[CPMD-list] Metadynamics

Axel Kohlmeyer akohlmey at cmm.chem.upenn.edu
Mon Feb 18 02:34:27 CET 2008


On Sun, 17 Feb 2008, Nora Mohamed wrote:

NM>                 Dr. Kohlmeyer,

nora,

NM>    Thank you very much for your very quick and extremely clarifying 
NM> response, I did not expect you to reply to my questions that fast. I 

well, you basically have the choice between fast or not at all.

NM> managed to get the two books you recommended for me and started 
NM> reading the first one besides the one I am already reading 
NM> "Essentials of Computational Chemistry". I also noticed that neither 
NM> of the books mention metadynamics, is there any textbook you know of 
NM> that explains the concept and algorithm of metadynamics, the way the 
NM> algorithm of Monte Carlo and MD were explained in the first 
NM> textbook?

none that i know of. MD is around for over 40 years, car-parrinello
for over 25 years. i guess it takes a while until method get 
transformed into textbooks. for the time beingm, you'll have to live
with the available papers and tutorials.

i won't worry too much, since it takes a while to master 
regular cp dynamics well enough.

NM> Also, your answered my second question as follows:
NM>  
[...]
NM> 

NM> I don't think I understand this part fully. In your tutorial you 
NM> wrote an input file to calculate the electronic structure of an 
NM> isolated hydrogen molecule and set the symmetry to one. if it is 
NM> isolated hydrogen molecule, then what does it mean that we set the 
NM> symmetry to simple cubic? also, you said that we do single molecule 
NM> calculations if we were interested in gas phase properties. If I am 
NM> interested in gas phase properties, shouldn't I set the symmetry to 
NM> zero and it would be meaningless to set the symmetry value to any 
NM> other value except zero.  Am I getting it right?

in CPMD you _never_ do a real isolated system calculation. the
difference between symmetry 0 and symmetry 1 (actually 8 for which
ibrav=1 is a special case) is that with symmetry 0 you run an 
additional step to "electrostatically decouple" periodic images
by solving the poisson equation. for a (non-polar) hydrogen molecule 
this has essentially no effect other than requiring more memory,
more time and imposing additional constraints on the unit cell size.
for non-polar molecules it is freuqently better to do a periodic 
calculation, since the long range interactions are negligible (as
you already have heard, dispersion is very underrepresented in
current DFT implementations.

hope that helps, 
   axel.

