[gmx-developers] [gmx-users] Wrong Nose-Hoover integrator

[Berk Hess]

Michael Shirts wrote:

>>Apparently I cut some other link before pasting...
>>Sorry, here is the correct link:
>>http://www.gromacs.org/pipermail/gmx-users/2005-January/013917.html
>>
>>Also I should slightly rephrase my integration scheme statement.
>>The proper solution is Velocity Verlet, as you would have reversible
>>Nose Hoover integration using the temperature at the full time-step.
>>This probably requires an extra communication step in parallel.
>>    
>>
>
>Right, that's what I would have figured.  I have velocity verlet
>working correctly (at least on a subset of the possible test cases) on
>the Folding at Home copy of 3.1.4.  I can't speak for the extra
>communication that might be needed, as I never really looked at that 
>I'm happy to start working to make velocity verlet an option in
>GROMACS, but from the feedback I got, it seems like that should wait
>until after the 3.3 branch is set for correct?
>
>  
>
Currently I am having a long discussion with Michel Cuendet.
I have figured out that for the coupling we should use the temperature
at t-1/2 and then we can have proper, reversible leap-frog Nose-Hoover 
integration,
I already have this working, so this can make it into 3.3 (see the link to
Michel Cuendet original post for the algorithm and reference).
There are even better, symplectic, integrators, but we can consider this 
later.

My statement about velocities in the previous mail was correct, but I 
incorrectly
draw conclusions about Ekin from this. Ekin is 4 times more accurate when
evaluated at half steps.
We will have to change all the Ekin output to 0.5(Ekin(t-1/2) + 
Ekin(t+1/2)).

Proper NPT with constraints is much more complicated and might require
velocity verlet with iteration of  constraint calculations.
This will probably have to wait till after release.

Berk.