cp2k-rs

Rust and Python bindings for CP2K, the electronic-structure and ab initio molecular-dynamics package.

Documentation · Examples · Changelog · Issues

pip install "cp2k-rs[ase]"

from ase.build import molecule
from ase.optimize import BFGS
from cp2k_rs.ase import CP2K

atoms = molecule("H2O")
atoms.center(vacuum=4.0)

with CP2K(nproc=4) as calc:           # CP2K on 4 MPI ranks
    atoms.calc = calc
    BFGS(atoms).run(fmax=0.05)        # each step reuses the force environment

No CP2K installation, no BLAS/LAPACK/FFTW/ScaLAPACK to build, no mpi4py in your script.

What it gives you

One command to install. The wheel is ~0.6 MB; a pre-built, statically linked worker binary carrying CP2K and its whole numerical stack is fetched on first use. The default CPU worker needs only your C library and host MPI; CUDA/ROCm workers additionally need their matching runtime, and the opt-in Skala worker needs a compatible PyTorch installation.

MPI without mpi4py. CP2K runs in a separate worker process. Rank 0 talks to your Python process over a Unix-domain socket and broadcasts commands to the other ranks, so an ordinary script drives a fully parallel calculation.

Errors you can catch. Most libcp2k functions return void and call MPI_ABORT on failure. cp2k-rs isolates CP2K, detects the abort and raises a typed exception instead of taking your process down with it.

More than energies and forces. Eigenvalues, HOMO–LUMO gaps, Mulliken and Hirshfeld charges, the electron density on the real-space grid, MO coefficients, and the AO overlap and Kohn–Sham matrices — dense or sparse, at Γ and at arbitrary k-points.

For advanced calculations, Python can also render a complete typed CP2K input tree. The package includes GauXC and SIRIUS plane-wave DFT, plus optional static backends such as DFT-D4, tblite, PLUMED, GreenX, TREXIO and libvdwxc.

Supported platforms

Linux on x86-64 (glibc), with MPI on the host. macOS and Windows are not supported: the worker links CP2K, OpenBLAS, ScaLAPACK, ELPA, SIRIUS and about twenty other libraries statically against glibc and the GCC Fortran runtime.

sudo dnf install openmpi && module load mpi/openmpi-x86_64   # Fedora / RHEL
sudo apt install openmpi-bin                                 # Debian / Ubuntu

Published worker variants: v3 (AVX2, default), v4 (AVX-512), CUDA, ROCm, and Skala for machine-learned functionals. See Installation.

Using it from Rust

use cp2k_rs::{init, finalize, ForceEnv};

init()?;
let mut force_env = ForceEnv::new("input.inp", "output.out")?;
force_env.calc_energy_force()?;
println!("{:.10} Ha", force_env.get_potential_energy()?);
finalize()?;

cp2k-rs = { version = "0.3", features = ["mpi", "extended", "build-cp2k"] }

See Using cp2k-rs from Rust.

Building from source

Only needed to develop cp2k-rs or to target an architecture with no published worker — pip install cp2k-rs is the supported path for using it.

module load mpi/openmpi-x86_64
cargo build --release --features mpi,extended,build-cp2k

The first build compiles CP2K and every dependency from source: ~8 GB of disk and 10–15 minutes. Full instructions, build-control variables and GPU builds are in Building from source.

NOTICE lists every statically linked library with its version and upstream source.

Contributing

See CONTRIBUTING.md and the test guide. Bug reports and merge requests are welcome on GitLab.

Citing

See CITATION.cff. Please cite CP2K itself as well — cp2k-rs is an interface, and the science is CP2K's.

License

GPL-2.0-or-later, the same as CP2K. Bundled dependencies keep their own licenses; see NOTICE.