Energy Jump

[Axel]

On Jun 25, 6:55 am, cbeck <beck... at osu.edu> wrote:
> Hello,
>
> I have been running geometry optimizations on an orthorhombic ice Ih
> unit cell with 8 water molecules (DZVP basis).
>
> We are interested in obtaining the energy as a smooth function of the
> two independent lattice parameters.  Using mgrid cutoff of 300 we find
> a large (~.005 hartree) jump in the energy as a particular lattice
> parameter value is crossed.

this is normal behavior when representing density
and potential on an FFT grid.

first off, those grids have to be integer, so there would
have to be jumps unless you use an infinitely large
density cutoff. those jumps are larger, since the FFT
only allows grids that can be factorized into small
prime numbers (for efficiency reasons).

the way to deal with this is to either use a larger
density cutoff to reduce the magnitude of the energy
jumps or fit the energies to an equation of state
and use that for your smooth function.

strictly speaking this behavior corresponds to
having a different effective cutoff for different
lattice parameters.

cheers,
    axel.


>
> We suspect this has something to do with the set of density basis
> functions selected by the mgrid cutoff changing at this point.  The
> following two data sets demonstrate the energy jump.  Notice how one
> of the PW_GRID total bounds changes by a large amount at the jump.
>
> Lattice a (Angstrom)      Lattice c (Angstrom)    Energy (Hartree)
> Bounds (1,2,3)
> SET 1
> 7.9                               7.5                               -137.58021900        {90., 48., 80}
> 7.96                              7.5                               -137.58014954        {90., 48., 80}
>                                 JUMP
> 8.02                              7.5                               -137.58580546           {90., 54., 80}
> 8.08                              7.5                               -137.58488193        {90.,
> 54., 80}
>
> SET2
> 7.9                              7.5                               -137.58021990            {90., 48., 80}
> 7.9                              7.643                     -137.58019671         {90., 48., 80}
> 7.9                              7.786                     -137.58034585         {90., 48., 81}
>                                  JUMP
> 7.9                             7.929                      -137.58523364          {90., 48., 90}
> 7.9                             8.072                      -137.58366580             {90., 48., 90}
>
> Has anyone else run across this?  Is there a way to obtain smooth
> output?  (like constraining the density basis if the hypothesis is
> correct)
>
> Corey