[gpaw-users] Changes in the new PAW setups bundle of 26/10 ("gpaw-setups-0.9.9672")

[Marcin Dulak]

Hi,

On 11/23/12 09:46, Erik R. McNellis wrote:
> Dear All,
>
> I only just now noticed that a new setups bundle was released recently.
> I have looked for a description of precisely how the new setups were
> generated on the GPAW website, but found nothing.
as for any pseudopotential some magic was involved.
The new setups were fitted against ELK LDA decomposition energies of 
fictitious, spinpaired dimers/trimers/tetramers.
The main difference wrt the old 0.8 version ones is that they use 
f-potential option
(which makes logarithmic derivatives to look better for f states), and 
include more valence electrons (called semi-core).
Including semi-core states seemed more important and f-option usually 
just allowed me
to reduce the cutoff radii and avoid ghost states.
For some elements i was not able to improve such decomposition energies 
in this way (the most tricky ones being La, Cr).
I also did not release the setups that do not show significant 
improvement compared to the old ones.

Setup parameters are very sensitive and depend in principle on the 
functional used, so later i used several other
functionals (also generating non-relativistic setups) to adjust the 
parameters (only cutoff radii) to avoid nodes or ghost states.
I think if one tries a functional that is sufficiently "far" from LDA 
one may still discover problems with some setups.
At the end so called Fourier-filtering was adjusted (again setups are 
very sensitive to that), in order to reduce the eggbox.

The setups were generated around 2 years ago, but only after publication 
of http://molmod.ugent.be/DeltaCodesDFT i decided to release them.
There are also tests of light elements on G2-1 set 
https://wiki.fysik.dtu.dk/gpaw/setups/g2_1.html and some bulk oxides
decomposition energies that i have not commited yet.
Of course more tests are needed, and work on the completely new 
generator2 will continue.

A negative outcome from the way new setups are generated is that most of 
them require now eigensolver='cg' or 'dav' to achieve convergence.
The new Oxygen setup is also a bit harder, so requires finer grid.

Best regards,

Marcin
>
> I would therefore be very grateful for a short description of the
> changes made in the new version (0.9.9672) compared to the previous
> version (0.8.7929), in particular with respect to third row and heavier
> elements.
>
> Thank you very much,
>
> Erik McNellis
>


-- 
***********************************
  
Marcin Dulak
Technical University of Denmark
Department of Physics
Building 307, Room 229
DK-2800 Kongens Lyngby
Denmark
Tel.: (+45) 4525 3157
Fax.: (+45) 4593 2399
email: Marcin.Dulak at fysik.dtu.dk

***********************************