Date: Tue, 30 Apr 2002 09:10:46 -0700 (PDT)
From: Jung-Hsing Lin
Subject: Re: NTT=5?



Thank you so much for the reply.

On Mon, 29 Apr 2002, David Case wrote:

> The general recommendation now is to set ntt=0 during production periods,
> only using thermostats to establish a target temperature during
> equilibration. With the default Ewald scheme, energy conservation
> should be good enough to alow the thermostats to be turned off (for all
> but the longest runs); alternatively, setting tautp to a fairly large
> value (like 5.0) also can avoid the "hot solvent/cold solute" problem.

The fluctuation in a microcanonical ensemble is different from that of a
canonical ensemble, and the thermodynamic relationships will also be
different. Apparently, in microcanonical ensemble energy is constant
and temperature will fluctuate, while in canonical ensemble temperature is
constant and energy will fluctuate. The probability densities of phase
points of these two ensembles are by definition different, and the
expectation value of physical observables will also be different. Thus I
personally feel it is not proper to use the NVE option for long
simulations, even if numberical errors could be minimal. I guess most
AMBER users want to simulate at ambient temperatures, so that they can
compare with experimental results.

If the combination of NTT=1 and Tauts=5.0 (or larger) works and can avoid
the "hot solvent/cold solute" problem, it would be nice that this could be
added in the AMBER users manual. However, academic users might still need
to cite a paper which has tested this issue systematically.

> The old ntt=5 scheme probably could have been kept if we had had a
> better scheme for deciding what was "solute" and what was "solvent".
> Earlier Amber codes progressively became more and more complex and
> arcane, with several different schemes and variables being used to try to
> identify "solvent", but none of them very good for anything but TIP3
> water.
>
> We are working on a revamped scheme to identify groups of atoms (no more
> rgroup()!). I had hoped this would be ready for Amber7, but it was not.
> Once that is in place, it should be much easier to go back and insert
> separate temperature baths if one wants. As with many things, the code
> itself is trivial -- it's designing the user interface, and then
> documenting that, that always seems to take more time than one can
> possibly believe....
>

This, I am afraid, would not be a very elegant approach. (Sorry if I have
offened you). The most pessimistic situation would be that one probably
needs to couple independent thermostat to each individual degree of
freedom. This will, in turn, violates the spirit of the weak coupling
method, which tries to minimize the local disturbance by a gloabal
coupling. It has been spread by many text books that Berendsen thermostat
does not generate a well-defined ensemble, but it seems to me the problem
can be remedied if the fluctuation formula in the article by Morishita (J.
Chem. Phys. vol. 113, pp. 2976-2982, cited in AMBER user's manual) are
applied. However, this article doesn't address the "hot solvent/cold
solute" problem, and many people attribute this problem to the use of
Berendsen thermostat.

Please correct me if I understand anything wrongly.


Thank you,

Jung-Hsin