vasp-mace 2.0.0

A VASP-like interface for running ML potential calculations with MACE

vasp-mace

VASP-like interface for structure relaxation, molecular dynamics, and energy calculations using MACE machine-learning potentials

vasp-mace emulates VASP for fast, low-cost atomistic simulations using pretrained MACE interatomic potentials, with optional empirical dispersion corrections (DFT-D3). It reads VASP-style inputs (POSCAR, INCAR) and produces VASP-compatible outputs (CONTCAR, OUTCAR, OSZICAR, XDATCAR, vasprun.xml), enabling seamless integration with existing VASP workflows and post-processing tools.

---

Features

• Single-point energy, force, and stress evaluation (NSW = 0)
• Geometry relaxation of atomic positions and/or unit cell, driven by MACE potentials
• Molecular dynamics (NVE, NVT Langevin/Nosé-Hoover/Andersen, NPT Langevin) with XDATCAR output
• Nudged Elastic Band (NEB): minimum-energy path and transition-state search via ASE's MDMin optimizer; optional climbing-image NEB (LCLIMB = .TRUE., VTST convention)
• Phonon calculations: Γ-point force constants and frequencies via finite differences (IBRION = 5); symmetry-reduced displacements via phonopy (IBRION = 6), with VASP-compatible DYNMAT and OUTCAR output
• Selective dynamics: per-atom coordinate fixing from POSCAR, preserved in CONTCAR
• DFT-D3 dispersion correction via IVDW in INCAR (zero-damping and Becke-Johnson variants, with optional three-body ATM term)
• Multiple ISIF modes: positions-only, full cell relaxation, constant-volume shape relaxation, volume-only
• Force-based (EDIFFG < 0) and energy-based (EDIFFG > 0) convergence criteria
• Target pressure support via PSTRESS (ISIF = 3)
• VASP-compatible output written to the run directory: CONTCAR, OUTCAR, OSZICAR, XDATCAR, vasprun.xml
• ASE trajectory and log written to ase_files/ subdirectory, keeping the run directory clean with only VASP-like output files

When would you use this?

For most workflows, a pure Python/ASE script is the better way to run MACE (more flexible, easier to customise, no file overhead). vasp-mace is not trying to compete with that.

It exists for the cases where VASP-style files are what you already have or what you need:

• You have a set of POSCAR/INCAR files from a previous VASP project and want a quick MACE relaxation without rewriting any input.
• You are using an external code that reads VASP output (e.g. vasprun.xml for ShengBTE, CONTCAR for a downstream workflow) and you want MACE to slot in transparently.
• You are comparing MACE results against VASP calculations and prefer to keep the input/output format identical to reduce variables.
• You are simply too fond of VASP files to let go, and that is a perfectly valid reason.

Whatever brings you here, enjoy vasp-mace.

---

Installation

conda create -n vasp_mace_env python=3.11 -y
conda activate vasp_mace_env
conda install -c conda-forge dftd4
pip install vasp-mace

DFT-D3 dispersion (IVDW tag) requires dftd4, which is best installed via conda before pip install vasp-mace. If you do not need dispersion corrections, the conda step can be skipped.

Development install (includes examples)

To get the example input files or contribute to the code, clone the repository instead:

git clone https://github.com/rgraucrespo/vasp-mace.git
cd vasp-mace
conda create -n vasp_mace_env python=3.11 -y
conda activate vasp_mace_env
conda install -c conda-forge dftd4
pip install -e .

Model checkpoint

Download a pretrained MACE model checkpoint, for example:

wget https://github.com/ACEsuit/mace-foundations/releases/download/mace_mp_0/2024-01-07-mace-128-L2_epoch-199.model

Point vasp-mace to it by setting the environment variable (e.g. in your .bashrc or .zshrc):

export MACE_MODEL_PATH=/path/to/2024-01-07-mace-128-L2_epoch-199.model

---

Usage

Place POSCAR and INCAR in your working directory, then run:

vasp-mace

The mode (single-point, relaxation, NEB, or MD) is determined automatically from the INCAR tags.

---

INCAR parameters

Only the tags relevant to vasp-mace are parsed; all others are silently ignored.

