Strange increase in the energy after complete relaxation

[boris]

Hello,

I performed an ISIF=3 calculation (complete relaxation of the cell), which was a restart job from a ISIF=2 calculation (relaxation of atomic positions only).

I expected the total energy in the ISIF=3 case to be lower than the ISIF=2 case, and it happens to be sometimes higher.

I restarted the calculation using only the CHGCAR file obtained in the ISIF=2 calculation.

This issue has occured twice, for two different defects in my structure.

How come the total energy is higher for a complete relaxation than for a relaxation of atomic positions only ?

Thank you for your help.

[forsdan]

It depends on how large energy differences you're talking about. Do you have differences in the eV or meV range?

The most likely reason if the energy difference is large is that the accuracy isn't high enough to provide correct volume relaxations (read the Pulay stress section in the manual and the "Accuracy and Validation of results" in the hands-on sessions). The forces are simply wrong and therefore the system may sometimes relax to a groundstate with higher energy. Increasing the cutoff by 30% is the usual approach. Other tags connected to the accuracy of the forces could also affect the results as well as your EDIFFG settings. If the energy difference is smaller than the latter you might consider decreasing it.

Also check how large drift you have in your forces. It gives a hint on if your fft-mesh is sufficient. The drift should always be smaller than the forces acting on the atoms.

In the case of spin-polarized calculations: another source could be that the ISIF=3 converge to another magnetic structure with higher energy, even if I doubt it. How large is the energy difference between the last ionic step with ISIF=2 and the first ionic step with ISIF=3?


Best regards
/Dan Fors

<span class='smallblacktext'>[ Edited Mon Mar 24 2008, 04:45PM ]</span>

[boris]

Dear Dan,

Thanks for your answer.

The energy differences I am dealing with are rather in the range of the meV/atom, sometimes less.

My energy convergence criterium is set to 1.0E-7 eV. I guess this value is low enough to get proper converged forces during relaxation.

Considering magnetism, this is a real issue in my case, and I'm constantly checking it. It sometimes happens that ISIF=3 leads to a different magnetic configuration than ISIF=2, but it usually gets a lower energy.

Anyway, this is not the case here.

I have set the Precision tag to 'Normal', as suggested in the VASP manual. As I'm not calculating any phonons or second derivatives, I didn't set it to 'Accurate'.

[quote="'smallblacktext'>[ Edited Fri Mar 28 2008, 09:22AM "]</span>

[forsdan]

The drift appears to be quite small, so I don't think this would be the main contribution to a large energy difference. I would suggest you try to increase the cutoff energy (at least 30%) for the two calculations and see if the difference still persists. This should give a more reliable stress tensor.

The total energy difference per supercell would however be a more appropriate measure to use (in my opinion). In my cases the difference between ISIF=2 and ISIF=3 have been in the order 0.1 eV / 100 atoms for vacancy defects, so the differences are typically small, but of course it depends on the systems.

If you're not happy with the drift in forces you might consider to change the ROPT-settings or use 'Accurate' instead. I doubt that it will be necessary to increase the accuracy for the drift here but I can't be absolutly sure. It depends on how large the energy difference is and how accurate calculations you want to accomplish. You got to test to see how much it affects the results.

Best regards,
/Dan Fors


<span class='smallblacktext'>[ Edited Tue Mar 25 2008, 04:14PM ]</span>

[admin]

Dear Boris and Dan,
PREC should be increased to accurate to get reliable force tensors, it determines the default-settings of NGx, NGxF, and ROPT, altogether (so you don't have to bother for the choice of ROPT).
Just to be sure, please let me add 2 more points that may have been overlooked (though I don't think this is the case):
1) please check if each of the ionic steps is fully converged electronically to the requested EDIFF limit, or if (one of them) just stopped after the default 60 electronic steps, without being actually onverged. If the latter has happened, the forces are not reliable and you might end up in geometry which does not correspond to the true equilibrium
2) only if you run the jobs parallel, on different numbers of CPUs: please check whether NBANDS is the same for both runs, if the the number of non-occupied bands differs, this also slightly influences the total energies (in the 0.xx meV range)