[CPMD-list] 143 steps too long?

[Axel Kohlmeyer]

On Tue, 23 Aug 2005, Matt wagner wrote:

matt,

MW> I am doing  single point calculation for density minimization on BH4, 
MW> cation.   Everything worked fined, but took 143 steps to complete.  Is this 
MW> too long?  Incorrect?  I first used ODIIS and it took more than 1000 steps 

too long really depends on the system, that you are looking at.

MW> for wavefunction optimization.  When I used PCG Minimizer, that is when I 
MW> got it down to 143 steps.  This was all done using the Vanderbelt 
MW> psuedopotential.  The manual says that steps larger than 100 is to be 
MW> considered slow. Does slow mean inaccurate in this case?  Any suggestions 

in some cases this is really difficult to tell. you can have a 
seemingly converged wavefunction, do an MD for quite a while
and then suddenly your system drops into a different (lower) 
state. but this is a really rare event.

in my personal experience, most of the time, however, slow 
convergence is due to one or some of the following reasons:

- you have a 'problematic' element (e.g. many d-metals)
  -> here you have not much of a choice, sometimes switching
     to a different type of pseudopotential or switching
     between different methods to optimize the wavefunction
     (PCG, ODIIS, ANNEALING, LANCZOS) helps.

- your converged wavefunction is very different from the
   initial (atomic) guess.
  -> you can try either using INITIALIZE WAVEFUNCTION RANDOM, or 
    modify the initial guess using an ATOMIC CHARGES block in the 
    &ATOMS section. 

- your starting structure is very 'bad'.
  -> if you just want to start an MD from the wavefunction,
     you can limit the number of wavefunction optimization steps.
     then use the not too well optimized wavefunction for a few
     steps of MD. then restart from that wavefunction and reoptimize
     to a tighter convergence. the same works, of course, for
     a geometry optimization. if you need the single point energy
     for _exactly_ your starting configuration, then you still 
     could do this procedure and then, provided your atoms did
     not move too much, in the final step restart only from the
     wavefunction and _not_ from the coordinates, so that you can
     use the 'improved' wavefunction as initial guess.

- you pick up a lot of numerical noise in the gradient correction.
     this is usually worst when using a low plane wave cutoff
     and for systems with a lot of vacuum.
   -> in this case you can change the GC-CUTOFF value from the
     default of 1.0e-8 to 1.0e-6 or what ever is needed. not that
     if you raise it too much, you will basically turn off the 
     gradient correction, so you want to keep this number rather low.
   -> also in case of very low plane wave cutoffs, you need to 
     increase the density cutoff. this can be done either directly via 
     the CUTOFF DENSITY keyword or indirectly via the DUAL keyword 
     (which defines  the ratio between them). the defaul is a ratio of 4, 
     but for a plane wave cutoff of 25 you should try raising it to 5 or 6.
   -> finally, you can reduce the noise by adapting the shape of the
     box and realigning the atoms in a way that there as little vaccuum 
     area as possible, within the requirements of the poisson solver.

- something else is wrong in your input.

MW> would be greatly appreciated.  Below is my input file:

a few more remarks about your input file.

MW> &SYSTEM
MW> SYMMETRY
MW> 0
MW> CHARGE
MW> 1.0

hmmm. with a charge +1 you have an odd number of electrons
in your system. if that is intentional, you have to use LSD.
this may contribute to your convergence problems.

MW> CELL
MW>   40.0000000    1.0000000    1.0000000    0.0  0.0  0.0

from the coordinates it seems, that your molecule is much 
longer in x-direction than in the other directions, so using
an orthorhombic box, might help cutting down the unneeded 
vacuum area, e.g. with

  CELL
   40.0  0.8 0.8 0.0 0.0 0.0

and you can probably shrink the whole box a little, 
too, without losing accuracy (e.g. go from a=40 to a=38).

MW> CUTOFF
MW> 80
MW> &END


since you don't have any problematic elements in your system
i would expect an input with those two suggested modifications
in place to converge within 50 steps or less. if you switch back
to ultrasoft pseudopotentials with the associated additional changes, 
it may take a few steps more, but there should be no problems either.

best regards,
	axel.

-- 

=======================================================================
Dr. Axel Kohlmeyer   e-mail: axel.kohlmeyer at theochem.ruhr-uni-bochum.de
Lehrstuhl fuer Theoretische Chemie          Phone: ++49 (0)234/32-26673
Ruhr-Universitaet Bochum - NC 03/53         Fax:   ++49 (0)234/32-14045
D-44780 Bochum  http://www.theochem.ruhr-uni-bochum.de/~axel.kohlmeyer/
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.