Large error in changing from Low to Accurate precision

[hess8]

As best I can tell this is a bug (Vasp 5.3):

I'm running supercells of atoms on a graphene lattice, at low and high precision. That works fine.

When I run the pure graphene alone as a reference check, in supercell form (up to 20 of them) they all work fine, except one: graphene 3x1.

When I run this particular structure (and on none of the others), I get a huge error in going from low to high precision.

Energy should be about -9.2 eV per atom in graphene.

Low precision: -9.2
Accurate precision: -6.7 eV. Huge error

I've tried changing every parameter I can think of to get the accurate precision run to work. If someone could run these same files, and switch between low and accurate precision to confirm the bug, I'd appreciate it. Better yet if you can fix it!

Thanks!

---------------------------------------------------------------------------

POSCAR:
graphene 3x1 supercell
1.0
4.26257700 0.00000000 0.00000000
6.39386550 3.69150000 0.00000000
0.00000000 0.00000000 15.00000000
6
Cartesian
1.42085900 0.00000000 0.00000000
2.84171800 0.00000000 0.00000000
3.55214750 1.23050000 0.00000000
4.97300650 1.23050000 0.00000000
5.68343600 2.46100000 0.00000000
7.10429500 2.46100000 0.00000000

INCAR:
IBRION=2 # #(changing to 1 does not help)
ISIF=4 #(changing to 2,3 does not help)
NSW=400
Algo=Normal #(changing to veryfast does not help)
PREC=Low or Accurate
EDIFF=0.0003
ISMEAR=0 #(changing this to 1 does not help)
ISPIN=1 #(changing this does not help)
SIGMA=0.1 #(changing this to other values does not help)

KPOINTS:
Automatic mesh
0
Gamma
6 6 1 #(changing this to higher numbers does not help)
0 0 0

POTCAR: PAW_PBE C 08Apr2002

[admin]

You are in a numerical trap. Have a look at your atomic coordinates
in OUTCAR (direct coordinates). All atoms are in highly symmetric
positions. This is a problem for relaxation procedure. Move at least
one atom slightly away from the symmetric position and it will work.
Note that the precision compatibly changes in series:
normal-single-accurate and low-medium-high

[hess8]

Thank you so much!

[hess8]

Hopefully a future version of VASP could have its own check and solution of this issue.