Restarting finite differences calculations

Mind: Available as of VASP 6.6.0

Important: This feature requires HDF5 support.

Since VASP 6.6.0, it has been possible to restart finite difference calculations using IBRION = 6 and CHECKPOINT. The displacements are written to a vaspcheck.h5 file. For details of a general finite difference calculation, see the phonons from finite differences HowTo. In this HowTo, we will concern ourselves with restarting and splitting finite difference calculations.

Important: This can only be done using IBRION = 6 currently. We recommend using this generally over IBRION = 5.

Input

There are several options for the CHECKPOINT tag:

CHECKPOINT = LEGACY - finite difference calculation are performed as those in VASP 6.5.1. No vaspcheck.h5 is written.
CHECKPOINT = RESET (default) - The vaspcheck.h5 file is overwritten or created, and the displacements are written to it.
CHECKPOINT = CONTINUE - used for restarting a finite differences calculation from a vaspcheck.h5 file.
CHECKPOINT = PREPARE - used for splitting a finite differences calculation.
CHECKPOINT = SINGLE - used for running a single displacement after splitting with CHECKPOINT = PREPARE.

We will describe the restart procedure and splitting a calculation below. As an example, we take a 3x3x1 graphene supercell POSCAR file from the phonon tutorials:

C18
1.0
   7.3521657209830806    0.0000000000000000    0.0000000000000000
  -3.6760828604915403    6.3671622872044793    0.0000000000000000
   0.0000000000000000    0.0000000000000000    8.0000000000000000
C
18
direct
   0.1111111111111133    0.2222222222222200    0.0000000000000000 C
   0.1111111111111133    0.5555555555555532    0.0000000000000000 C
   0.1111111111111133    0.8888888888888866    0.0000000000000000 C
   0.4444444444444466    0.2222222222222200    0.0000000000000000 C
   0.4444444444444466    0.5555555555555532    0.0000000000000000 C
   0.4444444444444466    0.8888888888888866    0.0000000000000000 C
   0.7777777777777801    0.2222222222222200    0.0000000000000000 C
   0.7777777777777799    0.5555555555555532    0.0000000000000000 C
   0.7777777777777799    0.8888888888888866    0.0000000000000000 C
   0.2222222222222200    0.1111111111111133    0.0000000000000000 C
   0.2222222222222200    0.4444444444444466    0.0000000000000000 C
   0.2222222222222199    0.7777777777777799    0.0000000000000000 C
   0.5555555555555532    0.1111111111111133    0.0000000000000000 C
   0.5555555555555534    0.4444444444444466    0.0000000000000000 C
   0.5555555555555532    0.7777777777777799    0.0000000000000000 C
   0.8888888888888866    0.1111111111111133    0.0000000000000000 C
   0.8888888888888866    0.4444444444444466    0.0000000000000000 C
   0.8888888888888866    0.7777777777777799    0.0000000000000000 C

along with a 4x4x1 k-mesh in our KPOINTS file:

K points
 0
Gamma
4  4  1
0  0  0

and PAW C_s 04May1998 POTCAR. We can use the following INCAR file:

SYSTEM = graphene
ENCUT = 400

# electronic
PREC = Accurate
NELMIN = 5
EDIFF = 1e-8
ISMEAR = -1
SIGMA = 0.2
LREAL = .FALSE.
LWAVE = .FALSE.
LCHARG = .FALSE.

# ionic (finite differences)
IBRION = 6
POTIM = 0.015

Calculations

Restarting a finite difference calculation

We assume that this calculation has failed after a certain point or accidentally been cancelled:

DAV:   9    -0.181350430125E+03   -0.25600E-06   -0.38452E-08   848   0.118E-03    0.774E-04
DAV:  10    -0.181350430581E+03   -0.45554E-06   -0.20686E-08   864   0.738E-04    0.208E-04
DAV:  11    -0.181350430676E+03   -0.94857E-07   -0.22346E-09   704   0.354E-04    0.207E-04
srun: Job step aborted: Waiting up to 32 seconds for job step to finish.
slurmstepd-test01: error: *** JOB 254054 ON test01 CANCELLED AT 2026-01-19T16:25:49 ***
slurmstepd-test01: error: *** STEP 254054.0 ON test01 CANCELLED AT 2026-01-19T16:25:49 ***

If you look in the directory, you will see the vaspcheck.h5 file. Inside, it will contain the displacement calculations that have been completed up to the point of the crash:

h5ls vaspcheck.h5

data-1                   Group
displacements            Group
symmetry                 Group

The calculation can then be restarted by adding CHECKPOINT = CONTINUE to the INCAR:

SYSTEM = graphene
ENCUT = 400

# electronic
PREC = Accurate
NELMIN = 5
EDIFF = 1e-8
ISMEAR = -1
SIGMA = 0.2
LREAL = .FALSE.
LWAVE = .FALSE.
LCHARG = .FALSE.

