AMBER Archive (2007)

Subject: RE: AMBER: SCF convergence issues

From: Ross Walker (ross_at_rosswalker.co.uk)
Date: Fri Feb 23 2007 - 13:00:21 CST

Hi Steven,

A cutoff of 6.5 angstroms for MM is dangerously small and a cutoff of 3.5
angstroms for the qm is very very small. I would never reduce this below 8
angstroms. I know this can give problems with distributed QM regions in
small periodic boxes but it is there for a reason. The cutoff cannot extend
beyond half the box or you get periodicity artifcacts and you can't make it
less than 8 angstroms or you seriously truncate the interactions. Thus the
only really acceptable solution is to make the box bigger.

> It seems one set of SCF iterations converges successfully, and it is
> then reported that the forces are returned (which get dumped to
> forcedump.dat) but then several more SCF iteration cycles are run, one
> of which fails to converge.

Most of the scf cycles converge properly. It is only one of them that don't.
However, the difference in energy between two steps is huge. E.g.

QMMM: SCF Energy = -1325.32549345000734 KCal/mol, -5545.16186459483106
KJ/mol
QMMM: SCF Energy = -1298.86881677272322 KCal/mol, -5434.46712937707434
KJ/mol

Is this system minimized? It looks to me that this system may be a long way
from equilibrium and that is causing the problems. Did you minimize it (even
100 steps of steepest descent would be good). I would then try some simple
MD and see how it behaves.

The above problem could also be caused by atoms moving in and out of the
cutoff. With such a small cutoff you will get huge jumps in energy as an
atom crosses the cutoff and this will lead to instabilities.

I would also try turning pseudo_diag back on as this can often get you away
from bad initial densities quickly. The issue above is that for the SCF set
that doesn't converge the initial density matrix is very bad. This is likely
caused by the system moving a long way from the previous step. I.e. several
atoms moved outside the small cutoff causing a large change in the field and
the difference was so severe that the scf routine couldn't recover.

On another note I don't think I ever tested dumpfrc with QM/MM. This tries
to decompose the energies and forces and this is not possible with QM/MM,
since you can't pairwise decompose things and so this could also be the
route of your problem. Why are you trying to decompose the force array? If
you want to test the accuracy of the forces you can use do_debugf=1 and
atomn=1,2,3 etc. You probably want to set scf_conv=1.0D-8 and tight_p_conv=1
though to obtain decent gradients. 10-6 is probably not tight enough. I'd
also just try some simple MD first.

All the best
Ross

/\
\/
|\oss Walker

| HPC Consultant and Staff Scientist |
| San Diego Supercomputer Center |
| Tel: +1 858 822 0854 | EMail:- ross_at_rosswalker.co.uk |
| http://www.rosswalker.co.uk | PGP Key available on request |

Note: Electronic Mail is not secure, has no guarantee of delivery, may not
be read every day, and should not be used for urgent or sensitive issues.

-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber_at_scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu