Wild guess: does this code increase the step size without bound every time
an iteration leads to a lower energy? If so, limiting the max step size
could be a solution as I've seen this type of algorithm (even in steepest
descent) lead to pathologies in energy minimization of systems that start
far enough from a minimum that the step size becomes unwieldy large at some
point.
On Thu, Mar 11, 2021 at 2:18 PM David A Case <david.case.rutgers.edu> wrote:
> On Thu, Mar 11, 2021, Adrian Roitberg wrote:
>
> >> After the first initial steps using steepest descent, CG energies and
> >> forces exhibited frequent jumps of several orders of magnitude as the
> >> minimization progressed.
>
> But it looks like you are still getting lower energies.
>
> FWIW: the conjugate gradient code in runmin is about the oldest code
> around in Amber: I think it even precedes molecular dynamics. The entire
> runmin subroutine will be gone from future releases of sander, but not for
> AmberTools21. I don't know how we will handle the situation for
> pmemd.cuda.
>
> >> If the xmin minimizer is used with PRCG then the issue disappears.
>
> Indeed. Anyone who wants a good minimization should use xmin; just be sure
> to turn off SHAKE. (The latter is done automatically in development
> codes.)
>
> ....dac
>
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Mar 11 2021 - 21:00:02 PST
