The short version
- Relax both end points with identical settings and the same atom order; then interpolate the images (IDPP is better than linear).
- Stock VASP does a plain NEB with
IMAGESandSPRING = -5, using IBRION = 1 or 3 (not 2). The climbing image needs VASP built with the VTST code. - Under VTST, LCLIMB is on by default: run a first, plain NEB with
LCLIMB = .FALSE., then switch it on. - The number of MPI ranks must be a multiple of IMAGES.
- A saddle point has exactly one imaginary frequency: check it.
Stock VASP or VTST
| Feature | Stock VASP 6.6.1 | VASP + VTST code |
|---|---|---|
| Plain NEB (IMAGES, SPRING) | yes | yes, with an improved tangent |
| Climbing image (LCLIMB) | no | yes, on by default |
| Force-based optimizers (IOPT: LBFGS, CG, quick-min, FIRE) | no | yes |
| Cell changing along the path (solid-state NEB) | no | yes (LNEBCELL) |
| Dimer method | yes (IBRION = 44) | yes (ICHAIN = 2) |
To see whether your build has VTST and what it read: grep -E "NEB:|OPT:" 01/OUTCAR. A stock build ignores the VTST tags.
The workflow
- End points. Relax the initial and final states with identical settings: ENCUT, PREC, smearing, KPOINTS, POTCAR, functional, ISPIN and MAGMOM, LREAL. Any error in them shifts the whole path. For spin-polarized r²SCAN do not use VASP 6.6.0.
- Atom order. Every atom must have the same index in both end points; a swapped pair of identical atoms gives a nonsense path that no tool flags. Compare per-atom displacements (
diffcon.pl) and look for atoms that jump. - Interpolate. IDPP (VTST
nebmake.py, ASE, pymatgen) avoids the overlaps that linear interpolation (nebmake.pl) can create. Use the periodic minimum-image path, so atoms do not fly across the cell. - Inspect. Make a movie of the images (
nebmovie.pl 0) and check the shortest distances: atoms almost on top of each other crash VASP. - Plain NEB with a loose force criterion, then a climbing-image NEB from its CONTCARs with a tight one.
- Confirm the saddle with a frequency calculation on the climbing image (one large imaginary mode), or refine it with the dimer method.
Set up and run
dist.pl ini/CONTCAR fin/CONTCAR # total displacement between the end points
diffcon.pl ini/CONTCAR fin/CONTCAR # per-atom displacements: spot ordering or wrapping problems
nebmake.py ini/CONTCAR fin/CONTCAR 5 # 5 images, IDPP by default; nebmake.pl for linear
cp ini/OUTCAR 00/ ; cp fin/OUTCAR 06/ # the analysis scripts need the end-point OUTCARs
nebmovie.pl 0 # movie.xyz of the starting images: look at it
srun -n 160 vasp_std > vasp.out # from the parent folder; ranks = a multiple of IMAGESFolders 00 to 06 hold POSCAR (and WAVECAR, CHGCAR if any); INCAR, KPOINTS and POTCAR stay in the parent folder. Start with few images (one to five) and add more once the band has converged.
INCAR recipes
ENCUT = 520 ! the same settings as the end-point relaxations
PREC = Accurate
EDIFF = 1E-6
ISMEAR = 0 ; SIGMA = 0.05
IMAGES = 4 ! MPI ranks must be a multiple of 4
SPRING = -5 ! the default; negative switches nudging on
IBRION = 1 ! or 3; not 2 (conjugate gradient usually fails for NEB)
POTIM = 0.5 ! tune it; IBRION = 1 is sensitive to the step
ISIF = 2 ! the cell stays fixed
NSW = 200
EDIFFG = -0.03 ! tighten to -0.01 for production
LCHARG = .FALSE.ENCUT = 520
PREC = Accurate
EDIFF = 1E-6
ISMEAR = 0 ; SIGMA = 0.05
IMAGES = 5
ICHAIN = 0 ! NEB
LCLIMB = .FALSE. ! the VTST default is .TRUE.
SPRING = -5
IBRION = 3 ; POTIM = 0 ! required by the VTST optimizers
IOPT = 3 ! quick-min (7 = FIRE) while forces are large
EDIFFG = -0.1 ! a loose first pass
NSW = 200
ISIF = 2ENCUT = 520
PREC = Accurate
EDIFF = 1E-6
ISMEAR = 0 ; SIGMA = 0.05
IMAGES = 5
ICHAIN = 0
LCLIMB = .TRUE.
