[gmx-developers] Re: [Fwd: Re: results]

[Michel Cuendet]

If I'm entitled to say anything here...

>the algorithms should be there because of choice, and users should
>be encouraged to take an informed decision.
>
I agree

>Second, the Nose-Hoover thermostat gives the correct ensemble for your
>system plus "something else". In fact it therefore does not give you the
>correct ensemble for what you are interested in. 
>
I'm not sure about what you mean exactly. As I understood from the very 
informative (Tuckerman, JCP115 : 1678,2001), a little bit of that good 
math shows that this "something else" nicely integrates out, to give the 
exact canonical distribution for your physical variables...

>Moreover, small errors in e.g. the temperature coupling
>algorithm are negligible compared to systematic bias due to force field,
>finite system size etc.
>
Well, I'm not so sure about that. We just saw an example where a little 
twist in the temperature calculation gives me a fair 3500 KJ/mol error 
at the end of the game. Also I reported to the user list before that the 
integrator of 3.2.1 makes the thermostat suck much more energy out of 
the system than required, with short time constants (the hack according 
to Holian et al. fixed that). Here again the uncontrolled energy flow 
was of the order of 1e3 KJ/mol.

Like you say, in MD, there are so many parameters, and so few checks to 
constrain these parameters. If we have the chance to detect that 
something might be a source of error, we should correct it. Good 
integrators are available
already in the literature (see references in a previous mail).

A little cautious pointer: The Nose-Poincare thermostat (Bond et al, J 
Comp Phys 151 : 114, 1999; Leimkuhler and Sweet, JCP 121(1) : 108, 2004) 
might become the method of choice.  It is hamiltonian like the Nose 
thermostat, but cures the problem of the uneven time sampling. Due to 
it's hamiltonian structure, it allows to build formally smplectic 
integrators, which have all the nice properties one might dream of  (eg. 
mathematcally ;-) bounded energy drift).

Best,

Michel

-- 
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Michel Cuendet, Ph.D. student
Laboratory of Computational Biochemistry and Chemistry
Swiss Federal Institute of Technology in Lausanne (EPFL)
CH-1015 Lausanne						
Switzerland                            Phone : +41 21 693 0324
lcbcpc21.epfl.ch/~mitch	                       +41  1 633 4194
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++