[Pw_forum] how to fix the occupations in DFT+U (Matteo Cococcioni)

simone marocchi simone.roz at gmail.com
Sun Nov 9 20:31:33 CET 2014


Dear Matteo Coccocioni, I am working with a molecule of point group
symmetry D_4d, the TbPc2. Unfortunately the symmetry operations implemented
so far in Quantum Espresso can not handle screw rotations and glide
reflections as in a standard Quantum Chemistry package. So I am trying to
impose such occupations as an approximation to the real symmetry of the
system. Hereafter I will try also your "dirty trick" suggesting a
starting_ns_eigenvalue bigger than one. Using the mixing_fixed_ns >
electron_maxstep I obtained big differences in the occupations imposed and
obtained, interestingly although I did not end in the predetermined states
the system fell always in the same strange occupation. The convergence was
quite nice (tens of iterations) up to 10^-6 Ry. Perhaps, as you pointed out
I have a different symmetry than the code finds for my system.
Unfortunately, although I would like to try to develop some new parts
codes, I do not feel enough trained and comfortable with the timetable of
my current project. Thanks a lot for your suggestion !

S

Dear Simone

I try to answer your questions below.

On Fri, Nov 7, 2014 at 3:32 PM, simone marocchi <simone.roz at gmail.com>
wrote:

> Dear all, I am simulating compounds with rare earths within a collinear
> calculation. I tried to suggest some occupations of the f orbitals for the
> Tb atom, using the starting_ns_eigenvalue(m,ispin,I).
>


why do you need to do that? if you want to force different values of
occupation on states that are equivalent by symmetry you will not be
successful. If this is the case you have to do something that makes the
system loose that symmetry operation connecting the two states.



> Unfortunately also with a small value of electronic mixing and big values
> of U, the imposed occupations was lost during the iterative cycle.
>


this can happen. a dirty trick I learned recently is to suggest a
starting_ns_eigenvalue(m,
ispin,I) bigger than one in input (1.2 or 1.3 maybe). This does not make
any physical sense of course. However the Hubbard potential becomes more
attractive for the specific eigenvector of the occupation matrix you want
to fill completely and the code takes more time to "come back" to a
physical value. If that is a state it likes (at least a local minimum of
the energy) it might fall into it. Of course you have to check at the end
that the occupation has gone back to a value <= 1.



> So I used also mixing_fixed_ns > electron_maxstep. Also in this case after
> I obtain the total energy convergence the density matrix of the last
> iteration is different to the one imposed in the input_file.
>


this is (possibly) strange. How different is it? How well are you
converging?
The fact that it is different is not surprising: the routine that prints
the ns, always prints the ones that are computed from the KS states just
obtained from the new diagonalization. These can be different from the ones
e.g. used in contructing the Hubbard potential.
However if your calculation is well converged this difference should not be
big: both KS wfcs and their occupations should be converged reasonably well.
If this does not happen and you still see a significant difference maybe it
means that the values you are trying to impose is not consistent with what
the system wants to do (e.g., you have less symmetry than the code finds
for your crystal)




> Can someone of you kindly explain me how the mixing_fixed_ns works ? Do it
> uses a sort of Lagrange multipliers to force the occupations or is more
> like a penalty function ? Finally, is it possible to work around the
> problem, converging to a determined occupation ?
>


no with the current version of the code. you could implement some (e.g.,
quadratic) constraint and try. I think I once tried (long time ago) and
seem to remember problems in convergence.

Best,

Matteo


-- 
Simone Marocchi

S3 Center, Istituto Nanoscienze, CNR
via Campi 213/A, 41125, Modena, Italy
Tel: +39 *0592055585 <0592055585>*;  Skype: jacobi84
URL: http://www.nano.cnr.it
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20141109/2e854432/attachment.html>


More information about the users mailing list