why MAX SCF=1 when tests the cutoff and rel cutoff for the convergence in the tutorial?

[Leopold Talirz]

Dear Jingyun,

just in order to make it clear to other readers -- you are referring to the 
tutorial http://cp2k.org/howto:converging_cutoff on the CP2K website 
entitled "How to converge the CUTOFF and REL_CUTOFF".

I think you raise a valid point here that would deserve attention from one 
of the many experts on this forum (and maybe a small update of the 
tutorial). 

One problem I see is that there are two possible meanings of "convergence" 
here, which may lead to misunderstandings.
Most users will be interested in converging physical properties of their 
system up to a given accuracy. These properties will of course be based on 
self-consistently obtained calculations.
However, this tutorial seems to concentrate more on converging the 
representation of the Gaussians on the real-space grid, which is defined 
completely by the basis set/pseudo potentials.

As far as I can see, the total energy is calculated here based on the 
initial "atomic guess" for the wave functions.
In order to link the convergence observed for this energy to the 
convergence of the self-consistently obtained energy, one would have to 
assume that no significant shift between "sharp" and "smooth" gaussians 
occurs during the self-consistency iterations.
While this may be the case, it should certainly be mentioned explicitly at 
some point in the tutorial.
One may also ask, why it is necessary in this case to use 8 Si atoms 
instead of just a single one.

Similar questions apply for the definition of the convergence criterion. In 
the tutorial, the choice of 250Ry is motivated as follows:

[...] *the total energy is converged, and the distribution of Gaussian 
functions on the grids are reasonable: it is the lowest cutoff energy where 
the finest grid level is used, but at the same time with the majority of 
the Gaussians on the coarser grids.*

In this particular case, only 2.5% of all Gaussians are on the finest grid. 
Taking this as a criterion for a calculation involving carbon with the 
TZV2P-MOLOPT-GTH basis set and GTH-PBE-q4 pseudo potential, I would need to 
go to a cutoff of more than 1500 Ry (with 60Ry relative cutoff) in order to 
have this few Gaussians on the finest grid.

While this may indeed be necessary to properly represent the sharpest 
Gaussians in the basis set and to converge the total energy to a level of 
10^-8 Hartree, many real-world applications might not require this level of 
accuracy.

Best,
Leopold



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20140423/bca2c013/attachment.htm>