LeMat-Rho

What LeMat-Rho is

LeMat-Rho is a computational dataset containing 69,117 inorganic crystal structures evaluated with density functional theory (DFT) using the r2SCAN meta-GGA (regularised-restored strongly constrained and appropriately normed meta-generalised gradient approximation) functional. Developed under the LeMaterial framework, it provides electronic charge densities, Bader partial charges, interatomic forces, stress tensors, and total energies.

The dataset unifies inorganic structures imported from three primary repositories: AFLOW (38,925 structures), the Open Quantum Materials Database (OQMD, 29,611 structures), and the Materials Project (581 structures). It supplies high-fidelity electronic structure outputs designed for training machine learning potentials and charge-density prediction models.

How it was made

Each crystal structure is processed through a standardised four-stage pipeline executed in the Vienna Ab initio Simulation Package (VASP). The workflow begins with PreStaticMaker for an initial single-point run on the unrelaxed cell, followed by two consecutive geometry relaxation stages (RelaxMaker_1 and RelaxMaker_2) using r2SCAN. The final stage, StaticMaker, runs a single-point calculation without ionic motion (NSW=0) to generate the reported energies, forces, stress tensors, and volumetric charge densities (CHGCAR, AECCAR0, AECCAR1, and AECCAR2).

Bader charges and atomic volumes are computed using the Henkelman group Bader charge analysis algorithm, using the sum of the core (AECCAR0) and valence (AECCAR2) charge densities as the reference. Charge densities are provided either as lossy, Gaussian-smoothed arrays compressed via pyrho or as unsmoothed arrays resampled directly from raw CHGCAR files using Fourier interpolation onto a fixed or adaptive spatial grid.

Key facts

What you can use it for

Charge density prediction

Train machine learning models to infer real-space electronic charge densities from crystal lattice geometry and atomic numbers.

Machine learning interatomic potential development

Parameterise and benchmark force fields and neural network potentials using uniform r2SCAN total energies, forces, and stress tensors.

Atomic charge and oxidation analysis

Investigate electron transfer, oxidation states, and local bonding across diverse crystal chemistries using Bader net charges and volumes.

Explore the data

The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.

Strengths and limits

Questions

How to cite it

Curtarolo et al., Computational Materials Science 58, 218–226 (2012); Saal et al., JOM 65, 1501–1509 (2013); Kirklin et al., npj Computational Materials 1, 15010 (2015); Jain et al., APL Materials 1, 011002 (2013); Henkelman et al., Computational Materials Science 36, 354–360 (2006); Furness et al., J. Phys. Chem. Lett. 11, 8208–8215 (2020); Shen, J.-X. et al., Sci Data, 9(661), 1–7 (2022), doi: 10.1038/s41597-022-01746-z