NequIP and Allegro

NequIP and Allegro are frameworks designed for developing interatomic potentials for molecular dynamics simulations using deep equivariant neural networks. The methodology and recent high-performance upgrades are described in detail in the literature (Batzner2022, Musaelian2023, Tan2025).

The CP2K interface is compatible with models trained and compiled with NequIP version >= 0.7.0, and it is consistent with the LAMMPS pair_nequip_allegro (v0.7.0) integration.

Input Section

Inference in CP2K has been unified and is configured entirely through the NEQUIP section within the &NONBONDED forcefield parameters.

An example of the input configuration:

&FORCEFIELD
  &NONBONDED
    &NEQUIP
      MODEL_TYPE  NEQUIP # possible choices are NEQUIP or ALLEGRO
      ATOMS H O
      POT_FILE_NAME NequIP/waterscan-neq0.16.nequip.pth
      UNIT_ENERGY eV
      UNIT_FORCES eV*angstrom^-1
      UNIT_LENGTH angstrom
    &END NEQUIP
  &END NONBONDED
&END FORCEFIELD

  • MODEL_TYPE: Specifies the architecture of the loaded model (NEQUIP or ALLEGRO).

  • ATOMS: Expects a list of elements/kinds.

  • POT_FILE_NAME: The path to the NequIP/Allegro model.

  • UNIT_*: These tags explicitly define the units for the model’s internal lengths, energies, and forces.

Full example input files demonstrating production-ready molecular dynamics setups can be found in the CP2K regression tests directory:

  • tests/Fist/regtest-nequip/water-bulk.inp

  • tests/Fist/regtest-allegro/water-bulk.inp

Compiling CP2K with LibTorch

Running NequIP or Allegro requires compiling CP2K with the LibTorch library. Versions of LibTorch (2.4 through 2.7) are supported.

For CP2K binaries running on CPUs, installing the toolchain using the flag --with-libtorch is sufficient.

To benefit from GPU acceleration, either compile LibTorch from scratch or download the precompiled LibTorch library for CUDA from PyTorch and provide the appropriate path to the toolchain script:

./install_cp2k_toolchain.sh --with-libtorch=<path-to-libtorch-cuda>

Validation & Reproducibility

  • Comparison with LAMMPS: We have verified that this implementation numerically reproduces the results of the LAMMPS pair_nequip_allegro plugin.

  • Data: The training datasets, model files inside data/NequIP and data/Allegro, input scripts, and parity plots used for validation are available on Zenodo: doi:10.5281/zenodo.18848354.

Further Resources

For additional references on NequIP, Allegro, and equivariant neural networks (e3nn), see: