SCF and NBANDS onvergence Problems in Atomic Systems

[juhL]

Hi everyone!

Lately I ran into trouble during calculations of single atoms for some reference calculations. The problem is, that (magnetic) single atoms in vacuum seem to have extremely bad/or now SCF convergence, i.e., single-points don't converge within a large number of steps or, if they do, for different calculations, using different settings for NBANDS, one optains different energies.

Considering, e.g., a single Ru atom, the setup is always the following:

INCAR
---------------------------------------------------------------------
PREC = Accurate
POTIM = 0.5
GGA = PE
VOSKOWN = 1

Electronic Relaxation

EDIFF = 0.1E-06
IALGO = 38
ISMEAR = 1

Write Flags

LWAVE = .FALSE.
LCHARG = .FALSE.

NPAR = 1

ENCUT = 276
NSW = 0
NELM = 300
IBRION = -1
ISPIN = 2
LORBIT = 11
SIGMA = 0.002
ISYM = 0
MAGMOM = 4
NBANDS = XXX (different values tested)


POSCAR
---------------------------------------------------------------------
single_atom
1.0
19.0000000000000000 0.0000000000000000 0.0000000000000000
0.0000000000000000 20.0000000000000000 0.0000000000000000
0.0000000000000000 0.0000000000000000 21.0000000000000000
Ru
1
Selective Dynamics
Direct
0.5000000000000000 0.5000000000000000 0.5000000000000000 T T T

KPOINTS:
---------------------------------------------------------------------
Only the Gamma-point

POTCAR: Standard PAW
PAW_PBE Ru 04Feb2005
8.00000000000000
...
---------------------------------------------------------------------

Using this setup I obtain (VASP 5.3.3)

NBANDS | E0
8 | no convergence
16 | -2.4722524
30 | no convergence
32 | -2.4723109
34 | no convergence
38 | -2.4722766
42 | -2.4722737
46 | no convergence
64 | no convergence


In the cases of no convergence, the SCF energy oscillates, with no sign of convergence within 300 steps. The "usual" tricks like adjusting the MIXING Parameters, increasing NELMDL, or switching IALGO don't yield improved convergence.

My problem is no twofold:
1) general convergence would be nice beyond a minimal threshold for NBANDS
2) after testing a few different atoms, there seems to be no clear setting for a proper NBAND value (dependent on NELECT and NIONS)

Am I overlooking something or what is the reason for these convergence issues?

[admin]

did you do this on a single CPU or parallel?

[juhL]

The data given was calculated parallel. I know this sounds inconvenient, but I don't see why this should cause these troubles, do you?

In fact, I rechecked this for Ru on a single node, ending up with the same convergence issues.

[samith]

In the cases of no convergence, the SCF energy oscillates, with no sign of convergence within 300 steps. The "usual" tricks like adjusting the MIXING Parameters, increasing NELMDL, or switching IALGO don't yield improved convergence.

[juhL]

?