[gmx-developers] double precision minimization accuracy limits.
David.Osguthorpe at UCHSC.edu
David.Osguthorpe at UCHSC.edu
Wed Jan 23 00:33:30 CET 2008
>
>Unless I'm missing something, this is a users-list question. Energy
>minimization is going to get stuck in a local minimum.
>
But the local minimum should be a minimum down to machine accuracy - at least thats
achievable with force field/MD programs I have written and used in the past - unless there are problems with
derivatives due to discontinuities (hence the use of switching functions) or derivative errors (ie. there are numerical errors in the
derivatives which are due to programming errors or approximations used in computing the forces
from the coordinates and their correspondence to the energies).
Hence the developers list - I dont know how most users are going to know the exact numerical coding
of the derivative computation functions.
>The acid-test of
>a minimization is if the resulting structure can be used to start an MD
>run.
Not true - energy minimization is required for example to compute normal modes
where good convergence is needed for isolated molecules to ensure the 6 lowest frequency
modes are close to zero - such structures need to be far more converged (ie. the maximum derivative
should be less than 0.10 kJ nm-1 - we used to use < 0.001 kcal angstrom-1)
to produce valid results whereas MD can be started from much larger derivatives.
To me a maximum derivative of 232 kJ nm-1 is too big - a double precision program
should be able to get them lower than that even for large protein systems.
The definition of a local minimum is all derivatives are zero
to infinite precision - otherwise its not a local minimum or minimum of any kind.
For limited precision computers we accept a definition of say the maximum derivative < 0.01 kJ nm-1.
>I don't see why you couldn't grompp your next energy minimization using
>-t .trr -e .edr but again, that should not be necessary.
thanks - I missed that possibility.
David
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-developers/attachments/20080122/2bda69c5/attachment.html>
More information about the gromacs.org_gmx-developers
mailing list