[CPMD-list] Energy components in PIMD

[Axel Kohlmeyer]

On Tue, 27 May 2008, ukasz Walewski wrote:

LW> Dnia 2008-05-14, ÿÿro o godzinie 12:35 -0400, Axel Kohlmeyer pisze:
LW> 
LW> Dear Axel,

dear lukasz,

LW> I did the BO dynamics based partially on your settings (see attached
LW> input file). The time step of 20 a.u. seems too big ÿÿto me (see the red

yes. 20 a.u. is very aggressive.

LW> plot on the attached ECLASSIC graph). The same run with 10 a.u. time
LW> step (the black plot) looks much more stable, however it also exhibits a
LW> drift of 6.3e-6 Ha/ps. Do you think one can accept this ?

that drift is quite small. i would perhaps go for 15 a.u. timestep
and hope for a bit more cancellation of the trends. BO-md can 
add a bit of friction when not converging extremely tightly. 
this is mostly removed by using a wavefunction extrapolation scheme. 

LW> 
LW> I wonder what can be the origin of the suspicious drop in conserved
LW> energy at the beginning of the 20 a.u. run ? After some 140 steps I

that could be due to the extrapolation not fully working yet and 
not converging the wavefunctions tightly enough. at second order
extrapolation the result tends to be worse than not using the 
extrapolation at all.


LW> restarted the simulation for some reason or another with the attached
LW> 3restart.inp file.
LW> 
LW> Turning back to CP dynamics: my best result (VDB, emass=100 a.u., dt=2
LW> a.u.) has a drift of 5.5e-5 Ha/ps in fictitious kinetic energy of the
LW> electrons. This means that during 10 ps run the electronic subsystem
LW> would gain 0.34 kcal/mol. Can it be assumed acceptable ? What do you
LW> think of quenching the electrons every 10 ps in the case of longer
LW> runs ?

the typical error for DFT calculations can be easily in the 1-2 kcal/mol
regime, so that seems tolerable although not perfect.
i'd rather use a thermostat on the electrons (and a small
overexcitation of the electrons may help to lower artifacts 
when actual reactions happen, so i'd rather go for the EKINC
value at the end of the test run than at the beginning as
target temperature for the electron thermostat.).

i am certain, that there are _many_ cpmd trajectories in the published
literature that have more drifting, but it does not show because of
the thermostat on the fictitious electron dynamics. at the very end
you always have to make a choice between numerical accuracy, 
algorithmical accuracy and statistical convergence. it is very important
to quantify the amount of error that you are going to introduce, but
then you have to weigh it against the systematic error of the method
(DFT, pseudopotentials, CP-dynamics, ergodicity) and make a balanced choice.

cheers,
   axel.

LW> 
LW> Best 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.