convergence problem with single atoms in a cell

[Danny]

hello,

I have a quite nasty system, that refuses relaxation. It consists of 4 atoms in a large cell (it's a partial calculation of a larger system) and for some reason it seems almost impossible to get the energy converged.

I'm using the GGA (PW91) potentials (it need to be these because I'm comparing a set of systems) for the C and O atoms.

here are the INCAR

Code: Select all

general:
   SYSTEM = GGA_CONW1t3
   ISTART = 0   ! start van 1: voorgaande 0: scratch
   ICHARG = 2   ! 0 from WAVECAR  1 from previous 2 superposition
   ISMEAR = 0   ! Gaussian smearing--> paar die gek doen
   SIGMA = 0.1  ! kleine breedte
   EDIFF = 1.0E-6
   PREC = High  ! om ROPT = 2.0E-4 te krijgen bij LREAL=Auto
   LWAVE = .FALSE.   ! geen wavecars nodig
   LCHARG = .TRUE.
   LVTOT = .TRUE.
   LREAL = .FALSE. ! grote cel
   ENCUT = 400  ! alle NW berekeningen zijn hierbij, geen d-e=trage conv
   NBANDS = 16   ! extra bands tegen vodden
dynamic:
   IBRION = -1   ! fix atomen
   NSW = 0       ! enkel een SC run

and POSCAR

Code: Select all

GGA_calculation_NW1t3
  5.7800000
     1.0000000000000000    1.0000000000000000    0.0000000000000000
    -2.0000000000000000    2.0000000000000000    0.0000000000000000
     0.0000000000000000    0.0000000000000000    5.7500000000000009
 2   2
Selective dynamics
Direct
  0.9776660572356373  0.4444409243090476  0.9065013708861110   F   F   F
  0.0223339427643630  0.5555590756909594  0.0934986291138889   F   F   F
  0.9939830853783936  0.2624464533081692  0.2836266776693664   F   F   F
  0.0060169146216061  0.7375535466918303  0.7163733223306344   F   F   F

files.

When I was looking at these systems using LDA potentials this system also gave a problem, but there I managed to solve it through using a ridiculous number of bands (64, where 16 normaly is enough).
I have tried the usual
1) different k-point sets, 1x1x1, 8x4x1, 9x5x1
2) increasing NBANDS (up to 64 and there is no change in the behaviour whatsoever)
3) different smearing (and smearing width)
4) changing the ENCUT
nothing even gives a hing of which direction to proceed...

can anyone, think of other options?
The job doesn't crash it just runs against the maximum of 60 electronic steps, while the energy oscillates during the entire calculation over 2 orders of magnitude.
e.g.:

Code: Select all

       N       E                     dE             d eps       ncg     rms          rms(c)
DAV:   1     0.119447572510E+03    0.11945E+03   -0.27392E+03    48   0.306E+02
DAV:   2     0.549814670627E+02   -0.64466E+02   -0.61385E+02    56   0.776E+01
DAV:   3     0.409174377037E+01   -0.50890E+02   -0.47168E+02    48   0.528E+01
DAV:   4    -0.133826716225E+02   -0.17474E+02   -0.17064E+02    48   0.356E+01
DAV:   5    -0.140564183251E+02   -0.67375E+00   -0.67366E+00    40   0.958E+00    0.173E+01
DAV:   6    -0.826887544374E+02   -0.68632E+02   -0.39859E+02    32   0.296E+01    0.506E+01
DAV:   7    -0.799781322852E+01    0.74691E+02   -0.15034E+01    32   0.165E+01    0.431E+01
DAV:   8    -0.100827648781E+02   -0.20850E+01   -0.10294E+02    40   0.569E+00    0.179E+01
DAV:   9    -0.730580120028E+01    0.27770E+01   -0.18520E-01    40   0.258E+00    0.183E+01
DAV:  10    -0.504979761162E+01    0.22560E+01   -0.66308E-02    32   0.125E+00    0.179E+01
DAV:  11    -0.406018739639E+01    0.98961E+00   -0.14238E-02    40   0.460E-01    0.177E+01
DAV:  12    -0.885235822070E+01   -0.47922E+01   -0.10727E-01    32   0.128E+00    0.182E+01
DAV:  13    -0.187124481714E+02   -0.98601E+01   -0.35395E-01    32   0.235E+00    0.191E+01
DAV:  14    -0.182557275235E+02    0.45672E+00   -0.49768E-03    40   0.402E-01    0.191E+01
DAV:  15    -0.180523604362E+02    0.20337E+00   -0.16128E-04    32   0.656E-02    0.190E+01
DAV:  16    -0.181064152676E+02   -0.54055E-01   -0.14730E-05    40   0.186E-02    0.190E+01
DAV:  17    -0.112015292192E+02    0.69049E+01   -0.17312E-01    32   0.168E+00    0.184E+01
DAV:  18    -0.192661371934E+02   -0.80646E+01   -0.89277E+01    32   0.474E+00    0.418E+01
DAV:  19    -0.498430361135E+02   -0.30577E+02   -0.50118E+02    32   0.186E+01    0.229E+01
DAV:  20    -0.578619273604E+01    0.44057E+02   -0.10177E+01    32   0.137E+01    0.409E+01
DAV:  21    -0.443889192166E+02   -0.38603E+02   -0.44150E+02    32   0.145E+01    0.220E+01
...

thanks
Danny

<span class='smallblacktext'>[ Edited ]</span>

[yuxiao_zihan]

You could try to add ISPIN=2 in INCAR, which I guess will help.

[Danny]

Thank you for the hint.
I have just tried it, but no improvement of the situation.

Danny

[tlchan]

If you're sure that the geometry is what you want to calculate, you should try to decrease the amount of mixing for the self-consistent iteration. Your geometry nearly consists of isolated atoms. Systems with broken bonds usually are harder to converge.

If this calculation is for comparing with other calculations, the energy cutoff and smearing should be the same with the other calculations, and the k-point set should be compatible with the other calculations as well. I don't think it is a good idea to adjust these values for self-consistency loop.

[Danny]

I hadn't really thought about the mixing...I'll have a look into that.

I know that for comparison I should not change any of my parameters(and with the exception of this case, it is never neccessary. However due to the huge oscillations (sometimes more than 3 orders of maginitude) I am more concerned about convergence first, if I succeed at that I can again try to move back to the parameters needed.

Danny

[Danny]

OK, it seems like the last tip was the charm, changing the mixing and getting rid of NSIM=8 makes the job converge (slowely, it takes 119 steps with only Gamma), but that is the least of my problems. I'm now moving back to my old parameters and it still shows resonable results.

Thanks for the good advice yuxiao_zihan and tlchan

Danny