[gmx-developers] Shift for MARTINI force field?

[Berk Hess]

Luca Monticelli wrote:
> XAvier Periole wrote:
>> Also for your information. The derive of temperature reported by Moritz
>> et al. (linked by Michael) when using MARTINI with increasing time step
>> is actually due to (the old story reported by van Gunsteren et al.)
>> the use
>> of the average of the half-time-step velocities to get the on-step
>> temperature instead of the average of the half-time-step kinetic energy.
>> We have written a comment on that paper ... nothing wrong with
>> MARTINI. :))
>
> That is also my understanding, thanks Xavier for clarifying this.
>
> By the way, I think it would be useful if we were able to choose how
> to calculate the kinetic energy (either from average half-step
> velocities, as was done until v3.2.1, or from half-step kinetic
> energies, as done in v3.3 and newer). At the moment, one needs to use
> the (much slower) v3.2.1 to look at the effect of calculating the
> kinetic energy in 2 different ways. Would it be a lot of work to add
> an implementation of the "old way" in Gromacs 4? (and then allow the
> user to choose between the two)
> I did notice large differences (in the density of liquids, for
> example) between v3.2.1 and v3.3 with MARTINI (with atomistic
> simulations the differences are less important since the time step is
> much smaller); it's clear that this is related to the different way of
> calculating kinetic energy, but it's not entirely clear to me why the
> new method would be best. I suppose Berk has done some work on this,
> but I haven't seen any paper published - did I miss it?
>
> Cheers,
>
> Luca
The situation is clear:
the error in the full step kinetic energy is exactly 4 times as high at
the half step kinetic energy.
Therefore the half step kinetic energy is always better than the full
step one.
I don't see any reason for having the option to output the full step
kinetic energy,
except for exact comparison against more incorrect results.

Also note that GROMOS had (and as far I as know still has) the issue of
inconsistent temperature
calculation. The half step temperature is used in the temperature
coupling, whereas the full step
one is reported in the output.

Unfortunately this half step/full step issue will soon be reintroduced
in Gromacs,
because the velocity verlet integrator uses full time step velocities
for coupling.
Micheal Shirts together with the group of Tuckerman has been looking
into using
the half step Ekin for coupling, but has not succeeded (yet).
So we will have a more accurate Nose-Hoover/Parrinello-Rahman integratore,
but with less accurate temperatures.

The nasty aspect of this Ekin inaccuracy is that different modes get
affected differently,
it is not a simply uniform scaling of the temperature.

Berk