Convergence of energy for single atom gas phase system

[RChattopadhyay]

Hello all,

I am trying to calculate the energy of a single Mg atom in gas phase. To set an appropriate simulation box size, I tried to determine the drifts in energy when different simulation box sizes are used. I find that the energy changes by around 102meV/atom from 8 Ang to 10 Ang box followed by a large jump at 12 Ang (plot attached). My aim is to have energies converged to around 0.1meV/atom. Am I missing some important settings here or are there some fundamental reasons for these large energy drifts (8-10 Ang should be good enough to avoid interactions with periodic images) ?

I have attached the input files and the output file for the 12 Ang calculation. I thank you in advance for your help.

[andreas.singraber]

Hello!

I was able to reproduce the behavior you observed and I think it is a convergence issue. For some reason unknown to me the it gets stuck in some local minima when larger boxes are used. I tried a lot of settings and found that using a different choice for ALGO can help. Commenting out ALGO = Fast which effectively selects ALGO = Normal, or, setting ALGO = All, often leads to reasonable convergence (close to zero TOTEN, continuing the trend from smaller to larger box). Interestingly, removing PREC = Accurate and lowering ENCUT (e.g. to 150eV, note the ENMAX = 126.143eV in your POTCAR) also often helped to converge. However, scanning the entire box length scale from 6 to 20 Angstrom I could not find a setup which worked for all runs.

So I tried switching the POTCAR file to Mg_pv (instead of Mg_GW you were using, see hints here). I set ENCUT = 525 and used ALGO = Normal, everything else I kept the same as in your INCAR. This way I was finally able to scan all box lengths:

energy.png

Even if this now seems to work I unfortunately do not have a nice explanation yet for the apparent convergence problems. I will ask my colleagues, maybe I can collect some further insight whether such behavior is expected for METAGGA = R2SCAN. Stay tuned...

All the best,
Andreas Singraber