Hello Mani,
I'll add my two cents worth to this discussion. Diffusion coefficients, when you are in the hydrodynamic regime are mass independent. For simulations of molecules larger than water to the microsecond time scale, you are in the hydrodynamic regime. Thus changing the masses will not gain you anything at all. Most dynamics, even that of smallish molecules translate in the hydrodynamic mode (and to calculate it you must use slip boundary conditions in classical hydrodynamics), whenever the solute is simply larger than any void volumes in the solvent. This certainly covers your molecules of interest in water (because you are in water, you would need the hydrodynamic stick boundary conditions to calculate transport in classical hydrodynamics). The underlying reason for this is that the motion is overdamped, so the accelerations average to zero and thus mass disappears from the picture.
When you use a water model such as TIP3P, the viscosity is already about a factor of two too small and dynamics is already accelerated. This is a defect of the water model, but it comes in handy in improving the speed of simulations when global solute dynamics is not intended to be extracted from the trajectory; only local, nonhydrodynamic dynamics is correct in that water model.
One possible way to reach very long simulation times is to try Brownian dynamics simulations. Another is to generate a coarse grained model of your system so that you have fewer effective atoms to move. To preserve chemical accuracy, you would need to train your models on shorter atomistic simulations.
A third method is the brute force method. We published the first long scale simulation (.32 microseconds) of an antibody (Trastuzumab) using ff99SB in TIP3P in 2010. The work took two years of computation because we didn't have GPUs at that time. It was still worth it. People are doing microsecond long simulations on GPUs now, so you may resign yourself to spending the time doing it right.
Cheers,
Sergio Aragon
Professor of Chemistry
San Francisco State University
-----Original Message-----
From: Manikanthan Bhavaraju [mailto:manikanthanbhavaraju.gmail.com]
Sent: Tuesday, May 13, 2014 10:05 AM
To: AMBER Mailing List
Subject: Re: [AMBER] Scaling of atomic mass of water molecules
I taught of lowering the masses by half. However, your explanation makes sense. I will go through the paper that you have suggested. Thanks for the information.
mani
On Tue, May 13, 2014 at 11:50 AM, Adrian Roitberg <roitberg.ufl.edu> wrote:
> Hi
>
> Well, first, tell us if you are going to increase or decrease the masses...
>
> This has a long history in the field, and for the most part, it has
> never worked well.
>
> Basically, there are arguments about lowering masses to increase
> diffusion coefficients. This seems ok, but then all vibrational
> frequencies go up and you need a much short time step to make it work.
> Hence, no gain.
>
> You can increase all masses, and then you get to increase your time
> step. That seems reasonable, but simple stats mech will show you that
> what you have messed up with the velocities. If you are still running
> at the same temperature as before, let's say 300K, that correlates
> with the kinetic energy. So, if your kinetic energy is the same as
> with regular masses, your velocities must have been reduced as the
> square root of the mass change. So, things will move slower, and you loose...
>
> There is something called HMR, hydrogen mass repartitioning. We are
> about to submit a paper on this, but essentially consists of taking
> local chemistry (e.g. CH2RR') and increasing the H mass and decreasing
> the H-partner mass in such a way that the 'local' mass has not
> changed.
>
> This means that the total mass of the system is still the same as
> before the repartitioning, and things are much better from a stat mech
> point of view.
>
> Once you do this, you can push your time step to 4 fs and gain a
> factor of 2x in overall speed.
>
> This is an EXPERT mode... We tested it a lot in amber, with different
> observables, and it work great. Some other people within Amber have
> also tried it and it works for them. Again, things break sometimes, so
> the use does not come with a warranty...
>
>
> Some of these ideas were expressed in "Improving efficiency of large
> time-scale MD simulations in hydrogen rich systems. Feenstra, Hess,
> Berendsen. J Comp Chem 1999"
>
>
> On 5/13/14 10:59 AM, Brian Radak wrote:
> > I don't think this kind of "trick" has been used in AMBER very
> > often, if
> at
> > all. Most classical static quantities of interest in the canonical
> ensemble
> > are mass invariant, so your plan is likely sound, but I don't have
> > any ideas as to how sampling will be affected. Perhaps that is why
> > it is not done very often?
> >
> > Regards,
> > Brian
> >
> >
> > On Tue, May 13, 2014 at 10:51 AM, Manikanthan Bhavaraju <
> > manikanthanbhavaraju.gmail.com> wrote:
> >
> >> Dear All,
> >>
> >> I am interested to study the mechanism of protein aggregation using
> >> an explicit solvent simulations. As a model system we have taken a
> >> amyloid protein ~ 100 residues. Initially, I would like to analyze
> >> the affect
> of
> >> scaling of atomic mass of the water molecules vs. standard
> >> simulation techniques. My assumption is that scaling the atomic
> >> mass of the
> solvent
> >> molecules can decrease the viscosity of the system, avoid heating
> >> (performing minimization and production runs), take larger time
> >> steps,
> and
> >> the system to some extent will be able to quickly cross the energy
> barriers
> >> on the potential energy surface.
> >>
> >> On the other hand, I want to conduct a regular explicit solvent
> simulation
> >> (minimization, heating, equilibration, and production) and then
> >> compare both the results. Can you anyone give me some suggestions
> >> or their
> expert
> >> opinion on this kind of analysis?
> >>
> >> Thanks,
> >>
> >> mani
> >> _______________________________________________
> >> AMBER mailing list
> >> AMBER.ambermd.org
> >> http://lists.ambermd.org/mailman/listinfo/amber
> >>
> >
>
> --
> Dr. Adrian E. Roitberg
>
> Colonel Allan R. and Margaret G. Crow Term Professor.
> Quantum Theory Project, Department of Chemistry University of Florida
> roitberg.ufl.edu
> 352-392-6972
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
--
Manikanthan Bhavaraju
_______________________________________________
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 Tue May 13 2014 - 11:00:08 PDT
