition
- I have reviewed mesh 10 times and more for bad mesh quality. Deleted elements, take modal analyse for retest RBM ...
- I know this "..+ a million of different other reasons why a nonlinear analysis might fail.." unfortunately.

The modell with COQUE_3D works different to DKTG (rest of comm are absolut the same only in COMPORTEMENT different). (the basic mesh is a tria3 and for COQUE_3D I modify it to tria6 and in comm than to "tria6_7".

So i will next search and test.

Thanks and greeting Markus

Last edited by m_golbs (2019-05-16 16:26:35)


The superfluous chase, is to miss the essential.
Jules Saliège

Offline

#7 2019-05-22 11:47:33

m_golbs
Member
From: Germany
Registered: 2009-11-09
Posts: 116

Re: STAT_NON_LINE basic convergence question

Hello,

I have test many entity. The actual result is "it is a mesh problem, but no by individual degenerate elements". The mesh is a local partially dirty complex industrial stage of development 1.000.000 nodes and 500.000 COQUE_3D. Now I have create a academic mesh related size,  related basic structure, but clearly structured., the same *.comm file. It works absolut correct. To create a cleanly mesh by a industrial stage of development cad modell is the challenge, short time...

Can mesh structures in picture are give numeric problems? I have fade in only a drop of elements an nodes. Basic are node with very low distance and on this nodes the elements....

Thanks and greeting Markus

Last edited by m_golbs (2019-05-22 12:53:06)


Attachments:
2019-05-22-bad-mesh_.png, Size: 36.25 KiB, Downloads: 13

The superfluous chase, is to miss the essential.
Jules Saliège

Offline

Board footer