SPRING = -5
IBRION = 3 ; POTIM = 0
IOPT = 1 ! LBFGS (or 2 = CG) once the forces are accurate
EDIFFG = -0.03 ! -0.01 for a well-converged saddle
NSW = 300
ISIF = 2ENCUT = 520
PREC = Accurate
EDIFF = 1E-7
ISMEAR = 0 ; SIGMA = 0.05
IBRION = 5 ! finite differences; honours selective dynamics
NFREE = 2
POTIM = 0.015
NSW = 1For the frequency check, fix the atoms far from the event with selective dynamics. In OUTCAR, imaginary modes are marked f/i=: expect one large one; tiny ones are numerical noise.
Monitor and analyse
nebef.pl # per image: max force, energy, energy relative to image 00
nebresults.pl # neb.dat, spline.dat, exts.dat, mep.eps, movie.xyz (then gzips the OUTCARs)
vfin.pl stage1 # archive a stage and move CONTCAR over POSCAR in every image
neb2dim.pl # set up a dimer run at the highest pointAfter nebresults.pl the OUTCARs are gzipped: use zgrep or unzip them. With pymatgen, NEBAnalysis.from_dir(".") reads the barrier from the same folders.
| Tag | What it does | Default | What to use | Watch out |
|---|---|---|---|---|
IMAGES | Number of images between the two end points of an NEB, one folder each (01, 02, …). | 0 | Start small (a few images) and add more once the band has converged. | The number of MPI ranks must be a multiple of IMAGES. |
SPRING | Spring constant between images; a negative value switches on nudging. | −5 | Keep −5. | Very large values slow convergence. |
ICHAIN VTST | VTST method: 0 nudged elastic band, 2 dimer, 1 dynamical matrix, 3 Lanczos. | 0 | 0 for NEB. | |
LCLIMB VTST | VTST climbing image: the highest image climbs to the saddle point. | .TRUE. (VTST) | .FALSE. for a first, plain NEB; then .TRUE. to converge the saddle point. | The VTST default is already .TRUE.: set .FALSE. explicitly for the first stage. Stock VASP has no climbing image. |
IOPT VTST | VTST optimizer: 1 LBFGS, 2 conjugate gradient, 3 quick-min, 7 FIRE (0 = use IBRION). | 0 | Quick-min or FIRE while forces are large; LBFGS or CG near convergence. | Set IBRION = 3 and POTIM = 0 so VASP's own optimizer stays idle (nothing checks this); a positive EDIFFG makes VTST stop. |
LNEBCELL VTST | VTST solid-state NEB: the cell changes along the path too. | .FALSE. | With ISIF = 3 and IOPT = 3 (or 7). | Needs a cutoff high enough for cell changes. |
Magnetic systems
- End points and images should sit on the same magnetic state. If one image drops into another spin state, converge a single-point run for that image and start the NEB from its WAVECAR and CHGCAR (each image folder can hold its own).
- Keep MAGMOM in the INCAR: on restarts it sets the magnetic symmetry.
- Before VASP 6.6.0, negative MAGMOM values could give too few bands by default: set NBANDS.
Common mistakes
- End points relaxed with different settings, or a different atom order (no tool reports it).
- Linear interpolation without the minimum-image convention: atoms cross the whole cell.
- Expecting a climbing image from stock VASP, or forgetting that VTST climbs by default.
- IBRION = 2 for the NEB; IOPT without IBRION = 3 and POTIM = 0; a positive EDIFFG with IOPT (VTST stops).
- A rank count that is not a multiple of IMAGES.
- Too many images at the start, which slows convergence.
- Different cells at the two ends in a fixed-cell NEB: use the same cell, or solid-state NEB.
- pymatgen's NEB input set switches DFT+U off unless told otherwise; check LDAU.
- Reporting the dimer's DIMCAR energy as the saddle energy: take it from OUTCAR.
The physics behind this sheet
- Lecture 32: The nudged elastic band method (in preparation)
Further reading
- VASP Wiki, “Nudged elastic bands”. · www.vasp.at
- Henkelman group, VTST tools, “Nudged elastic band” (climbing image, optimizers and scripts). · theory.cm.utexas.edu
- VASP Wiki, “Improved dimer method”. · www.vasp.at