[Pw_forum] new bfgs: strange behavior doing vc-relax
Stefano de Gironcoli
degironc at sissa.it
Tue Apr 19 11:03:41 CEST 2011
dear Max Popov
have you bothered looking at the output ? has the bfgs procedure
converged ?
Notice that in recent versions of vcrelax one additional scf
calculation is performed at the final configuration reinitializing
reciprocal lattice vectors .
this has nothing to do with the bfgs procedure. it's a check to
verify what is the impact of the deformation of the cutoff "sphere"
induced by the cell-shape change.
if the cutoff is large enough the change should be very small. 0.5
mRy looks small to me.
Stefano de Gironcoli
SISSA & DEMOCRITOS
On 04/18/2011 09:35 PM, Laurence Marks wrote:
> The latest BFGS has a curvature trap which can change the behavior
> particularly if you have too few k-points. Can you please do
>
> grep -e "! tot" -e bfgs WHATEVER.out
>
> (where WHATEVER is relevant) and send it to me, perhaps at my private
> email (or the full .out file). I am not an expert with QE, but might
> be able to say something.
>
> 2011/4/18 Максим Попов<max.n.popov at gmail.com>:
>> Dear QE developers,
>>
>> comparing vc-relax calculations between BFGS(4.2.1) and the latest BFGS(from
>> CVS) I faced with a following problem:
>> the newest QE calculates one extra point after achieving convergence
>> criteria (etot_conv_thr,force_conv_thr,press_conv_thr),
>> which leads to slightly higher total energy.
>>
>> Here are the two latest energies from outputs:
>> v.4.2.1 CVS(18.04.2011)
>>
>> -242.66342170 Ry -242.66342267 Ry
>> -242.66342336 Ry -242.66284562 Ry
>>
>> One can see that there is a good agreement between the last energy by
>> v.4.2.1 and next to the last by CVS-version, whereas
>> the last energy by CVS is higher by ~ 0.5 mRy.
>>
>> I tried relax (cell shape fixed) also - it works fine, no extra point
>> calculation.
>>
>> --
>> Best regards, Max Popov
>> Ph.D. student
>> Materials center Leoben (MCL), Leoben, Austria.
>>
>> _______________________________________________
>> Pw_forum mailing list
>> Pw_forum at pwscf.org
>> http://www.democritos.it/mailman/listinfo/pw_forum
>>
>>
>
>
More information about the Pw_forum
mailing list