<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html>
<head>
<meta content="text/html; charset=ISO-8859-1"
http-equiv="Content-Type">
</head>
<body bgcolor="#ffffff" text="#000000">
<font face="Calibri">Dear Hongwei,<br>
<br>
As Nicola mentions, input parameters are explained in the user
guide. Regarding conv_noise_*, from p22:<br>
<br>
"If conv_noise_amp > 0, once convergence (as defined above) is
achieved, some random noise f is added to the search direction,
and the minimisation is continued until convergence is achieved
once more. If the same value of Omega as before is arrived at,
then the calculation is considered to be converged. If not, then
random noise is added again and the procedure repeated up to a
maximum of conv noise num times. conv_noise_amp is the amplitude
of the random noise f that is added to the search direction: 0
< |f| < conv_noise_amp. This functionality requires
conv_window > 1. If conv window is not specified, it is set to
the value 5 by default. If conv_noise_amp <= 0, then no noise
is added (default)."<br>
<br>
The idea here was to have a means by which to be able to escape
from local minima. As an example in which is is useful, imagine
you seek the 8 MLWFs corresponding to the four valence and four
lowest conduction eigenstates of of silicon (as described in the
Souza-Marzari-Vanderbilt paper of 2001). If you start with an
initial projection of one s and three p orbitals on each Si atom,
the minimisation of the spread will rapidly plateau after a few
tens of iterations at a local minimum, as shown in the figure
below (SD=steepest descents; CG=conjugate gradients; this plot is
now at least five years old), and you may be forgiven for thinking
the calculation has converged. But, if one continues to persevere,
the global minimum (four sp3 orbitals on each Si) can eventually
be reached (more quickly with CG than with SD, as can be seen
below). Random noise using conv_noise_* can be used to accelerate
this process.<br>
<br>
Hope that helps,<br>
<br>
Arash<br>
<br>
</font><img src="cid:part1.01070704.08000305@imperial.ac.uk" alt=""><br>
<font face="Calibri"><br>
</font>
<pre class="moz-signature" cols="72">--
Dr Arash A Mostofi
Senior Lecturer and RCUK Fellow
Depts of Materials & Physics
Imperial College London
London SW7 2AZ, United Kingdom
T +44 (0)207 594 8154
F +44 (0)207 594 6757
E <a class="moz-txt-link-abbreviated" href="mailto:a.mostofi@imperial.ac.uk">a.mostofi@imperial.ac.uk</a>
W <a class="moz-txt-link-freetext" href="http://www.cmth.ph.ic.ac.uk/people/a.mostofi">http://www.cmth.ph.ic.ac.uk/people/a.mostofi</a>
</pre>
<br>
On 14/03/2011 16:33, Nicola Marzari wrote:
<blockquote cite="mid:4D7E435B.5060004@mit.edu" type="cite">
<pre wrap="">
Dear Hongwei,
guiding centers were used to suggest how to properly choose the branch
cut in the evaluation of complex logarithms - the logic is explained
inside wann_phases
(basically making sure that roughly every contribution that you add up
in evaluating
the position of the wannier center has a branch cut that is closest to
giving that position),
but these days you'd be better off by choosing carefully your
projections, and forgetting
about the guiding cetners (keywords are nevertheless explained in the
user guide).
With regard to conv_noise... , it must be something Arash or Jonathan
coded? Looks like
a way to get unstuck from very shallow minimizations (e.g. as in the
rotation of the d-orbitals
in a system where localization is almost invariant with respect to
rotation).
nicola
On 3/14/11 4:12 PM, Hongwei Wang wrote:
</pre>
<blockquote type="cite">
<pre wrap="">Dear Wannier90 users
I am a Ph. D student in the Key Laboratory of Quantum Information, University of Science and Technology of China. I am beginner in using the Wannier90 code, I can't well understand the Key word "guiding_centres". I wonder if you can tell me its physical picture and the situation in which we should use guiding_centres? In addition, I want to know how to set the Key word num_guide_cycles£¬num_no_guide_iter£¬ conv_noise_amp£¬ conv_noise_num reasonablely.
can you tell me some experience to set them.
Thanks for any of your help£¡
Sincerely,
Hongwei Wang
_______________________________________________
Wannier mailing list
<a class="moz-txt-link-abbreviated" href="mailto:Wannier@quantum-espresso.org">Wannier@quantum-espresso.org</a>
<a class="moz-txt-link-freetext" href="http://www.democritos.it/mailman/listinfo/wannier">http://www.democritos.it/mailman/listinfo/wannier</a>
</pre>
</blockquote>
<pre wrap="">
</pre>
</blockquote>
</body>
</html>