Spin spiral calculation and SCAN metaGGA

[mateusz_domanski]

Dear VASP developers,

I did spin spiral calculations with q-vector = (0 0 0) for the CaCuO2 system with metaGGA SCAN functional (primitive cell, should result in ferromagnetic solution).
It yields a total energy of -45.3594 eV, but when doing collinear calculations (without setting NUPDOWN) it yields a total energy of -45.3204 eV (per 1 Cu).
I expect the total energy to be the same for both cases - what may cause this difference?
More importantly, in the same system, a spiral (0.5 0.5 0) that corresponds to a C-type AF solution gives a much higher energy of -44.5841 eV than the (0 0 0) spiral above. This means the AF arrangement of CaCuO2 is not preferred in the spin-spiral & metaGGA calculations, despite being energetically preferred for collinear metaGGA calculations.
Are the meta-GGA functionals compatible with spin-spiral calculations?
Or is there anything wrong with the INCAR settings?

I use VASP version 5.4.4.

[svijay]

Dear mateusz_domanski,

I do not see any issues with your INCAR file for the non-collinear calculation. Could you please share your INCAR file for the collinear calculation and also the OUTCAR files so that I can try to reproduce the issue?

Sudarshan

[mateusz_domanski]

Dear Sudarshan,
In the .zip file below I am sending inputs & outputs for both spiral and collinear calculations that I wrote about.

Mateusz

[svijay]

Dear Mateusz,

Thanks for the files. Here are some suggestions after I was able to reproduce your error: (1) could you please try switching off LSPIRAL and see if the energies are consistent with those from a vasp_std calculation? (2) increase ENINI to 520 (what you had for ENCUT)?

Sudarshan

[mateusz_domanski]

Dear Sudarshan,
Ad (1) yes, I have tried it, the energies of standard non-collinear calculations are pretty much consistent with collinear calculations.
The results are:
for (0,0,0) spiral solution in eV/Cu: collinear -45.320404 ; non-collinear -45.320593 ; spiral -45.359392
for (0.5,0.5,0) spiral solution in eV/Cu: collinear -45.390423 ; non-collinear -45.389018 ; spiral -44.669547
So the differences for non-collinear results are rather small: -0.2 meV/Cu difference for (0,0,0) spiral solution and +1.4 meV/Cu difference for (0.5,0.5,0) spiral solution.
But for spiral calculations, the differences vs collinear results are significant: -39.0 meV/Cu for (0,0,0) spiral and +720.9 meV/Cu for (0.5,0.5,0) spiral.

Ad (2) after increasing ENINI to 520 the issue still occurs as it yields -44.551631 eV/Cu for (0.5,0.5,0) spiral, so the difference is even bigger.

Mateusz

[svijay]

Just to confirm before I move to non-zero QSPIRAL: if you set

Code: Select all

ENINI=520

do you get consistent results between non-collinear and collinear calculations for

Code: Select all

QSPIRAL = 0.0 0.0 0.0

[mateusz_domanski]

do you get consistent results between non-collinear and collinear calculations for QSPIRAL = 0.0 0.0 0.0

No, for spiral (0,0,0) with ENINI=420 eV I get -39.0 meV/Cu (vs collinear), while for ENINI=520 I get -8.8 meV/Cu difference.

Mateusz

[svijay]

Right, I see that you are using a different POSCAR for each run (the collinear and non-collinear calculation). Re-running the collinear calculation with the same structure and

Code: Select all

ENINI=520

gives me identically the same energy for zero QSPIRAL. Are you able to reproduce this behavior?

Sudarshan

[mateusz_domanski]

Right, I see that you are using a different POSCAR for each run (the collinear and non-collinear calculation).

I use a supercell for non-spiral calculations (LSPIRAL = F) so that it enables calculating an equivalent to (0.5, 0.5, 0) spiral solution.

Are you able to reproduce this behavior?

Yes, with a primitive cell for (0,0,0) spiral solution (LSPIRAL = T) and ENINI=520 eV, I get the same value of energy as in the non-collinear run for the primitive cell (difference around EDDIFF value). The energy difference with collinear calculation stays the same (some -0.2 meV).

Mateusz

[svijay]

Okay great thanks, so how I understand it now is: a collinear and non-collinear calculation give the same results when the the non-collinear case has LSPIRAL=.TRUE. QSPIRAL = 0.0 0.0 0.0 and ENINI matchining ENCUT for the same structure (not the supercell).

Now for QSPIRAL != 0.0, have you run a calculation where you set ENINI to 520 and ENCUT to 620 (commensurate with the +100 requirement in wiki/index.php/Spin_spirals)?

Sudarshan

[mateusz_domanski]

Now for QSPIRAL != 0.0, have you run a calculation where you set ENINI to 520 and ENCUT to 620 (commensurate with the +100 requirement in wiki/index.php/Spin_spirals)?

Yes, I did such a calculation - with (ENCUT 620, ENINI = 520) it yields -11.7 meV/eV vs collinear one with the same ENCUT (and the same POSCAR).

Mateusz

[svijay]

Right, but in this case we would not expect a match between the collinear and non-collinear calculations (i.e. for QSPIRAL != 0)?

Sudarshan

[mateusz_domanski]

Dear Sudarshan, that was just FYI - I would expect these to match, but they may not match each other as these are two different computational methods - that is not really a problem.
You have asked me about zero spirals, but I do not see a big problem there, these results are pretty consistent.
However, I expect the energy differences between (0,0,0) and (0.5, 0.5, 0) spirals from spiral calculations and equivalent collinear calculations (in a supercell, I sent you these in my 2nd post) to be generally equal.
The issue is they are not.
In collinear calculations (0.5, 0.5, 0) spiral equivalent has -70.0 meV/Cu energy compared to (0,0,0) spiral equivalent.
In spiral calculations (0.5, 0.5, 0) spiral has +689.8 meV/Cu energy compared to (0,0,0) spiral.
That is a huge mismatch between these two results. The result for spiral calculations with SCAN (especially 0.5, 0.5, 0) does not seem rational (as AF groundstate is expected).

For GGA+U calculations these two energy differences match each other - with the collinear approach, the difference is -185.9 meV/Cu, while for the spiral method it is -181.4 meV/Cu.
That is a level of mismatch that I would expect.

Since there is such a large error in spiral calculations with metaGGA I am asking:

Are the meta-GGA functionals compatible with spin-spiral calculations?

Mateusz