General

Tag	Default	Description
NSW	0	Max ionic steps. 0 = single-point calculation
IBRION	-1	-1 = use --optimizer CLI flag; 0 = MD; 1 = LBFGS; 2 = BFGS; 3 = FIRE; 5 = phonons (no symmetry); 6 = phonons (symmetry-reduced via phonopy). For NEB (IMAGES > 0): 1 and 2 both map to MDMin (the recommended NEB optimizer)
EDIFFG	-0.05	Convergence criterion. < 0: max force (eV/Å); > 0: energy change per ion (eV)
ISIF	2	Degrees of freedom to relax (see table below)
PSTRESS	0.0	Target hydrostatic pressure in kBar, applied when ISIF = 3
IVDW	0	Empirical dispersion correction (see table below)

ISIF modes

ISIF	Positions	Cell shape	Cell volume	Note
0, 1, 2	relaxed	fixed	fixed	MD: NVT. 0 and 1 are treated as 2 (in VASP they differ only in how much of the stress tensor is computed; vasp-mace always computes the full stress)
3	relaxed	relaxed	relaxed	MD: NPT (MDALGO=3)
4	relaxed	relaxed	fixed	Relax only
7	fixed	fixed	relaxed	Relax only
8	relaxed	fixed	relaxed	Relax only

IVDW (DFT-D3 dispersion)

IVDW	Method
0	None (default)
11	D3(zero-damping)
12	D3(Becke-Johnson)
13	D3(zero-damping) + ATM three-body
14	D3(Becke-Johnson) + ATM three-body

Nudged Elastic Band (IMAGES ≥ 1)

Triggered when IMAGES > 0 in INCAR. No top-level POSCAR is used; instead, place endpoint and (optionally) intermediate images in numbered subdirectories:

00/POSCAR   ← reactant (fixed endpoint)
01/POSCAR   ← intermediate image 1  (optional; generated by IDPP if absent)
…
NN/POSCAR   ← product  (fixed endpoint)   NN = IMAGES + 1

If intermediate POSCARs are absent, all images are generated automatically by IDPP interpolation.

Tag	Default	Description
IMAGES	0	Number of intermediate NEB images. IMAGES ≥ 1 triggers NEB mode
SPRING	-5.0	Spring constant for NEB (eV/Ų). Use negative values (SPRING < 0, VASP convention for NEB); the spring constant is `k =
LCLIMB	.FALSE.	Enable climbing-image NEB (CI-NEB). Not a native VASP tag — borrowed from the VTST Tools convention (see note below)

LCLIMB and VTST convention: In VASP with the optional VTST extension, CI-NEB is activated by LCLIMB = .TRUE.. Native VASP (without VTST) does not recognise this tag and always runs plain NEB. vasp-mace follows the VTST convention so that INCAR files from VTST-enabled VASP work without modification.
SPRING follows the VASP sign convention: negative values (SPRING < 0) indicate NEB, and the spring constant is k = |SPRING|. Positive values correspond to the non-nudged elastic band method and are not supported by vasp-mace. CI-NEB is controlled exclusively by LCLIMB, not by the sign of SPRING.

Phonon calculations (IBRION = 5 or 6)

Finite-difference second derivatives (force constants and phonon frequencies) at the Γ-point.

Tag	Default	Description
POTIM	0.015	Displacement amplitude (Å). In phonon mode this sets the finite-difference step, not the MD timestep. VASP uses 0.02 Å; either value is acceptable
NFREE	2	Displacements per degree of freedom: 2 = central differences (±δ, recommended); 1 = forward differences (+δ only)

IBRION = 5 computes all N × 3 × NFREE displacements (no symmetry).
IBRION = 6 uses phonopy to reduce the number of displacements via crystal symmetry. Install with pip install vasp-mace[phonons] or pip install phonopy. Falls back to IBRION = 5 if phonopy is not installed.

