hybrid monte carlo and cp2k

Matt McGrath obfis... at gmail.com
Wed Jun 29 08:57:32 CEST 2011


Hello Alin.

Ah, I am starting to see now.  I agree with the thoughts on HMC; this
is actually why I put it into CP2K.  Previously, I used pre-biasing
with an approximate potential to make MC more efficient (you can see
any of my papers about first principles Monte Carlo on water and
hydrogen fluoride for the details).  That worked somewhat, but I was
hoping HMC would be better (doesn't seem like it is, according to some
constant pressure simulations we've run, which surprises me...maybe
because the pre-biasing potential is "free" compared to the QS
calculation and the displacements are larger...on a per QS calculation
basis, there doesn't seem to be a significant speedup at the the
moment, although the MD energy calculations can be made less expensive
with various extrapolation techniques).

As for this calculation, it seems quite interesting.  From that paper,
it appears almost like you're doing a reverse HMC; using an MC run to
sample part of an MD run, instead of the other way around.  Since this
is a free energy technique with MD as the driver, I think it would be
better to treat it like metadynamics, and create a section in
&FREE_ENERGY for it.  Ideally, at least.  In principle, because it
seems like an MD run that calls MC which calls MD again, maybe it
would be more useful (and faster to have working) if you have your own
MD code and just treat HMC as a library call?  Have you considered
this at all?  I would hesitate to add another RUN_TYPE option, since
it seems like the basic elements (MD and MC) are already in CP2K as
high level options.

