thermoatoms/CuAu

Cu-Au DFT Training Data

DFT total energies and forces for ~96,000 small periodic Cu, Au, and Cu-Au structures (1–8 atoms per cell), computed with three exchange-correlation functionals. Intended as training/reference data for machine-learning interatomic potentials.

Each functional is a separate subset (config): pbe (default), lda, and r2scan. The same structure-generation protocol underlies all three; they differ in the functional used for the reference calculations.

Subsets

config	structures	total atoms	Cu-only	Au-only	Cu-Au	E/atom range (eV)
pbe	32,228	178,269	8,844	8,971	14,413	−3.73 … +6.64
lda	31,931	176,630	8,736	8,928	14,267	−4.69 … +5.43
r2scan	31,892	176,358	8,728	8,923	14,241	−4.25 … +6.10

All structures are fully periodic (pbc = TTT) with general triclinic cells. The datasets deliberately include strongly distorted and high-energy configurations for robust potential fitting.

Plane wave cut off was 550eV, kspacing 0.1/Angstrom, converged to at least 10^(-6) eV with Methfessel-Paxton smearing at 0.2eV. Calculated with VASP.

Fields

field	type	description
name	string	structure identifier encoding composition and generation family, e.g. CuEverything_Cu8VolMinAllMinRandom_structure_885
symbols	list[string]	chemical symbols, length N
numbers	list[int]	atomic numbers (29 = Cu, 79 = Au)
positions	list[list[float]]	Cartesian positions, Å, shape (N, 3)
cell	list[list[float]]	cell vectors as rows, Å, shape (3, 3)
energy	float	DFT total energy, eV
forces	list[list[float]]	atomic forces, eV/Å, shape (N, 3)
number_of_atoms	int	N (1–8)

Usage

from datasets import load_dataset

ds = load_dataset("thermoatoms/CuAu", "pbe", split="train")  # or "lda", "r2scan"
row = ds[0]

Rebuild an ASE Atoms object from a row:

from ase import Atoms

atoms = Atoms(
    numbers=row["numbers"],
    positions=row["positions"],
    cell=row["cell"],
    pbc=True,
)
energy = row["energy"]   # eV
forces = row["forces"]   # eV/Å

Structure generation

Structures were generated with the ASSYST method (Poul, Huber & Neugebauer, npj Comput. Mater. 11, 174 (2025)): symmetry-random candidate cells with up to 8 atoms sampled across space groups (841 Cu, 841 Au, 1335 Cu-Au samples), put through successive volume and full relaxations, then augmented with randomly rattled (σ = 0.5 Å) and hydrostatically/shear-strained variants of the minima for vibrational and elastic sampling. Structure families are encoded in the name field (e.g. VolMinAllMinRandom, VolMinTrace, VolMinAllMinTrace). The small cells make it efficient to relabel the same structure pool with several functionals, which is how the three subsets were obtained.

Citation

If you use this dataset, please cite:

@misc{Menon2026,
      title={Computing binary alloy phase diagrams with explicit configurational and vibrational entropy}, 
      author={Sarath Menon and Marvin Poul and Tilmann Hickel and Jörg Neugebauer and Ralf Drautz},
      year={2026},
      eprint={2607.14795},
      archivePrefix={arXiv},
      primaryClass={cond-mat.mtrl-sci},
      url={https://arxiv.org/abs/2607.14795}, 
}