[gpaw-users] LCAO mode for Zeolite

[Ask Hjorth Larsen]

Hi

2012/7/13 Juho Arjoranta <juho.arjoranta at helsinki.fi>:
> Lainaus "Ask Hjorth Larsen" <asklarsen at gmail.com>:
>
>> Dear Baldissin
>>
>> 2012/7/12 baldissin <baldissin at yahoo.it>:
>>> Dear gpaw users
>>>
>>> I am calculating the site preference of Al substitution in zeolite.
>>> Unit cells are massive ( 20x20x13 Ang ). My idea is using LCAO mode.
>>> I am wondering if the method is accurate enough to discriminate energy
>>> with difference of some meV (5-20).
>>> any opinion?
>>>
>>> thanks a lot
>>> Gael
>>>
>>> PS:   calc = GPAW( mode='lcao' ,  h=0.18, nbands=-5, xc='PBE',
>>> txt='atoms.out', kpts=[(0, 0, 0)], usesymm=True , charge=-1,
>>> maxiter=150   )
>>
>> Energies in LCAO mode are usually off quite significantly.  However
>> trends are reproduced very well in my experience.  You may need to do
>> it the hard way and compare to some grid-based calculations.
>>
>> Be sure to follow the advice for optimizing the performance of the
>> Poisson solver in such a large system:
>> https://wiki.fysik.dtu.dk/gpaw/documentation/lcao/lcao.html#notes-on-performance
>>
>> (Don't reduce the solver's tolerance, but make sure the number of grid
>> points is divisible with a high power of 2).
>>
> When parallelizing over domain should the number of grid points be
> divisible with a high power of 2 in the whole system or in the boxes
> that the system is divided into?

I was referring to the total number of grid points.

Although I have also been wondering about that, and I never completely
figured it out.  Empirically it appears pretty universal that if one
just ensures that the total number of grid points is divisible by 8,
or 16 in larger systems, then it can use enough multigrid levels.  I
have only seen one case where it didn't work as I expected (32 CPUs
parallelizing along a direction with less than 256 points).  By the
way, the number of multigrid levels is written to the log.

Note: As long as things are "reasonably fine" it doesn't matter much.
I just see a lot of cases where people end up with horrible
performance or even convergence trouble.

Regards
Ask