On Sat, Jun 14, 2014 at 11:53 PM, Kshatresh Dutta Dubey <kshatresh.gmail.com
> wrote:
> Okay Sir, I will do it, but following Gaussian output may, perhaps, make
> clear what I want to say, this is output of Gaussian for normal termination
> of every sander minimization step:
> *Maximum Force 0.129720 0.000450 NO*
> * RMS Force 0.019740 0.000300 NO*
> * Maximum Displacement 0.095943 0.001800 NO*
> * RMS Displacement 0.021493 0.001200 NO*
> * Predicted change in Energy=-5.123724D-02*
>
> and Gaussian adds this energy to sander output. So how one can believe
> that it is optimizing the system when proper convergence is not achieved.
>
​Do these values never change? Or is the "NO" part what you find
concerning? I'll repeat what others have said here -- Gaussian does NOT
have any control over coordinates when used by sander -- it is ONLY used to
compute energies and forces (the gradient of the potential) on the QM
region in the presence of point charges representing the atoms of the MM
region. In fact, Gaussian does not even compute all of the energies and
forces for the atoms in the QM region -- MM atoms interact with QM atoms
via the standard Lennard-Jones potential, so Gaussian does not even know
about all forces acting on the QM atoms. Therefore, Gaussian should _not_
be doing any minimizations (since only sander has all of the information).
Most importantly, what this means is that Gaussian, in general, *will not*
ever think an optimized geometry is reached, *even when it is* (unless the
entire system is QM -- in which case just use Gaussian). This is because
Gaussian does not know about all of the interactions taking place. Like
others have said, sander just asks Gaussian to do what is necessary to
obtain the information sander needs to perform its tasks (which is
basically a set of energies and forces for atoms in the QM region in a
field of MM charges). sander is the master that is contracting out work to
Gaussian (and geometry optimization is _not_ in Gaussian's job description
here).
I will say that in general, the minimizers in sander are worse than the
minimizers in ab initio programs like Gaussian. So it will take longer to
achieve "convergence" with sander than it will if you are strictly using
the Gaussian optimizers. If an optimized geometry is important, look into
better minimizers (like the XMIN minimizers, described in the manual). If
all you want to do is prepare to run dynamics, reaching a true minimum is
not important.
​​
> Is it Ok, and can I trust that Gaussian is doing well here?
>
You shouldn't blindly trust results you get from any program. They should
make sense to you. If they don't make sense, it's either because: a) some
assumption you made when predicting what could happen is wrong. b) You made
a mistake in your input. or c) there is a bug in the program. It is safer
to assume either (a) or (b), although (c) can happen on occasion, too. But
it's time to trust that Gaussian does not -- and should not -- ever do
geometry optimizations when used in the sander QM external interface. And
even better, understand why that is true.
HTH,
Jason
--
Jason M. Swails
BioMaPS,
Rutgers University
Postdoctoral Researcher
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Sat Jun 14 2014 - 22:30:02 PDT
