This should be perfectly alright and as long as you start from the correct CONTCAR each time the system should eventually relax completely.

However, I would advise you to choose NSW in such a way that the single job will finish in less than 2 hours so that the WAVECAR and the CHGCAR are written and you can restart your job with already converged wavefunctions and magnetic moments. Of course you still will have to copy CONTCAR to POSCAR, but do not set ISTART = 0. If a WAVECAR exists, VASP will set ISTART = 1 by itself only if you do not force 0 in the INCAR.

If the number of electronic steps vary a lot between ionic steps and so it is hard to estimate how large NSW should be, you could also monitor the progress closely and then write a STOPCAR (with LSTOP = .TRUE.) at the right moment in your workdirectory to allow VASP to complete the job in an orderly fashion and give you converged WAVECAR and CHGCAR files to start from again.

In the long run I would however discuss this matter with the administrator of your computational resources and ask for the possibility of longer jobs, maybe in another queue with less priority, since this is not a very time efficient solution.

Cheers, Michael

Hello Michael, thank you for your reply. The Vasp Wiki has the following line in the description of the ISTART tag:

"... always restart with ISTART=1 unless you really know what you are looking for.

There is only one exception to this general rule: all volume/cell shape relaxation algorithms implemented in VASP work with a constant basis set, so continuing such jobs requires to set ISTART=2 to get a 'consistent restart' with respect to the previous runs."

This confuses me. Also, why do they consider ISTART = 1 as a restart with constant energy cut-off when we are allowed to give a new value for the energy cut-off in the INCAR for the restarted job?

On a side-note: In Quantum Espresso, do you know why do they have different cut-offs for plane waves and charge density (and potential)?

Hi again.

Sorry, I have somehow ignored the fact that you where relaxing the cell also, which of course you mentioned by stating that you have set ISIF to 3.

In that case you are completely right that ISTART=1 (as automatically selected when a WAVECAR is present) is not the way to go, and instead you should use ISTART=2 in the INCAR.

As mentioned on the wiki page an by you in the last post, this is only different when relaxing the cell volume, as you want the same plane-wave basis in each run then. With constant cutoff, the number of plane waves will change if the cell volume changes.

Mybe ISTART=1 will overrule the ENCUT flag in the INCAR, but I have not tested that. However, I am having a hard time figuring out a reason why you would want to start from an existing WAVECAR and change ENCUT...

To summarize: I would suggest you do what I recommended in my last answer with the exception of setting ISTART=2 in the INCAR if the volume changes in your relaxation.

For the side note: I am not 100% sure. I guess the ecutrho setting of QE corresponds roughly to ENAUG in VASP, but PREC and NGX(Y,Z) also play a role there. I guess in both cases it is helpful to play around with these settings if you want to get the best speed at a certain accuracy...

Hope that added not to much to the confusion,

Michael

Although this is not directly related to your question, if you need very precise results, it may be better to do several calculations with ISIF=4 and different scale factors and use a state equation (like Birch-Murnaghan) to find the optimal volume. See the VASP manual example on calculating an hcp structure and also the section on Pulay stresses.

I have found that in a 32-atom cell, the difference in the calculated energy can be up to 40meV.

Hello Michael Wolloch, thank you for the elaborate reply. Do you think it might take more time for the system to write the extensive WAVECAR file as compared to starting with the CONTCAR with ISTART = 0?

As for the calculations, we're luckily getting a new faster cluster soon.

Hello Martin Baeker, yes I am familiar with this approach. It circumvents the error from Pulay stresses since the volume remains constant. It's just a little too expensive when compared to one single ISIF 3 calculation, or maybe not.

How do we qualify if our fit to the EoS is better than the minimization algorithm that VASP uses? Also, though it seems to be unlikely, is there a chance that the difference in energy might be system dependent?

What I did to understand this is to do ISIF=4 with 5 different volumes, then fit to the minimum, and compare this to ISIF=3.

In my tests (all done with nickel-base alloys or nickel phases like delta/eta), ISIF=3 usually gives a smaller final cell volume and errors of up to 1-2meV/atom.

Note that the volume variation can be made a lot cheaper if you always use the final CONTCAR of one calculation as the starting point for the next (with changed scale-factor, of course).

In systems like nickel, you may also need to watch out for changes in magnetic moments which may change abruptly if you change the volume.

If you want very precise results (or need DOS etc.), after doing the volume variation, it is best to add a single static simulation with ISMEAR=-5 at the minimum position (copy the CONTCAR of the run that is closest to the minimum). Usually, the minimum energy calculated this way is close to that predicted by the EOS; deviations are a few meV for a 32-atom cell, so this is only needed if you really need very precise numbers or if you want to do a static calculation anyway.

As I said, all this was done with nickel (and a few things with titanium) alloys; if you have a different system, I suspect that effect size may be different.