[gmx-users] Free energy calculation, parallel or sequential lambda runs ?
Ignacio Fernández Galván
jellby at yahoo.com
Sat Dec 16 10:13:41 CET 2006
--- David Mobley <dmobley at gmail.com> wrote:
> What I've shown in my work is that at some particular lambda values,
> there maybe worse sampling problems than at any other lambda value.
> This means that you may need to run longer at that particular lambda
> value in order to get adequate statistics. Unless you stop and
> each of your lambda values before going on, noticing this later would
> mean you would need to go back and re-start the simulation of all of
> the subsequent lambda values, since the first time you didn't, you
> didn't use the correct ending structure.
I still don't see it. I mean, I don't think I would need to rerun
anything (except the particular lambda value). The fact that I'm using
the final structure of a lambda run to start the equilibration of the
next lambda is not really so important, I mean it doesn't *have to* be
the final structure, it's only some structure that would probably be a
better starting point than the global initial structure (lambda=0).
Say, for instance, I want to run a TI with lambda=0, 0.1, 0.2 ... 1.0.
I optimize, equilibrate and run with lambda=0; then I take the final
structure, set lambda=0.1, equilibrate and run; take the final
structure, set lambda=0.2, equilibrate and run; etc. Then I may find I
need more sampling, and I decide I will run a simulation at
lambda=0.15, I would take the final structure from lambda=0.1,
equilibrate and run. But I would not change anything in the rest of the
simulations, the one with lambda=0.2 is already done and it's fine. Is
it really important that the initial structure prior to equilibration
comes from lambda=0.1 or lambda=0.15? And if all I do is running a
longer simulation at lambda=0.1, the effect would probably be even
lower, I'd just leave the simulation at lambda=0.2 as it is.
You seem to mean that, if I create this new lambda=0.15 I would have to
rerun 0.2 with a new starting structure (still, with an equilibration
period). Sure, the results would be different, but not significantly so
(it would only be a different trajectory), and not worse than starting
with a structure obtained with lambda=0. Am I still missing something?
Try the all-new Yahoo! Mail. "The New Version is radically easier to use" The Wall Street Journal
More information about the gromacs.org_gmx-users