NM> 
NM> Thank you very much Dr. Kohlmeyer. 
NM> Best Regards,
NM> Nora M. Hassan
NM> 
NM>  
NM>  
NM> 
NM> 
NM> 
NM> Axel Kohlmeyer <akohlmey at cmm.chem.upenn.edu> wrote: On Sun, 10 Feb 2008, Nora Mohamed wrote:
NM> 
NM> NM> Hi CPMD users.
NM> 
NM> hi nora,
NM> 
NM> NM> This is my first encounter with molecular dynamics and so I have few 
NM> NM> elementary question.
NM> 
NM> this is fine, but it would be _much_ easier to answer your 
NM> question if you gave us a little bit of information about
NM> what your background is and what you want to do.
NM> 
NM> one more thing that is important to note. CPMD stands for
NM> a software package and a method.the CPMD code implements
NM> this CP-MD method, but a lot of other types of calculations
NM> on top of that. the common denominator is that it you do
NM> quantum chemical calculations from first priciples with
NM> density functional theory and use plane waves as basis set
NM> and pseudopotentials.
NM> 
NM> since you have elementary questions, i think it is also in
NM> place to recommend a couple of textbooks which should cover
NM> a lot of what you want to know (and then some). 
NM> 
NM> on molecular dynamics (and monte carlo): 
NM> d. frenkel, b. smit, understanding molecular simulation
NM> 
NM> on electronic structure:
NM> j. kohanoff, electronic structure calculations for solids and molecules
NM> 
NM> NM> 1. I am confused when to do bulk systems simulations and one 
NM> NM>    molecule simulations. in the tutorials I read, I found examples 
NM> NM>    that simulate only one water molecule and others that simulate 
NM> NM>    many water molecules in the simulation cell. when I use either?
NM> 
NM> please note, that many examples in the tutorials and on the web are
NM> designed to be fast and simple and help understanding the method, but
NM> do not necessarily show best practices or overly useful applications.
NM> with that in mind, the answer to your question is very simple:
NM> you do single molecule calculations if you are interested in gas
NM> phase properties, you simulate clusters when you want to learn 
NM> about clusters and bulks systems if you are interested in the bulk.
NM> each of those systems have their own set of challenges to get them 
NM> right and produce meaningful results.
NM> 
NM> NM> 2. also, what does it mean when I have only one molecule in the 
NM> NM>    simulation cell and have the SYMMETRY keyword is set to fcc or 
NM> NM>    bcc. As I said this is my first encounter with simulation and so 
NM> NM>    have a difficulty in imagining the supercell that is simulated.
NM> 
NM> as far as the CPMD program (or any electronic structure code) is 
NM> concerned, there are no molecules. just electrons and the potential
NM> they are in. molecules is something that is defined by analysis
NM> of the result of the electronis structure calculation (or sometimes
NM> common sense). however, if inputs specify fcc or bcc symmetry they
NM> usually are meant for cystals. please check the dimensions of the 
NM> unit cell. it could just be that the two atoms in that input define
NM> the crystal structure (cf. diamond, silicon).
NM> 
NM> NM> 3. If I wanted to enter the primitive cell vectors instead of the 
NM> NM>    SYMMETRY keyword, do I still have to enter the atoms coordinates? 
NM> NM>    wouldn't that be redundant since primitive vectors define the 
NM> NM>    lattice completely?
NM> 
NM> the symmetry keyword only defines the bravais lattice, _not_
NM> the positions of the atoms in the primitive cell. this only
NM> seems redundant, if you have a simple lattice with just one
NM> atom in the primitive cell.
NM> 
NM> NM> 4. I want to make sure of an information I read, is it right that in 
NM> NM>    the &SYSTEM section, I define the symmetry and cell section 
NM> NM>    according to the conventional cell and in the ATOMS section I put 
NM> NM>    the atoms coordinates according to the primitive cell?
NM> 
NM> yes. there is a simple way to confirm this "experimentally".
NM> do a wavefunction optimization calculation with a small cutoff
NM> and MAXSTEP set to 1 and add the RHOOUT keyword. the resulting
NM> output (DENSITY) has to be converted to a .cube file and can
NM> then be visualized (e.g. VMD, xcrysden) and when looking at
NM> the electron density you'll see the initial guess and how CPMD
NM> understands the cell that you entered.
NM> 
NM> NM> 5. in the examples I read, I got confused by the way the atoms 
NM> NM>    coordinates were defined. in some examples I find negative 
NM> NM>    numbers like in the example below from Dr. Axel's website and 
NM> 
NM> CPMD _always_ wraps atom coordinates back into the principal
NM> unit cell for the electronic structure calculation. so if they
