[Wien] Deeper DFT question

[Gerhard Fecher]

Hi Laurence,
I cannot follow completely youre observation.
The things below are not a real counter example, but they also do not a confirm that a distorted structure needs more cycles to converge.
To me it seems that other parameters influence the convergence behaviour more, also without knowing why but at best from experience,
indeed for a restricted amount of systems.
Indeed the calculation itself will take much longer if using a "wrong" structure with  inappropriate symmetry
(mostly because of much more k-points etc.).

One strange thing appears, that I observed in many calculations, when changing some parameter and using dstart before restarting the calculation
then one suddenly needs (sometimes very much) more cycles to converge if the case.scf was not deleted (cleanlapw is not enough) !
Therefore, one has to take much care if comparing things.

For all calculations with Wien2k_11 (latest version the RMT's were 2.5, case.inst without magnetisation, no shifts of kmesh.
(Note: if using the same directory with changed parameters e.g. k-points one needs to delete the scf file,
clean_lapw and a new dstart is not enough !!
If you don't delete the case.scf you may suddenly need more cycles to converge, it seems that Wien takes some history
of the energies and charges from the scf file !)

First
making a fast test with Li one needs 9 cycles for the experimental lattice parameter (3.51 from webelements)
but only 7 cycles for the otimized lattice parameter, 
(8000 k-points,  Rkmax=7, -ec 0.00001 -cc 0.001)
this seems to confirm youre observation

Second
making a Volume optimization (-12% to 8% in steps of 2% from 3.51 A)
using dstart (NOT charge extrapolation)
(16000 k-points, Rkmax=8, -ec 0.00001 -cc 0.001)
The numer of cycles to converege were
5, 6, 5, -5-, 6, 5, 6, 6, 6, 5, 6
the one in dashes (-5-)  is the closest to the minimum energy, 
there is no real dependence of the number of cycles and the lattice parameter

(here the intersting thing appears, if you first converge a calculation and then start the optimization
without deleting case.scf, then you will need 9 cycles to converge the first structure, independent which one
you use as the first one.)

Unfortunately all my saved optimzations were using charge interpolation and the number of cycles
Maybe someone has optimizations using dstart for all structural changes and can tell
about the number of cycles needed to converge more complicated systems.

Third
I distorted the structure by displacing the center atom of the bcc structure using a tetragonal cell
with the center atom at 1/2, 1/2, 1/2 and distorted at 1/2, 1/2, 0.45 what corresponds to a 10% shift.
I did not use the new structure found by sgroup!
(for both cases: 8000 k-points, Rkmax=7, -ec 0.00001 -cc 0.001)
both variants needed 6 cycles
(for both cases: 16000 k-points, Rkmax=8, -ec 0.00001 -cc 0.001)
the correct one 1/2, 1/2, 1/2  needed 6 cycles
the distorted one 1/2, 1/2, 0.45 needed 5 cycles
In contrast to First this seems to be a counterexample.

Forth, in addition to Third I increased c by 10% that is I have a "real" tetragonal distortion
(for both cases: 16000 k-points, Rkmax=9, -ec 0.00001 -cc 0.001)
both variants, with center and off center atom) took 6 cycles

This may be a quick and dirty oversimplification of your problem, indeed, I calculated many systems that needed much more cycles
to converge but I can not follow anymore whether or not it was depending on the structural parameters. Mostly those were rather
complicated magnetic systems and it was critically depending on the initial magnetisation rather then on the structure.


Ciao
Gerhard

====================================
Dr. Gerhard H. Fecher
Institut of Inorganic and Analytical Chemistry
Johannes Gutenberg - University
55099 Mainz
________________________________________
Von: wien-bounces at zeus.theochem.tuwien.ac.at [wien-bounces at zeus.theochem.tuwien.ac.at]" im Auftrag von "Laurence Marks [L-marks at northwestern.edu]
Gesendet: Samstag, 23. Juli 2011 17:54
Bis: A Mailing list for WIEN2k users
Betreff: [Wien] Deeper DFT question

I am posting this email in the hope that someone can give me a hint
about a "well-known" empirical DFT result, but one which I cannot
rationalize/justify. Maybe someone can provide a counter example,
which would prove that the result is not general.

The empirical result can be phrased by looking at two cases:
a) An scf run with the atomic positions at the equilibrium (i.e.
lowest energy) positions
b) An scf run with the atomic positions displaced from equilibrium,
perhaps rather a lot.

The empirical result is that a) converges faster, perhaps much faster
than b) -- assuming of course equally good/bad starting points for the
electron density (e.g. atomic densities).

N.B., this result, if correct, implies something about the
convexity/radius of convergence of the scf problem, and relates to the
performance of MSR1a.

--
Laurence Marks
Department of Materials Science and Engineering
MSE Rm 2036 Cook Hall
2220 N Campus Drive
Northwestern University
Evanston, IL 60208, USA
Tel: (847) 491-3996 Fax: (847) 491-7820
email: L-marks at northwestern dot edu
Web: www.numis.northwestern.edu
Chair, Commission on Electron Crystallography of IUCR
www.numis.northwestern.edu/
Research is to see what everybody else has seen, and to think what
nobody else has thought
Albert Szent-Gyorgi
_______________________________________________
Wien mailing list
Wien at zeus.theochem.tuwien.ac.at
http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien