LeMat-Rho: High-Fidelity Charge Density Dataset for Atomistic Materials Modeling

Mathilde L. D. Franckel¹², Richard Tran², Martin Siron², Daniel Speckhard², Georgia Channing³, Guilherme  Penedo³, Ali Ramlaoui², Alexandre Duval², Jonathan Schmidt⁴

¹Department of Materials, Imperial College, London, UK

²Entalpic, Paris, FR

³Hugging Face

⁴Department of Materials, EPFL, Lausanne CH

Ab initio datasets have significantly advanced atomistic modeling by enabling the development of accurate machine learning interatomic potentials. Atom-centered models predict atomic energies and forces with high precision, as these values are derived from charge densities calculated using the Kohn-Sham equations. Access to extensive, high-accuracy charge density datasets provides additional physical insights for atomistic machine-learning approaches.

As part of the LeMaterial Open Science initiative, this is the initial release of LeMat-Rho, an r²SCAN charge-density dataset encompassing 70,000 materials. This dataset provides the research community with high-quality quantum-mechanical data describing electron distributions in materials. The r²SCAN meta-GGA exchange-correlation functional provides a more accurate representation of bonding and energetics than generalized-gradient approximations such as PBE, as illustrated in the Figure below (Fig. 1).

The workflow automates high-throughput density functional theory (DFT) calculations, beginning with input crystal structures from LeMat-Bulk and proceeding through a configurable workflow builder to the VASP ab initio code. The pipeline achieves robust convergence and high-fidelity electronic structure data using a staged PBE-to-r²SCAN workflow. This process starts with a PBE static pre-calculation that aims to produce reasonable initial wavefunctions at a low computational cost. This is followed by two consecutive r²SCAN relaxations and a final static evaluation. Indeed, as the cell volume equilibrates, the basis set must be reset to ensure optimal convergence.

The approach produces converged electronic-structure outputs, including the charge density, wavefunction, energies, and forces. All simulation results are systematically curated and stored in AWS Open Data. The dataset is now publicly available on Hugging Face, serving as an open resource for materials discovery, electronic structure analysis, and machine learning research.

The charge densities are fed into a Bader analysis code to provide information on the charge localization and chemical bonding. Subsequently, we have formatted the dataset for ease of use and published it publicly on HuggingFace.

To maximize computational resources, we used the high-throughput tools Atomate and JobFlow. The HuggingFace dataset is populated using an array of tools from the pymatgen Python library.

The resulting dataset is enriched in lithium- and cobalt-containing compounds because the initial crystal structures were sourced primarily from repositories developed for battery materials exploration, as shown in Figure 3. To maximize computational throughput and maintain consistency across calculations, the workflow preferentially selected structures with small primitive cells, thereby reducing the representation of larger, more compositionally complex oxide systems.

The workflow employed strict convergence criteria to ensure high-quality DFT results. Electronic self-consistent field (SCF) calculations were converged to an energy tolerance below 1 × 10⁻⁶ eV, while ionic relaxations were considered converged when the maximum residual force magnitude fell below 0.02 eV/Å. Overall workflow robustness was high. Across the completed dataset, 98% of structures achieved residual forces below 0.0182 eV/Å and residual stresses below 4.19 kbar, indicating strong structural and mechanical convergence as shown in Figure 4.

To facilitate machine-learning applications and large-scale data distribution, volumetric charge densities were transformed into a standardized 15×15×15 representation using the mp-rho framework. This compression preserves the global spatial distribution of electron density while reducing storage requirements sufficiently for Parquet-based datasets and efficient access through Hugging Face.

To enable broader analysis of the compounds' chemical properties, we provide Bader charges and Bader atomic volumes for each structure. The goal is to quantify the redistribution of electron density among atoms, serving as a measure of charge transfer and bonding characteristics. The accompanying figure (Fig.5) shows the distribution of Bader charges for three representative elements, highlighting the diversity of chemical bonding in the dataset. These distributions indicate how atomic charge states vary between compounds and may serve as valuable descriptors for understanding structure–property relationships.

This dataset of charge densities provides a compact and scalable representation of electronic structure information across a wide range of materials. We hope it will support the development of new machine learning models capable of learning complex material properties directly from electron density data, while also enabling faster initialization of DFT calculations.

By releasing this dataset, we hope to contribute to current accelerations in computational materials science.

Download and use within Python

from datasets import load_dataset

dataset = load_dataset('LeMaterial/LeMat-Rho')

# convert to Pandas, if you prefer working with this type of object:
df = dataset['train'].to_pandas()

Data fields

Notes

elements_ratios

Unlike the original schema, this dataset stores elemental fractions as a sequence rather than a dictionary. The values correspond to the ordering in the elements field.

Example:

{
  "elements": ["Be", "N", "Tb"],
  "elements_ratios": [0.25, 0.50, 0.25]
}

Charge Density Fields

The fields compressed_charge_density, compressed_aeccar0, compressed_aeccar1, and compressed_aeccar2 store compressed volumetric electronic density data. These fields can be used to reconstruct charge density distributions and related electronic structure properties.

To support different machine-learning and data-analysis workflows, LeMat-Rho is distributed in two compressed charge-density formats on Hugging Face.

Fixed Grid Dataset

The default LeMat-Rho dataset stores charge densities on a standardized 15×15×15 grid generated with pyrho for each material. This representation provides a uniform tensor shape across the entire dataset, simplifying batching, storage, and training of machine-learning models.

Adaptive Grid Dataset

For applications where preserving a consistent spatial resolution is more important than maintaining a fixed tensor shape, we additionally provide an adaptive-grid version of the dataset. This representation better preserves the spatial features of the original charge density and is often preferable for physical analysis

In this representation, the number of grid points along each lattice direction is determined from the corresponding lattice-vector length:

grid_points[i] = max(5, ceil(lattice_length[i] / 0.2))

This produces approximately one grid point every 0.2 Å along each crystallographic direction.

Examples:

5 Å lattice vector → 25 grid points 10 Å lattice vector → 50 grid points

Atomic Charge Analyses

The bader_charges, bader_atomic_volume, and ddec6_charges fields contain atom-resolved properties. The ordering of values matches the ordering used in:

• species_at_sites
• cartesian_site_positions
• forces
• magnetic_moments

S3 Bucket Structure and Access

The complete LeMat-Rho workflow outputs are publicly available through the AWS S3 bucket:

s3://lemat-rho

The bucket contains one directory per material, identified by its source database ID (e.g., Alexandria, Materials Project, or OQMD):

lemat-rho/
├── agm001987721/
│   ├── LeMatRhoPreStaticMaker/
│   │   ├── OUTCAR.gz
│   │   └── vasprun.xml.gz
│   ├── LeMatRhoRelaxMaker_1/
│   │   ├── OUTCAR.gz
│   │   └── vasprun.xml.gz
│   ├── LeMatRhoRelaxMaker_2/
│   │   ├── OUTCAR.gz
│   │   └── vasprun.xml.gz
│   └── LeMatRhoStaticMaker/
│       ├── CHG.gz
│       ├── AECCAR0.gz
│       ├── AECCAR1.gz
│       ├── AECCAR2.gz
│       ├── OUTCAR.gz
│       └── vasprun.xml.gz
├── mp-1234/
└── oqmd-5678/

The final charge-density outputs are located in the LeMatRhoStaticMaker directory:

Accessing Files

Individual files can be downloaded directly using the AWS CLI:

aws s3 cp s3://lemat-rho/agm001987721/LeMatRhoStaticMaker/CHG.gz .

or via HTTPS:

https://lemat-rho.s3.amazonaws.com/agm001987721/LeMatRhoStaticMaker/CHG.gz

The Hugging Face dataset contains compressed and processed representations of these charge densities, while the S3 bucket provides access to the original VASP outputs and volumetric charge-density files used to generate the dataset.

Software and Infrastructure

LeMat-Rho was generated using an automated high-throughput computational workflow built on the open-source LeMaterial ecosystem. Workflow execution and job management were orchestrated using FireWorks, Atomate, and JobFlow, enabling robust scheduling, monitoring, and recovery of large-scale density functional theory calculations.

Materials representations, structure analysis, and post-processing operations were performed using pymatgen. To facilitate machine-learning applications and large-scale data distribution, volumetric charge densities were compressed into a standardized 15×15×15 representation using the pyrho package.

All calculation outputs were curated and stored through the AWS Open Data infrastructure and subsequently distributed through the Hugging Face Datasets platform, providing efficient public access to the resulting dataset.

Citation Information

We are currently in the process of creating a pre-print to describe our methods, the materials fingerprint method and the dataset. For now however the following can be cited:

@misc {LeMat-Rho_2026,
    author = {{Mathilde L. D. Franckel}, {Richard Tran}, {Martin Siron}, {Daniel Speckhard}, {Georgia Channing}, {Guilherme  Penedo}, {Ali Ramlaoui}, {Alexandre Duval}, {Jonathan Schmidt}},
    title = {LeMat-Rho: High-Fidelity Charge Density Dataset for Machine Learning and Atomistic Materials Modelling},
    year = 2026,
    url = { https://huggingface.co/datasets/LeMaterial/LeMat-Rho},
    doi = {},
    publisher = { Hugging Face }
}

Acknowledgements:

• LeMaterial is the Open Science initiative hosted by Entalpic
• This project was supported by HuggingFace
• The dataset was made possible by the AWS Open Data program
• Original LeMat-Bulk structures were selected from Alexandria, MaterialsProject, and OQMD

License

This database is licensed by Creative Commons Attribution 4.0 License.