[CPMD-list] Energy components in PIMD

[Axel Kohlmeyer]

On Tue, 29 Apr 2008, ukasz Walewski wrote:

dear lukasz,

LW> Oh, yes! I forgot to mention: of course I did the CPMD (non-PIMD) NVE
LW> (http://bioexploratorium.pl/ljw/cpmd/nve/) and NVT
LW> (http://bioexploratorium.pl/ljw/cpmd/nvt/) test simulations (input files
LW> together with the plots).

did you do those in sequence? all your NVT inputs don't read 
in the velocities from the previous runs. if you have your 
system already working in CP-dynamics, you don't want to 
"rock the boat" but with each restart and re-initializing of 
the velocities you do exactly that and get further away from 
the BO-surface.

and from the NVT run you can see that your fictitous dynamics
is picking up significant energy (at what length of timestep?
1.0 a.u.?) so there is something fishy going on. so as i was
implying in my previous mail. the NVE runs work by accident
since you don't excite all DOFs. the massive ion thermosts
will enforce equipartitioning and takes that away and thus
creates some lack of adiabacity. in my experience with USPPs
the can be somewhat alleviated by increasing the density cutoff,
but that comes at a performance and memory penalty that usually
makes norm-conserving potentials somewhat competetive (and much
less of a hassle to boot. if you had any elements that were
"difficult" and are better represented by USPPs it would make
more sense...).

LW> > so before doing anything else you should verify the following:
LW> > 
LW> > can your input conserve energy for a non-PIMD run?
LW> > ...and then you have to check whether you can maintain
LW> > adiabacity with the massive thermostat on the ions only.
LW> 
LW> This is exactly what I did. The NVE CPMD simulations do conserve total
LW> energy very well (at least for 1ps runs). Thermostatting the ions makes
LW> the fictitious kinetic energy of the electrons drift after 3ps of
LW> simulation (http://bioexploratorium.pl/ljw/cpmd/nvt/ekinc.png) and this
LW> is what I realized already after performing some PIMD tests (the fault
LW> of my impatience).

yep, and that kind of drift you have to get rid of
before you can move to PIMD and/or put a thermostat on
the electrons that would hide it.

LW> > quite a few times, the no-thermostat CP-dynamics is only
LW> > meta-stable and when you switch to massive N-H chains you
LW> > excite DOFs that will make the CP dynamics lose it.
LW> > 
LW> > then you should try PIMD without a thermostat on the 
LW> > electrons. that has to be reasonably stable as well.
LW> 
LW> This is what I also did already (again - sorry not to mention this
LW> beforehand). The plots here http://bioexploratorium.pl/ljw/pimd/tests/
LW> contain three panels: bottom - no thermostats, different electron mass
LW> (bottom.inp); middle - thermostat on electrons (middle.inp); top -
LW> thermostat both on electrons and the ions, different time steps
LW> (top.inp).

ok, so if you don't have a thermostat on the electrons you get
the same drift and it seems to happen regardless of fictitious
mass and with a small time step. this should make all your
alarm bells ring. something is going south here that is none
of the usual. all other tests are useless unless you have 
sorted this out. as you can see from the other runs, the
electron thermostat does trick you into believing that everything
is fine (for a while), but as you noticed it is not.

LW> > ...and ultra-soft pseudopotentials.
LW> > i doubt that this will conserve energy
LW> > well even for classical atom cores.
LW> > at the very least you have to crank
LW> > up the density cutoff (or dual) to
LW> > have a tight enough real space grid
LW> > to accommodate the augmentation charges.
LW> 
LW> This is what I did not know. I performed a lot of single point tests
LW> with different cutoff and cell size and decided that 30 Ry is a "safe"
LW> value for electronic structure calcs with USPP's. Is it all different
LW> when it comes to dynamics ?

that is true, as far as the "wavefunction part" (= g-space) of the 
calculation goes. whenever you have to consider the (full) density
(r-space) however you are subject to severe "ripples", i.e. the fact
that you cannot resolve the density well enough with the grid you
are using and the plane waves that fit into it. this will show as
a drifting EKINC in CP-dynamics or difficult convergence in geometry
optimizations or BO-dynamics. the solution is to increase the density 
cutoff, at the expense of a significant performance penalty...

programs better adapted to USPPs than CPMD use either a fine 
"box grid" around the atoms or have a dual grid approach to
use the dense grid only where necessary. CPMD actually has
some infrastructure to support dual grids, last time i checked
it was not functional (any volunteers?).

so i would suggest you repeat the NVE tests (and NVT without 
a thermostat on the electrons) with a DUAL of 6, 8 and 10
and see if that improves the situation and to what degree.

i would also do one test with norm conserving potentials to
see how they compare.

good luck!

cheers,
   axel.



LW> 
LW> Regards,
LW> Lukasz
LW> 

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