The ABINIT Users Mailing List ( CLOSED )

[Matteo Giantomassi <gmatteo@pcpm.ucl.ac.be>]

On Thu, 27 Sep 2007, deyulu@yahoo.com wrote:

> Dear Abinit users:
>     I have a question regarding the convergence on the unoccupied states in 
> GW calculations. In the GW tutorial, I saw two different ways to compute KSS 
> files.
>
> 1. compute all the bands within the SCF (e.g.: tgw_1.in), and save to KSS.
> 2. compute ground state density in the 1st dataset and perform a 
> non-self-consistent calculation using the converged density to obtain the KSS 
> file.
>    In the second aproach, one can normaly set "tolwfr" to control the 
> convergence. In the first one, however, I couldn't find a criteria to control the 
> non-self-consistent steps in SCF. After a look in vtowfk.F90, it seems to me that 
> within one SCF, the number of non-self-consistent loop (nnsclo_now) is often to be 1 or 
> 2.
>    If it is true, does it mean that we should NOT use the 1st approach to 
> generate KSS file, since the unoccupied states are unlikely to be converged?

Dear Deyu,

The first algorithm you mention (tgw_1.1.in) corresponds to the choice 
kssform=1 (default value)
In this case the KSS file is generated by performing a full or partial diagonalization of the Hamiltonian matrix 
expressed in reciprocal space. The KS hamiltonian is constructed starting from the GS density which requires a 
SCF calculation, but the diagonalization does not require any kind of self-consistency. 
To obtain accurate wavefunctions, both for occupied and empty states, you have only to be sure that 
the occupied bands, at the end of the GS run, are well converged. This conditions indeed assures a converged density.

The second method (kssform=3) consists, instead, in performing a SCF calculation for both filled and empty states. 
The number of states to be explicitly included during the GS run is much larger than that required by the first method; 
the advantage consists in the fact that there is no need to store and diagonalize the large Hamiltonian matrix. 
Furthermore the GS calculation can be parallelized over k-points thus reducing both the CPU time and the memory allocated.
In this case, as you wrote, one can easily control the convergence of the 
wavefunctions by using tolwfr as stopping criterion.

>    There is another issue related to this. I'm testing the band/FFT parallelization of 
> Abinit5.4.3 on powerpc-ibm-aix5.2.0.0 machines. While the it works fine for the SCF calculations, 
> "segmentation fault" errors occurred in non-self-consistent calculations with 
> "iscf=-2". Below are my hostname.ac file and input file.
Concerning this point, I'm afraid I can't help you. You have to wait for the 
reply of one of the developers who are working on this part.
Maybe it's better if you open a new post with a different title, just to be 
sure that people interested in band/FFT
parallelism will read the mail.

Best Regards, 
Matteo Giantomassi