Differences in convergence behaviour of ionic relaxation. How to get better force convergence?

[garethb]

Hi,

This is my first forum post so please let me know if I need to provide more information on my system to ensure that the question is up to community standards.
I am investigating hydrogenated amorphous carbon systems where I have prepared POSCAR files via a liquid quench technique. Each system has 128 total atoms composed of Carbon and Hydrogen, ranging from 0 to 64 Hydrogen atoms (0 to 50 at%). My end goal is to simply obtain the Bulk Modulii of the systems after a ionic relaxation (isif=3) using the method of energy vs volume Equation of State fitting and isif=4 with 20 ionic steps. I am using version 5.4.4. I have done convergence testing with pure amorphous carbon systems and have determined the appropriate settings for INCAR:
---------
#electronic minimzation
ISTART=0
ICHARG=2
PREC=Accurate
ENCUT=520
EDIFF=1E-8
EDIFFG=-0.02
NBANDS=322 #this keeps NBANDS constant when hydrogen is added into the system and is based on pure carbon system value

LREAL=Auto
LPLANE=.TRUE.
NCORE=16
LSCALU=.FALSE.
NSIM=4

LCHARG=.FALSE.
LWAVE=.FALSE.

#ionic minimization
NSW=999
IBRION=2
ISIF=3

#functional
GGA=PE
ISMEAR=0
SIGMA=0
---------
KPOINTS convergence testing showed 3x3x3 mesh is sufficient for pure carbon system.

For most pure carbon systems, I obtain convergence in easily under 300 ionic steps and the behaviour of the force convergence is always steadily 'getting better' meaning I see more particles having normed forces less than EDIFFG setting until converged (all have normed force below EDIFFG) (see fig 1).

I realize I shouldn't expect the same behaviour for completely different systems (adding in hydrogen), but I have found that most systems do not show the same behaviour of force convergence and instead 'struggle' to find the optimal adjustments necessary to reach the force convergence criterion. More specifically you can see in fig 3 (pure carbon) and fig 5 (hydrogen added) that the number of particles that are not converged increases back up to the maximum value (128), indicating an issue that the system cannot find the optimal arrangement. Runs in this fashion either stop due to exceeding NSW of 999 or take too long to converge and are terminated due to timing on a cluster.

I am not versed well enough in DFT yet to know if this is an byproduct of an algorithm (IBRION, IALGO settings), the functional (PBE), both, or neither.

My question follows from looking at the difference in behaviour of the convergence of some systems with respect to the force criteria. The ideal behaviour is that the system persists to lower the forces on each particle and eventually reaches the convergence criterion slowly but steadily. Why is it that in some of my systems (fig 5), it shows the system 'reverting' to a state where the forces become larger (worse)? Is this because it is still able to decrease the system energy? Is there a INCAR tag or any setting that I could implement to hopefully see better force convergence?

Thanks

[fabien_tran1]

Hi,

It looks like if the total energy is converging steadily but slowly. In order to have a closer look, could you please upload the OUTCAR files of the three calculations?

[garethb]

Hi there and thank you for your response.
I have the OUTCARs but they are each much larger than the size limit allowed by this forum even with ZIP compression. Is there some other way I can get you the information from them?

[fabien_tran1]

Hi,
grep for TOTEN in OUTCAR.

[garethb]

Here are the grep'd TOTEN text files.

[fabien_tran1]

In the three calculations, the total energy obtained at each ionic step is going down steadily as it should, and this is the most important. There is no reason to worry about the forces that increase again, as long as they are below EDIFFG at the end of the calculation.

The number of ionic steps to achieve convergence depends a lot on how the starting geometry differs from the final optimized geometry. For the calculation in fig5-aCH, a NSW larger than 999 should have been used. Note that when restarting a calculation you should copy CONTCAR to POSCAR as mentioned at https://www.vasp.at/wiki/index.php/CONTCAR.

Besides, you are using ISTART=0 and ICHARG=2. In principle using ICHARG=0 should reduce the number of steps during the electronic minimization, see https://www.vasp.at/wiki/index.php/ICHARG.

[garethb]

Thank you that makes sense. I have two follow up questions.
1) When restarting a calculation, am I forced to set ICHARG=0? i.e. do I have to have a CHGCAR and WAVECAR as well as do the CONTCAR->POSCAR switch when restarting from the last ionic step? Or can I just use ICHARG=2 and only switch CONTCAR->POSCAR?

2) Is there anything else you might be able to add with respect to using different tags for IBRION or IALGO? For example, would the long plateaus of the same energy value disappear or shorten (thus speeding up convergence)? Or is this too wishful thinking and instead I should accept that these plateaus are normal?

[fabien_tran1]

1) When restarting a calculation, am I forced to set ICHARG=0? i.e. do I have to have a CHGCAR and WAVECAR as well as do the CONTCAR->POSCAR switch when restarting from the last ionic step? Or can I just use ICHARG=2 and only switch CONTCAR->POSCAR?

ICHARG has an influence on the electronic optimization, while CONTCAR->POSCAR is about the geometry optimization. If a WAVECAR file is present (which is the case after each ionic step), the default ICHARG and ISTART are preferable (but not mandatory), since this may reduce the number of steps during the electronic optimization. You can experiment yourself the effect of ICHARG on a small system.

2) Is there anything else you might be able to add with respect to using different tags for IBRION or IALGO? For example, would the long plateaus of the same energy value disappear or shorten (thus speeding up convergence)? Or is this too wishful thinking and instead I should accept that these plateaus are normal?

The electronic and geometry convergences are sometimes problematic and this is difficult to give a general recommendation.

[garethb]

Thank you so much for the help over the last few days!