[Wannier] Omega minimum search divergence in wannier90

[Jonathan Yates]

On Fri, 22 Aug 2008, Nicola Marzari wrote:


> Thanks Alex - Arash or Jonathan might comment better - here
> are my observations
>
> 1) when the optimal parabolic step length is much larger than
> the trial step length, it would be safer to choose a smaller step
> (e.g. 2 or 3 times the trial) and also reset to steepest
> descent the algorithm
>
> 2) it would be good to have a check that if the functional has
> not gone down, U and M are restored to the previous step, we
> go back to steepest descent, and the trial step length is halved
> (so that even a large step now will be 2 or 3 times the half trial).
>
> Alex, maybe you could even look into the code and fix these,

Alex,

Something I always try when the minimiser gives me very large spreads is
  guiding_centres = T
(see comments about branch cuts in the orignal papers for the meaning of 
this keyword).
  At the moment this requires projections to be defined - however, a quick 
way to do this is to set
  begin projections
  random
  end projections

With this set the minimisation goes down-hill very quickly - after 100 
steps the spread is converged to 10^-7 giving
          Spreads (Ang^2)       Omega I      =    17.485221718
         ================       Omega D      =     0.010329097
                                Omega OD     =     3.222348850
     Final Spread (Ang^2)       Omega Total  =    20.717899666


  - so, in this case, the upward jump in the spread is a bit misleading - 
but Nicola's suggestions are still worthwhile in the general case. Also 
note the keywords fixed_step and step_length.


>> I
>> suspect it may also be a compiler issue, since I couldn't get wannier90
>> to run with gfortran on my machine, but it seems to run with no problem
>> with ifort.

  In the past I've had w90 working fine with gfortran on linux and osx. 
I'd be interested to know what the problem was, and the combination of 
compiler version, platform and libraries. (assuming this is the latest 
stable gfortran)


  Jonathan


-- 
Dr Jonathan Yates         |    Theory of Condensed Matter Group
Corpus Christi College    |    Cavendish Laboratory
Cambridge, CB2 1RH, UK    |    Cambridge, CB3 OHE, UK
email jry20 at cam.ac.uk     |    Tel +44 (0)1223 337461