[gmx-users] Free energy of transfer with accuracy

[chris.neale at utoronto.ca]

Hi David,

I found this comment quite interesting:

"Also, if your C60 is hollow, I can imagine that it might
exclude solvent atoms from the inside, in which case you are going to
have a big convergence issue as solvent begins to interpenetrate with
the C60, which may lead to "noise" problems when this starts to
happen, whatever soft core functional form you use. This just means
you will need to run longer simulations in this regime."

Which makes sense. Do you know a reference for this phenomenon?

Thanks,
Chris.

-- original message --

Eudes,

On Thu, Dec 11, 2008 at 8:27 AM, Eudes Fileti <fileti at ufabc.edu.br> wrote:

[Hide Quoted Text]
Hello David, thanks for the reply!
I am simulating the free energy of transfer of the C60, from ethanol to
benzene.
The softcore parameters, I used the are the same values of Tieleman et.al.
(nature nanotechnology, 3, 363).
At figure, each curve refers to a different solvent; one of them for ethanol
and another for benzene.
For lambda=0 I have the Coulomb and LJ parameters turn on and for lambda=1,
turn off.
I have found a large difference when simulating my system using sc_power=2.
See in this link ( http://cromo.ufabc.edu.br/~fileti/web/dgdl-scpower.jpg )
the comparison between
plots from sc_power=1 and sc_power=2 for solute in benzene.
Near 0 and 1, I believe that problem cancels out at difference, as can seen
at
( http://cromo.ufabc.edu.br/~fileti/web/dgdl-transfer.jpg ) . Then the main
problem, in my view,
is in the intermediate region, where the plot presents a deep (-500 units)
and noise minimum.
Observe that in lambda=0, sc_power=2 reduce largely the derivative, but the
critical region,
one that makes the difference in calculation of the transfer free energy,
remains the same,
ie, the derivative still oscilates a lot and is not smooth.
Are you believe that changing the value of alpha from 0.47 to 0.50 can
resolve this?
OK, there are several issues here.
1) The "noisy" minimum in dV/dlambda
2) The overall shape of dV/dlambda
3) Other potential problems

On (1), the noisiness of the minimum you refer to really means there
are convergence problems for some reason -- probably correlation times
are particularly long there. This may be helped by using the soft core
parameters I mention, but I am not sure. See item #3.

On (2), in general, the overall shape of dV/dlambda seems to be
smoothest and best behaved with the particular soft core parameters I
mention. We did a lot of testing on this (i.e., many values of alpha
at each of several sc-power values).

On (3), there are other thing that can go wrong when using soft core.
See http://www.alchemistry.org/wiki/index.php/Best_Practices for some
discussion, and see rule #2 in particular. If you have any partial
charges on atoms in your C60 this could be a potential source of error
and noise. Also, if your C60 is hollow, I can imagine that it might
exclude solvent atoms from the inside, in which case you are going to
have a big convergence issue as solvent begins to interpenetrate with
the C60, which may lead to "noise" problems when this starts to
happen, whatever soft core functional form you use. This just means
you will need to run longer simulations in this regime.

Anyway, long story short, I suggest two things: (a) Make sure you turn
off partial charges on C60 (without using soft core) before turning
off the LJ using soft core, and (b) use the soft core functional form
I suggest.

David