Dear Julien,
a possibility for what you see could come from the pseudopotentials - as
you increase the charge density cutoff your pseudopotential becomes
different (for the case of ultrasoft, say, with augmentation charges -
I guess similar considerations apply to PAW), so you are not comparing
the same exact external potential at increasing number of plane waves.
For convergence, you could switch to the CP code. It uses a variational
algorithm, rather than an iterative one, so, even if less efficient,
it is bound to converge. If the system is metallic, you should switch to
ensemble dft, but I'm not sure it's really implemented in full there
(especially with US or PAW).
nicola
On 28/02/2019 07:27, Julien Barbaud wrote:
Thanks very much to everyone for their answers,
I have tried the different fixes suggested (higher ecuts, trying out
different smearings, raising mixing_ndim, changing the mixing_mode,
etc...). Unfortunatley, none of them worked...
Following Pietro's recommandation, I also designed an input geometry
perfectly cubic with no distortion and the MA+ ion oriented along the [1
1 1] direction for higher symmetry. I based the quantitative part of the
geometry on other cif files available (i.e. for the lattice constant,
bond lengths, etc...)
Benefitting from the high symmetry, the algorithm worked faster, but was
still unable to converge to a threshold of 1e-7Ry in 500 iterations...
I finally relaxed the condition to 1e-6Ry in order to reach a scf
convergence, and see if I could get more informations on the problem
from it.
Under those conditions I plotted the scf total energy (converged to
1e-6Ry) as a function of ecut. I got the curve attached. This seems
surprising to me because the curve is raising past 45Ry
I thought that the total energy would only have a monotonic trend,
decreasing with higher ecut, as a consequence of the variational
principle (indeed, the energy of the ground state should be a minimum
for all possible energies, so a truncated version of the ground state
wavefunction should yield a higher value of energy). Here it seems to
break that rule, and to "converge" to a value that is not the minimum
among all possible wavefunctions. This can not be attributed to the
accuracy of the calculation, because the scf is converged down to
1e-6Ry, and the augmentation observed is significantly above that threshold
Did I misunderstand something, or is it another sign that something is
seriously wrong with the calculation ?
(by the way, the curve stops after 65Ry, because scf failed to reach
convergence again at ec=70Ry, even for a threshold at 1e-6 and 500
iterations).
What other options should I try to solve this problem ?
Julien
Le 22/02/2019 à 16:45, Pietro Delugas a écrit :
Dear Julien
even if the scf loop converges you have still to check that the
k-point sampling and the plane wave basis set guarantee you an
accurate result.
obviously before worrying about accuracy you would like to have a
converged density.
You could try to start with a more symmetric cell, use a cubic cell
without distortions and align the molecule along one of the
diagonals of the perovskite box.
On 22/02/19 08:22, Julien Barbaud wrote:
Thank you Pietro for your experienced advices,
I had tried to increase the kmesh size before but only up to sizes
of 7x7x7. Reading your suggestions, I ran additional tests up to
10x10x10 but this did not show any sign of improvement on 70
iterations. As shown in file kmesh.png, the estimated accuracy is
still stagnating after a while and the 10*10*10 is actually giving
arguably worse results than the 9*9*9 although this is most likely
not significant. Actually, some papers report DFT simulation of
MAPbI3 using 6x6x6 kmesh
(https://aip.scitation.org/doi/full/10.1063/1.4864778), or even
single gamma-point calculation
(http://people.bath.ac.uk/aw558/publications/2013/aplm_perovskite_13.pdf),
so I guess this should not be the obstacle to convergence here.
Regarding the orientation of MA, I definitely agree with you, but I
don't think it can prevent the system from converging ? Sure enough,
it can have an important influence on the precision of the results in
later uses. But I would like to achieve convergence on this simple
single cell first, before building up supercells to take more complex
effects into account. A crystal with perfectly aligned MA might not
reflect the true experimental system, but it should still be a
possible configuration that the QE code should be able to compute, am
I wrong ?