NM> are "outside" they are just periodic replica of the corresponding
NM> atom "inside". please check in the MD literature about periodic
NM> boundary conditions.
NM> 
NM> NM>    from another tutorial I find the coordinates are defined from 0 
NM> NM>    to 1 or to the value of "scale". Are there any rules to use 
NM> NM>    either or I have complete freedom in my choice.
NM> 
NM> you are free to use whatever you want. output will (almost) always
NM> be in cartesian coordiantes though.
NM> 
NM> [...]
NM> 
NM> NM> 6. In the example below ( from the Institute of Physical Chemistry 
NM> NM>    (Group of Prof. Dr. Jürg Hutter), I found the information given 
NM> NM>    in the SYSTEM section inconsistent with that given in the ATOMS 
NM> NM>    sections which shows that there must be some information that I 
NM> NM>    am missing. In the SYSTEM section I have the symmetry set to fcc 
NM> NM>    and the cell size to 20.53 au and scale to 2 which is consistent 
NM> NM>    with the information given in kittel (5th edition). according to 
NM> NM>    Kittel p25 the silicon crystallizes in (diamond structure) an fcc 
NM> NM>    lattice with two silicon atoms for the basis with conventional 
NM> NM>    cell lattice constant a= 5.43 A = 10.263102 au half the value 
NM> NM>    given under the CELL keyword. however, according to kittel p16 
NM> NM>    and p17 the primitive cell as shown takes up only one fourth of 
NM> NM>    the volume of the conventional cell, and in the ATOMS section the 
NM> NM>    primitive cell takes up the whole volume of the conventional 
NM> NM>    cell. there must be something I am missing but I don't know what 
NM> NM>    it is. Any help?
NM> 
NM> have you considered that the input is representing a supercell
NM> built from 8 primitive cells? this input is the primitive cell 
NM> doubled into each direction. there is a good reason to do this:
NM> if you compare the resulting total energy per atom with a 
NM> similar calculation that has only the primitive cell and two 
NM> atoms and second calculation with a 2x2x2 k-point grid (and
NM> larger k-point grids). you can see that enlarging the cell is 
NM> equivalent to using k-points and that doing a gamma point 
NM> calculation of the minimal primitive cell is not sufficient.
NM> for more details, please see the kohanoff textbook.
NM> 
NM> [...]
NM> 
NM> NM> 7. How to choose my pseudopotential? I know I have to test each 
NM> 
NM> pseudopotentials are a bit of a black art (same as force fields
NM> in classical molecular dynamics). you can trust the person that
NM> provides you with a potential (dangerous), you can check in the
NM> literature for publications that list pseudopotential parameters
NM> (which is rarely done these days), you can run your own tests.
NM> 
NM> NM>    pseudopotential for each atom in my simulation. for example if I 
NM> NM>    am simulating ch4, then I have to test the pseudopotential for 
NM> NM>    hydrogen alone and the pseudopotential for carbon alone. however, 
NM> 
NM> you can tests against whatever is relevant for your project.
NM> you do not need to test each single pseudopotential seperately,
NM> but this way it is sometimes easier to determine its adequacy.
NM> if a pseudopotential cannot reproduce a lattice constant of the
NM> element (eg. diamond or graphite) in the case of carbon) within
NM> the limits of the functional used, it has to be discarded. but
NM> that does not guarantee transferability, i.e. applicability 
NM> where its oxidation state is different. so you'll have to test 
NM> that as well. there are some tests, that can also be run on
NM> the pseudopotential directly with an atomic code, e.g. check
NM> the eigenvalue spectrum of the psp against all-electron calculations
NM> for different occupationa and oxidation patterns. finally one
NM> has to check for "ghosts" (i.e. unphysical bound states) and
NM> the required basis set size (= cutoff).
NM> 
NM> NM>    should I test the pseudopotential with bulk carbon or with one 
NM> NM>    carbon molecule. I also want to know what the pseudopotential 
NM> NM>    accuracy depends on. does it depend on my system (i.e my 
NM> NM>    computer, operating system and the set up in general? I got that 
NM> NM>    impression since I read everywhere that I must test the 
NM> NM>    pseudopotential first)
NM> 
NM> some of the pseudopotential libraries available come with a list
NM> of references that document the use of those pseudopotentials
NM> (e.g. mauro boero's collection). if you trust (see above) the 
NM> publications to be correct and applicable for your purposes, 
NM> then you can get away with some minimal testing. please note,
