ABINIT Discussion Forums

I'm doing a GaN calculation and I have already carried out the convergence study as well as electronic/band structure calculation over the primitive cell. Convergence includes ngkpt, ecut, acell, as well as atomic positions. I used the simplest form of XC and psp though. For psp, I did not pay much attention and just googled 'abinit pseudopotentials' and ended up using 31ga.pspnc and 7n.pspnc from norm-conserving Troullier-Martins psps. For XC, I did not change the default value of ixc so I assume the LDA functional in the Teter parametrization was used.

Now I need to make multiple orthogonal changes to my system:

- Change primitive to a supercell calculation, followed by replacement of one or more of the atoms with an impurity atom.

- User PAW instead of NC (I would need to go through the PAW tutorial for this).

- Instead of LDA, use GGA (PBE) for convergence work, and preferably HSE06 for final band structure calculation.

Today I made only one change, replaced the primitive cell with a 2x2x2 supercell (so a 32 atom calc instead of a 4 atom calc) and the ngkpt convergence step is running for the last 3 hours and has only done (2 2 2), (4 4 4), (6 6 6), (8 8 8) datasets (the total energy so far looks far from converged, though the relative change is very small, i.e., less than 1 Ha when total is -4000 Ha). The primitive runs happened in a matter of minutes (sometimes less than a minute).

So I guess I have a couple of questions. Do I have to do a ngkpt convergence study even if I only go from primitive to supercell and don't change anything else? Are there any tips to either speed up calculations, or be able to use past convergence information?

As an extension to the question above, is this a good way to be efficient going from primitive cell to supercell to supercell with impurity or defect:

- do kpt, ecut, convergence on primitive cell
- use the resulting output (DEN? WFK? something else?) to process it into something that would be produced by a given supercell, e.g., a 2x2x2 supercell
(can this be done? like "fold" DEN into a 2x2x2 supercell?)
- use the "folded" output from previous step as a starting point for convergence study for supercell
- use the output of the supercell run as a starting point for a supercell with impurity or defect

Is this doable? more efficient? or any similar scheme?

Today the ngkpt convergence study on 32 atoms took 8 hours, and ecut convergence run is still running for the last many hours. I'm not sure this is an efficient way to do it.

Another thing I just noticed, for the primitive cell I was using kptopt 1 with

nshiftk 1
shiftk 0.0 0.0 0.5

am I supposed to keep it the same with the supercell for which I have symmetry checking turned off (chkprim 0).

Also for ecut I'm using getwfk -1 but it doesn't seem to speed things up at all.

Dear fhssn1,
First of all, regarding norm-conserving pseudo, use the ones from pseudodojo:
http://www.pseudo-dojo.org/
And the related paper:
https://www.sciencedirect.com/science/article/pii/S0010465518300250
They are carefully optimized, needs lower ecut than the TM and are all tested in different situation to guaranty a good transferrability.

Otherwise, regarding your convergence question, you can do a convergence of ecut and kpt on the unit cell level, the ecut will be the same for any supercell size, the grid of k-point can be more or less divided by the number of unit cell you have in each direction. However, once you add a defect, depending of what this defect is going to induce in your system, it is possible that you need to re-check the convergence on both ecut and kpt when the deffect is included.

Best wishes,
Eric

Eric,

Thanks for the response. Regarding the PseudoDojo resource you shared, I have two questions:

- Are they equivalent or better then PAW? (if not then maybe I should jump straight into PAW?)

- Are they compatible with GGA (PBE) and HSE06 XC functional approximations? Since I would probably move towards those (although I'm also weighing the option of going the LDA+U route, just not sure how LDA+U compares with GGA (PBE), I guess it's probably not as good as HSE06 for sure?)

I would like to add that I just looked at pseudodojo and it doesn't have psp's available for some of the elements I plan to try as dopants.

It also doesn't have psp's for hybrid functionals.

fhssn1 wrote:- Are they equivalent or better then PAW? (if not then maybe I should jump straight into PAW?)

- Are they compatible with GGA (PBE) and HSE06 XC functional approximations? Since I would probably move towards those (although I'm also weighing the option of going the LDA+U route, just not sure how LDA+U compares with GGA (PBE), I guess it's probably not as good as HSE06 for sure?)

PAW vs Norm-converving depends on what you want to do and what is available in Abinit for these two approaches. For example, at the present status of the implementation of the code, LDA+U is implemented only for PAW. Hybrids work for Norm-conserving (forces+stress) and is still under dev for PAW (only forces).
In the v0.4, the PseudoDojo are done for GGA PBE, LDA and PBEsol are coming (they are available in v0.3). Hybrids are not yet in the list, though you can use GGA pseudos to run Hybrids.

fhssn1 wrote:I would like to add that I just looked at pseudodojo and it doesn't have psp's available for some of the elements I plan to try as dopants.

Well if you want elements with Z>86, then yes, this is not available (well, you can try Og ). The rare earth are present with f-in-the-core only. You have more choice in PAW in this respect. Otherwise all the other elements are present.

All the best,
Eric