Convergence of MGGAs
I have found that it is very challenging to converge even simple systems (e.g. the H2 molecule) with MGGAs in GPAW. I thought I'd open this issue to track these problems.
Ways of improving the convergence of MGGAs have been discussed in a few papers.
- http://dx.doi.org/10.1063/1.2800011 where it is suggested that a singularity in the derivative of the factor D_\sigma (eq 1a in the paper, a factor of the equal-spin correlation) can be removed by rewriting D_\sigma as in eq 6 of the paper
- Another way of improving convergence is discussed in the VASP MGGA implementation paper, http://dx.doi.org/10.1103/PhysRevB.84.035117 . The change that is suggested is a regularization of the z=tau_W/tau parameter (eq 19 in the paper). Update: this regularization is actually already implemented, but not turned on by default (https://gitlab.com/gpaw/gpaw/blob/master/gpaw/xc/mgga.py#L109)
- Johnson, 2009 (10.1063/1.3177061) pointed to a very strong mesh dependence for MGGAs. If the mesh is not fine enough, oscillations appear in the potential. Do you need an extra fine mesh (manually set h lower) for MGGAs to converge?
Edited by Jens Jørgen Mortensen