Ab initio MD & metadynamics with PLUMED
Mei
Is there anyone tried quantum ab initio Meta-dynamics with PLUMED? I
have some questions about metadynamics with ab initio simulation, hope
I can get some advice here. Thanks in advance!
I've used CPMD to do some metadynamics calculations before, the
problem is I cannot get a convergent energy, or I don't know when step
adding the Gaussian... Then I tried PLUMED with DLPOLY for some
system, it's working very well, and I can understand that the energy
is convergence after a period. Well-tempered looks a good approach to
get the convergent energy.
But in my system, the problem is that, I'm not confident with the pair
potential in DLPOLY (or any other classical MD). Ab initio simulation
based on plane wave and DFT is a safe way. I'm trying to do espresso
with PLUMED, but still cannot get QE working... According to my
current knowledge, I have some questions about PLUMED and the methods
in this code. I appreciate for any useful information about that.
1: Well-tempered is a good method to get convergent energy. But how
long does it take to get the convergent energy? I learnt from PLUMED
+DLPOLY that, it takes about 50ns (with timestep ~0.25ps ) to get
convergent minima energy (about 6 minima points ). But for ab initio
simulation, the maxima time I can calculate is about 30 ps (with
timestep ~ 0.1fs ), that's even more expensive in terms of CPU time.
Is that long enough to get convergent energy? Or is there any trick to
get a good sampling and free energy surface in a short simulation
time?
2: Is there any way to use well-tempered manually in some other
codes? For example, I'm using CPMD now, do you know is that possible
to do that? Thank you very much!
Cheers Mei
-------------------------------------------------------------
Ms Yuan Mei
PhD candidate
School of Earth and Environmental Sciences
The University of Adelaide, AUSTRALIA
Ph : +61 0401071189
EMAIL : yuan...@adelaide.edu.au
Fabio Pietrucci
the total time required to get a converged free-energy surface from
ab-initio MD depends strongly on the system. In my experience a
"simple" chemical reaction (say a bond breaking or bond formation
process), or the diffusion of an atom to nearest neighbor position in
a crystal, may take around 10-20 ps of simulation. This is typical of
a simple case in which you have two well defined initial and final
states and one or two transition states in between (typical barriers
of the order of 1 eV). You speak of 6 minima, which looks like not so
simple, can you give some more details? Could you split the simulation
into several simpler transformations like A -> B, then B -> C, etc.?
Also, which specific free-energy calculation technique is best depends
on the characteristics of your system and it is hard to make general
statements.
Note that we are going to release plumed 1.3 in which cpmd will be supported.
have a nice day
fabio
> --
> You received this message because you are subscribed to the Google Groups "PLUMED users" group.
> To post to this group, send email to plumed...@googlegroups.com.
> To unsubscribe from this group, send email to plumed-users...@googlegroups.com.
> For more options, visit this group at http://groups.google.com/group/plumed-users?hl=en.
>
>
--
____________________
CECAM Lausanne
homepage: http://sites.google.com/site/fabiopietrucci
yuan mei
Thank you very much for your suggestion. I tried to calculate the
free energy of zinc chloride complexes. For example:
Zn++ + Cl- = ZnCl+
Zn++ + 2Cl- = ZnCl2
…
In the plumed metadynamics I used coordination number as the
collective avariable, and I used 2 CVs (Zn-H2O and Zn-Cl), so the
observed sampling goes from 0-6 H2O and 4-0 Cl, actually more than 6
minima… I'm not sure whether I can spilt the simulation if using
coordination number as CVs.
In this case, it costed about 2 ns to sample all the possible
structures and about 9 ns to get convergent energy (adding very small
gaussion).
Does the simulation temperature effect the sampling? The case
mentioned is at 600K, but for the same system, if calculate that at
300K, the sampling is very slow. After 4 ns, still didn't explore all
the possible structures (only CN_Cl=0-2). I have the same problem with
CPMD. At 300K, it only sampled CN_Cl=1-2 in more than 30ps (250000 md
steps, 700 mtd steps).
