[Pw_forum] nscf calculation failure~

[Axel Kohlmeyer]

On Sat, 11 Nov 2006, lan haiping wrote:


HP> I browsed the mail list, it seems this error is related to ultrasoft
HP> pseudopotential's bad behavior, as pointing out by paolo "S is the
HP> overlap matrix in the Ultrasoft pseudopotential formalism. It has to
HP> be positive definite. Likely one (or more than one) of the
HP> pseudopotentials is bad "

nevertheless, it may simply be that the pseudopotial is only
'marginally bad' or that different algorithms are less sensitive
to the problematic pseudopotential(s). i know from personal experience,
that a few of the pseudopotential generated from parameter files shipped 
with the vanderbilt atomic code have problems when re-creating them with 
the current version of the code, most likely due to changes in the 
fitting algorithm after the pseudopotential parameters were added. 

this seems to mainly affect rather heavy elements and pseudopotentials
that were created as part of the dacapo project (or recreated with a
different functional but using the same parameters). those potentials
were conceived with an older version of the atomic code (7.0.0, see the
dacapo homepage), where those problems may not have appeard. the
potentials still test ok with the internal tests, so one really has to
look through the (not exactly user friendly) log files. :-(

one can usually correct for the problems (there are reports of
'negative densities' in the pseudopotential generation log) by
adding a few additional constraints to the pseudopotential fits,
but of course after that you need to re-verify the pseudopotential.
it still is more convenient that having to build a pseudopotential
completely from scratch...

BTW. if somebody is interested in looking into this in more detail,
for example by checking and (if needed) improving the pseudopotential 
files on www.pwscf.org, i should have a few example inputs with 
constraints (the explanation in the docs can be confusing) and can 
give a few (empirical!) hints on where to add those constraints.

HP> But no CRASH is occured in 'scf'  or 'relax' calculation, it's too strange
HP> to belevie the problem of pseudopotential,

please check for the indicators listed above, or let us know
what pseudopotential files you are using. if you can prove that
the pseudopotentials you use do not suffer from the problem i
mentioned, it would be a much higher incentive to fix the code.

thanks and best regards,
    axel.

HP> I also try to use  'cg' method for diagonalization instead of  'david '
HP> method, and no such CRASH would be occured.
HP> 
HP> Regards,
HP> hai-ping
HP> 

-- 
=======================================================================
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.