From: Thomas Evangelidis (tevang3_at_gmail.com)
Date: Tue Sep 22 2015 - 05:31:34 CDT
And this:
---------- Forwarded message ----------
From: Luca Maragliano <Luca.Maragliano_at_iit.it>
Date: 18 December 2014 at 15:35
Subject: RE: string method with swarms-of-trajectories - NAMD implementation
To: Thomas Evangelidis <tevang3_at_gmail.com>
Hi Thomas,
I don't know if the same code can run on GPU's, I have no experience on
that.
Most importantly however the code was done serially because the system is
super-simple (the AD
in vacuum). There is no point in doing the string method serially
otherwise, it would be a great
loss of time for any reasonable size system.
The evolution is just a steepest descent in the space of CVs, where you use
an instantaneous
approximation of the mean force rather than the true one (which otherwise
would be estimated
via an average on the x at z fixed).
the slowing down recovers for the use of instantaneous rather then average
and it is obtained by simply
scaling the time step of z motion wrt that of the x (which is internal in
NAMD).
This dynamics is performed at each image and can be sketched as
- evolve x with NAMD + harmonic force from potential
.5*k*(theta(x)-z)^2 and its proper tstep
- evolve z with force k*(thetha(x)-z) and scaled tstep
If the CV is non-linear then the mean force on z is multiplied by the
metric factor M.
Every few steps of this descent you reparametrize the values along the
curve (an example code in fortran
is also in the bundle). Reparametrization does not require MD but does
require communication of the CV
values of all images.
So in general you have
- few steps of descent in CV space, at each image independently of the
others
- reparametrization call
How many steps in between rep calls depends on the system. Too few would
require too much communication
and too many would break the curve integrity (but you can cycle the
reparametrization a few times, if needed).
Again please feel free to ask at any time
ciao,
Luca
------------------------------
*From:* Thomas Evangelidis [tevang3_at_gmail.com]
*Sent:* Thursday, December 18, 2014 2:07 PM
*To:* Luca Maragliano
*Subject:* Re: string method with swarms-of-trajectories - NAMD
implementation
Hi Luca,
Thank you for the clarification
> For the implementation of the otf version, look at table 1 of our JCTC
> paper (page 527).
> The otf method is similar to the mean forces method except that steps 1-3
> are
> substituted by the concurrent evolution of the CVs and the physical
> variables, coupled
> via the harmonic potential. In other words, rather than doing step 1 with
> fixed CV values
> you evolve the CVs as well but on a different, slowed-down time-scale,
> i.e. you give them
> a high mass or friction. More details are in appendix B.
> In terms of implementation, this amounts to modify the restraint dynamics
> of step 1 by
> introducing a line that evolves the CV values.
>
>
Do those files with the "otf" keyword in their name need the modification
you mentioned? I need to read more carefully your paper to understand how
program the restr_dih_otf.tcl script to evolve the CVs slowly. Right now
things are hectic, so I will probably come back with more questions. I just
have one simple question, I saw you evolve the image trajectories serially.
Does this mean the same code car run on GPUs?
thank you for your help!
Thomas
This archive was generated by hypermail 2.1.6 : Tue Dec 27 2016 - 23:21:20 CST