I used following parameters for temperature and Well Tempered bias:
WELLTEMPERED SIMTEMP 300 BIASFACTOR 10
The well tempered TEMP equals to the simulation TEMP and the
biasfactor is 10. According to the well tempered paper (2008), TEMP
will effect the sampling etc. Normally how to choose these two
parameters?
By the way, what do you mean about "two well defined initial and final
states "? I guess the initial states are defined in the input files,
but how about the final states?
That's very good news that CPMD will be supported. Since there's
metadynamics in CPMD, what's the main difference between PLUMED and
the original CPMD? What I concern is the well tempered sampling.
Anything else?
Thank you again for your help!
Cheers Mei
Fabio Pietrucci
do you really need to explore also states with 0 waters coordinated
to Zinc or with 4 Cl- ions coordinated? I ask because if they are very
unlikely, by putting walls around your CV-space you can restrict your
system to a smaller number of (relevant) states, accelerating the
sampling. By "well-defined initial and final states" I just meant
minima in the free-energy surface. The more you have, the slower the
convergence, of course. Another trivial way to accelerate is to
enlarge the gaussians or depose them more frquently. About the choice
of the CVs, maybe you want to give a look to the paper Geissler
J.Phys.Chem.B 103-3706 (1999) (link:
http://dx.doi.org/10.1021/jp984837g) and to those citing it.
I cannot comment much about well-tempered features, I'm not expert of it.
With respect to the original CPMD implementation of metadynamics,
Plumed is richer and more up-to-date with recent methods and CVs, and
it allows you to setup exactly the same free-energy calculations on
different codes (e.g. same input and output). In my experience, it is
also easier to modify, but this is a personal opinion:)
bye
Fabio
http://dx.doi.org/10.1021/jp984837g
Giovanni Bussi
in general, metadynamics can be used with two different spirits:
Qualitative approach: To see that "something happens", and to have a
rough physical guess of the preferred path for a reaction.
Quantitative approach: To really converge free energies.
Now, in my experience, metadynamics can be used quantitatively only
with very long simulations (nanoseconds at least). For complex
systems, even microseconds of simulation could be needed. Thus, it is
very difficult to achieve this goal in ab initio MD (a part maybe for
small systems and very simple transitions). Most ab-initio simulations
with metadynamics are thus aimed at highlighting a reaction path, not
at converging a free energy barrier (please: people with more
experience in ab-initio, correct me if I'm wrong).
Well tempered metadynamics has a sense only for the "quantitative"
application, so probably is not very useful in CPMD, and certainly
should not be a reason to migrate from CPMD di PLUMED.
On the other hand, if you compare PLUMED with CPMD, keep in mind that
many people are now using and developing PLUMED, so it is likely that
you will find more modern methods and collective variables in PLUMED.
This will become more and more true in the future, likely.
Giovanni
On Thu, Mar 3, 2011 at 10:03 AM, Fabio Pietrucci
Davide Branduardi
As I recently did some (and still doing) some calculations using electronic
structure methods (not cpmd) I can complement with something from my recent(so, probably not that deep) experience.
I found very convenient to use walls:
try not to let the system go in CV space where you will never sample enough
(ie. in full solution if you use a coordination number) and maximize the exploration
in interesting places: concentrate where the reaction happens. For all the other cases
probably common molecular mechanics will be enough and there are also smart ways
to include entropic effects and mass law effect into your game (see Arieh Warshel abundant literature...).
As Giovanni is suggesting you have two flavors of metadynamics: explorative and quantitative.
My opinion is that one could first go for the explorative way (there is a nice paper from Iannuzzi et al) and
then try to look to the basins and see wether one can further reduce the manifold (do you have any unuseful cv or
unuseful space) . Make use of the symmetry of your system to further decrease your exploration of the space.
If it happens that 2d dimensions could suffice then you can go for a umbrella sampling
or, I tried recently , the gradient fitting procedure from "single sweep" method (very cheap and nicely working if your surface is smooth enough).
If more than 2d are needed then metadynamics is an option. Once again, unsing walls to reduce the space to be explore will help you out a lot!
Remember that you also have multiple walkers and you can exploit this feature as well (see the paper from Raiteri et. al)
Good luck.
D