Materials Science · LeMaterial
LeMat-Traj
A large-scale dataset of bulk material molecular dynamics and relaxation trajectories for machine learning potentials.
What LeMat-Traj is
LeMat-Traj is a multi-database repository of bulk materials trajectories created to facilitate the training of machine learning interatomic potentials. Developed by LeMaterial, it pools and standardises structural relaxation and molecular dynamics trajectories from the Materials Project, Alexandria, and the Open Quantum Materials Database (OQMD).
A frequent challenge in materials informatics is that individual databases calculate quantum-mechanical properties using varying density functional theory (DFT) settings, making cross-source training unreliable. LeMat-Traj directly targets this issue by systematically curating subsets based on DFT cross-compatibility, checking parameters such as pseudopotentials, Hubbard U corrections, and spin polarization.
The dataset spans over 100 million structural frames containing energies, site forces, and stress tensors aligned with the OPTIMADE standard. It provides a standardized data substrate for physical scientists, chemists, and computational researchers designing universal crystal potential models.
How it was made
Trajectories were pooled across the Materials Project, Alexandria, and OQMD, then standardised into OPTIMADE-compliant schema. Frames were evaluated and excluded if self-consistent field (SCF) calculations failed to converge, if energy or force fields were missing, or if a frame deviated by more than 1 eV/atom from the final state of its trajectory. Trajectories were not downsampled by length or structural similarity.
To enable cross-database learning, a compatibility evaluation was performed across electronic structure parameters. Calculations were aligned according to pseudopotential choices, leading to the removal of incompatible elements (such as Ca, Ti, V, Cr, Mn, Ru, Rh, Ce, Eu, and Yb from OQMD, alongside deprecated Yb materials from the Materials Project). Mismatched Hubbard U parameters led to the exclusion of specific transition metal oxide and fluoride systems from OQMD. Non-spin-polarized entries containing solely sp-elements were also filtered out from OQMD to ensure physical consistency across the magnetic states.
Key facts
- Data format
- Parquet
- Size category
- 100M to 1B entries
- Supported functionals
- PBE, PBESol, SCAN, r2SCAN
- Source databases
- Materials Project, Alexandria, OQMD
- Periodic dimensions
- 3 (bulk systems only)
- Data schema
- OPTIMADE compliant
What you can use it for
- Machine Learning Interatomic Potential Training
- Train neural network potentials and foundational interatomic models using forces, stress tensors, and total energies calculated across periodic crystal systems.
- Relaxation and Dynamics Benchmarking
- Evaluate model capabilities across initial-to-relaxed trajectories, double relaxation calculations, and multi-step optimization runs.
- Multi-Functional Energy Modeling
- Compare structural energies and stress responses between different exchange-correlation functionals including PBE, PBESol, SCAN, and r2SCAN.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from LeMaterial/LeMat-Traj at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 4,566,000 rows of compatible_pbe/train
cartesian_site_positions
list- min
- 1
- median
- 14
- mean
- 19.24
- max
- 444
chemical_formula_anonymous
string_text- min
- 1
- median
- 6
- mean
- 5.95
- max
- 18
chemical_formula_descriptive
string_text- min
- 1
- median
- 12
- mean
- 12.61
- max
- 34
chemical_formula_reduced
string_text- min
- 1
- median
- 9
- mean
- 8.52
- max
- 23
cross_compatibility
bool- True100%
dimension_types
list- min
- 3
- median
- 3
- mean
- 3
- max
- 3
elements
list- min
- 1
- median
- 4
- mean
- 3.77
- max
- 9
elements_ratios
list- min
- 1
- median
- 4
- mean
- 3.77
- max
- 9
energy
float- min
- -2420.44
- median
- -76.93
- mean
- -111.08
- max
- 2.57
energy_corrected
float- min
- -2475.88
- median
- -81.38
- mean
- -118.07
- max
- 2.57
3.4% empty
forces
list- min
- 1
- median
- 14
- mean
- 19.24
- max
- 444
functional
string_label- pbe100%
id
string_text- min
- 13
- median
- 19
- mean
- 18.87
- max
- 21
immutable_id
string_text- min
- 5
- median
- 12
- mean
- 11.54
- max
- 12
last_modified
datetimelattice_vectors
list- min
- 3
- median
- 3
- mean
- 3
- max
- 3
nelements
int- min
- 1
- median
- 4
- mean
- 3.77
- max
- 9
nperiodic_dimensions
int- min
- 3
- median
- 3
- mean
- 3
- max
- 3
nsites
int- min
- 1
- median
- 14
- mean
- 19.24
- max
- 444
relaxation_number
int- min
- 0
- median
- 0
- mean
- 0.46
- max
- 2
relaxation_step
int- min
- 0
- median
- 24
- mean
- 60.10
- max
- 741
space_group_it_number
int- min
- 1
- median
- 14
- mean
- 44.49
- max
- 230
species
string_text- min
- 138
- median
- 558
- mean
- 525.88
- max
- 1,254
species_at_sites
list- min
- 1
- median
- 14
- mean
- 19.24
- max
- 444
stress_tensor
list- min
- 3
- median
- 3
- mean
- 3
- max
- 3
Strengths and limits
Good at
- Filters records explicitly by DFT calculation parameter cross-compatibility, preventing model corruption from mismatched pseudopotentials and Hubbard U values.
- Preserves full trajectory pathways without artificial downsampling for structural diversity or length.
- Standardised to the OPTIMADE specification with tabular Parquet formatting compatible with modern data loaders like Dask and Polars.
Watch out for
- Restricted strictly to bulk structures with periodic boundary conditions in all three dimensions, lacking surfaces, interfaces, and isolated molecules.
- High-energy configurations and potentially non-convergent force values are retained in the compatible splits and must be manually filtered by users if not desired.
- Species mass information is not currently filled in the species metadata field.
Questions
Can I look inside LeMat-Traj without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is LeMat-Traj under?
Its authors declare cc-by-4.0. Read the terms on the original page before reusing the data.
What is the difference between the compatible subsets and the 'all' subset?
The compatible subsets (such as compatible_pbe) exclude configurations that conflict in exchange-correlation setup, pseudopotentials, Hubbard U corrections, or spin polarization. The 'all' subset contains every trajectory regardless of parameter consistency, requiring users to inspect and filter data manually.
How are structural boundary conditions defined?
All structures are bulk materials having periodic boundary conditions along all three dimensions (nperiodic_dimensions=3 with dimension_types=[1, 1, 1]).
Are site positions provided in fractional or Cartesian coordinates?
Site positions are stored as Cartesian coordinates in the cartesian_site_positions field, matching the exact ordering of atomic forces and magnetic moments.
How to cite it
@misc {lematerial_2025_traj, author = { {Martin Siron}, {Inel Djafar}, {Ali Ramlaoui}, {Lucile Ritchie}, {Etienne Du-Fayet}, {Amandine Rossello}, {Leandro von Werra}, {Thomas Wolf}, {Alexandre Duval} }, title = { LeMat-Traj Dataset }, year = 2024, url = { https://huggingface.co/datasets/LeMaterial/LeMat-Traj }, doi = { }, publisher = { Hugging Face } }
Written from LeMat-Traj's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.