Re: Metadynamics Problems

From: Giacomo Fiorin (giacomo.fiorin_at_gmail.com)
Date: Thu Mar 08 2012 - 13:15:33 CST

Yes, that's exactly the reason why hills near or beyond the grid boundaries
are still recorded, because not doing so would lead to ignoring them and
generating a spurious artificial minimum right beyond the repulsive wall.
 If you set upperBoundary beyond upperWall, you have the annoyance that
metadynamics will include the wall itself in the PMF, but at least you can
keep the simulation always fast.

About (2). yeah, it would be best if you can reproduce what you observed.
 Quite often if the PMF doesn't make much sense it is not because of
configuration errors on your part, but because of the fact that the
collective variables are not describing the system well.

My advice is, instead of just the PMF have a closer look at the trajectory
(both of your variable and of the entire system) to see what's going on.

Giacomo

On Thu, Mar 8, 2012 at 12:59 AM, Aron Broom <broomsday_at_gmail.com> wrote:

> Hi Giacomo,
>
> Thanks for the clarification about the hills, I see now that all the
> explicit hills are indeed past the boundary of the colvars. In this
> particular case it was a distance for a ligand to a protein, so the only
> real "boundary" is that set in the colvar. I guess then an intelligent
> thing to do is to increase the strength of the boundary wall constant? The
> case where it was most obvious was a fairly unusual case where there was
> very rapid diffusion because there was no solvent, and in most cases I
> don't actually notice a difference, but I just wondered if something was
> off. I see now though that if this wasn't the case, when the biasing
> potential from the metadynamics became stronger than the parabolic wall,
> the system would fall into an artificial minimum at the position of the
> wall.
>
> In terms of (2): what you mention is what I initially thought when I saw
> the PMF - that setting the hill height lower had somehow caused all the
> previous hills to be reinterpreted as having the same lower height - but
> this isn't the case. If if I restart and only run for a short time, the
> PMF does not change appreciably. The change only become apparent after
> running for longer times. I haven't had the opportunity to do a thorough
> test to see if I can understand what is happening (such as what happens
> when the hill height is actually increased). Perhaps best if I find some
> time for that shortly and then report back with more detailed and clear
> information on the problem, I just thought maybe someone had seen it before.
>
> Thanks for the rapid response!
>
> ~Aron
>
>
> On Wed, Mar 7, 2012 at 7:02 PM, Giacomo Fiorin <giacomo.fiorin_at_gmail.com>wrote:
>
>> Hello Aron, and thank you for the compliments :-)
>>
>> (1) is a fully normal behavior. The issue is that whenever you go beyond
>> the boundaries of the grid, the biasing forces suddenly disappear. For
>> this reason, as a safety precaution the metadynamics bias stores all those
>> gaussians that are near the boundary, so that they can be still used when
>> you're off the grid. In your case, it looks like you keep reaching the
>> boundaries of the grid very very often.
>>
>> Also, the fact that your simulation slows down (and not just the startup
>> phase when the file is read) suggests that you not only get close to the
>> boundaries, but you are going outside of them as well. At that point, the
>> advantage of the grids disappears.
>>
>> You should expand the boundaries to the full range of values of your
>> variable(s), plus a buffer the size of a Gaussian.
>>
>> If you're in the situation where values close to the boundary are visited
>> often but the variable never goes beyond it (e.g. a RMSD variable around
>> zero), you can safely delete those gaussians from the state file, because
>> they're never going to be used.
>>
>> But as I said, if the step performance is slow, it seems that you are
