[CPMD-list] geometry optimization - isotopes and convergence doubt

[Mark Kosmowski]

On 4/1/08, Axel Kohlmeyer <akohlmey at cmm.chem.upenn.edu> wrote:
> On Tue, 1 Apr 2008, Mark Kosmowski wrote:
>
>
> dear mark,
>
> MK> AK> by themselves do not provide. i.d. using DFT, pseudpotentials
> MK> AK> and finite (fft compatible) grids.
> MK> AK>
> MK>
> MK> Right - our group compares theoretical results to experimental result
> MK> to discuss the accuracy of current calculations.
>
> well, i'd like to remark that for the kind of systems you seem to be
> running, this is a bit unfair w.r.t. the calculations. plane-wave
> pseudopotential calculations are best suited for (dynamical) bulk
> systems, where you benefit most from good parallelization, lack
> of BSSEs and no linear dependency problems, that programs with gaussian
> based atom centered basis sets are facing.
>
> for isolated systems, those problems are much less pronounced and
> with CPMD you have to spend a lot of additional effort to compute
> "vacuum" and to decouple the interactions between periodic images.
>

The isolated thing was, I think, my misunderstanding of something - I
had proposed doing geometry / vibrational convergence studies on an
isolated system to determine what cutoff is required for that element.
 My thought was to then have known minimum cutoffs required for good
results with various elements.  As I recall from a response to this it
would be better to do the cutoff selection study (by parameter, i.e.
geometry convergence) on a real system.

The actual systems we are studying are molecular solid state systems.
We compare calculations to experimental inelastic neutron scattering.

> MK> AK> optimization which in turn depends on the (final) convergence for the
> MK> AK> wavefunction.
> MK> AK>
> MK>
> MK> I am using the default convergence parameters for wave function and
> MK> geometry.  (I watch for 10-4 in the geometry steps to know how close
> MK> the run is to being done.)
>
> the default convergence in CPMD is chosen to give a reasonable starting
> point for MD runs. since you are using (excessively?) high cutoffs
> and you want to have consistent and "accurate" results, you have to
> crank up the convergence on the wavefunction, or else you won't have
> accurate forces. in CP-dynamics it does not matter how accurate your
> starting forces are since you are oscillating around the ground state
> anyways. for BO-dynamics already you'll have to go to _much_ tighter
> convergence (which can be relaxed a bit again, when using wavefunction
> extrapolation) to conserve energy.
>
> the major improvement that you gain from the high cutoff is the
> reduced "ripple-effect", i.e. artefacts in the potential from
> finite number of plane waves in combination with the spline
> interpolation on the grid. the rule of the thumb is, that you need
> to converge a wavefunction at least two orders of magnitude tighter
> than the forces. i.e. you'd need 1.e-6 convergence when your after
> 1.e-4 accuracy in the forces.
>

So, before going into trying to choose a converged cutoff (wrt
geometry and vibrations), I should do a parameter convergence study of
wavefunction convergence criteria at one cutoff?  Can this value then
be used at other cutoffs or does each cutoff choice need a new
wavefunction convergence criteria study?

> MK> The only difference in the inputs was the isotope definition.  I have
> MK> a bash script that runs the deuterated case first, then copies to a
> MK> new directory and runs a sed script to change the ISOTOPE section from
> MK> 2H to 1H.  The current run is with light main group elements using LDA
> MK> DFT and pade pseudopotentials from the Goedecker directory (downloaded
> MK> from the CPMD website).
>
> hmm... that reminds me of another potential source of error:
> you are treating the hydrogen and deuterium atom cores both as
> classical particles. particularly for hydrogen that may not be
> sufficient and introduce additional systematic errors beyond
> the numerical accuracy.

If hydrogen behavior is important, are the Goedecker pseudopotentials
not reccomended?

>
> MK> Once the LDA work is done, I plan on also using at least the PBE
> MK> functional and associated pseudopotentials.  Are the pseudopotials of
> MK> different functionals different enough to require a full convergence
> MK> study by cutoff and k points as I am already doing for LDA / pade?
>
> convergence behavior should be similar.
>
> [...]
>
> MK>      MK> I am assuming that this just means that 50 Ry is not a high enough
> MK>      MK> cutoff energy for my system regardless what the parameter vs. cutoff
> MK> AK>
> MK> AK> that is more likely because of an instability in the lanzcos
> MK> AK> procedure and you'll have to tweak the lanzcos related parameters
> MK> AK> to get it to run more stable. it looks also like the code detected
> MK> AK> that and resetted the diagonalization.
> MK>
> MK> Thank you for explaining this.
>
> BTW: these calculations should become much faster with the next
> CPMD release as alessandro curioni and his group have extended k-point
> support for direct wavefuntion optimization (e.g. via ODIIS instead
> of LANCZOS) to geometry optimizations and BO-MD.

This will be nice - one of my systems is taking several days just for
the lanczos force initialization.

>
> cheers,
>   axel.
>
>
> --
> =======================================================================
> Axel Kohlmeyer   akohlmey at cmm.chem.upenn.edu   http://www.cmm.upenn.edu
>   Center for Molecular Modeling   --   University of Pennsylvania
> Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
> tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
> =======================================================================
> If you make something idiot-proof, the universe creates a better idiot.
>