3
$\begingroup$

I am a bit perplexed to see substantive differences in forces (as taken from OUTCAR with ASE) computed when I take AIMD frames and run static electronic relaxations on each one (no additional ionic steps, NSW=0, EDIFFG does not matter). All conditions are the same: EDIFF=1e-4, ALGO=VeryFast. All other parameters are identical. Electronic convergence is reached in both cases. The calculation is collinear spin-polarized, but all magnetic moments are rigorously zero (only Ca, O, and H atoms are in the calculation). VASP 6.3.0, r2SCAN+rVV10, 2x2x2 k-points, again everything identical.

This occurs independently of the thermostat (Langevin / npt and Nosé / nvt). Changing POTIM for the electronic relaxation does not matter. Changing the algorithm introduces only minute differences and does not change the overall picture.

Here is an example plot of the difference in forces (30 frames, 66 atoms each, NVT / Nosé):

forces example

Here is another example (also 30 frames, 66 atoms, but NPT / Langevin and double the POTIM):

enter image description here

I can probably live with the bulk of it when the differences are below 5 meV/angstrom. But the outliers towards 20 meV/angstrom and above are scaring me. Context: the forces are fed to training ML potentials.

What could be the reason(s) for this discrepancy in forces?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
3
$\begingroup$

This is a bad EDIFF for what you are trying to do I suspect. Your forces will not be converged to the same value that your energies are converged to since the forces are the gradient of energy. If you want to get 1e-4 force accuracy then you will need roughly 1e-9 ediff by the "2n+1 theorem" (I suspect this has a better name but I don't recall it at the moment).

By solving for an n value, given "4=2n+1", we get n=1.5 corresponding to a force error of roughly 31 meV (1e-1.5). It may be that it is better as you see in some cases, but you can't expect agreement better than that in general. If you perform the MD with better EDIFF, you should see better agreement between this and the single points.

Improve this answer
$\endgroup$
4
  • $\begingroup$ Interesting, thank you! I'll test this out with the EDIFF. I was not familiar with the "theorem", but from googling it, it seems like you are referring to Wigner's 2n+1 rule? $\endgroup$
    – Andrey Poletayev
    Dec 14, 2022 at 14:13
  • $\begingroup$ And this may need to become a separate question. Between an AIMD frame and a relaxation of the same frame, should either be trusted over the other at a given level of precision, or are they both uncertain to the same extent? $\endgroup$
    – Andrey Poletayev
    Dec 14, 2022 at 14:38
  • 1
    $\begingroup$ I think they are uncertain to the same extent, the MD frame has some history of wavefunction / charge density but the single point doesn't...no way to know if either is actually better. I think its Wigners $\endgroup$
    – Tristan Maxson
    Dec 14, 2022 at 18:32
  • $\begingroup$ Tested this out, and the EDIFF matters ... sometimes. I tested on two systems and only for one did improving the precision (EDIFF) improve the discrepancies. On top of that, there are always outliers. And if using spin-polarized calculations, magnetizations need to be very much in agreement (within 0.01), or else the forces are really off. (For the plots in the question, the calculations were not spin-polarized) $\endgroup$
    – Andrey Poletayev
    Dec 18, 2022 at 20:37

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .