Suspicious behaviour of the local potential
It would seem VASP does not compute the exchange and correlation energies for negative densities. This is particularly apparent when large vacuum regions are present in the unit cell (e.g. slabs or isolated molecules/atoms). The densities in the vacuum can be very small and still have a very pronounced effect on the local potential (order of e-1 eV has been observed).
- Be pragmatic: only incorporate the coulomb energy in the local potential (LVHAR = .TRUE.). This also circumvents the problem that the density tails of very slowly. The LVHAR tag has only been available since VASP.5.2.12. Previous versions did not incorporate the exchange-correlation energy by default.
- If, for some reason, the exchange and correlation contributions must be included, the only solution is to impose a very strict energy convergence (EDIFF = e-8 or e-9). Increasing computational accuracy (KPOINTS or ENCUT), changing smearing schemes or increasing the vacuum size do not seem to improve the situation.