Output files produced (VASP-compatible format):

File	Description
DYNMAT	Force-constant matrix in VASP DYNMAT format (central-difference half-forces)
OUTCAR	Phonon eigenvalues and eigenvectors (modes ordered high → low frequency)
OSZICAR	One energy line per displaced configuration
XDATCAR	Initial + all displaced configurations
CONTCAR	Initial (equilibrium) structure
ase_files/force_constants.npy	Force constant tensor C[i,α,j,β] in eV/Ų (shape N×3×N×3)
ase_files/phonopy_params.yaml	Phonopy parameters file (IBRION = 6 only)

Note: IBRION = 5/6 performs single-point force evaluations only (no ionic relaxation). The structure should be pre-relaxed to a minimum before running phonon calculations.

ISIF is ignored for phonon runs. In VASP, ISIF ≥ 3 with IBRION = 5/6 additionally computes the elastic tensor and internal strain tensor. vasp-mace currently computes force constants only, regardless of ISIF. Elastic constant support is planned for a future release.

Molecular dynamics (IBRION = 0)

Tag	Default	Description
MDALGO	3	1 = VelocityVerlet: NVE if ANDERSEN_PROB = 0, NVT Andersen if ANDERSEN_PROB > 0; 2 = NVT Nosé-Hoover; 3 = NVT Langevin (ISIF=2) or NPT Langevin (ISIF=3)
TEBEG	0.0	Starting temperature (K). Velocities initialised from Maxwell-Boltzmann distribution
TEEND	-1	Ending temperature (K) for linear ramp; -1 = same as TEBEG (constant temperature)
POTIM	0.5	MD timestep (fs). Use ≤ 1.0 fs for systems containing hydrogen
NBLOCK	1	Write XDATCAR frame and trajectory snapshot every NBLOCK steps
ANDERSEN_PROB	0.0	Collision probability for Andersen thermostat (MDALGO = 1)
LANGEVIN_GAMMA	10.0	Friction coefficient(s) (ps⁻¹) for atoms in Langevin MD (MDALGO = 3). A single value applies to all atoms; multiple space-separated values are assigned per species in POSCAR order. Also reads from SMASS if LANGEVIN_GAMMA is missing
LANGEVIN_GAMMA_L	10.0	Friction coefficient (ps⁻¹) for the lattice in Langevin NPT (MDALGO = 3, ISIF = 3)
PMASS	0	Piston mass (amu) for Langevin NPT (MDALGO = 3, ISIF = 3). 0 = automatic (N × 10000 amu)
SMASS	-3.0	Nose-Hoover mass or Langevin friction. Values > 0 are used if LANGEVIN_GAMMA is missing

NVE — microcanonical ensemble

Use MDALGO = 1 with ANDERSEN_PROB = 0.0 (no collisions → pure VelocityVerlet integrator).

IBRION        = 0
MDALGO        = 1
ANDERSEN_PROB = 0.0
NSW           = 5000
TEBEG         = 300      # sets initial velocity distribution only
POTIM         = 1.0
NBLOCK        = 10

NVT — canonical ensemble

Three thermostat options are available:

Andersen thermostat (MDALGO = 1, ANDERSEN_PROB > 0): stochastic velocity rescaling at each step. Simple and robust; ANDERSEN_PROB controls how frequently velocities are reassigned from the Maxwell-Boltzmann distribution (typical range: 0.01–0.1).

IBRION        = 0
MDALGO        = 1
ANDERSEN_PROB = 0.05
NSW           = 5000
TEBEG         = 300
POTIM         = 1.0
NBLOCK        = 10

Nosé-Hoover thermostat (MDALGO = 2): deterministic extended-system thermostat; generates a correct NVT ensemble. SMASS > 0 sets the coupling time (ps); the default (SMASS ≤ 0) uses a period of 40 MD steps.

IBRION = 0
MDALGO = 2
SMASS  = 1.0
NSW    = 5000
TEBEG  = 300
POTIM  = 1.0
NBLOCK = 10

Langevin thermostat (MDALGO = 3, ISIF = 2): stochastic friction + random force; well-suited for systems with slow equilibration. LANGEVIN_GAMMA accepts a single value (all atoms) or one value per species in POSCAR order.

IBRION         = 0
MDALGO         = 3
ISIF           = 2
LANGEVIN_GAMMA = 10.0 20.0   # per-species: species1=10, species2=20 ps^-1
NSW            = 5000
TEBEG          = 300
POTIM          = 1.0
NBLOCK         = 10

NPT — isothermal-isobaric ensemble

Use MDALGO = 3 with ISIF = 3. The Langevin barostat controls the cell volume; LANGEVIN_GAMMA_L sets the lattice friction. Set PSTRESS to the target pressure in kBar (0 = ambient pressure). PMASS sets the barostat piston mass in amu (default: N × 10000 amu).

