Re: [AMBER] NaN error in .rst files

From: Ross Walker <ross.rosswalker.co.uk>
Date: Wed, 26 Jan 2011 22:46:15 -0800

Hi Peker,

> That was actually what i did. no ig option with ntt=3. but there was
> nothing about this, in tutorial. also in the manual it says, if ntx is
> less than 3 and initial temp does not equal to 0 velocities will
> depend on this seed. but my ntx is 7 and my initial temp is 300.0. on
> the other hand it effects the Andersen coupling ?? and that is not
> given in the tutorials ? this is confusing really. Should i start from
> beginning ? I mean i don't what i should say or ask...

Welcome to the world of research... This is a classic case of this not being
known about at the time the tutorials were written. Indeed none of the major
developers even realized this. However, since then things have changed. We
have attempted to update some of the tutorials and manuals but alas lots of
other demands get in the way. It is also somewhat of a historic situation
where people used to use Berendsen for the thermostat (ntt=1) which does not
depend on the random number generator. Advice then switched to using
Langevin (ntt=3) but nobody, at least initially, considered the subtle
complexities this method introduced. My vote is to make ig=-1 the default
but this needs someone to volunteer to go through and update all the test
cases. Note earlier versions of AMBER, including pmemd v10 did not support
ig=-1 hence why it is not in a lot of the tutorials. Although many make
mention of the fact that you should manually set ig to a new value for each
run.

With regards to what is affected.

The random number generator is used to both control a run depending on the
thermostat and to seed the initial velocities if irest=0. In the case of
seeding the initial velocities my guess is that it makes little difference
since you only do it once and then you restart each time so the velocities
do not get seeded again. You then have the following:

ntt=0, irest=0 - Random number generator seeds initial velocities only.
ntt=0, irest=1 - Random number generator is not used.
ntt=1, irest=0 - Random number generator seeds initial velocities only.
ntt=1, irest=1 - Random number generator is not used.
ntt=2, irest=0 - Random number generator seeds initial velocities and
controls the random reset of temperature to the target.
ntt=2, irest=1 - Random number generator controls the resetting of
temperature to the target.
ntt=3, irest=0 - Random number generator seeds initial velocities and is
used 3xnatom times on 'every' time step to generate the random forces as
part of the langevin equation.
ntt=3, irest=1 - Random number generator is used 3xnatom times on 'every'
time step to generate the random forces as part of the langevin equation.

It is these last two cases where the random number generator performance is
CRITICAL. The problem occurs on restarts with the same random seed because
what you do is repeat the same sequence of random numbers and thus 'random'
forces for each new restarted run. It can be shown analytically that this
results in two uncorrelated trajectories actually becoming correlated during
a Langevin MD simulation. This is of course BAD! - Hence while ntt=3 is good
equilibrating the temperature of a system it has to be used VERY carefully
to avoid artefacts...

This is my argument for why thermostats should not be used in production
simulations at all. But this is another argument for another day over a
different choice of beverage.

All the best
Ross

/\
\/
|\oss Walker

---------------------------------------------------------
| Assistant Research Professor |
| San Diego Supercomputer Center |
| Adjunct Assistant Professor |
| Dept. of Chemistry and Biochemistry |
| University of California San Diego |
| NVIDIA Fellow |
| http://www.rosswalker.co.uk | http://www.wmd-lab.org/ |
| Tel: +1 858 822 0854 | EMail:- ross.rosswalker.co.uk |
---------------------------------------------------------

Note: Electronic Mail is not secure, has no guarantee of delivery, may not
be read every day, and should not be used for urgent or sensitive issues.





_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Jan 26 2011 - 23:00:02 PST
Custom Search