Strange features in IP-RPA epsilon_2 for large system
Posted: Tue May 05, 2015 9:46 pm
Dear all,
I am computing the independent-particle optical properties (yambo -o
c) of a metal slab with adsorbates.
The system relates to the one of a problem with RAM/CPU cost:
viewtopic.php?f=13&t=1007
System size is 20-30GB of SAVE folder according to specs.
Now I am reporting about the strange results I obtain in terms of
IP-RPA dielectric function. I anticipate that I started by getting
the problem with yambo-3.4.1, then this was reproduced with the new
version yambo-4.0.0. This one I am reporting here in detail.
In principle, at the IP-RPA epsilon_2 should increase monotonically
with the number of bands. This is what I get for a small number of
K-points (here, 10KP): the pw.x scf/nscf was computed with 593 bands,
and if I ask to use 563 indeed epsilon_2 is lower as expected. This
is best seen in log scale in the upper panel of the figure.
However, if I increase the number of K-points (here, 54KP) with the
same number of bands, I obtain nasty peaks in the region 5-9 eV as
shown in the lower panel. Additionally, the position and magnitude of
these peaks depends on the number of bands included, in a
non-monotonic way. Let me call this "the problem".
The input-outputs and additional information are collected in a zip
which can be downloaded from:
https://www.sendspace.com/file/xglz57
(beware of ads)
Remarks:
1) The problem was reproduced with:
- yambo 3.4.1 and yambo 4.0.0
- various other related systems, also with different unit cell sizes
- different machine (galileo.cineca.it / in-house linux cluster)
2) In all cases, I'm moving out the KB files to avoid computing the
Vnl contributions at this stage.
3) if I do nscf with M<N bands and run yambo with these M bands, I
obtain the same result as with the nscf with N bands and yambo with M
bands (set in input in BndsRnXd).
4) apparently if I choose a Gamma-centered or shifted kpoint mesh, it
does not matter for the appearance or not of the problem.
5) if I edit the code so to include only few specific k-points when
summing over the transitions, the problem persists. In particular, if
the problem is present with many KP and I include only transitions
from one KP, the problem is still there (although a single kpoint
would not produce a problem).
6) if I edit the code so to include only the bottom M bands out of a
yambo run with N bands [ie ic<M], then [when the problem exists] I obtain
different results than directly asking M bands in input (BndsRnXd)
7) I may see strange peaks also if I include in the summation over
transitions only bands which should produce excitation energies
e_c-e_v lower than the energy at which the peaks appear.
I look forward some suggestion.
On the one side one could avoid the problem by splitting the nscf calculation
in various small sets of KP, but this does not look satisfactory...
Thank you for your consideration.
Kind regards,
Guido Fratesi
I am computing the independent-particle optical properties (yambo -o
c) of a metal slab with adsorbates.
The system relates to the one of a problem with RAM/CPU cost:
viewtopic.php?f=13&t=1007
System size is 20-30GB of SAVE folder according to specs.
Now I am reporting about the strange results I obtain in terms of
IP-RPA dielectric function. I anticipate that I started by getting
the problem with yambo-3.4.1, then this was reproduced with the new
version yambo-4.0.0. This one I am reporting here in detail.
In principle, at the IP-RPA epsilon_2 should increase monotonically
with the number of bands. This is what I get for a small number of
K-points (here, 10KP): the pw.x scf/nscf was computed with 593 bands,
and if I ask to use 563 indeed epsilon_2 is lower as expected. This
is best seen in log scale in the upper panel of the figure.
However, if I increase the number of K-points (here, 54KP) with the
same number of bands, I obtain nasty peaks in the region 5-9 eV as
shown in the lower panel. Additionally, the position and magnitude of
these peaks depends on the number of bands included, in a
non-monotonic way. Let me call this "the problem".
The input-outputs and additional information are collected in a zip
which can be downloaded from:
https://www.sendspace.com/file/xglz57
(beware of ads)
Remarks:
1) The problem was reproduced with:
- yambo 3.4.1 and yambo 4.0.0
- various other related systems, also with different unit cell sizes
- different machine (galileo.cineca.it / in-house linux cluster)
2) In all cases, I'm moving out the KB files to avoid computing the
Vnl contributions at this stage.
3) if I do nscf with M<N bands and run yambo with these M bands, I
obtain the same result as with the nscf with N bands and yambo with M
bands (set in input in BndsRnXd).
4) apparently if I choose a Gamma-centered or shifted kpoint mesh, it
does not matter for the appearance or not of the problem.
5) if I edit the code so to include only few specific k-points when
summing over the transitions, the problem persists. In particular, if
the problem is present with many KP and I include only transitions
from one KP, the problem is still there (although a single kpoint
would not produce a problem).
6) if I edit the code so to include only the bottom M bands out of a
yambo run with N bands [ie ic<M], then [when the problem exists] I obtain
different results than directly asking M bands in input (BndsRnXd)
7) I may see strange peaks also if I include in the summation over
transitions only bands which should produce excitation energies
e_c-e_v lower than the energy at which the peaks appear.
I look forward some suggestion.
On the one side one could avoid the problem by splitting the nscf calculation
in various small sets of KP, but this does not look satisfactory...
Thank you for your consideration.
Kind regards,
Guido Fratesi