Re: energy drift in NVE

From: David Hardy (dhardy_at_ks.uiuc.edu)
Date: Mon Aug 20 2007 - 22:12:54 CDT

Hi Peter,

Back when the NAMD paper was written, there was no known way to
conserve both energy and momentum with PME (or any other grid-based
evaluation method). This has been mentioned in the literature as one
of the primary deficiencies of PME. The good news is that the small
force correction made by "zeroMomentum=on" not only conserves energy
(as shown in the paper) but also permits some latency in its
application, particularly when using multiple time stepping, so you
should experience no appreciable performance degradation when using
this.

Regards,
Dave

On Aug 20, 2007, at 2:15 AM, Peter Freddolino wrote:

> Hi Dave,
> thanks for the correction; for some reason I'd thought based on the
> namd paper that the zeroMomentum fix lead to energy changes. Learn
> something new every day :)
> Thanks,
> Peter
>
> David Hardy wrote:
>> Hi,
>>
>> A small correction to Peter's comments. Use of zeroMomentum=on
>> will correct momentum for PME while still conserving energy. (See
>> preprint http://bionum.cs.purdue.edu/SkHP07.pdf for details.)
>>
>> Energy drift can also result from incorrect switching/shifting
>> options for nonbonded potentials (i.e. config options switching,
>> switchdist, and pairlistdist). Generally speaking, a
>> discontinuity in the force will cause gradual heating of the system.
>>
>> Regards,
>> Dave
>>
>>
>> On Aug 19, 2007, at 3:53 PM, Peter Freddolino wrote:
>>
>>> Hi Victor,
>>> please have a look at the section on integrators in the NAMD paper
>>> (Phillips et al., JCC 26:1781--1802, 2005), as these issues are
>>> discussed there in some detail. In brief, the Verlet integrator does
>>> conserve energy on paper, although there will always be *some*
>>> numerical
>>> error and very small energy drift (smaller than what you
>>> observed) even
>>> with a good integrator. As discussed in that paper, the following
>>> factors can cause larger energy drift:
>>> -Multiple time stepping (anything more aggressive than 1/1/3 for
>>> simulations without rigidbonds or 2/2/4 with rigidbonds will not
>>> conserve energy well, and 1/1/1 is preferrable)
>>> -PME with momentum corrections
>>> -any applied forces or constraints (obviously)
>>>
>>> In practice, NVE simulations with suitable conservative timestepping
>>> should show no discernable energy drift. If you do use a thermostat,
>>> slightly more efficient timestepping schemes (eg, 1/2/4 for
>>> simulations
>>> without rigidbonds) can be used. Energy conservation should not (in
>>> principle) be affected by SHAKE, although I've never tried this
>>> myself.
>>>
>>> I didn't see a script snippet at the bottom of your email, so I
>>> can't
>>> comment on the parameters you're using.
>>>
>>> Best,
>>> Peter
>>>
>>> Victor Ovchinnikov wrote:
>>>> Dear all
>>>>
>>>> I my NVE simulations of a solvated protein system done at 298K
>>>> (about
>>>> 110,000 atoms), the temperature increases by 4 degrees over 12
>>>> ns; K.E.
>>>> increases by 700 kcal/mol; total energy increases by 2200 kcal/mol;
>>>>
>>>> a snippet from my script is at the end of this message;
>>>>
>>>> 1)The integrator used in NAMD is Verlet, which is energy
>>>> conservative,
>>>> correct?
>>>> 2)How is the total energy expected to vary in practice?
>>>> 3)Is conservation affected by rattle/shake,
>>>> 4)Is conservation affected by Timestep splitting?
>>>> 5) Which of the script parameters could affect stability the
>>>> most in
>>>> this case (dt, stepspercycle, fullElectFrequency) ?
>>>>
>>>> Thanks for any feedback,
>>>> Victor
>>>>
>>

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:46:49 CST