IBRION           = 0
MDALGO           = 3
ISIF             = 3
LANGEVIN_GAMMA   = 10.0 20.0   # per-species: species1=10, species2=20 ps^-1
LANGEVIN_GAMMA_L = 10.0
PMASS            = 50000
PSTRESS          = 0.0
NSW              = 5000
TEBEG            = 300
POTIM            = 1.0
NBLOCK           = 10

---

CLI options

Option	Default	Description
--model PATH	$MACE_MODEL_PATH	Path to MACE .model checkpoint
--device	auto	auto (→ cuda if available, else mps, else cpu), cpu, cuda, mps
--dtype	auto	auto (→ float64 on CPU, float32 on GPU/MPS), float32, or float64
--optimizer	BFGS	Fallback optimizer: BFGS, FIRE, or LBFGS. Overridden by IBRION if set in INCAR

GPU acceleration

vasp-mace supports GPU-accelerated inference via PyTorch. Every energy and force evaluation is a neural-network forward pass, so a GPU can deliver 10–100× speedup over CPU depending on system size.

NVIDIA (CUDA)

Install a CUDA-enabled PyTorch build before installing vasp-mace (see pytorch.org/get-started for the right command for your CUDA version), then run:

vasp-mace --device cuda

Or simply run vasp-mace without --device — it will auto-detect CUDA.

Apple Silicon (MPS)

vasp-mace --device mps

Precision

GPU mode defaults to float32, which is faster and sufficient for MACE inference. Use --dtype float64 to override if higher precision is needed.

NEB memory note: NEB runs load one calculator per image. With many images and a large model, this multiplies GPU memory usage. If memory is tight, reduce the number of images or use --dtype float32 (which is already the default on GPU).

---

Differences with respect to VASP

While vasp-mace aims for a high degree of compatibility, there are important technical differences to keep in mind:

• Electronic Steps: Since MACE is a machine-learning potential, there are no "electronic steps" in the DFT sense. For compatibility with tools like vasprun.xml, a single dummy electronic step is recorded per ionic step.
• Nosé-Hoover Coupling: In VASP, SMASS directly sets the thermostat mass ($Q$). In vasp-mace, if SMASS > 0, it is treated as a characteristic damping time in picoseconds ($t_{damp} = \text{SMASS} \times 1 \text{ ps}$). The default SMASS = 0 (or $\le 0$) correctly maps to an oscillation period of 40 time steps, matching VASP's default behavior.
• Langevin NPT Algorithm: The NPT implementation in vasp-mace (MDALGO = 3, ISIF = 3) uses the stochastic barostat algorithm of Quigley and Probert (2004). This correctly samples the NPT ensemble but may fluctuate differently than VASP's internal implementation.
• Piston Mass: The lattice "piston mass" for NPT defaults to N × 10000 amu, but can be set explicitly via the PMASS INCAR tag (in amu, matching VASP's convention).
• Optimizers: Relaxation (IBRION = 1, 2, 3) uses ASE's robust optimizers (LBFGS, BFGS, and FIRE) rather than VASP's internal RMM-DIIS or conjugate gradient routines.
• NEB Optimizer: NEB calculations always use ASE's MDMin optimizer (regardless of IBRION). MDMin projects velocities along the force direction and resets them when the velocity and force point in opposite directions, which prevents the divergence that plagues BFGS and FIRE under non-conservative spring forces. VASP with VTST uses a different quasi-Newton method.
• LCLIMB tag: Not a native VASP tag. It originates from the VTST Tools extension package for VASP. Native VASP (without VTST) does not recognise LCLIMB and always runs plain NEB.
• Phonon calculations: IBRION = 5/6 produce DYNMAT, OUTCAR (eigenvalues + eigenvectors in VASP format), OSZICAR, XDATCAR, and CONTCAR. The OUTCAR phonon section is VASP-compatible (modes ordered high → low frequency; f = for real, f/i= for imaginary). VASP's electronic-iteration lines in OSZICAR are omitted (replaced by a single energy summary line per configuration). vasprun.xml is not written for phonon runs.

