LeMaterial/LeMat-Bulk

Motivation: check out the blog post https://huggingface.co/blog/lematerial to hear more about the motivation behind the creation of this dataset.

Download and use within Python

from datasets import load_dataset

dataset = load_dataset('LeMaterial/LeMat-Bulk', 'compatible_pbe')

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

Data fields

Feature name	Data type	Description	Optimade required field
elements	Sequence[String]	A list of elements in the structure. For example a structure with composition Li2O7 will have [”Li”,”O”] in its elements. Notes: Currently not necessarily sorted but future iteration will be sorted by alphabetic order.	✅
nsites	Integer	The total number of sites in the structure. For example a structure with an un-reduced composition of Li4O2 will have a total of 6 sites.	✅
chemical_formula_anonymous	String	Anonymous formula for a chemical structure, sorted by largest contributing species, and reduced by greatest common divisor. For example a structure with a O2Li4 un-reduced composition will have a anonymous formula of A2B. “1”’s at the end of an element composition are dropped (ie not A2B1)	✅
chemical_formula_reduced	String	Reduced by the greatest common divisor chemical composition. For example a structure with a un-reduced composition of O2Li4 will have a reduced composition of Li2O. Elements with a reduced composition of 1 have the “1” dropped. Elements are sorted by alphabetic ordering. Notes: Not using the same method of Pymatgen’s composition reduction method which takes into account certain elements existing in diatomic states.	✅
chemical_formula_descriptive	String	A more descriptive chemical formula for the structure, for example a fictive structure of a 6-fold hydrated Na ion might have a descriptive chemical formula of Na(H2O)6, or a Titanium chloride organic dimer might have a descriptive formula of [(C5H5)2TiCl]2. Note: this field is absolutely not standardized across the database. Where possible if available we scrapped as is from the respective databases. Where not possible this may be the same as the chemical formula reduced.	✅ Note: not standardized in naming approach.
nelements	Integer	Total number of different elements in a structure. For example Li4O2 has only 2 separate elements.	✅
dimension_types	Sequence[Integer], shape = 3x1	Periodic boundary conditions for a given structure. Because all of our materials are bulk materials for this database it is [1, 1, 1], meaning it is periodic in x, y, and z dimensions.	✅
nperiodic_dimensions	Integer	The number of repeating periodic boundary conditions, because all our structures in this database are bulk structures, they are repeating in x, y, and z dimensions and thus they have 3 periodic dimensions.	✅
lattice_vectors	Sequence[Sequence[Floats]], shape = 3x3	The matrix of the structures. For example a cubic system with a lattice a=4.5 will have a [[4.5,0,0],[0,4.5,0],[0,0,4.5]] lattice vector entry.	✅
immutable_id	String	The material ID associated with the structure from the respective database. Note: OQMD IDs are simply integers, thus we converted them to be “oqmd-YYY”	✅
cartesian_site_positions	Sequence[Sequence[Floats]], shape = Nx3	In cartesian units (not fractional units) the coordinates of the species. These match the ordering of all site based properties such as species_at_sites, magneitc_moments and forces. For example a material with a single element placed at a fractional coordinate of [0.5, 0.5, 0.5] with a cubic lattice with a=2, will have a cartesian_site_positions of [1, 1, 1].	✅
species	JSON	An Optimade field that includes information about the species themselves, such as their mass, their name, their labels, etc. Note: we have not currently filled out the mass portion of the species. Additionally, none of our inputted structures should be solid solution thus the on-site concentration for all our species should be [1]. This is an Optimade field.	✅
species_at_sites	Sequence[String]	An array of the chemical elements belonging to each site, for example a structure with an un-reduced composition of Li2O2 may have an entry of [”Li”, “Li”, “O”, “O”] for this field, where each species should match the other site based properties such as cartesian_site_positions.	✅
last_modified	Date/time	The date that the entry was last modified from the respective database it was pulled from. Note: we could not find this information in OQMD so we used the date of the latest database release as the input for this field.	✅
elements_ratios	Dictionary	The fractional composition for a given structure in dictionary format. For example a structure with an unreduced composition of Li2O4 would have an entry of {’Li’:0.3333, ‘O’:0.6667}	✅
stress_tensor	Sequence[Sequence[Float]], shape = 3x3	The full 3x3 vector for stress tensor in units of kB. Note: for OQMD stress tensor were given in Voigt notation, and were converted to the full tensor.
energy	Float	The uncorrected energy from VASP in eV.
magnetic_moments	Sequence[Floats]	The magnetic moment per site given in µB.
forces	Sequence[Sequence[Floats]], shape = 3xN	The force per site, in the proper order of the sites based on other site specific fields for each site in the x, y and z directions, given in eV/A.
total_magnetization	Float	The total magnetization of the structure in µB. Note: the sum of the magnetic moments is not always the total magnetization of the structure reported.
functional	String, either ‘pbe’, ‘pbesol’ or ‘scan’	What functional was used to calculate the data point in the row.
cross_compatibility	Boolean	Whether or not this data can be mixed with other rows from a DFT calculation parameter perspective. More information on our approach below.
entalpic_fingerprint	String	Results of initial version of materials fingerprint function as described in [blogpost]. Code release to come soon

