[gmx-developers] rlist larger than L/2 in simple neighbor search

[Lee-Ping Wang]

Hi there,

 

I will definitely keep you informed of my progress.  In QTPIE, the screened
Coulomb integral is actually an overlap integral between two s-type GTOs.  I
found yesterday that this makes implementation of the Ewald summation
especially easy, because the interaction between a Gaussian-blurred charge
and a periodic lattice of Gaussian-blurred charges should be exactly equal
to the long-range part of the Ewald interaction energy with a certain choice
of the Gaussian width.

 

You are right that I do need to perform one periodic summation for each of
the matrix elements in the charge-transfer problem Mq=x (M contains Coulomb
integrals, x contains overlap integrals).  For now, efficiency
considerations aside, I plan to adopt the Ewald-summation code already in
GROMACS and see if it works.

 

I ran some dynamics yesterday with an infinite-sized cutoff and my system
appeared to be frozen; this is probably the "ordering" artifact you're
talking about.  I'm working on implementing the Ewald summation today and
will give you more updates.

 

-          Lee-Ping 

 

From: gmx-developers-bounces at gromacs.org
[mailto:gmx-developers-bounces at gromacs.org] On Behalf Of Gerrit Groenhof
Sent: Wednesday, October 13, 2010 2:34 AM
To: Discussion list for GROMACS development
Subject: Re: [gmx-developers] rlist larger than L/2 in simple neighbor
search

 

On 10/12/2010 06:49 PM, Lee-Ping Wang wrote: 

Hi there,

 

For the past few months, I've been adding a charge equilibration model to
GROMACS (to be specific, QTPIE - a refinement of QEq).  For the model to
work properly, each atom must "see" each other atom in the electrostatic
interaction (it is a screened Coulomb interaction).  This is easy enough to
do with pbc = no; I just set all the cutoffs to zero, and add a new
neighbor-list that includes excluded atom pairs.


Since we'd be interested in that, I would appreciate if you can keep us
informed of your progress.




 

Now I would like to run the simulation with PBC turned on, which means
either implementing Ewald summation or using cutoffs for the electrostatic
interaction.  When cutoffs are turned on (with or without PBC), the model
goes haywire even with very long cutoff lengths.  I haven't implemented
Ewald summation for the screened Coulomb interaction, and I'm not sure if
it's worth the effort if an easier fix is possible.


If I recall correctly, the screened Coulomb here means the overlap integral
between s-type Slater functions. This interaction behave as a normal coulomb
at large distances, so that coudl explain the issues with increasing the
cut-off. But an Ewald expression will be difficult as well, since  the all
pairs-wise interactions are required to compute the charge flow parameters.
I have so far only seen it applied to molecules. For non bonded system like
a box of water, a direct charge flow between atoms far apart is unlikely
though. Could one not simply scale also the charge transfer variables to
zero with distance as well?



 

As a quick fix, I changed the simple neighbor-searching code to set rcut =
1000000 when QTPIE is turned on.

 

I am worried about the other possible consequences of my change; will this
adversely affect the molecular dynamics or create artifacts in some
unexpected way?  

You will get the ordinary artefacts caused by cutoff, like ordering etc.



 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-developers/attachments/20101013/54781d17/attachment.html>