---

Output files

Relaxation (NSW > 0, IBRION ≠ 0)

File	Description
CONTCAR	Final structure in VASP format (preserves Selective Dynamics if present in POSCAR)
OSZICAR	Per-step energy, ΔE, and Fmax
OUTCAR	Minimal OUTCAR with lattice, stress tensor, forces, and per-step energies
XDATCAR	Trajectory of ionic positions (one frame per ionic step)
vasprun.xml	Minimal XML with energies and final structure
ase_files/mace.traj	Full ASE binary trajectory
ase_files/opt.log	ASE optimizer log

Molecular dynamics (IBRION = 0)

File	Description
CONTCAR	Final structure
XDATCAR	Trajectory in fractional coordinates (written every NBLOCK steps)
ase_files/mace.traj	Full ASE binary trajectory
ase_files/md.log	ASE MD log (step, time, energy, temperature)

NEB (IMAGES ≥ 1)

Written per image directory (00/, 01/, …, NN/):

File	Description
CONTCAR	Final image geometry
OSZICAR	Per-step energy for this image
OUTCAR	Forces and energies for this image
vasprun.xml	Single-point or relaxation XML for this image

Shared output in ase_files/:

File	Description
ase_files/neb_opt.log	MDMin optimizer log
ase_files/mace.traj	Converged NEB band as ASE trajectory (one frame per image, reactant → product)

No XDATCAR is produced for NEB runs.

Single-point (NSW = 0)

File	Description
OUTCAR	Lattice, stress tensor, and forces
OSZICAR	Single-line energy summary
vasprun.xml	Full single-point XML compatible with ShengBTE and Phonopy

Phonon calculations (IBRION = 5 or 6)

See the Phonon calculations INCAR section above for a description of output files.

---

Examples

Ready-to-run examples are provided in the examples/ directory. Copy an example folder to your working directory and run vasp-mace inside it.

Example	System	Description
example01_MgO/	MgO (rock salt, conventional cell)	Variable-cell relaxation (ISIF = 3), no dispersion
example02_hBN_D3-dispersion/	h-BN (hexagonal)	Variable-cell relaxation with D3(BJ) dispersion (IVDW = 12)
example03_CsPbI3_MA_MD/	Cs₆₃MA·PbI₃ perovskite (4×4×4, 327 atoms)	NVT Nosé-Hoover MD at 500 K with one methylammonium defect
example04_PbTe_pressure/	PbTe (rock salt)	Variable-cell relaxation under 15 kBar target pressure (PSTRESS = 15)
example05_Si_NEB/	Si (diamond cubic)	CI-NEB for Si interstitial migration (LCLIMB = .TRUE., 4 intermediate images)
example06_Pt_NEB/	Pt (fcc-001 surface)	CI-NEB for Pt adatom collective jump (LCLIMB = .TRUE., 3 intermediate images)
example07_PbTe_MD/	PbTe (rock salt, 512 atoms)	Sequential NVT → NPT Langevin MD; per-species LANGEVIN_GAMMA and explicit PMASS
example08_PbTe_phonons/	PbTe (rock salt, 8 atoms)	Phonon calculation with symmetry reduction (IBRION = 6, NFREE = 2). Requires pip install phonopy

example01 — MgO variable-cell relaxation

NSW    = 100
ISIF   = 3
EDIFFG = -0.01

example02 — h-BN with DFT-D3 dispersion

NSW    = 100
ISIF   = 3
EDIFFG = -0.01
IVDW   = 12

example03 — CsPbI₃ perovskite MD with methylammonium

4×4×4 supercell (327 atoms) of cubic CsPbI₃ with one Cs site replaced by methylammonium (CH₃NH₃⁺). NVT Nosé-Hoover thermostat at 500 K.

IBRION = 0
MDALGO = 2
NSW    = 200
TEBEG  = 500
POTIM  = 0.5
NBLOCK = 5
SMASS  = 1.0

example04 — PbTe under pressure

NSW     = 100
ISIF    = 3
EDIFFG  = -0.01
PSTRESS = 15

example05 — Si interstitial migration (CI-NEB)

4-image CI-NEB for a Si self-interstitial hop in the diamond cubic lattice. Intermediate images are provided as starting POSCARs (previously converged); IDPP interpolation is used automatically if they are absent.

