On Tue, Apr 26, 2005, Wu, Xiongwu (NIH/NHLBI) wrote:
>
> I did Langevin dynamics simulations with an alanine dipeptide molecule
to
> examine the effect of stopping COM motion. A box of 60*60*60 was set to
> avoid the molecule from flying away. With gamma_ln=10/ps, temp0=400K,
> IPS=1, and nstlim=100,000,000, I get the following average temperature:
> nscm(steps) 0 10 30 100
> 300 1000
> Average temperature(K): 399.65 404.67 411.53 421.19
> 425.82 427.51
Although at first glance, this seems very weird, I think I understand what
is
going on.
Note that "ndfmin" is set in mdread to reflect the number of degrees of
freedom removed from the system. So, in the table above, ndfmin=0 for
nscm=0,
but ndfmin=6 for the other columns. This affects how the temperature is
calculated, but does not affect the nature of the random forces or the
friction. So, the "real" temperature is lower than that reported by Amber
by
a factor of (3*nat-ndfmin)/(3*nat). If the alanine dipeptide has 22
atoms,
this factor is 0.909, which roughly explains the difference between the
nscm=0
and nscm=1000 values: by 1000 steps, the system has effectively
reequilibrated
to a system with 66 degrees of freedom, but Amber is calculating the
temperature as if there were only 60 degrees of freedom. So, the reported
temperature is wrong, by a noticeable amount. For values of nscm between
0
and 1000, you essentially have a system that is trapped between the two
degrees of freedom.
Note that this is not really a problem with the trajectories: it is a
problem
with how temperature is calculated and reported. All other properties of
the
ensemble should be correct, and the "true" average (internal) temperature
is
probably 400K in all the above examples. Amber is just not calculating
the
temperature correctly, because it does not know how many effective degrees
of
freedom there are.
My prediction (tm) is that if you did something similar for a larger
system,
the temperature differences in these sorts of runs would be much smaller.
But you are correct in one sense: setting nscm>0 with the current version
of
Amber leads to inconsistent results with Langevin dynamics.
Question: does the system actually "fly away" with nscm=0 and long
Langevin
dynamic runs? If not, we should probably not require nscm=0 when Langevin
is selected.
Thanks for running these tests!.....dac
Received on Wed Apr 05 2006 - 23:49:57 PDT
