Parameter choice for GAPW

[Axel]

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
that ALHPA0_HARD and ALHPA0_SOFT, LMAXN0, LMAXN1 in %FORCE_EVAL%DFT
%QS
seem to be the most important global options, and that LEBEDEV_GRID,
RADIAL_GRID, and HARD_EXP_RADIUS in %FORCE_EVAL%SUBSYS%KIND seem to be
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
RHO0_EXP_RADIUS
be understood in a similar way as the ultrasoft cutoff radius for
USPPs?

thanks in advance,
axel.

[Marcella Iannuzzi]

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 parameters.

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 10

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

in KIND

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  implementation.

ciao

 Marcella

[Axel]

On Mar 10, 8:35 pm, Marcella Iannuzzi <marcella.iannuzzi-ma...@psi.ch>
wrote:
> Hi Axel,

hi marcella,

thanks a _lot_ this is _very_ helpful.
i'll have to make a few extra checks on
some of the parameters i'm using, but it
looks that most of my "gut-feeling" guesses
based on my experience with USPPs were pretty
close. seems like GAPW has a few more knobs
that in USPPs are automatically determined. ;-)

[...]

> - if you use all electron and OT, convergence is significantly
> improved by setting the preconditioner to FULL_ALL

hehe. i found that one already. with the preconditioner that was
good for GPW the convergence was indeed really bad for GAPW.

> I am available for further clarifications on the method and its
> implementation.

thanks again. i'll have to keep my tests running and
see if they produce something meaningful. so far, so lucky.

ciao,
axel.

>
> ciao
> Marcella
>

[Teodoro Laino]

Axel,

just a warning:
There's a "serious" bug in GAPW, that Manuel discovered 1 year ago
and I identified clearly a couple of months ago, when your system is
interacting with the periodic replica (i.e. more than 0 order cell
interactions). This of course depends on the cell size (usually must
be quite small to pop up this error) and the system type/size.

Keep it in mind in case you should observe strange behaviors
(compared with larger simulation boxes) for small cells (though this
is not the usual kind of run I would say).

Moreover, for all people willing to use GAPW: don't expect it to be
as easy as running GPW. If, for GPW, you need to keep a constant eye
on your numbers, for GAPW you may want to keep the alert status
continuously ON.
The main problem is that the method is far from being numerically
stable and you may easily find systems where an ACCURATE tuning of he
whole set of the parameters mentioned by Marcella is mandatory in
order to get realistic numbers.

Just keep it in mind..
Ciao,
Teo

[janos...@gmail.com]

Dear All,

For many systems with with hard electron density GAPW seems to be a very promising approach. On the other hand side,

I have heard and read about its instability (such in this old post from Teo). I would like to ask the developers if these past issues are considered solved by now 

or if it is still not routine to run GAPW.

Best regards,

Janos