[CP2K:757] Parameter choice for GAPW

Marcella Iannuzzi marcella.ia... at psi.ch
Tue Mar 11 00:35:53 UTC 2008

Hi Axel,

Playing with the GAPW parameters requires some insight in the formalism.
I will try to summarize briefly the role of the most relevant GAPW  
You are right they are grouped in DFT%QS and in KIND
In QS section
- EPSFIT: this parameter is used to determine which primitive  
Gaussians are considered hard and which one soft. With the default  
values exponents >= 4 are considered to be hard.
- EPSISO, EPS_SVD and MAX_RAD_LOCAL are parameters used in the  
construction of the GAPW projectors. Playing with them might help in  
difficult cases that turn out to have instabilities. Otherwise they  
can be safely left to default values
- in GAPW local compensation densities are needed in order to compute  
correctly the hartree contributions that involve the local densities.  
In the original scheme we have a hard compensation density, expanded  
in Gaussians with fixed exponent ALPHA0_HARD, and a soft compensation  
density that has the same multipole expansion as the hard one and it  
is use also on the fft. Having two different compensation densities  
introduces some additional terms in the expression of the total  
energy. These terms  are rather expensive to be computed because they  
involve three center neighbor lists. By default  ALPHA0_HARD is set to  
and ALPHA0_SOFT=ALPHA0_HARD, so that all the complex 3-center terms  
disappear.  Since we noticed very small differences by having or not  
having these terms, by default we set equal hard and soft exponents ?  
IF FULL_GAPW is set to true, all the terms are computed. If the  
exponents are not given from input, they are automatically set to  
meaningful values.
- LMAXN0 and LMAXN1 are maximum angular momenta used in the expansion  
of compensation charge and local charges. The maximum possible value   
is 2 times the maximum angular momentum present in the basis set.  
Having LMAXN0 or LMAXN1 smaller than this can improve the performance,  
but might have an effect on the accuracy.
-QUADRATURE  this variable defines the algorithm used to assign the  
points on  the local radial grid

HARD_EXP_RADIUS is the radius of the region around the atom where the  
local densities are defined. The  spherical regions centered on  
neighboring atoms do not have to overlap, therefore the radius cannot  
be too extended..

LEBEDEV_GRID RADIAL_GRID are parameters that determine se size of the  
local grids, The optimal choice  depends on the system and on the  
maximum angular momentum that need to be collocated. For not too  
complex systems  one can use about 100 points on the radial grids, and  
50 on the angular.

Other useful info:
- GAPW does not need a large PW cut-off, since the hard density is  
computed on the local atomic grids.
- Comparing GAPW and Gaussian calculations for small molecules you  
should find differences of a few microhartree. Due to the definition  
of the local densities, GAPW improves systematically by improving the  
basis set
- if you use all electron and  OT, convergence is significantly  
improved by setting the preconditioner to FULL_ALL

I am available for further clarifications on the method and its   


On 10 Mar 2008, at 18:35, Axel wrote:

> hi!
> i'd appreciate it, if somebody could shed some light on how to
> justify
> parameter choices for GAPW calculations. the input reference mostly
> contains statements of the kind XXX sets the value for 'xxx', which
> are
> probably of most use to the person who knows the implemention very
> well
> (which i obviously don't :-) ).
> from looking through the regtests and the input reference it seems to
> me
> %QS
> seem to be the most important global options, and that LEBEDEV_GRID,
> the most
> important per atom type parameters.
> are there any do's and don'ts associated with those parameters, e.g.
> what needs to
> be changed when going from a bulk water system to a compound with
> "heavy" elements?
> which of those parameters are related to higher angular momentum and
> thus need to
> be increased or decreased? can the HARD_EXP_RADIUS and its sibling
> be understood in a similar way as the ultrasoft cutoff radius for
> USPPs?
> thanks in advance,
>     axel.
> >

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20080311/b1ea30c3/attachment.htm>

More information about the CP2K-user mailing list