ABINIT Discussion Forums

Hello everybody,

I'd have a question regarding structure optimization performed by the following parameters:

optcell 1
ionmov 3
ntime 10
dilatmx 1.05
ecutsm 0.5

Germanium structure optimization done using LDA/GGA gives the results 5.46 / 5.6 [Angstrom] which doesn't seem right according to known values. Does anybody have any experience with obtaining correct lattice constants for germanium (or basically any other element)? Is that mismatch the normal Abinit behaviour or there's something I'm missing?

I would kindy appreciate any suggestions.
Best regards,
Michal

First, I would always use ionmov 2 even for simple systems. I've never had a problem with this choice.

Second, my understanding is that germanium is an (in)famous case where Kohn-Sham DFT fails, especially for LDA--that is, it predicts Ge to be a metal, rather than a semiconductor. What kind of pseudopotential are you using? Before you try to optimize the structure you should compute the band structure and see if you are getting it (at least qualitatively) correct. You may need to go beyond simple KS DFT and use exact exchange or something similar.

jzwanzig wrote:First, I would always use ionmov 2 even for simple systems. I've never had a problem with this choice.

Second, my understanding is that germanium is an (in)famous case where Kohn-Sham DFT fails, especially for LDA--that is, it predicts Ge to be a metal, rather than a semiconductor. What kind of pseudopotential are you using? Before you try to optimize the structure you should compute the band structure and see if you are getting it (at least qualitatively) correct. You may need to go beyond simple KS DFT and use exact exchange or something similar.

Thank you very much for your suggestions (sorry for the delay in replying). I've been using Hartwigsen-Goedecker-Hutter psp for Ge. I'll try it your way.

I was also going to see the result of running the problem with the new Tran-Blaha exchange-correlation functional (TB/MBJ). I was wondering if I can find an appropriate psp file for that job, but no luck so far.

The Tran-Blaha metaGGA is supported by libxc, and libxc can be linked into abinit as a plug-in. I'm pretty sure that PAW in abinit does not support metaGGAs, though--also, I know of no PAW generation package (at least not one compatible with abinit) that supports metaGGA's. The package that most of us in the abinit community use for PAW generation is atompaw, and this package does support the use of libxc but does not support any metaGGA's, only LDA and GGA's. I'm not sure about norm-conserving pseudopotentials, whether any available package compatible with abinit supports metaGGA's.

Having said all that I did play around with Ge using PAW and PBE, and found that the band gap vanishes (wrong answer, but expected), but that it can be opened up by using some portion of exact exchange (accessed in abinit through use of useexech, lexexch, and exchmix variables, and only in the PAW case). For my Ge PAW set the optimized cell constant is contracted by 1.7% compared to experiment (normally PBE leads to a small expansion, but note here I am not using PBE but a mix of PBE and exact exchange).

Joe

Hi,
my Ge band gap with the GGA-PBE vanishs too. This is confirmed in the band plot.
however in the log file one can read:



which I don't understand clearly.
thank you for comments.

Can you post your input file?

thank you for replying professor,
here is my input file (for scf calculation)



Actually, I think that it was just a problem of grid shift because when I give a gamma-centered grid with



the band gap disappears. This means that in the shifted grid I missed the minimum of the valence and conduction bands
which explains that the code gives a band gap as the maximum of the valence band is no longer equal to the minimum of the conduction band.

is this correct ?
thank you professor.

I think you are right, although for band gaps I never take them from the base output file but rather run a full band structure calculation. By the way your pawecutdg looks very high to me, is this set from a convergence study? This parameter needs to be checked for convergence but I never need to set it that high, and it chews up a LOT of memory.

although for band gaps I never take them from the base output file but rather run a full band structure calculation

the previous input was just the first step to obtain the self consistent density used as starting point for band structure calculations (though I missed the prtden 1)

By the way your pawecutdg looks very high to me, is this set from a convergence study? This parameter needs to be checked for convergence but I never need to set it that high, and it chews up a LOT of memory.

I need differences converged to 1 meV and I just want to avoid any new errors added up to those from ecut and k-point samplings ( I have even used values greater than this value). In fact convergence studies give a slightly smaller values.
For the memory it's really memory consuming and it was even the source of some buggy runs.

thank you professor for your instructive comments for this post an all other posts in this forum.