As to your suggestion on VdW corrections, I just gave it a try, but
unfortunately, this is unconclusive too. I report the accuracy at
each iteration in vdw.png. Again, the accuracy stops improving after
a while. Plese note that I had to change my pseudo-potentials to use
'xdm' correction (which only supports PAW PP). the input file for
this test is included as attached file
Julien
Le 21/02/2019 à 16:35, Pietro Davide Delugas a écrit :
Hi
Have you tried to increase the k_point mesh ? 4 4 4 seems a little
bit lax as mesh for MAPbI3.
If I remember well I am afraid that to get convergence you will need
something like 10X10X10.
As for the structure neighboring methylammoniums like to orient
differently one from the other, you should probably use a larger
cell. Also consider to add some correction for van der Waals
interactions see here (
https://www.quantum-espresso.org/Doc/INPUT_PW.html#idm45922794348896)
hope it helps
Pietro
On 02/21/2019 04:17 AM, Julien Barbaud wrote:
Dear users,
I am new to QE, and trying to run a simple scf calculation on a
CH3NH3PbI3 crystal (semi-conducting material). I am using ultrasoft
pseudopotentials based on the exchange-correlation functionnal PBEsol.
I set up a first input, with values of parameters inspired from
literature on the subject. However, I could not reach convergence
after 100 iterations. The estimated error was actually "exploding"
to very high values, indicating a serious problem. I tried several
changes but was unsuccessful:
* varying plane-wave cutoff energy does not solve the problem
(cf attached ecut.png, giving the estimated error as a function
of the number of iterations. It is shown here only on the first
15 iterations as the results pretty much only stall from there)
* varying cutoff energy for charge (cf ecutrho.png)
* taking larger k-point sampling (not shown)
* I also read that for metallic or "close to metallic
conductors", there might be problems with the first unoccupied
states that can be solved by adding a few empty bands. My
system being a semi-conductor, I tried adding additional bands
using a m-p smearing but no improvement was found (not shown)
The only change that I found effective was to reduce the
mixing_beta factor.
It effectively prevents the error from diverging to very large
values, but I still do not reach convergence, even after longer
iterations. I tried much smaller values of mixing beta which
improves the final value of the error, but I still cannot reach
convergence on 100 iterations. As shown in the mixbeta2_zoom.png,
the error reduces to smaller values around ~1e-5~1e-6, but it keeps
stalling after a while. I do not observe a well-converging
behaviour for any value.
I attached the "default version" of my script on which the various
modifications described above have been independently performed. I
obtained the geometry from a CIF file in literature and checked it
with visualization software; it seems perfectly ok as far as I can
tell.
Any insight on what I did wrong would be really helpful. I suspect
a shameful beginner mistake, but can not find it out.
Thanks in advance,
Julien barbaud
P.S: this is my first time posting on this user list. Please let me
know if my question is not suitable for it, or can be improved
either in its content or presentation. I will gladly take any
recommandation into account in order not to negatively impact the
quality of this user list !
_______________________________________________
users mailing list
[email protected]
https://lists.quantum-espresso.org/mailman/listinfo/users
_______________________________________________
users mailing list
[email protected]
https://lists.quantum-espresso.org/mailman/listinfo/users
_______________________________________________
users mailing list
[email protected]
https://lists.quantum-espresso.org/mailman/listinfo/users
_______________________________________________
users mailing list
[email protected]
https://lists.quantum-espresso.org/mailman/listinfo/users
_______________________________________________
users mailing list
[email protected]
https://lists.quantum-espresso.org/mailman/listinfo/users
--
----------------------------------------------------------------------
Prof Nicola Marzari, Chair of Theory and Simulation of Materials, EPFL
Director, National Centre for Competence in Research NCCR MARVEL, EPFL
http://theossrv1.epfl.ch/Main/Contact http://nccr-marvel.ch/en/project
_______________________________________________
users mailing list
[email protected]
https://lists.quantum-espresso.org/mailman/listinfo/users
