[gpaw-users] difference in convergence behavior between versions

[Marcin Dulak]

Hi,

On 07/17/2014 12:39 PM, Alfredo Palace Carvalho wrote:
> hi,
>
> I got this "problem" running in both versions of gpaw the same script, 
> using both a non-default mixer settings:
>
> Mixer(beta = 0.2, nmaxold = 5, weight = 100.0)
>
> and an explicit eigensolver choice:
>
> eigensolver = 'dav'
>
> I attach the script (very similar to the one for the PW problem in my 
> other message).
> Probably there are differences in the default settings of the davidson 
> eigensolver in the two versions, but would they be enough for such a 
> different convergence path?
Tristan's explanation looks very plausible, without seeing the actual 
CONTCAR structure.
This setting: convergence = {'energy': 1.0e-5},
is dangerous - the forces obtained from such calculation may be not 
accurate enough if SCF is stopped prematurely by this threshold.
The best way of using GPAW should be with defaults (i mean no setting of 
bands or convergence should be necessary for standard jobs).
Another thing: if the system is spin-polarized one should not use Mixer, 
but MixerSum or MixerDif.

Note that the mixing parameters has been changed only for periodic 
systems - i see

"Molecule single point FINISHED"

in your script. If you calculate a molecule use non-periodic cell.

Best regards,

Marcin

>
> Cheers,
> Alfredo
>
> On 2014-07-16 19:26, Tristan Maxson wrote:
>> Hello,  a major difference between gpaw-0.9 and gpaw-0.10 is the
>> default eigensolver changed.
>>
>> Try running your calculations with the new gpaw version but ensure the
>> following settings are set.
>>
>> Mixer : Use a non-default mixer for both,  or ensure the new gpaw is
>> using Mixer(0.1,3,50) (The gpaw-0.9 default)
>>
>>  Eigensolver : Ensure both are using the same mixer,  the 0.9 default
>> was RMM-DIIS and the 0.10 default is davidson.  If you are using
>> RMM-DIIS for both, import the RMM-DIIS solver from
>> "gpaw.eigensolvers.rmm_diis_old" in the 0.10 version as the
>> eigensolver was improved in 0.10 but the old one is still available.
>>
>> Try setting these and see what your results are.
>>
>> source:
>> https://wiki.fysik.dtu.dk/gpaw/devel/releasenotes.html#version-0-10-0
>> [3]
>>
>> On Wed, Jul 16, 2014 at 9:50 AM, Alfredo Palace Carvalho
>> <ajpalace at uevora.pt [4]> wrote:
>>
>>> Hi,
>>>
>>> An issue unrelated to my previous email, I don't really know if it
>>> is just the way it is/should be or if there's really something fishy
>>> in here:
>>> I have recently updated gpaw to the newest stable (0.10.0.11364),
>>> previously I had been using version gpaw-0.9.1.11146.
>>> A single point calculation which took, with the older version, 26
>>> SCF cycles to converge, now takes 50 iterations with the new
>>> version.
>>> The final results are close but not equal: -235.026624 vs.
>>> -235.026593
>>> At first I thought it could be due to different setups files, but
>>> using the new version with the new setups and the old ones I just
>>> get the same results.
>>> While the two versions of gpaw have been using also different
>>> versions of ASE, I think I have also excluded the influence of such,
>>> because running the old gpaw with the new ASE gives the same results
>>> (less iterations) than if using the old ASE.
>>> So, the difference has to be due to changes in gpaw.
>>> I would have accepted it as normal if there was only some slight
>>> difference. But, now, it takes the double of the previous number of
>>> iterations. I thought it was too much! What do you think? Is it not
>>> surprising?
>>>
>>> Cheers,
>>> Alfredo
>>> _______________________________________________
>>> gpaw-users mailing list
>>> gpaw-users at listserv.fysik.dtu.dk [1]
>>> https://listserv.fysik.dtu.dk/mailman/listinfo/gpaw-users [2]
>>
>>
>>
>> Links:
>> ------
>> [1] mailto:gpaw-users at listserv.fysik.dtu.dk
>> [2] https://listserv.fysik.dtu.dk/mailman/listinfo/gpaw-users
>> [3] 
>> https://wiki.fysik.dtu.dk/gpaw/devel/releasenotes.html#version-0-10-0
>> [4] mailto:ajpalace at uevora.pt
>
>
> _______________________________________________
> gpaw-users mailing list
> gpaw-users at listserv.fysik.dtu.dk
> https://listserv.fysik.dtu.dk/mailman/listinfo/gpaw-users


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listserv.fysik.dtu.dk/pipermail/gpaw-users/attachments/20140717/403ad927/attachment-0001.html>