[gmx-developers] re: Energy and Foreign Lambdas and Hamiltonian Replica

[hessb at mpip-mainz.mpg.de]

> Dear Berk,
>
>> I have implemented replica exchange for the free-energy lambda in the
>> CVS head branch,
>> based on linear extrapolation of dV/dlambda (not foreign lambda
>> energies).
>> I haven't tested it though.
>> Please tell me if it works or not.
>
> I'm fairly concerned about exchanging based on linear extrapolation of
> dV/dlambda. In my experience, dV/dlambda can be pretty nonlinear,
> especially in the case when soft core potentials are used -- even if
> lambda spacings are pretty close. I'm thinking the risks of doing
> things this way may outweigh the benefits, and that it would probably
> be better to hold off until foreign lambda energies are in that branch
> also.
>
> Thanks,
> David

I agree that this might often not be a good approximation.
But at least the exchange code is there now (although completely
invisible to the user). So people can experiment with it,
knowing the limitations.
Even without the foreign lambda code, one could just call do_force
with a different lambda at an exchange attempt.
But it might be better to wait till the foreign lambda code is ready.

One issue with the foreign lambda code is constraints.
When constraint lengths depend on lambda, one can not simply
determine the potential energy for a different lambda
than at which the actual state is.

Berk.