Available subsets

To better support the diverse communities that may utilize this dataset, we are providing the following subsets of our database:

• Compatible, PBE (default): This subset includes rows filtered to ensure cross-compatibility from a DFT perspective. For details on the filtering methodology, see the section below. Only PBE records are included. We designate this as the default subset to prevent accidental training of models on non-compatible data.
• Compatible, PBESol: Similar to the Compatible, PBE subset, but includes only PBESol data.
• Compatible, SCAN: Similar to the Compatible, PBE subset, but includes only SCAN data.
• All: This includes all records formatted as described above. Disclaimer: Researchers must carefully evaluate the suitability of individual rows for their specific applications.

Database	Number of materials	Number of structures*
Materials Project	148,453	189,403
Alexandria	4,635,066	5,459,260
OQMD	1,076,926	1,076,926
LeMaterial (All)	5,860,446	6,725,590
LeMaterial (Compatible, PBE)	5,335,299	5,335,299
LeMaterial (Compatible, PBESOL)	447,824	447,824
LeMaterial (Compatible, SCAN)	422,840	422,840

*Number of structures: only includes the output of resulting calculations from either a structure optimization for any available functional. For MP we do not consider all of their structures from the relaxation trajectory for instance, nor from tasks that are not structure optimization. For OQMD we only consider the output of structure relaxation as well, not accounting for any other calculations they performed.

Method for compatibility compliance

To ensure compatibility of rows from a DFT perspective, we implemented the following compatibility scheme:

• Pseudopotentials: Calculations were verified to use consistent pseudopotentials. Notably, most pseudopotentials were aligned between MP and Alexandria, except for vanadium (where Alexandria used V_sv and MP used V_pv) and cesium (where Alexandria used a later version of the generic pseudopotential). For OQMD, this resulted in incompatibilities across records involving the following elements: Ca, Ti, V, Cr, Mn, Ru, Rh, Ce, Eu, Yb. We note that at the time of this release Materials Project deprecated all Yb containing materials due to the use of a pseudopotential that led to different than expected results. Thus no Yb containing materials from MP are in our database.
• Hubbard U Parameters: To ensure uniformity in Hubbard U parameters, we excluded records containing oxygen (O) and any of the following elements: V, Cr, Mn, Fe, Ni, Cu, Th, U, Np, Pu, Mo, W. Similarly, records containing fluorine (F) and any of the following elements: Co, Cr, Fe, Mn, Mo, Ni, V, W were also excluded. This exclusion applied specifically to OQMD, which used different Hubbard U parameters compared to MP and Alexandria. However, records from OQMD containing O and Co were retained, as their Hubbard U parameter differed by only 0.02 eV.
• Spin Polarization: OQMD only considered spin-polarized calculations for structures with d or f electrons. While non-spin-polarized calculations are not inherently incompatible (as they represent higher-energy magnetic phases compared to the ground state), we decided to exclude non-spin-polarized calculations for this release. This led to the removal of structures containing only the following elements: H, Li, Be, Na, Mg, K, Ca, Rb, Sr, Cs, Ba, Fr, Ra, B, C, N, O, F, Ne, He, Al, Si, P, S, Cl, Ar, Ga, Ge, As, Se, Br, Kr, In, Sn, Sb, Te, I, Xe, Tl, Pb, Bi, Po, At, Rn from OQMD.
• Convergence Criteria: OQMD typically used a larger plane-wave cutoff but a less dense k-point grid. Despite these differences, we did not exclude records based on these parameters, assuming that OQMD, Alexandria, and MP operated within acceptable convergence zones for energy calculations. A similar approach was applied to other VASP parameters, though we welcome feedback on this assumption.
• Convergence: Across all databases, we identified numerous records with potentially non-convergent calculations or high-energy configurations, often evidenced by significant atomistic forces. We chose not to exclude these records, as users can filter them easily using the “forces” tag if needed.
• Energy Above the Hull: We opted not to filter materials with high energy above the hull, given the current scope of the dataset.

