[Wien] Ghost band come from large charge fluctuation

[L. D. Marks]

In addition to the points Peter made about using the original case.in1 and
not using -in1new until you are close to the solution (and not being too
concerned about ghostbands until you are close), let me add something
about mixing.

Please, DO NOT SIMPLY REDUCE THE MIXING FACTOR. A rather major change in
the mixing was made a couple of releases ago:
	1) Introduction of a plane wave scaling term (PW)
	2) Addition of an output of the plane wave convergence :PLANE
	3) Addition of a diagnostic parameter :DIRB
	4) Addition of an auto-restart of the Broyden cycle

In iterative numerical methods scaling is VERY important, and has a
huge effect. The UG is not very detailed about these (Peter: perhaps
should add something, perhaps even a FAQ). I've attached a file "Check"
which I suggest you put in your $HOME/bin and run every now and then to
monitor the progress.

Key points:
	a) :DIRB gives the modulus of the Broyden step the algorithm is
using, the modulus of the Pratt step (it does not use it) and the angle
between the two. If this angle is more than about 45 degrees in 99% of the
cases this means that your plane scaling term is too large. Reduce it,
remove *bro* files and continue. (I prefer a value of 0.25 for my initial
estimate of this, smaller than Peter's default of 0.5, not sure which is
"best".) If the angle gets really large, > 138 at present, the algorithm
abandons the previous Broyden history and restarts with a small Pratt
step.
	b) In most cases "sloshing" is due to poor convergence of the
plane waves (:PLANE). If the atomic-sphere charge convergence (:DIS) is
oscillating but the plane wave convergence is improving DO NOT WORRY, this
is OK. In my experience, it is only when :PLANE is less than about 0.01
that the iterations start to converge. I never use -in1new before
obtaining reasonable convergence of both :DIS and :PLANE.

Caveat: There is a pathalogical case which Peter found once, and I have
also found in a seperate system. At least in my case it appears to be
oscillations which can go unstable between some spin-up and spin-dn states
very close to the Fermi energy. These presumable contain both
atomic-sphere and plane wave components so a relative scaling is not
helping enough. If anyone has any ideas about how to handle such a problem
in terms of scaling let me know off-line.

N.B., testconv -p :PLA works to test the plane wave convergence, but at
present is not built into the runXYZ scripts but can easily be added.

-----------------------------------------------
Laurence Marks
Department of Materials Science and Engineering
MSE Rm 2036 Cook Hall
2220 N Campus Drive
Northwestern University
Evanston, IL 60201, USA
Tel: (847) 491-3996 Fax: (847) 491-7820
email: L - marks @ northwestern . edu
http://www.numis.northwestern.edu
-----------------------------------------------
-------------- next part --------------
grep -e :DIRB -e :ENE -e :DIS -e BRO -e PLANE *.scf $1 -e Chaos | tail -20