Get started
Work

Atoms that Matter: Largest Materials Datasets Are 99% Redundant, and Local Chemistry Can Guide the Required 1%

Year: 2026
Type: preprint
Abstract: Data accompanying "Atoms that Matter: Largest Materials Datasets Are 99% Redundant, and Local Chemistry Can Guide the Required 1%" This record holds the data needed to reproduce the results of the paper without re-running the expensive steps of the workflow: - environment tokenisation and subset selection; - model training; - the benchmark simulations; - the latent-space feature extraction. The scripts that produced these files, and that turn them into the numbers, tables and figures of the paper, are in https://github.com/em819/omat24-chemenv-subset. Each archive below names the workflow stage it replaces and the script that reads it. Not included, because they are publicly available elsewhere: - the OMat24 dataset (Meta FAIR, CC BY 4.0); - the MatPES test set; - the MLIP-Arena and LiTraj nebDFT2k reference data; - the MACE-OMAT-0 model. subsets.zip: OMat24 structure IDs of every training subset: EnvTok, DIRECT and four random subsets. It also contains the EnvTok descriptor counts the selection is based on. The training sets can be rebuilt from OMat24 with these IDs. models.zip: the trained MACE models and the models of the data-reduction study. benchmarks_matcalc.zip: MatCalc-Bench results (equilibrium, elasticity, phonon, softening) per model, including MACE-OMAT-0, and the data-reduction models. benchmarks_matpes.zip: per-atom force errors and per-structure errors on MatPES. benchmarks_md-stability.zip: MLIP-Arena NVT heating and NPT compression trajectories (parquet). benchmarks_nebdft2k.zip: LiTraj nebDFT2k Li-migration barrier predictions. latent-space.zip: per-descriptor metrics of the MACE latent space and descriptor counts in the training set.
DOI: 10.48550/arxiv.2610.14392
Language: English
Open Access status: green
OpenAlex ID: W7221081422
Cites: 11
Cited by:
Related to: 10