non-collinear spin vasp 4.6.28 vs 5.2.11

Message

Zeila · #1 Post by **Zeila** » Wed Jun 01, 2011 6:30 pm

Dear all,

I'm investigating non-collinear magnetization in BiFeO3.
I've tested my calculation with vasp version 4.6.28, 4.6.35 and 5.2.11 and I find that vasp 5.2.11 gives different value for the magnetization w.r.t. previous vasp version!

In my input BiFeO3 is in the rhombohebral cell, SAXIS 100 and a magnetization is expected along 'y' (for this choice of coordinates).

To summarize the creepy results:

vasp 4.6.28
magnetization 0.0000198 -0.0363467 -0.0000001

vasp 4.6.35
magnetization 0.0000198 -0.0363468 -0.0000001

vasp 5.2.11
magnetization 0.0000119 -0.0762098 -0.0000001

=> the magnetization along y is doubled in vasp 5.2.11!

vasp 4.6.35 and 5.2.11 have been compiled with the same options and run on the same machine.
I've started each calculation from the CHGCAR of a previous collinear calculation, so everything is the same, just the vasp version is changed.

I've also checked against vasp 4.6.28 compiled and run on another machine, and I've got confirmation of the 4.6.35 results.

The question is:
which vasp version should I trust?

Hopefully the latest one should be the better one, maybe some bug fixes have occurred. Can anyone comment on this?

Thanks.

All the best,
Zeila

boris · #2 Post by **boris** » Wed Jun 01, 2011 7:43 pm

Hi

I observed the same problem using the LDA+U method. There's a difference between the two versions and I couldn't figure out why.

It seems something was changed and not mentioned in the manual, as usual...

Maybe your problem is related to mine, so I will wait for an answer as well.

Regards
[ Edited Wed Jun 01 2011, 07:49PM ]

Zeila · #3 Post by **Zeila** » Thu Jun 02, 2011 3:20 pm

I'm running the calculation with LSDA+U !
(sorry, I've forgot to mention it...)
so, yes my problem is related to your.

thanks for giving a feedback,
Zeila

boris · #4 Post by **boris** » Fri Jun 03, 2011 12:15 am

Ok so this is a problem I've been working on for a while now.

When using the LDA+U, there's a difference in the total energy between the two versions. It seems that it's because the occupation matrices are calculated differently in the two versions, therefore yielding different magnetic moments.

It might be the cause of the discrepancy.

So maybe they changed the way to calculate the occupancies, or they changed something in the PAW subroutine, I have no idea....

Maybe we'll get an answer. But they should be a bit concerned by such a glitch in the code.

Boris

[ Edited Fri Jun 03 2011, 12:17AM ]

Zeila · #5 Post by **Zeila** » Fri Jun 03, 2011 3:18 pm

Indeed, I've repeated my calculation without "U" and now I'm getting the same results for both code versions.

Who can we ask to know which version is reliable?

Zeila

CMG · #6 Post by **CMG** » Sat Jun 04, 2011 2:52 am

[quote="boris"]Ok so this is a problem I've been working on for a while now.

When using the LDA+U, there's a difference in the total energy between the two versions. It seems that it's because the occupation matrices are calculated differently in the two versions, therefore yielding different magnetic moments.

It might be the cause of the discrepancy.

So maybe they changed the way to calculate the occupancies, or they changed something in the PAW subroutine, I have no idea....

Maybe we'll get an answer. But they should be a bit concerned by such a glitch in the code.

Boris

[/quote]

Dear Dr. Boris Dorado,
Sorry this is not directly related to this thread, but I just have a question about your methods for avoiding/locating local minima by controlling occupation matrices for DFT+U calculations using the VASP code.

You mentioned "As mentioned above, we implemented the possibility of defining initial occupation matrices and impose them during the calculation of the DFT+U potential in the VASP code. We
thus precondition the calculation of the potential which is then applied as a correction to the standard DFT potential. Occupation matrices are imposed during the first ten electronic steps. After these ten steps, the constraint is lifted and the calculation is left to converge self-consistently on its own." (Phys. Rev. B. 82, 035114 (2010)).

Could you please elaborate on how you define the initial occupation matrix and impose them during the first ten electronic steps in VASP? For example, maybe you do it by setting FERWE or FERDO tags in INCAR? Or you modified part of the source codes (e.g., the LDApU.F subroutine) to realize this implementation?

We are much interested in your invention and am trying to use your methods in our calculations using LDA+U. If you could kindly provide more details, we will appreciate much!

Cheers,
CMG

Zeila · #7 Post by **Zeila** » Sun Jun 05, 2011 4:53 pm

Hi everybody!
I've clarified the problem. I was using LSDA+U with the Durarev approach, which is buggy in vasp4.6.

I've re-did my calculations with the Liechtenstein approach and now vasp4.6 and vasp5.2 give the same results which, by the way, are the same as Dudarev in vasp5.2 (since, dudarev = liechestein with J=0).

Here are the results:

vasp 4.6.28, Dudarev: magnetization 0.0000198 -0.0363467 -0.0000001

vasp 4.6.35, Dudarev: magnetization 0.0000198 -0.0363468 -0.0000001

vasp 5.2.11, Dudarev: magnetization 0.0000119 -0.0762098 -0.0000001

Liechtenstein:
vasp4.6, U=4 eV, J=0: magnetization??? 0.0000119? -0.0762081? -0.0000001

vasp5.2, U=4 eV, J=0: magnetization??? 0.0000119? -0.0762096? -0.0000001

So, don't use vasp4.6 with dudarev LDA+U!!!

Cheers!
Zeila

boris · #8 Post by **boris** » Mon Jun 06, 2011 7:23 pm

@CMG: I will gladly help you do that. I guess we should talk about that via email. My email is: bdorado [at] lanl.gov

Feel free to contact me.

@Zeila: This is very interesting. I will try to run the calculation with the Liechtenstein approach to see if there is any difference. Thanks for noticing that

Boris

Edit : I still have the total energy difference even with the Liechtenstein approach:

VASP 4.6:
free energy TOTEN = 1542.041737 eV

VASP 5.2:
free energy TOTEN = 1542.038787 eV

However, I get the same results if I consider a d electron system instead of a f electron system.

Therefore, it seems that there is a difference in the two versions only with f electrons...
[ Edited Mon Jun 06 2011, 08:04PM ]

Alenushka · #9 Post by **Alenushka** » Tue Jun 21, 2011 8:30 pm

Hi everybody,
actually, besides all the discrepancies mentioned already above, I have difference in total energy even without U, if I calculate system with f-electrons with different VASP versions:
system UO22+, charged, so cubic unit cell, multipole corrections are switched on
XC = PW91
Total energy v 4.6 -4.72588
Total energy v 5.2 -4.74441

Waiting for the answer,
regards

boris · #10 Post by **boris** » Fri Jul 08, 2011 2:10 am

Hi

I also have a difference in the total energy without the U. So it seems that as soon as I consider f electrons, there is a difference between 4.6 and 5.2.

Something has changed, and I don't know what it is...

boris · #11 Post by **boris** » Tue Jul 12, 2011 11:29 pm

Hi everyone,

Here's the answer to our problems: the 4.6 version does not take into account the one-center multipoles up to L=6 (only up to L=4).

This has been corrected in the new 4.6 version available on the vasp FTP server.

Regards