simulating vibrational spectra in the solid state, help me!

[Glaucio]

I am using cp2k for simulating vibrational spectra in the solid state, but I'm optimizing the crystal structures. My question is about the convergence criteria RMS gradient. Default is 3x10-4, but I'm used 1x10-6. The problem is the delay in the calculation, I want to avoid the presence of imaginary frequencies.
Someone could indicate their experiences?

[Florian Schiffmann]

Hi,

as always, the answer to your question and providing help is much easier (and likely to happen) if input files are provided.
Cp2k is a bit special for vibrational spectra. The imaginary frequencies are not necessarily connected to a badly converged geometry. Using GTH pseudopotentials the density at the core decays to 0. This is a bit problematic in the xc potentials as it leads to very sharp features. These features can cause trouble in the FFT's (see Quickstep paper, smoothing).
Cutting a long story short, if you use the GTH pseudos (and an insufficient CUTOFF) as they are this problem leads to a slight translational invariance of the system as the result of the FFT depends on the relative position of the atom with respect to the grid. This problem is most obvious for 1st row transition metals and the elements B to Na. Also it depends on the functional as the pseudos have different parameters (BLYP pseudos are worst from my experience). I attach a graph with some PBE test on this behaviour and cutoffs needed to achieve a good translational invariance (<1E-5 Ha). Keep in mind using different functionals the necessary  cutoffs can be different.
At the moment there are two ways out either you have to use a hugh cutoff (600+ depending on your elements) or using NLCC pseudos ( http://arxiv.org/abs/1212.6011 ). The NLCC's are not yet in svn as they are not 100% tested with every feature (energies and derivs are fine) and they only exist for PBE and a limited number of elements (B - Cl). If you are intersted in trying them I can send you the CP2K format. I will add them to svn soon after testing is complete.

regards
Flo

Re

[Daria]

Hi,

I'm also working on vibrational spectra and I have found similar problem.

I'm intersted in trying your NLCC pseudos. Can I ask you to send me the CP2K format so I can do some test?

regards

Daria

[Florian Schiffmann]

Hi Daria,

I attached the NLCC pseudos. As said before they are only optimized for PBE and there is no warranty that they work for anything else (e.g. PBEsol, PBE0,...).
Previously I forgot to mention two other option to overcome that problem:
1) Use SMOOTING (splines) in the XC_GRID section. Personally I wouldn't recommend that option. It solves the riplle problem but has a hugh effect on the total energy and converges slower to the PW basis set limit. Furthermore I am not 100% whether the error due to smoothing is systematic or not.
2) Use the USE_FINER_GRID option in XC_GRID. This will be expensive for calculations dominated by FFT anyway. This uses a grid with 4*CUTOFF for the xc part. I am about to commit the option to specify the finer grid by the user but it isn't quite ready. With this option (PBE case) you should get away with a CUTOFF of 400.

In my oppinion the best options are NLCC's or the finer grid. NLCC's have the problem that they are hard to fit and there is no hope that there will be an extended set in the forseeable future. Therefore the FINER GRID provides an option for all calculations and has several advatages to the other options:
1) It really improves accuracy unlike smoothing
2) forces and 2nd derivatives are better compared to an increase of the overall cutoff. This can be understood as the energy ripples are mainly due to the xc bit and will be accurate to the limit of the Finer grid. The potential neverthelees gets mapped back on the total grid and thus the ripples appear on a larger grid spacing than in case of high total cutoffs. If you take now the numerical derivative of the grid 'ripples' the larger spacing significantly reduces the force contribution.

Hope that helps
Flo

[Daria]