# ionic (finite differences)
IBRION = 6
POTIM = 0.015
CHECKPOINT = CONTINUE

You can resubmit your calculation in the directory, e.g., with a job script: sbatch job.sh, and the finite differences calculation will continue from there after an SCF step has been done:

Continuing from previous run

The calculation will then continue as normal until all displacements have been completed and the phonon modes calculated:

h5ls vaspcheck.h5

data-1                   Group
data-2                   Group
data-3                   Group
data-4                   Group
displacements            Group
symmetry                 Group

Splitting a finite difference calculation

For some large structures, it may be easier to split the displacements into separate calculations. This is done in three steps:

Preparing the separate displacements - CHECKPOINT = PREPARE
Single displacement calculations - CHECKPOINT = SINGLE
Collected finite differences - CHECKPOINT = CONTINUE

Preparing the displacements

The displacements can be prepared using the CHECKPOINT = PREPARE tag:

SYSTEM = graphene
ENCUT = 400

# electronic
PREC = Accurate
NELMIN = 5
EDIFF = 1e-8
ISMEAR = -1
SIGMA = 0.2
LREAL = .FALSE.
LWAVE = .FALSE.
LCHARG = .FALSE.

# ionic (finite differences)
IBRION = 6
POTIM = 0.015
CHECKPOINT = PREPARE

This will create CONTCAR_disp-N files containing each of the displacements in the parent directory as you can see in the stdout:

Creating CONTCAR files for finite difference displacements

and the vaspcheck.h5 file:

h5ls vaspcheck.h5

displacements            Group
subdir_prefix            Dataset {SCALAR}
symmetry                 Group
total_count              Dataset {SCALAR}

Single displacement calculations

You can then create different directories for each of these CONTCAR_disp-N and run them separately. For this, you will need to set CHECKPOINT:

max=$(printf "%s\n" CONTCAR_disp-* | sed 's/.*-//' | sort -n | tail -1)

for i in $(seq 1 $max); do
  mkdir -p disp-$i
  cp CONTCAR_disp-$i disp-$i/POSCAR
  cp INCAR POTCAR KPOINTS CHG CHGCAR WAVECAR vasp.run disp-$i/
  sed -i 's/PREPARE/SINGLE/g' disp-$i/INCAR
done

Each INCAR file in the directories will then look like:

SYSTEM = graphene
ENCUT = 400

# electronic
PREC = Accurate
NELMIN = 5
EDIFF = 1e-8
ISMEAR = -1
SIGMA = 0.2
LREAL = .FALSE.
LWAVE = .FALSE.
LCHARG = .FALSE.

# ionic (finite differences)
IBRION = 6
POTIM = 0.015
CHECKPOINT = SINGLE

Important: You can set different NCORE settings in these calculations, offering parallelization that is not otherwise currently possible for finite differences.

In each subdirectory, you can see that a single displacement has been recorded in the stdout:

Computing single independent displacement for finite differences

and to the vaspcheck.h5 file:

h5ls vaspcheck.h5
 
data-1                   Group

Collected finite differences

Returning to the parent directory, you can combine these separate displacements into one finite difference calculation using the CHECKPOINT = CONTINUE tag:

SYSTEM = graphene
ENCUT = 400

# electronic
PREC = Accurate
NELMIN = 5
EDIFF = 1e-8
ISMEAR = -1
SIGMA = 0.2
LREAL = .FALSE.
LWAVE = .FALSE.
LCHARG = .FALSE.

# ionic (finite differences)
IBRION = 6
POTIM = 0.015
CHECKPOINT = CONTINUE

Warning: You can do this calculation in a different directory but you must include the vaspcheck.h5 from the prepare step there as it contains important metadata.

Running this calculation, you can see that each of the single displacements are combined into one calculation in the stdout:

 Combining displacements from subdirectories

and vaspcheck.h5 file:

h5ls vaspcheck.h5

displacements            Group
subdir_prefix            Dataset {SCALAR}
symmetry                 Group
total_count              Dataset {SCALAR}

You can find information about the computed phonon modes under Eigenvectors and eigenvalues of the dynamical matrix as usual.

Practical hints

The phonon frequencies will differ slightly between one run and the split calculation, on the order of the 4th or 5th significant figure. We do not expect this to be signficant.

You can can use this method for any electron-phonon calculations.

You can set different NCORE settings in these calculations, offering parallelization that is not otherwise currently possible for finite differences.

Make sure to check that you are using the correct vaspcheck.h5 file with each calculation, particularly the split calculation. If data is read from an inappropriate vaspcheck.h5 file, you will see it under "".