- Contemporary messages sorted: [ by date ] [ by thread ] [ by subject ] [ by author ] [ by messages with attachments ]

From: Jason Swails <jason.swails.gmail.com>

Date: Tue, 10 Mar 2015 14:06:49 -0400

On Tue, Mar 10, 2015 at 1:08 PM, Joseph Baker <bakerj.tcnj.edu> wrote:

*> Hi Ross,
*

*>
*

*> Thanks. We've been testing with increasing mcbarint (starting from the end
*

*> of our Hold simulations of our lipids), and with an mcbarint = 10,000 there
*

*> does not seem to be much noticeable improvement, however with an mcbarint =
*

*> 50,000 our POPE 128 lipid bilayer seems to be keeping a roughly constant
*

*> area per lipid of ~ 55.3 A^2 over the first 50 ns of simulation so far
*

*> (from my understanding, in the limit of very large mcbarint I am coming
*

*> closer to running an NVT simulation, and so this is just telling me that
*

*> the box size at the conclusion of the Hold simulations is pretty good?). Is
*

*> there anything to be wary of when using such a large mcbarint compared to
*

*> the default (mcbarint=100)?
*

*>
*

How long is your simulation? For an infinitely long simulation, the value

of mcbarint will have *no* impact on the resulting ensemble. It will,

however, impact how quickly your unit cell size and shape reaches a point

on the phase space (assuming it's out of equilibrium to start with).

If you are seeing differences between using 50,000 and 100, then the

minimum that you know is that the volume in your mcbarint=50K is not

converged. In fact, with anisotropic scaling convergence is even slower,

since the original move set for anisotropic volume adjustments was...

suboptimal (in terms of how quickly it converged; it is still

asymptotically "correct").

I'm inclined to avoid ever messing with mcbarint, except maybe to make it

smaller if it's taking too long to converge. Basically, I would place no

trust in MC barostat results in which you had to *increase* mcbarint to get

"better" numbers.

I think what's warranted here is an investigation as to why the MC

barostat and Berendsen barostat give different values for some observables,

and determine which one is in error and why.

The possibilities of the problems are:

1. There is a bug in the Berendsen barostat (either in computing the virial

or in the scaling factors).

2. There is a bug in the MC Barostat move evaluation (possible, but it's a

pretty simple barostat).

3. There is a fundamental problem with the underlying theory behind one of

the two barostats (Berendsen has its known problems, while the MC barostat

is more rigorously correct).

As part of the debugging process, you could always test the results against

an alternative implementation of the MC barostat (or Berendsen barostat).

OpenMM has a MC barostat with support for Amber files, which you can use to

compare against the results you get with pmemd.cuda's MC barostat. If you

look at the "OpenMM" command in ParmEd, it should show you how to run these

calculations the same way you would with pmemd.cuda.

Good luck!

Jason

Date: Tue, 10 Mar 2015 14:06:49 -0400

On Tue, Mar 10, 2015 at 1:08 PM, Joseph Baker <bakerj.tcnj.edu> wrote:

How long is your simulation? For an infinitely long simulation, the value

of mcbarint will have *no* impact on the resulting ensemble. It will,

however, impact how quickly your unit cell size and shape reaches a point

on the phase space (assuming it's out of equilibrium to start with).

If you are seeing differences between using 50,000 and 100, then the

minimum that you know is that the volume in your mcbarint=50K is not

converged. In fact, with anisotropic scaling convergence is even slower,

since the original move set for anisotropic volume adjustments was...

suboptimal (in terms of how quickly it converged; it is still

asymptotically "correct").

I'm inclined to avoid ever messing with mcbarint, except maybe to make it

smaller if it's taking too long to converge. Basically, I would place no

trust in MC barostat results in which you had to *increase* mcbarint to get

"better" numbers.

I think what's warranted here is an investigation as to why the MC

barostat and Berendsen barostat give different values for some observables,

and determine which one is in error and why.

The possibilities of the problems are:

1. There is a bug in the Berendsen barostat (either in computing the virial

or in the scaling factors).

2. There is a bug in the MC Barostat move evaluation (possible, but it's a

pretty simple barostat).

3. There is a fundamental problem with the underlying theory behind one of

the two barostats (Berendsen has its known problems, while the MC barostat

is more rigorously correct).

As part of the debugging process, you could always test the results against

an alternative implementation of the MC barostat (or Berendsen barostat).

OpenMM has a MC barostat with support for Amber files, which you can use to

compare against the results you get with pmemd.cuda's MC barostat. If you

look at the "OpenMM" command in ParmEd, it should show you how to run these

calculations the same way you would with pmemd.cuda.

Good luck!

Jason

-- Jason M. Swails BioMaPS, Rutgers University Postdoctoral Researcher _______________________________________________ AMBER mailing list AMBER.ambermd.org http://lists.ambermd.org/mailman/listinfo/amberReceived on Tue Mar 10 2015 - 11:30:02 PDT

Custom Search