The GW tutorial : addressing cross-convergence

Documentation, Web site and code modifications

Moderators: baguetl, routerov

Locked
gabriel.antonius
Posts: 58
Joined: Mon May 03, 2010 10:34 pm

The GW tutorial : addressing cross-convergence

Post by gabriel.antonius » Wed Feb 15, 2012 2:55 pm

Dear all,

There is some issue with the GW tutorial I would like to discuss. Although it does a good job at showing all the parameters of a GW calculation, it says nothing about the interdependence of these parameters during a convergence study. It might be worth mentioning that ecuteps, nband in sigma and nband in screening must be crossed-converged. I've come with the following procedure to ensure reliability of the converged parameters.


1) Produce a KSS file with high ecut (30 Ha) and number of bands (800).
Note that for such a high number of bands, one should carefully look at the warnings
to make sure the wave functions were converged.

2) Produce a SCR file with moderately high ecuteps (10 Ha) and number of bands (500),
keeping ecutwfn = ecut

3) 3.1) (optional) Make a convergence study on ecutsigx.
Note that this parameter does not depend on ecuteps or any parameter
used in the screening.

3.2) Make a convergence study on ecutwfn with a high number of band (600),
If the maximum value of ecutwfn you can pick (the ecut used in the kss)
turns out to be under-converged, go back to step 1), raising ecut..

4) Make a convergence study on nband used in sigma.
If nband in the kss turns out to be under converged, go back to step 1, raising nband.
If nband used at step 3.2 was under converged, go back to step 3.2, raising nband.

5) Make a convergence study on ecuteps.
If ecuteps used at step 4 was under converged, go back to step 4, using a screening
file with sufficient ecuteps.

6) Make a convergence study on nband used in the screening.
If nband in the screening used at step 4 was under converged, go back to step 4,
using a screening file with sufficient nband.
It will probably be smaller than nband used in sigma

7) (optional) Make a convergence study on ecutwfn used in the screening.
It will most probably be smaller than ecutwfn used in sigma.

8) Make a convergence study on the k-point grid used throughout the GW procedure.
This parameter is quite independent of the rest.


Note : If you are not using a plasmon-pole model, you must also do a convergence study on
the number of frequencies used, which has some interdependence with the
number of band used in the screening and ecuteps.


So the idea here was to start with high parameters in order to converge down the calculation,
and avoid having to reiterate the procedure.
Are these starting parameter too heavy for common systems?
Let me know what you think.
Gabriel Antonius
Université du Québec à Trois-Rivières

Locked