                                                     Cheers, Matt

On Jun 28, 4:49 pm, Alin Marin Elena <alinm... at gmail.com> wrote:
> Hi Matt,
>
> Thank you very much for you answer and for your tips on both HMC and the input
> file.
>
> here is what I have to do with hybrid MC. I have to run simulations "à la"
> metadynamics (the technique I am going to use is Temperature Accelered Monte
> Carlo, TAMC -- G Ciccotti, S Meloni, Physical Chemistry Chemical Physics,  
> 353, 1-15, (2011)) in which I deal with an extended system composed of the
> atoms + a set of collective variables. These two subset are coupled by a
> suitable potential. In TAMC, similarly than in metadynamics, these two subset
> are evolved together but the inertia of the collective variables, or other
> parameters controlling the timescale of the dynamics (e.g. the parameters of
> the thermostat), are selected such that the atoms are faster than the
> collective variables (adiabatic condition). At a variance with metadynamics, I
> use two thermostats at different temperatures for the atoms and the collective
> variables (it is the high temperature on the collective variables that allow
> to overcome the free energy barriers possibly present in the system). It can
> be demonstrated that the dynamical system described above samples the
> "marginal" probability density function that the collective variables assume a
> given value, and from this I can reconstruct the free energy of the system
> along these collective variables. in metadynamics (and Temperature Accelerated
> Molecular Dynamics, driven Adiabatic Free Energy Dynamics, etc.) the above
> sampling is performed running MD on both the atoms and the collective
> variables. However, if the collective variables are not analytical I cannot
> compute the force on the atoms coming from the "coupling potential" associated
> to them. This can be solved replacing MD for the evolution of the atoms with
> MC (in metadynamics and similar techniques we use MD only to sample the
> probability density function of the considered ensemble and so we can safely
> replace it with any sampling technique). My problem is that I need to do this
> by ab initio simulations (there are bonds broken and formed in the process I
> am investigating) and in this case the standard MC is quite inefficient: I
> move one atom and have to perform an SCF cycle (...it will converge faster BUT
> NOT N times faster, where N is the number of atoms in the sample). So, I
> intend to use hybrid MC, which moves all the atom together,  to improve the
> efficiency. This is what I need to do and why I have to combine HMC with
> quickstep.
>
> From my reading of the code I think I will have to add a new run mode to cater
> for my needs... am I right?
>
> regards,
> Alin
>
> On Thu 23 Jun 2011 23:15:31 Matt McGrath wrote:
>
>
>
>
>
>
>
> > Hi Alin.
>
> > On Jun 23, 7:19 pm, Alin Marin Elena <alinm... at gmail.com> wrote:
> > > Hi Matt,
>
> > > Thank you very much for you fast and detailed answer.
> > > I read the HMC part from the paper that you suggested, thank you, it
> > > made some things clearer on what is implemented in cp2k...
>
> > > > Good question.  HMC is something I've just recently implemented.
> > > >  I've
> > > > tested it to make sure it samples correctly, but I've noticed too
> > > > that
> > > > it's not the easiest to use, unless you happen to be the person who
> > > > wrote the MC routines (and even then, new issues pop up sometimes).
>
> > > they are not so bad... I just want to be sure that I do not miss things
> > > that can bite me later.
>
> > > Is there any pilot study you can point me to for the 0.7 acceptance rate
> > > for the HMC? I know that the MC schemes efficiency tends to be
> > > connected with the acceptance rate.
>
> > Unfortunately, I can't find any studies that say that.  It just seems
> > to be what is currently used in the literature.  You're right,
> > acceptance ratio really impacts your sampling.  The conventional
> > wisdom in the field is to aim for 50% for moves involving the
> > Metropolis acceptance ratio, 50% for volume moves (though 30-40 is
> > good for equilibration), and various guidelines for other move types,
> > although even these are not set in stone (I've seen arguments for
> > lower acceptance ratios in translations/rotations, for example, even
> > as low as 25%).  In the end, all you can do is monitor your system and
> > see when it's equilibrated (according to density, energy, or
> > whatever).  On the plus side, a bad acceptance ratio doesn't lead to
> > the wrong answer...it just might make you get there slower.
>
> > > What I want to sample is indeed a canonical ensemble. The MD part will
> > > be DFT, so now you may understand my worry on efficiency.
> > > My system is simple, 23 atoms I do not care about the molecules, so I
> > > created a topology file that puts all of them in one "macro-molecule".
> > > I attached my input files, so any feedback is much appreciated...
>
> > I took a look at your input file, and I can't see any obvious
> > problems.  I am, however, not the ideal person to comment on the DFT
> > parameters (I generally rely on other smart people for that :) ).  I
> > have had it recommended to me to use FULL_ALL as the preconditioner
> > and PS 3 to speed up the MD part, although the first calculation in
> > any sequence is always going to take more time if the HMC move was
> > rejected.
>
> > > Now what I try to do... I try to use a metadynamics run with CV and
> > > forces coming from DFT.
> > > The acceptance hamiltonian is the one coming from the all system
> > > (atoms+CVs) but the trajectories for MD are evolved using only the
> > > physical forces.
>
> > > I was wondering if that is possible to do it in the current state of
> > > cp2k or should I define my own type of dynamics?
>
> > I'm not seeing how that could work at the moment, theoretically or
> > practically.  It would take some time for me to understand
> > metadynamics (coincidentally, I am just starting a metadynamics
> > project), so I'm guessing that it's nothing that will happen soon
> > (bigger priority is to clean up the MC routines).  Is there a big
> > reason why you want to do this over regular MD with meta?  I'm not
> > seeing the advantage, honestly, and I don't see how you would include
> > metadynamics when your MD run is just a regular NVE (if I'm
> > understanding you correctly).
>
> >                            Cheers, Matt
>
> > > regards,
> > > Alin
> > > --
> > >  Without Questions there are no Answers!
> > >  _____________________________________________________________________
> > >   Alin Marin ELENA
> > >   Advanced Molecular Simulation Research Laboratory
> > >   School of Physics, University College Dublin
> > >   ----
> > >   Ardionsamblú Móilíneach Saotharlann Taighde
> > >   Scoil na Fisice, An Coláiste Ollscoile, Baile Átha Cliath
>
> > > -----------------------------------------------------------------------
> > > ---- --------http://alin.elenaworld.net
> > >   ______________________________________________________________________
>
> > >  hmc.tar.bz2
> > > 319KViewDownload
>
> > >  signature.asc
> > > < 1KViewDownload
>
> --
>  Without Questions there are no Answers!
>  _____________________________________________________________________
>   Alin Marin ELENA
>   Advanced Molecular Simulation Research Laboratory
>   School of Physics, University College Dublin
>   ----
>   Ardionsamblú Móilíneach Saotharlann Taighde
>   Scoil na Fisice, An Coláiste Ollscoile, Baile Átha Cliath
>
>   --------------------------------------------------------------------------- --------
>  http://alin.elenaworld.net
>   ______________________________________________________________________
>
>  signature.asc
> < 1KViewDownload


More information about the CP2K-user mailing list