Re: [AMBER-Developers] lmod xmin problems with sander and sqm

From: Andreas Goetz <agoetz.sdsc.edu>
Date: Thu, 10 Dec 2009 17:48:26 -0800

> Thanks for the info. Istvan and I spent all day today looking at
> minimizations and lmod in both sander and sqm. I'll be checking some things
> in soon. I think we have fixed the problems in sander (zero rms gradient,
> bad results for the lmod tests) and fixed up other printouts.

Thanks. Let me know when this is checked in and I will run the tests.

> There are still some things to look at in terms of taking too many steps
> in sqm. It may be a problem with slightly inaccurate gradients, even though
> one can take sqm minimizations down to very small values. Some molecules seem
> to work *much* better than others, and we are still trying things. But right
> now, this seems to be a matter of efficiency, not accuracy of the final
> result.

Yes, it is normal that the number of optimization steps strongly depends
on the system under investigation. The accuracy of the final
optimization results of sqm has been established (see below), so it is
only about efficiency. I compared three programs given identical
starting conditions with optimizations in the same coordinate space. It
is just the optimization algorithms that differ. Here is some updated
data for NMA (I started from a reasonably distorted structure) which
includes mopac2009 with L-BFGS (all AM1 data):

GRMS steps program algorithm E(final)

0.02 47 orca2.6.35 BFGS -1008.31272
0.02 90 mopac2009 EF -1008.31236
0.02 169 mopac2009 L-BFGS -1008.31236
0.02 186 sqm L-BFGS -1008.31227

0.05 47 orca2.6.35 BFGS -1008.31272
0.05 87 mopac2009 EF -1008.31236
0.05 109 mopac2009 L-BFGS -1008.31168
0.05 125 sqm L-BFGS -1008.31172

GRMS in kcal/(mol*A)
E(final) in eV (electronic energy)
EF = Eigenvector following

I used eV here to avoid any conversion errors (eV is the only energy
unit which is printed by all three programs output)

I would say we have excellent agreement (1 eV = 23 kcal/mol) and the
efficiency is just a matter of the algorithm. This is something we may
think of for future releases (including coordinate spaces which are more
appropriate for geometry optimizations than Cartesian).

> I'm glad to have the orca and mopac2009 results for comparison. Have you
> ever compared the SCF energies for a given structure in sqm vs orca vs
> mopac2009? (We did comparisons to Gaussian and Martin Field's programs in the
> original paper.)

AM1 Energies for the inital NMA structure (attached) I used:

program E/eV

mopac2009 -999.55240
orca2.6.35 -999.55279
sqm -999.55253

The SCF has been tightly converged in all cases. Not perfect but quite
good agreement I would say. I will see if I can get hold of atomic units
data since this would eliminate all potential differences that might
stem from the use of different physical constants.

All the best,
Andy


> ...thanks....dac
>
>
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers


   12
NMA distorted
 C -0.767000 3.380000 4.245000
 C -1.809707 2.570219 4.710852
 N -2.661673 2.646098 3.869189
 C -3.662846 1.855199 3.619926
 O -1.634379 2.116691 5.584678
 H 0.218271 3.164467 4.684075
 H -0.578839 3.210298 3.174581
 H -1.024346 4.421330 4.488732
 H -2.569216 3.204198 3.074324
 H -3.347225 0.824359 3.812610
 H -4.529359 1.998602 4.311686
 H -4.036138 1.906340 2.552627


_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Thu Dec 10 2009 - 18:00:02 PST
Custom Search