Relaxation (ISIF = 2,3,4) significantly differs from literature (both theoretical and experimental)

[trashcan7]

I'm trying to do a relaxation on both OsO2 and RuO2, which have the same crystal structure. However, resulting CONTCAR that I get is significantly different than the reported crystal structure that has been reported in literature, both experimental and theoretical.
For example, for RuO2, I have the POSCAR:

Code: Select all

RuO2?crystal
4.56
1.0?0.0?0.0
0.0?1.0?0.0
0.0?0.0?0.693
2?4????????????!?Ru?O
Direct
0.0?0.0?0.0
0.5?0.5?0.5
0.3068?0.3068?0.5
0.6932?0.6932?0.5
0.8068?0.1932?0.5
0.1932?0.8068?0.5

but it results in the CONTCAR

Code: Select all

RuO2?crystal????????????????????????????
???4.56000000000000?????
?????1.0954835323609240???-0.0003499940062927????0.0000000000000000
????-0.0003499940062876????1.0954835323609342????0.0000000000000000
?????0.0000000000000000????0.0000000000000000????0.5793475491135363
???2???4
Direct
??0.0000000000000000??0.0000000000000000??0.0000000000000000
??0.5000000000000000??0.5000000000000000??0.5000000000000000
?-0.2305121913043273?-0.2305121913043273??0.5000000000000000
??1.2305121913043275??1.2305121913043346??0.5000000000000000
??0.7692025139789147??0.2307974860210851??0.5000000000000000
??0.2307974860210851??0.7692025139789146??0.5000000000000000
?
??0.00000000E+00??0.00000000E+00??0.00000000E+00
??0.00000000E+00??0.00000000E+00??0.00000000E+00
??0.00000000E+00??0.00000000E+00??0.00000000E+00
??0.00000000E+00??0.00000000E+00??0.00000000E+00
??0.00000000E+00??0.00000000E+00??0.00000000E+00
??0.00000000E+00??0.00000000E+00??0.00000000E+00

My INCAR is rather simple, but I have tried varying parameters, including ISIF to 2 and 4 (which obviously don't result in the same CONTCAR as above, but are similarish), IBRION=1,2, ISMEAR=-5,1,2. I even varied KPOINTS using both Mankhorst and Gamma.

And strangely, it often converges to the same, but wrong, CONTCAR.

I also got incorrect results when trying to model OsO2.

Do you have any suggestions on what might be going on or what I could try to check?

[WolverBean]

trashcan,
I'm assuming your output above was from an ISIF=4 run, since the lattice vector lengths have changed?
What value did you use for NSW? When you do calculations that change the size or shape of the unit cell, you can run into problems introducing spurious forces that are basically caused by the finite size of your integration grid, and the fact that that grid has to change shape or size along with the unit cell. If you want to let a calculation take 20 steps, I'd recommend setting NSW=5, then running four consecutive calculations, each reading from the results of the last, rather than doing a single calculation with NSW=20. Each time a new calculation launches, it will create a new integration grid, so by doing four separate short calculations in a row, you'll rebuild your integration grid four times during the optimization, rather than just once at the outset. This should result in more accurate forces. (The other option is to use a huge value of ENCUT, which is a much more expensive way to treat the same problem.) You'll see the energies jump around quite a bit from run to run, but eventually they should converge.
Don't use ISMEAR=-5 for geometry optimization, especially for metals. It gives funky forces. ISMEAR=1 or 2 should be fine. What sigma value do you have? Something in the 0.2-0.5 range is probably good. (Note that the smaller sigma is, the larger your kpoint grid will probably have to be.)

How big a k-point grid are you using? I've used an 8x8x12 grid with good success for VO2, which has the same structure as RuO2. Too small a k-point grid might give an inaccurate structure, especially for metals.
Does your calculation need to be spin-polarized? I've found that for VO2, the non-spin-polarized calculation gives poor geometries and energies, and I need ISPIN=2, NUPDOWN=2 to get the right crystal structure and density of states. For MoO2, which has a distorted version of the same structure, the non-spin-polarized calculation works just fine. I'm not sure what you'll need with RuO2 and OsO2.

I'm not really sure what the source of your problem is, but those are the first things that come to mind.

[trashcan7]

Do you mean ISIF=3? This output is ISIF=3.
I used:
NSW=100
ENCUT = 520 (1.3*ENMAX)
SIGMA=0.2
IBRION = 1

The k-point grid for this run is 8x8x8. I've tried 11x11x11, but with similar results. Would it be better to use 8x8x12 instead of 8x8x8?

Thank you, I have rerun the calculations per your suggestions regarding both NSW and ISPIN. However, the spin-polarized calculations converged to the wrong answer, and splitting up the NSW seems to converge to a different wrong answer.
And I have tried both RuO2 and OsO2.

Additionally, I want to note that the resulting CONTCAR file results in the symmetry of the rutile structure being broken.

<span class='smallblacktext'>[ Edited Sun Jun 02 2013, 11:36AM ]</span>

[trashcan7]

Actually, I have fixed the problem. I realized that my original POSCAR was wrong! It should be (0.3068,0.3068,0) instead of (0.3068,0.3068,0.5). I ran it and got much closer results. Thank you for your help!

Do you have tips on running relaxations to get the accurate answer? For example, I increased KPOINTS to 11x11x11 to look for convergence and then started from the CONTCAR file.
This source recommends then switching to ISMEAR = -5, but would you recommend against it?