Phonon for Supercell

[ams_nanolab]

Is it possible to simulate the phonon dispersion/ phonon DOS of a small supercell say 10-12 atoms without the calculator internally repeating the structure in all the directions? I just want the phonon spectra / phonon DOS for the supercell with DFT and at the same time and want to save simulation time and resources.
We have multi-core Intel server running RHEL 6.3, for a long time we have been running ATK on these for years and usually most simulations seem to converge very well. But the problem seems with phonon spectra/ phonon DOS using DFT, where the simulation seems to go on for 2-3 days and still doesn't converge.
Could you give any solution for this? 

[Anders Blom]

The automatic supercell extension is not made for fun Without it, the results are simply not correct. It is possible to make a slightly smaller supercell, as an approximation, but one should be aware of the consequences.

[ams_nanolab]

  That's true but what about computing time involved... simulations running 2-3 days and still not finishing isn't a very good idea.... especially for those who do not have large clusters / supercomputers .... as far as I understand many groups regularly work on phonon spectra using VASP etc. even without having supercomputers ... could the code be made a little less computationally intensive so as to run on high-end standalone servers?

[Anders Blom]

2-3 days sounds a bit much, but it's hard to know anything without details on your system. Si bulk (2 atoms) takes about 5 hours on 6 nodes, indicating a serial speed of about 1 day. One thing to keep in mind, is that you rarely need any k-point sampling, since the supercell takes care of that, so you should go with 1x1x1 usually - that may be part of the explanation for your long runtime. However, the system used for a phonon calculation should also always be relaxed first (optimized positions and cell size) and that of course needs to be done with k-points. ATK is probably one of the most efficient codes available for doing these calculations - you wouldn't even attempt to run this in serial with VASP unless you make only a very small supercell, in which case the accuracy is questionable. And as mentioned, you can lower the accuracy in ATK too - here's how:

Code: python

dmp = DynamicalMatrixParameters(repeats=(5,5,5), atomic_displacement=0.01*Angstrom)
calculator = LCAOCalculator(...all usual stuff...,
    dynamical_matrix_parameters=dmp
    )

As always, it's necessary to have the right tools for the job. You wouldn't build a house out of concrete using a hand-powered drill, and if you are serious about computational physics, investing in a slightly better computer - or a second one, for parallelization - will be a sensible investment. This doesn't mean supercomputers, just like for VASP and other codes you can nicely get by with ATK on a high-power workstation or a small office cluster. But I think in your case, probably you need to revise the k-point sampling first, that may explain everything. Finally, you mention that your calculation "doesn't converge". Do you really mean that, or you just mean that it didn't finish yet? For a phonon calculation you need to do at least 3N+1 separate selfconsistent calculations, and that takes time, but "not converging" means that one of those calculations doesn't reach the tolerance (or doesn't do so unless it takes very many steps), and that's a whole other story.

[Anders Blom]

Another option you have in ATK (but not in VASP) is to use DFTB or a classical potential instead. With the Tersoff or Brenner potential, for instance, you can get phonon spectra of nanowires, graphene, nanotubes, etc in minutes.

[ams_nanolab]

How about I stick to DFT but reduce the number of displacements for phonon calculation? Is that possible? DFT with lesser number of displacements give a more accurate result than ATK-Classical etc. with more displacements? 

[Anders Blom]

You can't reduce the number of displacements, that wouldn't make any sense. The number of displacements is just 3N where N is the number of atoms in the original cell (without repetitions).

[Anders Blom]

I have a suspicion there is a factor 2 too much in the repetitions (so, 7x7x7 should really be 3x3x3). I will investigate - but it would mean you can just use the commands above to set 3x3x3 yourself, and the results should not be affected (but the calculation time will!).