My Community

Dear all,

I have been trying to calculate the band structure of a slab which is only metallic at the surface and insulating in the bulk. I tried using the normal procedure of 1) converge geometry -> 2) diagonalise the exact hamiltonian -> perform GW calculations.

I was having severe troubles in step 2 because it made the mateial a conductor instead of a wide bandgap insulator that it is known to be. So, instead of adding extra unoccupied bands at step 2) i added them all already in step 1 and then performed step 3 directly and the results did not give me any error messages and seem to be in mild agreement with respect to other studies. I want to know is their a physical/fundamental reason why we need the hamiltonian diagonalised if i'm performing scGW0. Is the approach I am taking fundamentally incorrect ?

MfG,
askhetan

The approach is fundamentally correct. The point is that you are working with far too small number
of unoccupied states. If you in your large system increase the number of unoccupied states
so that a good GW screening is guaranteed then your suggested approach is not possible
and you must do with one "exact" diagonalization in one step as suggested in the manual.
Note that in the test calculation on Si with 4 electrons the number of bands is 96.
http://cms.mpi.univie.ac.at/vasp/vasp/R ... tions.html
Cannot you modify your approach so that instead of "the large system and poor GW precision"
you will change to "the small system and good GW precision"?

Thank you very much for the reply. My system size is the bare minimum i can solve for. Its not a bulk but a slab system so there are all kinds of dependencies like vaccum gap and number of layers in the slab, apart from kpoints and number of empty bands. My system has 216 electrons (108 spin pol bands) and I add an additional 200 (spin pol) bands on top as a start. I added these bands already in the first step and then calculated GW and the results seemed to be error free at least.

I found that a major source of error was actually not the second step itself but using DFT+U in the second step. I now converge the geometry of this correlated system with DFT+U and then do the diagonalisation without DFT+U. Doing GW on this is not giving me errors. The approach of using DFT+U in the diagonalisation step is fundamentally incorrect, or so I learnt from some old texts.