[gmx-users] minimization -- bugzilla or general advice?

[David van der Spoel]

David Mobley wrote:
> Xavier,
> 
>>  I was surprised by the value you use for emstep=1.0e-8. This is the
>>  "maximum step size" and is in nm! I think that's where your problem is.
> 
> I added this recently in testing because I thought perhaps the problem
> could be due to using too large of steps. It does not help. So
> everything I send in my previous e-mail holds true regardless of which
> emstep value I use (default vs 1.0e-8).
> 
>>  The minimizer is probably incapable to find anything better for your
>>  system with this step size and therefore stops.
>>
>>  Did you try to keep the default value emstep=0.01 nm or something
>>  bigger than 1.0e-8?
> 
> Also, I think your argument is backwards: steepest descents should
> work better (but slower) the smaller the step size, I think. If steps
> are too big it won't be able to move down the gradients accurately
> enough and risks not being able to find the minimum. The smaller the
> steps are the more accurately it should be able to find the minimum.
> 
David,

if you add 1+1e-8 in float at least the result is 1. Since coordinates 
in typical systems are on the order of a few nm such a stepsize does 
indeed not make sense, but since you tried larger stepsize as well the 
problem is real. The stepsize is used as the maximum displacement along 
the force vector in the steep algorithm.

I encourage you to submit a bugzilla with a few examples.


> Thanks,
> David
> 
> 
>>  Best
>>  XAvier
>>
>>
>>
>>  > I often have problems with minimization in gromacs, but now I think I
>>  > have something a little more systematic to say.
>>  >
>>  >First, my system: I am simulating a short peptide in a mix of two
>>  > co-solvents. I use a program called packmol which can start pack
>>  > specified number of molecules into a box of specified size, with a
>>  > specified distance tolerance between atoms, to generate starting gro
>>  > files. I am simulating a number of variants of this peptide in various
>>  > mixes of the two co-solvents (one of which is water).
>>  >
>>  > The problem I am having is that, even in double precision, gromacs
>>  > fails to properly minimize some of the systems -- minimization using
>>  > any of the minimizers I've tried (including steepest descents, my
>>  > preference) often terminates early (i.e. after ~50 steps) even in
>>  > double precision, leaving really large forces on some of the atoms
>>  > (usually 10^10 or more). The largest forces are usually on one atom of
>>  > a water molecule. Hence, when I plug the minimized output into
>>  > equilibration, my simulations explode due to large forces.
>>  >
>>  > Small tweaks to the system can change whether minimization works
>>  > properly or not in some cases -- for example, switching from double
>>  > precision to single precision will make some systems which fail in
>>  > double minimize properly, as will changing cutoffs. I assume also
>>  > things like changing architectures will allow some systems to minimize
>>  > (I've seen this behavior in the past).
>>  >
>>  > The problem is, I can find no combination of settings which will allow
>>  > *all* of my systems to minimize properly. Inevitably some fail, and I
>>  > end up having to go in by hand, look at the minimization log file, see
>>  > which water molecule has the huge force on it, and tweak its position
>>  > slightly by hand in the input gro file. Typically then minimization
>>  > will proceed just fine.
>>  >
>>  > This brings me to my question: Are the minimization algorithms
>>  > truncating some of the forces at some threshold value or something?
>>  > That is, why does minimization often end even though forces are still
>>  > huge?
>>  >
>>  > Currently, I'm using steepest descents with constraints turned off,
>>  > double precision, and emstep = 1.0e-8. I'm using Gromacs 3.3.2, rlist
>>  > = 0.9 = rcoulomb, rvdw-switch = 0.75 and rvdw = 0.8, pme_order 4,
>>  > ewald_rtol = 1e-04, and fourierspacing = 0.12, DispCorr = AllEnerPres,
>>  > and switched vdw.
>>  >
>>  > Any tips will be appreciated. Having to tweak coordinates by hand (or
>>  > write a script to do it) to get things to minimize isn't really a
>>  > satisfactory solution for me.
>>  >
>>  > I could submit a bugzilla, but I imagine what will happen if I submit
>>  > the cases that don't work for me is that they'll work fine for whoever
>>  > addresses the bugzilla and so this won't be regarded as a real
>>  > problem. Alternatively I could submit a whole slate of starting system
>>  > configurations and someone can debug just the ones that fail for them.
>>  >
>>  > Thanks,
>>  > David
>>  > _______________________________________________
>>  > gmx-users mailing list    gmx-users at gromacs.org
>>  > http://www.gromacs.org/mailman/listinfo/gmx-users
>>  > Please search the archive at http://www.gromacs.org/search before posting!
>>  > Please don't post (un)subscribe requests to the list. Use the
>>  > www interface or send it to gmx-users-request at gromacs.org.
>>  > Can't post? Read http://www.gromacs.org/mailing_lists/users.php
>>
>>  -----------------------------------------------------
>>  XAvier Periole - PhD
>>
>>  NMR & Molecular Dynamics Group
>>  University of Groningen
>>  The Netherlands
>>  http://md.chem.rug.nl/~periole
>>  -----------------------------------------------------
>>  _______________________________________________
>>  gmx-users mailing list    gmx-users at gromacs.org
>>  http://www.gromacs.org/mailman/listinfo/gmx-users
>>  Please search the archive at http://www.gromacs.org/search before posting!
>>  Please don't post (un)subscribe requests to the list. Use the
>>  www interface or send it to gmx-users-request at gromacs.org.
>>  Can't post? Read http://www.gromacs.org/mailing_lists/users.php
>>
> _______________________________________________
> gmx-users mailing list    gmx-users at gromacs.org
> http://www.gromacs.org/mailman/listinfo/gmx-users
> Please search the archive at http://www.gromacs.org/search before posting!
> Please don't post (un)subscribe requests to the list. Use the 
> www interface or send it to gmx-users-request at gromacs.org.
> Can't post? Read http://www.gromacs.org/mailing_lists/users.php


-- 
David.
________________________________________________________________________
David van der Spoel, PhD, Assoc. Prof., Molecular Biophysics group,
Dept. of Cell and Molecular Biology, Uppsala University.
Husargatan 3, Box 596,  	75124 Uppsala, Sweden
phone:	46 18 471 4205		fax: 46 18 511 755
spoel at xray.bmc.uu.se	spoel at gromacs.org   http://folding.bmc.uu.se
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++