minimization and gradient trend

From: Francesco Pietra (chiendarret_at_gmail.com)
Date: Mon Oct 31 2011 - 03:57:10 CDT

Hello:

From: Jim Phillips (jim_at_ks.uiuc.edu)
Date: Thu Jul 22 2010 - 14:52:07 CDT :
The reported gradient value is along a non-normalized search vector, so
its magnitude is not physically meaningful. Even if it was, it probably
wouldn't be quite what you want, since it only takes one large force to
kill a simulation. Specifying that all net atomic forces are below a
certain value would accomplish that, although that would need some
additional control logic to stop minimization ahead to schedule.
My philosophy has always been that if a thousand steps of minimization
isn't enough to start equilibration, then a million won't be either.

Does that mean that equilibration should be attempted any time that
the gradient shows a good trend, irrespective of the final value
attained?

Usually, with NAMD2.8-CUDA multi GTX-580 gradient decreased below 5.
With a new system, now the gradient drops to ca 650 and stands there.
The start was

ENERGY: 0 12549.3777 3696.8817 797.9205
412.7359 -128440.2111 16996805.3527 0.0000
0.0000 0.0000 16885822.0573 0.0000
16885822.0573 16885822.0573 0.0000 3558449.8652
3560832.3956 428929.6100 3558449.8652 3560832.3956

MINIMIZER SLOWLY MOVING 92 ATOMS WITH BAD CONTACTS DOWNHILL
ENERGY: 1 12555.9917 3696.9878 797.9205
412.7359 -128560.4149 1350221.0316 0.0000
0.0000 0.0000 1239124.2526 0.0000
1239124.2526 1239124.2526 0.0000 624645.3412
626926.0315 428929.6100 624645.3412 626926.0315

rapidly finding no more atoms to move:

MINIMIZER SLOWLY MOVING 2 ATOMS WITH BAD CONTACTS DOWNHILL
ENERGY: 6 12571.7126 3696.4479 797.9205
412.7359 -128695.7186 27881.6581 0.0000
0.0000 0.0000 -83335.2435 0.0000
-83335.2435 -83335.2435 0.0000 8936.8431
10934.2709 428929.6100 8936.8431 10934.2709

MINIMIZER STARTING CONJUGATE GRADIENT ALGORITHM
LINE MINIMIZER REDUCING GRADIENT FROM 2.8723e+08 TO 287230
ENERGY: 7 12540.0058 3694.3472 797.9177
412.6201 -128703.1532 27637.7033 0.0000
0.0000 0.0000 -83620.5591 0.0000
-83620.5591 -83620.5591 0.0000 8778.1098
10763.9832 428929.6100 8778.1098 10763.9832

FINAL:

ENERGY: 10000 4862.2449 3817.6582 776.9748
22.6194 -175462.9400 12465.0669 0.0000
0.0000 0.0000 -153518.3758 0.0000
-153518.3758 -153518.3758 0.0000 -7851.2169
-7795.6746 428929.6100 -7851.2169 -7795.6746

thanks
francesco pietra

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:57:53 CST