[gpaw-users] New PAW setups - help needed!

[Jens Jørgen Mortensen]

Den 13-08-2013 15:08, Jens Jørgen Mortensen skrev:
> Here is an update of how things are going:
>
> Den 02-07-2013 09:47, Jens Jørgen Mortensen skrev:
>> Here is an approximate road map:
>>
>> I have now a set of PAW setups for all elements from hydrogen to 
>> radon that seem to perform quite well on our initial benchmark tests 
>> [1].  Those setups are generated with the new generator2.py which 
>> doesn't have all the features of the old generator, so I need to take 
>> a look at these issues:
>>
>> * Generating basis sets.  Should be quite simple.
>
> Done.
>
>> * Figure out if we really need to generate special GLLBSC setups. 
>> Maybe GLLBSC calculations can be based on a PBE setups?  A bit like 
>> vdW-DF is based on revPBE-setups and Meta-GGA's and PBE0 on PBE-setups.
>
> Haven't started looking at this, so please help if you can.
>
>> * Calculate kinetic energy densities for use by Meta-GGA calculations.
>
> Done.
>
>> * Test new strategy for Fourier filtering of setups (projector 
>> functions and zero-potential):  The new setups are unfiltered and 
>> filtering away high frequency components not representable on GPAW's 
>> real-space grid will now happen at runtime.
>
> Done.
>
>> With those four finished, I'm gonna try to get all tests (nightly + 
>> agts) to pass - this will be in the newsetups branch.
>
> This is almost done, but I could use some help with some of the 
> tests.  If you want to help, you should grab the code from the 
> newsetups branch and the setups from here:
>
> http://dcwww.camd.dtu.dk/~jensj/gpaw-stuff/gpaw-setups-0.10.1.tar.gz
>
> Some of the tests check the position of a peak in a spectrum, but in a 
> very fragile way that breaks after the slightest change of the 
> setups.  I have fixed this by using a slightly more robust findpeak() 
> function and by having more realistic tolerances.  See here for examples:
>
> https://trac.fysik.dtu.dk/projects/gpaw/browser/branches/newsetups/gpaw/test/aluminum_EELS.py 
>
> https://trac.fysik.dtu.dk/projects/gpaw/browser/branches/newsetups/gpaw/test/big/response/be_1ml_surf_response.py 
>
>
> There are still some of these left like this one:
>
> https://trac.fysik.dtu.dk/projects/gpaw/browser/branches/newsetups/gpaw/test/big/response/check_spectra.py 
>
>
> If you have written tests like that, could you please run them again 
> with the new setups and code and check that things are still ok - and 
> adjust the tests accordingly.
>
> lcao_projections.py:  Anybody remember what this one tests?
>
> muffintinpot.py and coreeig.py: Anybody know why these two fail and do 
> we still need them?
>
> simple_stm.py: Also fails.  Can we simply remove it and put our effort 
> into the STM code in ASE?

You can see some preliminary test results for the new PAW setups here:

   http://dcwww.camd.dtu.dk/~jensj/newsetups/newsetups.html

As you can see, for the first and second main group elements and the 
transition metals there are semicore versions of the setups also. Those 
give very good results, but the non-semicore ones for the late 
transition metals alkalis are also quite good.

Unfortunately, we made a mistake when calculating the rocksalt reference 
energies, so those numbers are not correct and need to be recalculated ...

Jens Jørgen

>
> Jens Jørgen
>
>> If this goes well, we'll do a 0.10 stable release where people can 
>> choose between the old 0.8 setups or the new ones (which will be 
>> default).
>>
>> What happens after that is impossible to predict, but I'd like to:
>>
>> * find a robust eigensolver-mixer combination that converges everything
>> * make a series of cheaper PAW setups with fewer electrons
>> * make a 0.11 release based on the stuff in the gpaw-aep1 branch 
>> which contains a new ASE interface build on top of new stuff in 
>> ASE-trunk [2] (will require Python 2.6)
>>
>> Comments, questions and/or help is much appreciated.
>>
>> Jens Jørgen
>>
>> [1] https://wiki.fysik.dtu.dk/gpaw-files/workshop13/a03.pdf (page 9)
>> [2] https://wiki.fysik.dtu.dk/ase/development/calculators.html
>>
>