Total energy difference between VASP4.6 et VASP5.2

[boris]

Hi,

I would like to know why, for the same system, VASP 4.6 and VASP 5.2 do not yield the same total energy. Hopefully, this is only a shift but does anyone know the origin of this difference?

It is not mentioned in the vasp manual...

Thank you

[forsdan]

I think you will have to be more specific. For my systems the VASP versions 4.6.31, 4.6.34, 4.6.36, 5.2.2, 5.2.8 give consistent results for the energy evaluation given the same settings. What kind of calculations do you perform?

Have you investigated different optimizations levels for the compiler?

Cheers,
/Dan
<span class='smallblacktext'>[ Edited Wed Nov 24 2010, 05:39PM ]</span>

[boris]

I'm performing DFT+U calculations on uranium dioxide.

Using the exact same INCAR file, I have two different total energies: -116.313 eV and -117.136 eV for U4O8.

I thought that it would be the same for everyone. Maybe it has something to do with the DFT+U then. Other groups in the world observed this difference for uranium dioxide, so I guess it is not due to the compiling flags.
<span class='smallblacktext'>[ Edited Wed Nov 24 2010, 05:52PM ]</span>

[alex]

Hello Boris,

I'd guess you caught different spin states. It's an f2-system, so you'll probably ended up with different solutions in your two calculations.

Check your magnetization, the DOS or the alpha-beta spin density (CHG and CHGCAR, 2nd block)

Cheers,

alex

[boris]

Hello Alex,

I've checked the spin states, they're the same. Plus, I can impose the occupancies of the f states, so it shouldn't reach different spin states.

I've performed a standard GGA calculation and there's also a difference in the total energy...

I'm puzzled...

Below is the result of a GGA calculation in the UC system:
VASP 4.6:
Free energy of the ion-electron system (eV)

---------------------------------------------------
alpha Z PSCENC = 130.37794423
Ewald energy TEWEN = -1603.86917202
-1/2 Hartree DENC = -138.20639013
-V(xc)+E(xc) XCENC = 49.69923935
PAW double counting = 1062.07597707 -1046.98705307
entropy T*S EENTRO = 0.00556046
eigenvalues EBANDS = -24.74236648
atomic energy EATOM = 1550.88688892
---------------------------------------------------
free energy TOTEN = -20.75937168 eV


VASP 5.2:
Free energy of the ion-electron system (eV)

---------------------------------------------------
alpha Z PSCENC = 130.37794423
Ewald energy TEWEN = -1603.86917202
-1/2 Hartree DENC = -138.24143522
-exchange EXHF = 0.00000000
-V(xc)+E(xc) XCENC = 49.70163040
PAW double counting = 1062.07192912 -1046.96677860
entropy T*S EENTRO = 0.00615455
eigenvalues EBANDS = -24.72854581
atomic energy EATOM = 1550.88688892
---------------------------------------------------
free energy TOTEN = -20.76138443 eV
<span class='smallblacktext'>[ Edited Thu Nov 25 2010, 05:12PM ]</span>

[alex]

Hello Boris, this is weird. Could you please post the input files? And make a "grep TIT POTCAR" and send the results, too.

Thx

alex

[boris]

Alex,

After further calculations, it appears that the problem occurs only with f electrons, such as 4f or 5f (CeO, UO2, UC). We performed calculations on other systems with only s, p and d electrons, and there was no difference in the total energy.

As for the "grep TIT POTCAR" for UO2:
TITEL = PAW_PBE U 06Sep2000
TITEL = PAW_PBE O 08Apr2002

Thanks
<span class='smallblacktext'>[ Edited Fri Nov 26 2010, 08:02PM ]</span>

[alex]

Hi Boris,

f-states are notoriously difficult to converge into the proper solution. I'd suggest you try a highspin solution with both versions for starters, i.e.
NUPDOWN = 8
for U4O8. To improve your convergence you may also enter
MAGMOM = 4*2 8*0
to distribute the electrons on U and O, respectively. Then, please check if the total energy is the same for both program version and you may check in CHG or CHGCAR if you have an f2-state on every U atom (I'd assume, but please make sure).

Hth,

alex

[boris]

Hi Alex,

I have tried the ferromagnetic and antiferromagnetic solution as you suggested, and in both cases, with different vasp versions, there's also a difference in the total energy.

With both versions, the states are identical, same cell parameters, same f occupancies, etc. but the total energy is different.

The difference in energy seems to be proportional to the number of bands...
<span class='smallblacktext'>[ Edited Tue Nov 30 2010, 03:10PM ]</span>

[alex]

Hi Boris,

You've got different number of bands? Well, that's it, I'd guess.
Try to set NBANDS to the same value in INCAR and reduce NPAR to 1 (just to make sure NBANDS isn't overwritten) and try again. Check for NBANDS in both cases. It's written at the beginning of OUTCAR.

Cheers,

alex

[peterklaver]

Some collegues of mine reported different results for the same input files for v4.6 and 5.x too. While looking for how that could be, I found this thread. I'll report my finding here, in case more people googling 'VASP 4.6 5.2 energy differences' end up here.......

Their system was a transiiton metal oxide, nothing fancy like f-elements. And the answer was simple too: VASP v4.6 and 5.x have different default settings for the PREC tag. So if you don't explicitly set PREC to the same value in the INCARs for 4.6 and 5.X, then they will run with different settings, yielding slightly different answers.

It is of course good practice to always include a setting for PREC (and any other parameter with different defaults, are there more?) in the INCAR anyway. But definitely if you are mixing different VASP versions.