LeMat-Rho

LeMat-Rho is a dataset of ~69,000 inorganic crystal structures computed with the r2SCAN meta-GGA functional in VASP, featuring charge densities, Bader charges, forces, stresses, and energies. It is part of the LeMaterial ecosystem.

Dataset Summary

Property	Value
Materials	69,117
Functional	r2SCAN
Sources	AFLOW (agm), OQMD (oqmd-), Materials Project (mp-)
Bader charge coverage	86.2% (59,613 materials)
Charge density grid	15×15×15 (lossy compressed via pyrho)
DDEC6 charges	Planned for future release

Sources

Source	Materials
AFLOW (agm*)	38,925
OQMD (oqmd-*)	29,611
Materials Project (mp-*)	581

Workflow

Each material was computed with a 4-step VASP pipeline:

1. PreStaticMaker — initial single-point on unrelaxed geometry
2. RelaxMaker_1 — geometry relaxation (r2SCAN)
3. RelaxMaker_2 — second relaxation pass (r2SCAN)
4. StaticMaker — final NSW=0 single-point; source of all reported properties

Forces, stresses, and energies are from the final static calculation. Charge densities (CHGCAR, AECCAR0, AECCAR1, AECCAR2) are from the StaticMaker output.

Bader charges were computed using the Bader charge analysis code (Henkelman group) with AECCAR0+AECCAR2 as the reference charge density.

Columns

Column	Type	Unit	Description
immutable_id	string	—	Unique material identifier (e.g. agm001234567)
elements	list[string]	—	Element symbols present
nelements	int	—	Number of distinct elements
nsites	int	—	Number of atoms in the unit cell
chemical_formula_reduced	string	—	Reduced formula (e.g. Fe2O3)
chemical_formula_anonymous	string	—	Anonymous formula (e.g. A2B3)
chemical_formula_descriptive	string	—	Full descriptive formula
lattice_vectors	list[list[float]]	Å	3×3 lattice matrix
cartesian_site_positions	list[list[float]]	Å	Atomic positions
species_at_sites	list[string]	—	Element at each site
functional	string	—	DFT functional (r2scan)
energy	float	eV	Total DFT energy (final static)
energy_corrected	float	eV	Energy with MP2020 corrections (equals energy for r2SCAN)
forces	list[list[float]]	eV/Å	Per-atom forces from final static
stress_tensor	list[list[float]]	kBar	3×3 stress tensor from final static
bader_charges	list[float]	e	Bader net charges per atom (positive = cationic); null if AECCAR files absent
bader_atomic_volume	list[float]	ų	Bader atomic volumes per atom
ddec6_charges	list[float]	e	DDEC6 net charges per atom (null in current release; planned)
compressed_charge_density	string	—	Lossy-compressed CHGCAR (15×15×15 grid, JSON-encoded via pyrho)
compressed_aeccar0	string	—	Lossy-compressed AECCAR0 (core charge density)
compressed_aeccar1	string	—	Lossy-compressed AECCAR1
compressed_aeccar2	string	—	Lossy-compressed AECCAR2 (valence charge density)
charge_density_grid_shape	list[int]	—	Grid dimensions [nx, ny, nz]
space_group_it_number	int	—	International Tables space group number
bawl_fingerprint	string	—	Structure fingerprint (BAWL hasher)
magnetic_moments	list[float]	μ_B	Per-atom magnetic moments (null if not spin-polarized)
total_magnetization	float	μ_B	Total cell magnetization
cross_compatibility	bool	—	Cross-compatibility flag with other LeMaterial datasets

Usage

from datasets import load_dataset

ds = load_dataset("LeMaterial/LeMat-Rho", split="train")

# Filter to materials with Bader charges
with_bader = ds.filter(lambda x: x["bader_charges"] is not None)
print(f"Materials with Bader charges: {len(with_bader)}")

# Access a material
row = ds[0]
print(row["chemical_formula_reduced"], row["bader_charges"])

Notes

• Charge density compression: The stored charge densities are lossy-compressed to a 15×15×15 grid using pyrho. This is suitable for qualitative analysis and ML feature extraction but not for quantitative integration (e.g. re-running Bader). The raw CHGCAR files are available on request.
• Bader coverage: ~13.8% of materials lack Bader charges because their AECCAR files are absent from the source S3 data. This is consistent across all three source databases and reflects an upstream calculation issue.
• DDEC6 charges: ddec6_charges is null in this release. A follow-up patch run is planned.
• r2SCAN functional: All calculations use the r2SCAN meta-GGA functional. MP2020 energy corrections are PBE-specific and do not apply; energy_corrected equals energy for all rows.

Citation

If you use this dataset, please cite the underlying databases:

• AFLOW: Curtarolo et al., Computational Materials Science 58, 218–226 (2012)
• OQMD: Saal et al., JOM 65, 1501–1509 (2013); Kirklin et al., npj Computational Materials 1, 15010 (2015)
• Materials Project: Jain et al., APL Materials 1, 011002 (2013)
• Bader analysis: Henkelman et al., Computational Materials Science 36, 354–360 (2006)
• r2SCAN functional: Furness et al., J. Phys. Chem. Lett. 11, 8208–8215 (2020)