[CP2K:781] Re: Parameter choice for GAPW

[Teodoro Laino]

Axel,

just a warning:
There's a "serious" bug in GAPW, that Manuel discovered 1 year ago  
and I identified clearly a couple of months ago, when your system is  
interacting with the periodic replica (i.e. more than 0 order cell  
interactions). This of course depends on the cell size (usually must  
be quite small to pop up this error) and the system type/size.

Keep it in mind in case you should observe strange behaviors  
(compared with larger simulation boxes) for small cells (though this  
is not the usual kind of run I would say).

Moreover, for all people willing to use GAPW: don't expect it to be  
as easy as running GPW. If, for GPW, you need to keep a constant eye  
on your numbers, for GAPW you may want to keep the alert status  
continuously ON.
The main problem is that the method is far from being numerically  
stable and you may easily find systems where an ACCURATE tuning of he  
whole set of the parameters mentioned by Marcella is mandatory in  
order to get realistic numbers.

Just keep it in mind..
Ciao,
Teo

On 11 Mar 2008, at 02:38, Axel wrote:

>
>
>
> On Mar 10, 8:35 pm, Marcella Iannuzzi <marcella.ia... at psi.ch>
> wrote:
>> Hi Axel,
>
> hi marcella,
>
> thanks a _lot_ this is _very_ helpful.
> i'll have to make a few extra checks on
> some of the parameters i'm using, but it
> looks that most of my "gut-feeling" guesses
> based on my experience with USPPs were pretty
> close. seems like GAPW has a few more knobs
> that in USPPs are automatically determined. ;-)
>
> [...]
>
>> - if you use all electron and  OT, convergence is significantly
>> improved by setting the preconditioner to FULL_ALL
>
> hehe. i found that one already. with the preconditioner that was
> good for GPW the convergence was indeed really bad for GAPW.
>
>> I am available for further clarifications on the method and its
>> implementation.
>
> thanks again. i'll have to keep my tests running and
> see if they produce something meaningful. so far, so lucky.
>
> ciao,
>    axel.
>
>>
>> ciao
>>   Marcella
>>
>
> >