BigDFT SCF convergence and parameters

[dcaliste]

I'm transfering here the mail of Ronaldo Giro sent to the mailing list:
«
I am trying to use BigDFT with Abinit and I have some doubts:

1) I do not need any convergence study, like for wvl_hgrid wich is the similar
as ecut in psp-normconserv.? I just need convergence study for k points if my
system is not a "molecule in a big box" ?

2) There is some gain in CPU time and hardware resources like memory if I use
BigDFT instead of plane wave formalism to deal with systems containing many
atoms and a vacuum layer?

Thanks in advance,

Ronaldo Giro
»

[dcaliste]

1) I do not need any convergence study, like for wvl_hgrid wich is the similar
as ecut in psp-normconserv.? I just need convergence study for k points if my
system is not a "molecule in a big box" ?

The current version of wavelets inside ABINIT (BigDFT 1.2) does not take into account the k points. This is a current development of BigDFT stand-alone and will be available with version 1.4. The inclusion of this future version inside ABINIT is not planned yet.

Regarding the convergence study, yes, wavelets requires also to check convergence. With respect to two main parameters, wvl_hgrid, which define the quality of the grid (like ecut defines the quality of the plane-wave projection) and wvl_crmult which is related to the spatial expansion of the grid. Typical values for hgrid varies from 0.50 Bohr to 0.25 Bohr, depending on your pseudo (silicon is happy with 0.45 while carbon prefers 0.35 for high quality results). The typical values for wvl_crmult are betwen 4 and 7 (multiplying factor).
An important point regarding to convergence study with wavelets is that, both wvl_crmult and wvl_hgrid should be improved together, otherwise the error raised by one will mask the improvement of the other one.

2) There is some gain in CPU time and hardware resources like memory if I use
BigDFT instead of plane wave formalism to deal with systems containing many
atoms and a vacuum layer?

Yes, mainly. The better improvements are definitely achieved for dipolar surfaces since the Green function treatment in the Poisson solver is exact in that case, while the vacuum part for super-cell approachs must be wide. Moreover, there no convergence study with respect to vacuum size to be done.
But, currently, wavelets in ABINIT are not using the surface conditions. This is nevertheless possible with BigDFT stand-alone.