My Community

Hi there,

I've been having an convergence issue when applying the Hubbard correction to a LaFeO3 which should be a semiconductor with a bandgap around 2-3 eV.

I've tried running this with three different smearing methods, Gaussian, Tetrahedron and Methfessel-Paxton.

For Gaussian, it does not achieve electronic convergence. The energies cycle around a set of values, as i understand this is due to electron hopping from degenerate partially filled orbitals

For Tetrahedron method, i can achieve convergence but i get the error messsage:
'Tetrahedron method does not include variations of the Fermi |
| occupations, so forces and stress will be inaccurate. We suggest |
| using a different smearing scheme ISMEAR = 1 or 2 for metals in |
| relaxations.

And for Methfessl-Paxton, it achieves electronic convergence and runs fine.


If i inspect the DOS from tetrahedron or Methfessel-Paxton i can see there are orbitals around the Fermi level even for high Hubbard U values (~6) which should have opened up the bandgap. As i can tell VASP is convinced this system is a metal but its actually a semiconductor.

Is there any advice here? Ive included the input files and the POTCARs are: Fe_GW, La_GW and O.

Thanks!

It seems like you want to describe an antiferromagnet. Yet in the INCAR file you do not set ISPIN or magnetic moments.

Martin,

Are you suggesting this error would be fixed if i included the spin polarisation?

i was trying to simplify the problem by removing the spin polarization

Could you explain a bit clearer what your issue is?

From your first post, I got that your issue was that the system is metallic instead of semiconducting. But you would expect this to happen if you force the filled majority and empty minority spin channel to have the same occupations.

Then you mention that you tried different smearing schemes and it seems that some work and others don't, which is also common for any a bit more complex system. Typically you can play with the convergence settings (ALGO, ISMEAR, SIGMA, AMIN, BMIX, ...) to find the optimal setup. For magnetic systems you can also play with the initial magnetic moments.

So to reiterate: Was your question about how to make your system a semiconductor or how to find the electronic ground state?

Okay i will try to elaborate further.

LaFeO3 as a material is a semiconductor, however the electronic structure results from these calculations show a metallic groundstate which is incorrect. This is especially surprising as i have applied a reasonable large Hubbard U value, which tends to open up bandgaps therefore leading to a semiconducting groundstate.

RE: smearing schemes. I would say none of them actually worked as they either failed to converge (gaussian) or converged on an incorrect ground state (tetrahedron and Methfessel-Paxton)

I can try adjusting the other convergence settings as suggested, but to be clear. The convergence issue is not a matter of failing to converge within the time limit. Instead the electronic convergence cycles around the same energy region which is indicative of the electron hopping between partially filled degerate orbitals close to the Fermi level. Which arguably is another sign on a metallic electronic structure.

I have tried to re-run this calculation using Gaussian smearing with spin polarization but i seem to be getting the same result which failure to converge.


My question is not how to make the system a semiconductor as such, instead i was stating that from the evidence above, it appears that VASP is treating this system as metallic and it shouldnt, and i am not sure why

VASP doesn't know whether the system should have a band gap or not, so it will treat the system as it "sees" it. Specifically it means that for a given density, we will calculate the eigenstates and then fill it according to a distribution (Gaussian, MP, Tetrahedron, ...). However, if the eigenstates in your calculation happen to be such that a metallic state is energetically favorable it will choose that state.

There is no guarantee that the DFT or DFT+U state accurately captures the ground state of your system. In fact, for systems with strong correlations you frequently end up with metallic solutions where insulating ones would be correct. With this in mind, there are three scenarios:
(1) There is an error in your setup preventing VASP to find the ground state.
(2) For your system, DFT+U is not accurate enough but VASP finds the ground state at this level of theory.
(3) DFT+U would be accurate enough to produce a gap, but VASP fails to find that particular solution.

Scenario (1) can be caused by using an inappropriate symmetry so that VASP will choose equal occupancies for specific states even though they should be split. To distinguish between the latter two cases, you can run multiple calculations changing the settings as I suggested. If you always end up in the same ground state (same energy, same band structure), then we are likely in scenario (2). If you get very different results depending on the initial state, we are in scenario (3) and you need to carefully tweak the parameters to go to the global minimum.

From what I read from you so far, it appears you are most likely in scenario (2). Then the solution would be to either switch to a higher level of symmetry (e.g. hybrid functionals) or to enforce that your system is semiconducting (check FERWE or I_CONSTRAINED_M tags).

ben_howchen1 wrote: ↑Thu Jun 24, 2021 9:50 amMy question is not how to make the system a semiconductor as such, instead i was stating that from the evidence above, it appears that VASP is treating this system as metallic and it shouldnt, and i am not sure why

So to answer your question: My hypothesis is that your system is a metal at the DFT+U level of theory. You can check slightly distorting your crystal to exclude scenario (1), but my suggestion would be to enforce a semiconducting state by forcing the occupations/magnetic moments and relaxing with that constraint. Then you can release the constraint at the end and see if the system remains semiconducting.

I wanted to post this just incase any others have come across this issue.

I found the same convergence issue with Hubbard U for three different iron containing materials LaFeO3, BiFeO3 and Ca2Fe2O5.

After changing the psuedopotential from PAW_PBE Fe_GW to either PAW_PBE Fe_sv or PAW_PBE Fe the problem is solved.

Unless i am mistaken there must be some bug with this pseudopotential

Can you post your final input files that reproduce this issue, i.e., the INCAR, KPOINTS, POSCAR that yield a nonconvergent calculation with the Fe_GW POTCAR but converge with the other two potentials?

Thank you.