NSW    = 100
EDIFFG = -0.01
IBRION = 1
ISIF   = 2
IMAGES = 4
SPRING = -5
LCLIMB = .TRUE.

example06 — Pt adatom collective jump (CI-NEB)

3-image CI-NEB for a collective Pt adatom jump on the fcc-Pt(001) surface. Endpoint and intermediate POSCARs are provided.

NSW    = 100
EDIFFG = -0.01
IBRION = 1
ISIF   = 2
IMAGES = 3
SPRING = -5
LCLIMB = .TRUE.

example07 — PbTe sequential NVT → NPT MD

512-atom PbTe supercell. Demonstrates per-species LANGEVIN_GAMMA (Pb and Te assigned different friction coefficients) and explicit PMASS. Run with the provided run.sh, which chains the two stages automatically and saves outputs to nvt_output/ and npt_output/.

Stage 1 — NVT equilibration (INCAR_NVT):

IBRION         = 0
MDALGO         = 3
ISIF           = 2
LANGEVIN_GAMMA = 10.0 20.0   # Pb: 10 ps^-1, Te: 20 ps^-1
NSW            = 500
TEBEG          = 300
POTIM          = 1.0
NBLOCK         = 1

Stage 2 — NPT production (INCAR_NPT, starts from NVT CONTCAR):

IBRION           = 0
MDALGO           = 3
ISIF             = 3
LANGEVIN_GAMMA   = 10.0 20.0   # Pb: 10 ps^-1, Te: 20 ps^-1
LANGEVIN_GAMMA_L = 10.0
PMASS            = 50000
NSW              = 500
TEBEG            = 300
POTIM            = 1.0
NBLOCK           = 1

bash run.sh --model /path/to/model

example08 — PbTe phonon calculation (IBRION = 6)

8-atom PbTe conventional cell. Symmetry-reduced phonon calculation: only 2 irreducible displacements are needed (vs. 48 for a brute-force IBRION = 5 run). Writes DYNMAT, OUTCAR (with frequencies and eigenvectors), OSZICAR, XDATCAR, CONTCAR, and ase_files/phonopy_params.yaml.

ISIF   = 0
IBRION = 6
NFREE  = 2
POTIM  = 0.02
NSW    = 1

Run with:

vasp-mace --model /path/to/model

Requires phonopy for symmetry reduction: pip install phonopy or pip install vasp-mace[phonons].

---

Development

pip install -e ".[dev]"   # installs pytest, black, ruff, mypy
ruff check vasp_mace/     # lint
black vasp_mace/          # format

---

License and citation

MIT License © 2025 Ricardo Grau-Crespo.

If you use vasp-mace in your work, please cite:

vasp-mace:

• Grau-Crespo, R. vasp-mace: a VASP simulator based on the MACE machine-learning interatomic potential (2025). Zenodo. https://doi.org/10.5281/zenodo.19479246

@software{graucrespo2025vaspmace,
  author  = {Grau-Crespo, Ricardo},
  title   = {vasp-mace: a VASP simulator based on the MACE machine-learning interatomic potential},
  year    = {2025},
  url     = {https://github.com/rgraucrespo/vasp-mace},
  doi     = {10.5281/zenodo.19479246},
}

MACE potentials:

• Batatia, I.; Kovács, D. P.; Simm, G. N. C.; Ortner, C.; Csányi, G. “MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields”. Advances in Neural Information Processing Systems (NeurIPS), 2022.
• Batatia, I. et al. “A foundation model for atomistic materials chemistry.” The Journal of Chemical Physics 163, no. 18 (2025).

VASP (if referring to specific VASP formats or comparing against VASP results):

• Kresse, G.; Furthmüller, J. “Efficiency of ab-initio total energy calculations for metals and semiconductors using a plane-wave basis set.” Computational Materials Science 6 (1996) 15–50.
• Kresse, G.; Furthmüller, J. “Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set.” Physical Review B 54 (1996) 11169–11186.