[CPMD-list] question about CUTOFF

[Mark Kosmowski]

Axel:

AK> mark,
>
> MK> I've been choosing cutoff by making an unoptimized single energy vs. cutoff
> MK> energy plot (I'll usually use 25 Ry to 250 Ry in 25 Ry increments) and
> MK> deciding how "flat" I want the curve to be at my cutoff of choice.  This is
> MK> done knowing that the calculation as a whole will take longer to perform.
>
AK> while this is a convenient way of checking for cutoffs and in most cases
AK> it will mean, that will err to the side of having too large a cutoff.
AK> but more importantly, i would still recommend to make 1-2 small checks
AK> w.r.t. the property of interest on your final system (or some simpler
AK> version of it) since the convergence behavior of different properties
AK> can be very different from the convergence of the total energy. as
AK> stated before forces (for MD and geometry optimization) converge in
AK> general better. for ultrasoft pseudopotentials, one also has to consider
AK> the relation between (wavefunction) CUTOFF and DENSITY CUTOFF (by
AK> default a factor of 4. see DUAL). in most cases good convergence
AK> and energy conservation with ultrasoft pseudpotentials require the
AK> density cutoff to be relatively larger, ranging from DUAL 5 up to
AK> DUAL 10 for a stress tensor calculation (if somebody would actually
AK> fix the calculation of the stress tensor for USPPs in CPMD, that is).

Is force convergence also affected by the nature of the system in
addition to the atomic species involved?  For example, would ethane
(H3C-CH3) and acetylene (HC-triple bond-CH) be expected to behave
differently wrt force convergence at the same cutoff energy?

Regarding the small checks - are you saying, for instance, that for
small cell Z=2 systems (that I'm presently working with) a lower
cutoff energy with a greater number of k points may be a better use of
computational resources than high cutoff and fewer k points?
(Specifically, I am looking at geometry optimization followed by
vibrational analysis.)

For a simpler version of a system, would it be reasonable (for common
organic atoms) to do large cell simulated gas calculations (not using
the ISOLATED MOLECULE keywords / symmetry - just a molecule in a big
box) for the diatomic gasses (and perhaps methane and water for carbon
and oxygen after looking at H2 - thus avoiding the problems of O2) to
investigate cutoff vs. geometry accuracy?
>
> MK> Since cutoff choice is based primarily on atomic species, I do not
> MK> repeat this analysis for a new species unless the new species
> MK> contains elements not previously examined in this fashion.
>
AK> right.
>
> MK> >From my limited experiences, if I am concerned about a 5 Ry cutoff
> MK> difference, I am probably not on the "flat" enough part of the curve.
> MK>
> MK> For the Goedecker pseudopotentials available from the CPMD website I
> MK> have been using a cutoff of 150 Ry as an acceptable compromise.
> MK> The systems I am looking at are small molecule solids.
>
AK> as you have pointed out yourself. the choice depends on the atomic
AK> species in question. so while 150ry seems like a good guess for
AK> your systems there are goedecker type potentials that work well
AK> enough at 60ry and less and others that require over 200ry. so
AK> _any_ statement about using a "standard" or default cutoff shows
AK> at best a lot of experience, confidence or trust in other people's
AK> work, and at worst ignorace, i.e. that there were no checks made.
>
AK> it always puzzles me, that while it is quite difficult to test for
AK> basis set completeness and balance with atomic (gaussian) basis sets
AK> and not to mention BSSE errors, people tend to do these tests,
AK> whereas with plane waves, where BSSE is not an issue and the test
AK> would be easy as you essentially only have a single parameter that
AK> can be systematically variied, people don't do this...
>
AK> if anybody can clue me in on why this happens, i would
AK> be very thankful as it would help enormously to address
AK> this issue and provide better guidelines to people to
AK> use CPMD correctly. there is nothing worse than having
AK> a powerful tool at hand and not using it right.
>
Basis set vs. parameter checking sounds like a good idea for a thesis
chapter, especially since I'm the first student in my research group
to be using plane wave methods.  If interested, I could send a draft
either to you or the list for comments with the aim of editting it for
inclusion in the CPMD manual.  I do not know when I will have a chance
to do this though, as I work full-time while studying toward a PhD
part-time.

You mentioned that one cannot quantitatively project the improvement
of a calculation by using "better" gaussian basis sets (i.e. 6-31G**
instead of 6-31G) while one can with plane waves.  I've noted that
some people who have extensive experience with atom-centered
computational methods are rather unimpressed by an energy (single
point, unoptimized geometry) vs. cutoff energy curve.  I think there
is an attitude of, "Of course the energy decreases as you improve the
basis set."  Thus the ability to quantitatively project accuracy
improvement by choosing a point on a continuum (cutoff energy) is not
well appreciated.

This is further complicated since many people with atom-centered
calculation experience are chemists coming from a single molecule
background and need to learn about the intracacies of solid state /
periodic calculation.  In my experience, the most vexing thing to come
to grips with for solid-state calculation are k points.  And there are
other significant differences between single molecule and periodic
calculations that need to be learned.  Maybe the power of basis set
analysis with plane waves is neglected in view of all of the other
things that one must become familiar with for periodic calculation.

Thank you for the reply,

Mark