NM> however, that what is adequate accuracy has changed from the
NM> time CPMD was conceived to now. some calculations that were a 
NM> big achievment 20 years ago, would now ne considered sloppy.
NM> the advances in high-performance computing have made it 
NM> possible that calculations which once took months on 
NM> supercomputers now complete in hours on a laptop.
NM> 
NM> NM> 8.  I got warned that cf4 is a molecular crystal, bonded by Van der 
NM> NM>    Waals forces between the molecules and so not well given by 
NM> NM>    present Density Functional Theory calculations.Does this mean 
NM> NM>    that I can't use cpmd to simulate cf4 or ch4 since my results 
NM> NM>    would be wrong or inaccurate at best.
NM> 
NM> the warnings that plain DFT with current functionals does not 
NM> represent vdW interactions well is correct. there are several
NM> ways how people try to cope with this. improving functionals,
NM> using hybrid functionals with hartree-fock exchanged mixed in
NM> (very costly with plane waves), using empirical corrections.
NM> whether CPMD is the right tool for your needs depends a lot
NM> on what you want to do and cannot be answered a priori.
NM> 
NM> NM> 9. I also read that when I do calculations under high pressure, I 
NM> NM>    must also watch that I do not have too much overlap between the 
NM> NM>    pseudising spheres of neighboring atoms and that 10 percent maybe 
NM> NM>    OK.Does anybody know how do I know that my pseudising spheres 
NM> NM>    overlapped or even how to calculate that ten percent? I need more 
NM> NM>    details on that matter please.
NM> 
NM> if you want to do high pressure systems, you will most likely need
NM> custom pseudopotentials with lower pseudization cutoff radii. 
NM> so it is probably advisable that you read up on pseudopotential 
NM> generation first. the fhi98pp code (see CPMD manual) comes with
NM> a nice tutorial on that, so you can gain practice.
NM> the cutoff radii for many CPMD pseudopotentials are listed 
NM> in their respective &INFO section.
NM> 
NM> NM> 10. I am really worried about the results I get. I know I have to 
NM> NM>     optimize my wavefunction at first and then do whatever molecular 
NM> NM>     dynamics I need. the problem is how to know that my results are 
NM> NM>     right; I mean is there any convention or procedure I should 
NM> NM>     follow to make sure that the results I got, either from 
NM> NM>     wavefunction optimization or molecular dynamics, are accurate? 
NM> NM>     what values I should keep an eye on to validate my results or 
NM> NM>     know that there is something wrong?
NM> 
NM> before doing CP dynamics, you should practice and understand 
NM> classical molecular dynamics. the CP hamiltonian puts an extra
NM> complication on that, namely the fictitious dynamics of the 
NM> electronic system. please read through the two textbooks suggested
NM> above and you most of your uncertainties will be gone, after 
NM> you had a little practice with actual examples where the results
NM> are known.
NM> 
NM> NM> 11. concerning Metadynamics, I don't know how to set the necessary 
NM> NM>     parameters. Mainly I want to study how a material undergoes 
NM> NM>     phase change under high pressure, I studied the example about 
NM> NM>     silicon and read the section about metadynamics in the cpmd 
NM> NM>     manual, but I think I should read more about this part. does 
NM> NM>     anybody know any paper or material to read concerning this part 
NM> NM>     or its theoretical background?
NM> 
NM> it is completely pointless to try using metadynamics in CPMD without
NM> being confident in doing regular CPMD calculations. meta-dynamics
NM> is essentially a completely independent technique from CPMD (it
NM> has been implemented into various other codes and used with other
NM> MD methods), and thus there is no need to complicate your life
NM> by looking into this right now. so please ask about this again, 
NM> when you have caught up on the previous 10 questions.
NM> this will be a _lot_ of work. CPMD is a very powerful tool, but 
NM> that also means that one needs to learn it with care (which you
NM> obvious have, as is evident from your questions) to not shoot
NM> yourself in the foot. 
NM> 
NM> with best regards,
NM>    axel.
NM> 
NM> 
NM> NM> 
NM> NM> Thank you.
NM> NM>        
NM> NM> ---------------------------------
NM> 
NM> 

-- 
=======================================================================
Axel Kohlmeyer   akohlmey at cmm.chem.upenn.edu   http://www.cmm.upenn.edu
   Center for Molecular Modeling   --   University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.


More information about the CPMD-list mailing list