"Overshooting" of the free energy

[salva]

Dear users and developers,
As mentioned in previous posts, we have been working with a magnetic system. We believe we are close to exhaust the flags related with the convergence of the Free energy: Flags ENCUT, ENAUG, AMIX,
BMIX, AMIX_MAG and BMIX_MAG have had different values. In all, we have performed about 20 calculations with the system, varying those flags. VOSKOWN has been set to 1. The system is a single molecule magnet. The calculation runs with a single k point. It has 88 atoms, and the species involved are Mn, O, C, and H. We have used hard PAW pseudos, and with semicore states when available.

A characteristic "overshooting" of the Free energy has occurred consistently in our calculations: The Free energy reaches a minimum and then it bounces back to slightly higher energies. We fixed the total magnetic moment, as well as the individual atomic magnetic moments. Those individual magnetic moments are consistent with published work. We are currently running this calculation with a slightly higher magnetic moment, as suggested in the manual. One incar file and its associated standard output are attached. We ask for advice on what could be the source of this problem, and on possible ways to solve it.
Best regards,
-Salvador.

===========================================================
SYSTEM1
ISPIN=2
ENCUT =600
ENAUG =1100
NGXF= 400; NGYF=400; NGZF=350

VOSKOWN=1

AMIX=0.1
BMIX=0.005
AMIX_MAG=0.4
BMIX_MAG=0.005

SYSTEM = MnIICN2DMF4
NWRITE = 2
PREC=High
ALGO=Normal
GGA=91
!LSORBIT = .TRUE.
!SAXIS = 0 0 1 ! direction of the magnetic field
NBANDS = 290
!NUPDOWN = 1
!IMAGES=1
LREAL=Auto
ISYM=0
EDIFF=1.0E-5
SIGMA=0.01
ISMEAR=0
ISTART=1
NELM=200
!200
NUPDOWN =20
MAGMOM= -3 -3 -3 -3 4 4 4 4 4 4 4 4 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

===========================================================
STANDARD OUTPUT:
===========================================================
> mpirun -np 12 ~/VASP/vasp.4.6/vasp
running on 12 nodes
distr: one band on 1 nodes, 12 groups
vasp.4.6.28 25Jul05 complex
POSCAR found : 4 types and 88 ions
LDA part: xc-table for Ceperly-Alder, Vosko type interpolation para-ferro
POSCAR, INCAR and KPOINTS ok, starting setup
FFT: planning ... 1
reading WAVECAR
WARNING: random wavefunctions but no delay for mixing, default for NELMDL
entering main loop
N E dE d eps ncg rms rms(c)
DAV: 1 0.585855752352E+04 0.58586E+04 -0.18253E+05 1452 0.124E+03
DAV: 2 0.248790062701E+04 -0.33707E+04 -0.32103E+04 1416 0.317E+02
DAV: 3 0.318062261544E+03 -0.21698E+04 -0.21085E+04 2088 0.189E+02
DAV: 4 -0.553683257739E+03 -0.87175E+03 -0.83638E+03 2028 0.138E+02
DAV: 5 -0.710246893696E+03 -0.15656E+03 -0.15337E+03 2004 0.706E+01 0.640E+01
DAV: 6 -0.663696779406E+03 0.46550E+02 -0.13432E+02 1980 0.357E+01 0.456E+01
DAV: 7 -0.632817593663E+03 0.30879E+02 -0.19732E+02 2052 0.583E+01 0.405E+01
DAV: 8 -0.629950532579E+03 0.28671E+01 -0.50020E+01 2028 0.305E+01 0.194E+01
DAV: 9 -0.630763040093E+03 -0.81251E+00 -0.10222E+01 2016 0.242E+01 0.191E+01
DAV: 10 -0.627238631606E+03 0.35244E+01 -0.15234E+01 2004 0.208E+01 0.118E+01
DAV: 11 -0.626695461812E+03 0.54317E+00 -0.46981E+00 1992 0.956E+00 0.838E+00
DAV: 12 -0.626327185522E+03 0.36828E+00 -0.84503E-01 2040 0.665E+00 0.609E+00
DAV: 13 -0.626198855579E+03 0.12833E+00 -0.73567E-01 1788 0.229E+00 0.380E+00
DAV: 14 -0.626082582361E+03 0.11627E+00 -0.58718E-01 2064 0.378E+00 0.146E+00
DAV: 15 -0.626073559803E+03 0.90226E-02 -0.10165E-01 1896 0.137E+00 0.792E-01
DAV: 16 -0.626072811445E+03 0.74836E-03 -0.50786E-02 1956 0.727E-01 0.880E-01
DAV: 17 -0.626067869070E+03 0.49424E-02 -0.85066E-03 1932 0.606E-01 0.431E-01
DAV: 18 -0.626067617097E+03 0.25197E-03 -0.11292E-02 2112 0.544E-01 0.255E-01
DAV: 19 -0.626067146134E+03 0.47096E-03 -0.32702E-03 1764 0.259E-01 0.154E-01
DAV: 20 -0.626067138509E+03 0.76257E-05 -0.43511E-04 1968 0.105E-01 0.141E-01
DAV: 21 -0.626066978427E+03 0.16008E-03 -0.12437E-03 2004 0.200E-01 0.480E-02
DAV: 22 -0.626067013199E+03 -0.34772E-04 -0.21304E-04 1620 0.442E-02 0.427E-02
DAV: 23 -0.626067007167E+03 0.60327E-05 -0.40996E-05 1032 0.262E-02
1 F= -.62606701E+03 E0= -.62606701E+03 d E =-.878861E-35 mag= 20.0000
writing wavefunctions

[salva]

This one further complements my previous post. Looking at the output files, we determined that if the VOSKOWN flag is set to zero, convergence is reached in a more smooth way. So we are about to try with PBE paw pseudos (we were using PAW PW91), but we have not been able to locate them in the servers. Could someone please indicate their location to us? (i.e., server, username to be used and, pssibly, path)

Thank you in advance!
Salvador.

[admin]

1) all paw pseudopotentials are located of the paw account of our ftp server. please ask the person who is responsible for VASP at your site to download the pseudopotentials you need for you.
2) VOSKOWN=1 only needs to be set for PW91 PPs, you are right, but please use the potpaw_GGA pseudopotentials for that
case, this should be better that LDA-PPs + GGA=91
3) concerning the relaxation behaviour: please note that the system starts to oscillate after the first mixing step (step 6).
It may help to increase this 'preconditioning' phase (eg, to NELMDL=-8)

[tjf]

I've also seen this type of behaviour. This raises the question: With the iterative Davidson algorithm, is there a variational principle in action before convergence?