Re: well-tempered metadynamics deforms the water box and the system becomes unstable..

From: Giacomo Fiorin (giacomo.fiorin_at_gmail.com)
Date: Tue Oct 20 2020 - 16:01:43 CDT

Hi Zeynab, thanks for looking at it in detail. Can you give some details
about how you are restarting the simulation? E.g. did you use either the
NAMD keyword/command colvarsInput or the "cv load ..." command?

I tested the Colvars versions bundled with the latest NAMD major versions:
2.12, 2.13 and 2.14. 2.12 and 2.13 include a minor bug (see here
<https://github.com/Colvars/colvars/pull/357> for the detailed fix), which
adds a new Gaussian at the first step, which coincides with the last step
of the previous run, if you are restarting. In your case, that would mean
an extra Gaussian hill every 1 ns. NAMD 2.14 will instead wait for a
"true" new step when restarting.

However, any of these versions do give you correctly Gaussian hills with
lower weight than the initial ones, following the well-tempered scaling.

Those plots suggest that either there is a bug that is triggered by
specific ways of restarting (hence my asking), or there is no restart
information given to the metadynamics code and it always starts over again
every ns.

Thanks again for checking this out,
Giacomo

On Tue, Oct 20, 2020 at 4:37 PM zeynab hosseini <hosseinizeynab93_at_gmail.com>
wrote:

> Hi Giacomo,
>
> I‌ really appreciate your help and the comprehensive explanation. Just to
> review the details of my well-tempered metadynamics simulation, in my case
> the initial hill weight is 0.1 kcal/mol and due to the limited wall time I
> ‌ can't perform one long-time continuous simulation. Then I ‌ divided one
> long simulation into several dependent/discontinuous short jobs each taking
> one nanosecond.
>
> I‌ expected the hill weight to get updated when restarting dependent
> simulations, but although the collective variable (CV) values are updated
> properly, the hill weights are not. If I understood correctly, you meant
> that since the newly added Gaussians are scaled based on the number of
> Gaussians already added at the CV‌ location, then starting the next
> nanosecond of the simulation again with the starting hill weight of 0.1
> kcal/mol doesn't matter, since if the CV is located at the place where
> already enough Gaussians are added (depending on shape of the free energy
> surface) then the hill weight would reduce automatically. I may be wrong,
> but if the length of each discontinuous/dependent simulation wasn't long
> enough, then the hill weight may not have enough time to reduce efficiently
> before the next job starts and this causes the simulation never converges--000000000000ff2d7705b2208b1e--

This archive was generated by hypermail 2.1.6 : Thu Dec 31 2020 - 23:17:14 CST