F L P R S   website, discussed version: 5.4.4 and 6.3.1

Credits: Sergey Matveev and Victor Vysotskiy (installation, scripts), Dmitry Aksenov (guidelines, scripts)

Installation

• Follow VASP instructions. Here are our makefiles.
• To generate input files locally unzip pseudopotentials to a desired program folder, I use C:/Sci/vasp. Note that individual files may be compressed – decompress them.
• Job submission scripts: Raven cluster, Magnus cluster
• PBS submission command: qsub vasp/_pbs -N test1
• SLURM submission command: sbatch -J test1 -c 16 -t 24:00:00 vasp/_pbs
  or: sbatch -J test1 --ntasks-per-node=16 --mem=32gb vasp/_pbs
• Sample input files: zip

Important notes

• On some clusters VASP does not recognize INCARS created in Windows, so they should be convereted to UNIX format.

Help

• Resources:
  • wiki
  • forum
  • tips by Byungki Ryu
• Electronic density (CHGCAR file) can be visualized by VESTA program.
• VASP has its own visualization tool: p4vasp.
• To continue interrupted ionic relaxation copy CONTCAR to POSCAR.
• In case of RHOSYG: internal error: stars are not distinct, set ISYM=0.
• For linear response LREAL=F is the only option.
• NPAR=1 is probably the best choice for single node shared memory run.
• To remove orbital occupations use this script (run trimoutcar.sh OUTCAR > OUTCAR_new, thanks to Dmitry Aksenov).
• For calculation of DOS use ISMEAR=-5 (it might be necessary to converge WAVECAR by other method).
• For calculation of band structure with hybrid functionals the approach described in VASP-wiki works only for the provided example. For other cells and finer grids it produces inaccurate eigenvalues at least those with zero weight. The main challenge is to converge eigenvalues (individual eigenvalues converge much more slowly than band energy). At the minimum, one has to increase precision (PREC, maybe PRECFOCK) and number of SCF steps.

SCF convergence

• Read VASP documentation: Electronic optimization.
• Look into existing convergent calculations in any internal (Data folder) or external database (e.g. Materials Project).
• Continue iterations (run VASP with existing WAVECAR).
• Start with good WAVECAR, e.g. from smal energy cutoff. Set ICHARG=0 otherwise VASP may take wrong charge density.
• Start with good CHGCAR, e.g. from spin nonpolarized calculations, small k-grid etc. If needed correct unit cell parameters in CHGCAR to match POSCAR otherwise VASP ignores it. Alternatively you can set the grid appropriately.
• Change algorithm: ALGO=Fast is reasonable default choice for insulators (KFeSO4F) but only for Gaussian smearing.
• Change density mixing (LiFePO4 by materialsproject).
• Do non-SCF steps: NELMDL=-9 (usually a number from -6 till -12).
• Set LMIXMAX=4 (or even 6 for f-elements) for LDA-U calculations, see details here.
• Try LSUBROT=T, see details here.
• Try U-ramping. Check occupations by LDAUPRINT.
• Shake geometry and take the lowest energy result.
• On some computing clusters (e.g. Pardus and Arkuda) blocked-Davidson scheme changes number of electrons, so ALGO=VeryFast seems to be the only option.