transferability of GW datasets

[raul_l]

In most papers using the GWA that I've read nothing is said about the transferability of the datasets/pseudopotentials used. Does this mean that the authors are simply using ground state datasets? In some papers the convergence studies go up to a few thousand empty bands. Do the authors really put effort into ensuring the datasets/pseudopotentials are perfectly transferable over a range of several dozen Ry? The only discussion about the quality of GW datasets that I've found is in Shishkin and Kresse, PRB 74, 035101 (2006), although they don't exactly show how much influence the high-energy states have on QP energies. In fact, this is what it says at the VASP webpage about the transferability of GW datasets: "We believe that this is important for GW and RPA calculations, although firm proof is still missing."

In my own system convergence is achieved by including states up to about 19 Ry while the highest state used in the convergence studies is at about 30 Ry. My datasets have perfect transferability up to 15 Ry and (I think) good transferability up to 20 Ry. Is this sufficient?

[Alain_Jacques]

Does this mean that the authors are simply using ground state datasets?

[Sh]It happens

Do the authors really put effort into ensuring the datasets/pseudopotentials are perfectly transferable over a range of several dozen Ry?

They should - or at least the provider of the pseudo should quote a "safe" range.

Is this sufficient?

When I build PAW pseudos, I add (at least one) unbound state for each angular momentum with a reference energy (a few Ry) higher than the highest expected energy and I check the presence of ghost states and the match between the exact and PAW dataset logarithmic derivatives on the energy range I expect good transferability. And then I compare results obtained from an all-electron code (Elk) and Abinit with the PAW pseudo on bulk / oxides test cases. So I would say that expecting good transferability up to 30 Ry with pseudo datasets tested at 15 Ry is a bit "risky" - it's so "easy" to check this on simple test cases before running complicated calculations

Kind regards,

Alain

[raul_l]

I did some tests by generating datasets with more projectors per angular momentum and concluded that the given datasets are accurate within 70 meV, which I think barely qualifies as acceptable for GW calculations [J. Phys.: Condens. Matter 26, 125503 (2014)].
I wouldn't say it is so easy to check the transferability with simple test cases. Comparison with an AE code only yields information about the quality of the datasets for ground state calculations (Elk, as far as I know, cannot do GW calculations).