[Q-e-developers] Problems converging anions in vacuum

[Stefano de Gironcoli]

I think that for some reason, real or spurious, ghost or negative 
affinity, there is a delocalized state close to the localized state that 
almost converges and then suddenly the new state gets occupied and self 
consistency goes bananas.
Notice however that the total energy of the first problematic iteration 
is LOWER than the one the system was converging to. so what happens is 
that the system suddenly finds a "better" solution and runs for it.

I would (sorry if you already did it)
- start from atomic+random wfcs
- compute more that the strict minimum nbnd
- apply martyna-tuckerman or counter charge correction to have the zero 
of the  potential properly set
- monitor eigenvalues at each iteration (iprint=1)
- study what happens as the cell size is increased.

stefano


On 05/12/2012 07:38 PM, Nicola Marzari wrote:
> Thanks to all! We'll keep looking into this.
>
> We are aware of DFT issues on self-interaction, loosely
> bound anions, and periodic-boundary interactions (these are
> actually research areas in the group).
>
> I think the problem here is different, although I'll double
> check with Oliviero that we are not on the verge of losing an
> electron to vacuum (I'm pretty sure this was checked).
>
> Namely, the calculations for all these anions are almost perfectly
> converged, and then in one iteration c_bands gives "1 eigenvalues not
> converged", and after that the calculation diverges/crashes (see below).
>
> If one looks at the charge density at the first iteration with c_bands
> not converged, one can see little ripples at high frequency appearing
> in the total charge density. Also the negative rho shoots up by at least
> one/two orders of magnitude. I presume these must be driven by the
> fact that one of the eigenstates in produced by c_bands is now
> meaningless - so the issue is why c_bands with CG close to convergence
> can fail - that would be my question.
>
> Oliviero and Davide noted that a different O pseudopotential
> does not give rise to this error, so I presume it could be due
> to a ghost state almost overlapping with the least bound electron,
> but it is surprising for this to happen in a large number of
> anions so close to perfect convergence.
>
> 			nic
>
>
>
>        iteration # 27     ecut=    30.00 Ry     beta=0.40
>        CG style diagonalization
>        c_bands:  1 eigenvalues not converged
>        ethr =  3.22E-10,  avg # of iterations =  6.0
>
>        negative rho (up, down):  0.462E-02 0.000E+00
>
>        add environment contribution to local potential
>
>        total cpu time spent up to now is     2842.4 secs
>
>        total energy              =     -94.77428965 Ry
>        Harris-Foulkes estimate   =     -94.77428964 Ry
>        estimated scf accuracy<        0.00000028 Ry
>
>        iteration # 28     ecut=    30.00 Ry     beta=0.40
>        CG style diagonalization
>        c_bands:  1 eigenvalues not converged
>        ethr =  3.22E-10,  avg # of iterations =  5.3
>
>        negative rho (up, down):  0.116E+00 0.000E+00
>
>        total cpu time spent up to now is     2921.7 secs
>
>        total energy              =     -94.78295397 Ry
>        Harris-Foulkes estimate   =     -94.77524064 Ry
>        estimated scf accuracy<        1.78307693 Ry
>
>
>> --
>> Goranka Bilalbegovic
>> Department of Physics, Faculty of Science,
>> University of Zagreb, Croatia
>>
>>
>> _______________________________________________
>> Q-e-developers mailing list
>> Q-e-developers at qe-forge.org
>> http://qe-forge.org/mailman/listinfo/q-e-developers
>