Datasets

As part of LeMaterial, we released and will maintain different datasets, unifying and standardizing data from existing databases:

• LeMat-Bulk

  • Released in December 2024
  • This dataset unifies data from Materials Project, Alexandria, and OQMD into a high-quality resource with consistent and systematic properties (6,7M entries, 7 material properties)
  • https://huggingface.co/datasets/LeMaterial/LeMat-Bulk

• LeMat-BulkUnique

  • Released in December 2024
  • This dataset provides de-duplicated material from Materials Project, Alexandria, and OQMD using our structure fingerprint algorithm. It is available in 3 subsets, for PBE, PBESol, and SCAN functionals
  • https://huggingface.co/datasets/LeMaterial/LeMat-BulkUnique

• LeMat-Traj

  • Released in August 2025
  • LeMat-Traj provides a large-scale dataset, aggregating over 120 million atomic configurations of ab-initio relaxation trajectories, curated from multiple sources (MP, Alexandria, OQMD) and simulation protocols. It enables training and benchmarking of MLIPs and trajectory-aware models (e.g. force regressors, uncertainty quantifiers).
  • https://huggingface.co/datasets/LeMaterial/LeMat-Traj
  • https://arxiv.org/pdf/2508.20875

• LeMat-Synth

  • Released in September 2025
  • LeMat-Synth is a multi-modal dataset that links materials, their synthesis procedures, and performance data. It was built by parsing scientific literature using VLM/LLM pipelines, and aims to support research on synthesisability prediction and planning. By analyzing over 80,000 open-access papers, LeMat-Synth builds one of the first large-scale datasets of material synthesis recipes, covering 35 synthesis methods and 16 material classes.
  • https://huggingface.co/datasets/LeMaterial/LeMat-Synth
  • https://www.arxiv.org/pdf/2510.26824

More datasets are under development, including surfaces, defects, and electron densities, and we welcome collaborators interested in extending or improving any of the above.