The “all” split does not contain any filtering based on this approach, so all records can be downloaded.

De-duplication method and our materials fingerprint

For our methods for finding duplicates across databases we creating a hasher function which works the following way:

• We compute bonds using the EconNN algorithm already built in Pymatgen
• We create a structure graph from this, encoding the species in the node
• We hash this graph using Weisfeller-Lehman algorithm
• We add symmetry and composition

Any structure which has a duplicate based on this method is dropped, only keeping the lowest energy structure. We benchmarked this to be robust to small gaussian noise on atomic positions, lattice vectors, and to respect detected symmetries in a structure. In searching for this method we tried to select one of the more sensitive bonding algorithms that would leave to the least amount of duplicates. We plan on releasing more information on this, as well as code to properly benchmark other fingerprint methods soon.

Check out these helpful spaces to understand the database

Materials Explorer	Phase Diagram
Let's you browse entries in our database, view the crystal structure and its associated properties. Built using Materials Project's dash components	Lets you generate binary and ternary phase diagram using various correction scheme. Built using the Phase Diagram Crystal Toolkit component by the Materials Project Disclaimer: the MP2020 correction scheme has not yet been uniformed across datasets, when using this correction scheme please be cautious about interpreting data. We will fix this in upcoming release!

Stay tuned for future updates

We plan to release very soon:

• Band gap information on all materials, including direct and indirect band gaps.
• Unification of energy corrections (currently a beta version of this is available for the purpose of the phase diagram application, but please see the disclaimer above).
• Bader charges for all Materials Project materials where possible and the addition of charge data from Alexandria and OQMD
• R2SCAN data from Materials Project

In the longer run we plan to release additional datasets including trajectories and surface, adsorbates, and molecules.

And more! Stay tuned.

Support

If you run into any issues regarding feel free to post your questions or comments on any of the following platforms:

• HF Discussions
• Github Issues

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 {lematerial_2024,
    author = { {Martin Siron}, {Inel Djafar}, {Lucile Ritchie}, {Etienne Du-Fayet}, {Amandine Rosselo}, {Ali Ramlaoui}, {Leandro von Werra}, {Thomas Wolf}, {Alexandre Duval} },
    title = { LeMat-Bulk Dataset },
    year = 2024,
    url = { https://huggingface.co/datasets/LeMaterial/LeMat-Bulk },
    doi = { 10.57967/hf/3762 },
    publisher = { Hugging Face }
}

CC-BY-4.0 (license used for Materials Project, Alexandria, OQMD) requires proper acknowledgement.

Thus, if you use materials data which include (”mp-”) in the immutable_id, please cite the Materials Project.

If you use materials data which include (”agm-”) in the immutable_id, please cite Alexandria, PBE or Alexandria PBESol, SCAN.

If you use materials data which include (”oqmd-”) in the immutable_id, please cite OQMD.

If you make use of Optimade in your research, please cite Optimade

Finally, if you make use of the Phase Diagram for visualization purposes, or the crystal viewer in the Materials Explorer, please acknowledge Crystal Toolkit.

License

This database is licensed by Creative Commons Attribution 4.0 License.

Disclaimer: it is made up of Alexandria, Materials Project and OQMD materials, which are all licensed by Creative Commons Attribution 4.0 License.