[Pw_forum] phonon accuracy and ecutrho for ultrasoft pseudopotential
nicola.marzari at epfl.ch
Tue Aug 1 18:08:25 CEST 2017
great that you did these tests - it's exactly the way to operate
if one wants to understand how things work.
> We can see that except for > ecutrho=75, dispersion for all other cases are quite close. The
biggest > deviation for ecutrho=100 to 450 is 8.2275cm^-1, while for
ecutrho=150 > to 450, the biggest deviation is almost as negligible as
0.7cm^-1. This > is quite surprising，because the tutorial link
recommended high > ecutrho=450 for converged calculation.
Well, not sure if the tutorial referred to the same pseudopotential, but
you should indeed trust your calculations.
> Besides, I extract X point frequency from the above dispersions and
> plot them with respect to ecutrho ( here is the image
> https://pasteboard.co/GDEqdDb.png )
> On the other hand, I can also calculate phonon frequency at X point
> (0,1,0) using ph.x directly with the same scf parameter as the above.
> Again plot them with respect to ecutrho (Here is the image
> https://pasteboard.co/GDDu3wt.png ).
> From these two plots, I got confused. The convergence behaviour of
> matdyn.x and ph.x at the same point is not the same. The convergence of
> dispersion is much quicker ( above ecutrho=200) then convergence of
> single point ph.x calc ( above ecutrho=400 ). Is it generally true?
when you do a matdyn calculation on the full 4x4x4 mesh you also impose
the acoustic sum rules on its fourier transform (i.e. on the real
space constants) - things like making sure that the interatomic force
constants are such that a translation leaves the energy in the
quadratic hamiltonian exactly unchanged.
when you do a single point ph.x calculation you cannot impose these
so it's all consistent - enforcing an exact condition makes your
convergence faster (to be honest, I do not have a proof for this -
but to give you an extreme example, if you had looked at the
3 lowest phonons at Gamma, those will always come out to be zero
from a mesh calculation with acoustic sum rule imposed (as they
should)), but they will typically take very high cutoffs to converge
in a single point calculations (and often they do not even converge
exactly to zero, due to numerical noise in the xc functional (I think)).
Prof Nicola Marzari, Chair of Theory and Simulation of Materials, EPFL
Director, National Centre for Competence in Research NCCR MARVEL, EPFL
More information about the users