[Wannier] convergence problem in construction of WF

[Jonathan Yates]

On 2 Jun 2013, at 11:51, Lin Xie wrote:

> Dear all
>    I'm using Fleur+Wannier90 and want to get the WF basis for the valence band of MoS2. The SCF is converged with a 7x7x2 M-P mesh and I have no problem in getting the maximally localized Wannier functions with a 5x5x2 mesh. Moreover, the Hamiltonian matrix is real for the 5x5x2 mesh, indicating that the WFs are optimized. However, when I try to get the MLWFs with a denser mesh, e.g. a 7x7x2 mesh, the spread is about 10% larger than that with the 5x5x2 mesh. Also, there is imaginary part in the Hamiltonian matrix with the 7x7x2 mesh. Can anyone tell how to overcome this problem? 

Lin Xie,

 Let me make a couple of general comments that might help.

First, note that our representation of the spread operator converges slowly with the density of the k-point grid. So you may well find that your Hamiltonian in the MLWF basis is well converged at 5x5x2 (ie the wannier interpolated bands are a good match to those directly from Fleur). The spread of the MLWF is not a good parameter to check k-point convergence with - unless you need to use the spread. In that case you may need to do something more advanced (see sections I.C.2 and II.F.2 of Rev Mod Phys 84, 1419 )

Second, as you increase the density of the k-point mesh the minimiser has to work harder to localise the MLWF. This might explain why there is some imaginary component to the MLWF for the denser mesh. Have you looked at the WF from the 5x5x2 calculation - do they look like the start guess? Can you use this information to improve the start-guess for the 7x7x2 case? Is the imaginary component quite small - if you simply ran for more iterations does it get smaller?

 Jonathan





-- 
Department of Materials, University of Oxford, Parks Road, Oxford, OX1 3PH, UK
tel: +44 (0)1865 612797                http://users.ox.ac.uk/~oums0549/