>> indeed using those gaussians and therefore should expand the grid to get
>> rid of them.
>>
>>
>> In (2) you seem do be doing something like well-tempered metadynamics,
>> that will be available soon.
>>
>> But I don't understand the results you got: are you saying that after
>> restarting while setting smaller hills, the smaller height affects not only
>> the new hills, but also the existing PMF up to that point?
>>
>>
>> On Wed, Mar 7, 2012 at 6:29 PM, Aron Broom <broomsday_at_gmail.com> wrote:
>>
>>> Thanks Axel,
>>>
>>> I'm using NAMD 2.8 (2011-05-31), the latest non-nightly build release.
>>>
>>> Oh I see, they contributed the entire colvars module, how wonderful,
>>> thanks a lot to them!
>>>
>>> ~Aron
>>>
>>>
>>> On Wed, Mar 7, 2012 at 6:22 PM, Axel Kohlmeyer <akohlmey_at_gmail.com>wrote:
>>>
>>>> aron,
>>>>
>>>> two items:
>>>>
>>>> 1) you forgot one *very* important piece of information:
>>>> what version of NAMD exactly are you using?
>>>>
>>>> 2) the documentation gives clear credit to the authors
>>>> of the module (giacomo fiorin and jerome henin).
>>>>
>>>> to add to the praise. i am currently interfacing the
>>>> colvars code to another MD code, and it has been
>>>> a breeze with only very minor issues to work out.
>>>>
>>>> cheers,
>>>> axel.
>>>>
>>>> On Wed, Mar 7, 2012 at 6:13 PM, Aron Broom <broomsday_at_gmail.com> wrote:
>>>> > Hello NAMD Users,
>>>> >
>>>> > I've been using the Metadynamics module in NAMD for some time now, as
>>>> it is
>>>> > quite simply fantastic. I have, however, become confused by a few
>>>> things
>>>> > and wondered if anyone could clarify.
>>>> >
>>>> > 1) When "usegrids on" is set, my impression is that all repulsive
>>>> gaussian
>>>> > hills added are to be descretized onto grids, such that no individual
>>>> hills
>>>> > are recorded, and hence, the speed of the simulation does not slow
>>>> down as
>>>> > the number of hills added increases. What I've found is that a few
>>>> hills
>>>> > still get added periodically, and over very long simulation times (or
>>>> many
>>>> > many hill additions) the simulation slows down considerably, and the
>>>> size of
>>>> > the .colvars.state file, which holds the grids (and now also these
>>>> spurious
>>>> > hills) grows to mammoth proportions (hundreds of MB). Is this
>>>> supposed to
>>>> > happen? Has anyone else encountered this?
>>>> >
>>>> > 2) If a simulation is restarted and none of the hill parameters are
>>>> changed,
>>>> > everything is fine. If, however, one changes the height of hills
>>>> after a
>>>> > restart (still when using grids), the new spurious hills (see above)
>>>> are
>>>> > correctly identified as having a small height, but the effect on the
>>>> > resulting PMF (and I presume the grids) is that these hills seem to
>>>> subtract
>>>> > from the aggregate biasing force, such that if one runs a simulation
>>>> for
>>>> > time t with hill height h, and then restarts with hill height h/10
>>>> and runs
>>>> > for 10*t, the resulting PMF is essentially flattened and one returns
>>>> to an
>>>> > unbiased run (rather than simply smoothing out the PMF as one would
>>>> have
>>>> > hoped). Is the changing of hill heights not something that can be
>>>> done, or
>>>> > is it something that just can't be done while using grids?
>>>> >
>>>> > Again, many thanks to whomever added this feature to NAMD.
>>>> >
>>>> > ~Aron
>>>> >
>>>> > --
>>>> > Aron Broom M.Sc
>>>> > PhD Student
>>>> > Department of Chemistry
>>>> > University of Waterloo
>>>> >
>>>>
>>>>
>>>>
>>>> --
>>>> Dr. Axel Kohlmeyer
>>>> akohlmey_at_gmail.com http://goo.gl/1wk0
>>>>
>>>> College of Science and Technology
>>>> Temple University, Philadelphia PA, USA.
>>>>
>>>
>>>
>>>
>>> --
>>> Aron Broom M.Sc
>>> PhD Student
>>> Department of Chemistry
>>> University of Waterloo
>>>
>>>
>>
>
>
> --
> Aron Broom M.Sc
> PhD Student
> Department of Chemistry
> University of Waterloo
>
>

This archive was generated by hypermail 2.1.6 : Mon Dec 31 2012 - 23:21:18 CST