AMBER Archive (2007)

Subject: Re: AMBER: Sander (Amber 9) crash on cortex

From: Thomas Cheatham (tec3_at_utah.edu)
Date: Thu May 10 2007 - 21:56:19 CDT


> simulation with cut=10 and it ran successfully. So, this is a fix. Now, my
> understanding is that the higher the cutoff value the more accurate the energy
> values for the simulation. However, I understand that PME does compensate for
> the cutoff. In your opinion, will the change from cut=10 from cut=20 make

This is not entirely correct, at least when talking about electrostatics
with Ewald or particle mesh Ewald; if you were to run a straight cutoff
simulation, then yes, one might think that making the cutoff larger will
lead to better behavior. This too is an erroneous assumption due to
fortitous cancellation of error. There are a number of earlier papers
that discuss these issues, a classic being Steinhauser on peptide
simulation.

If you are using a cutoff, then this should be smoothed or shifted in
terms of the energy or force; there are classic's by Brooks and Steinbach
and more recent ones by Daggett and others that provide the gory details.

With Ewald, it is all about the balance. The direct space (i.e. within
the cutoff) and the reciprocal (effectively what is missing for the true
periodic image) need to be balanced in term of their accuracy. It is not
formally correct to think of the direct as short range and the reciprocal
as long range, although this is often done. In each case, the term is
smoothed in some manner. In the case of the direct sum, within the
cutoff, this is smoothed by the erfc() which brings the terms close to
zero at the cutoff. In AMBER you can set DSUM_TOL to control how accurate
this is. The higher the accuracy and shorter the cutoff means that you
have to put more accuracy into the reciprocal sum to compensate; either
more lattice vectors in Ewald, or finer grid, higher interpolation in PME.
If you want to run a 20A cutoff fine, then you can run a not so accurate
reciprocal and get equivalent answers. The problem with this is that the
direct part costs significantly more in terms of computer time so you
are shooting yourself in the foot...

When Ross Walker mentioned 8-10A, this is how we have balanced both
accuracy and performance with PME with default parameters. We could
probably go to even smaller cutoffs (with better runtime performance)
however we need to worry about vdw too... These cannot be run at 6 A
cutoff accurately (and keeping two pairlists kills the speed advantage
somewhat); 8-10A is pushing it, although in AMBER you can apply a uniform
density approximation to partially correct...

Short summary: Bigger isn't necessarily better; you want balance between
the two terms. You can check accuracy by increasing DSUM_TOL and/or grid
density, interpolation order, etc. We tend to try to run with force
errors on the 10**-6 range, i.e. pretty small. A good test is energy
conservation (as you alter the variables in NVE simulation).

Good luck with your protein structures.

-- tec3

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