Meta Open Materials 2024 (OMat24) Dataset

Overview

Several datasets were utilized in this work. We provide open access to all datasets used to help accelerate research in the community. This includes the OMat24 dataset as well as our modified sAlex dataset. Details on the different datasets are provided below.

The OMat24 datasets can be used with the FAIRChem package. See section on "How to read the data" below for a minimal example.

Datasets

OMat24 Dataset

The OMat24 dataset contains a mix of single point calculations of non-equilibrium structures and structural relaxations. The dataset contains structures labeled with total energy (eV), forces (eV/A) and stress (eV/A^3). The dataset is provided in ASE DB compatible lmdb files.

We provide two splits - train and validation. Each split is comprised of several subdatasets based on the different input generation strategies, see paper for more details.

The OMat24 train and validation splits are fully compatible with the Matbench Discovery benchmark test set.

1. The splits do not contain any structure that has a protostructure label present in the initial or relaxed structures of the WBM dataset.
2. The splits do not include any structure that was generated starting from an Alexandria relaxed structure with protostructure lable in the intitial or relaxed structures of the WBM datset.

Train

Sub-dataset	Size	Download
rattled-1000	11,388,510	rattled-1000.tar.gz
rattled-1000-subsampled	3,879,741	rattled-1000-subsampled.tar.gz
rattled-500	6,922,197	rattled-500.tar.gz
rattled-500-subsampled	3,975,416	rattled-500-subsampled.tar.gz
rattled-300	6,319,139	rattled-300.tar.gz
rattled-300-subsampled	3,464,007	rattled-300-subsampled.tar.gz
aimd-from-PBE-1000-npt	21,269,486	aimd-from-PBE-1000-npt.tar.gz
aimd-from-PBE-1000-nvt	20,256,650	aimd-from-PBE-1000-nvt.tar.gz
aimd-from-PBE-3000-npt	6,076,290	aimd-from-PBE-3000-npt.tar.gz
aimd-from-PBE-3000-nvt	7,839,846	aimd-from-PBE-3000-nvt.tar.gz
rattled-relax	9,433,303	rattled-relax.tar.gz
Total	100,824,585	-

Validation

Models were evaluated on a ~1M subset for training efficiency. We provide that set below.

Sub-dataset	Size	Download
rattled-1000	122,937	rattled-1000.tar.gz
rattled-1000-subsampled	41,786	rattled-1000-subsampled.tar.gz
rattled-500	75,167	rattled-500.tar.gz
rattled-500-subsampled	43,068	rattled-500-subsampled.tar.gz
rattled-300	68,593	rattled-300.tar.gz
rattled-300-subsampled	37,393	rattled-300-subsampled.tar.gz
aimd-from-PBE-1000-npt	223,574	aimd-from-PBE-1000-npt.tar.gz
aimd-from-PBE-1000-nvt	215,589	aimd-from-PBE-1000-nvt.tar.gz
aimd-from-PBE-3000-npt	65,244	aimd-from-PBE-3000-npt.tar.gz
aimd-from-PBE-3000-nvt	84,063	aimd-from-PBE-3000-nvt.tar.gz
rattled-relax	99,968	rattled-relax.tar.gz
Total	1,077,382	-

sAlex Dataset

We also provide the sAlex dataset used for fine-tuning of our OMat models. sAlex is a subsampled, Matbench-Discovery compliant, version of the original Alexandria. sAlex was created by removing structures matched in WBM and only sampling structure along a trajectory with an energy difference greater than 10 meV/atom. For full details, please see the manuscript.

Dataset	Split	Size	Download
sAlex	train	10,447,765	train.tar.gz
sAlex	val	553,218	val.tar.gz

How to read the data

The OMat24 and sAlex datasets can be accessed with the fairchem library. This package can be installed with:

pip install fairchem-core

Dataset files are written as AseLMDBDatabase objects which are an implementation of an ASE Database, in LMDB format. A single **.aselmdb* file can be read and queried like any other ASE DB (not recommended as there are many files!).

You can also read many DB files at once and access atoms objects using the AseDBDataset class.

For example to read the rattled-relax subdataset,

from fairchem.core.datasets import AseDBDataset

dataset_path = "/path/to/omat24/train/rattled-relax"
config_kwargs = {} # see tutorial on additional configuration

dataset = AseDBDataset(config=dict(src=dataset_path, **config_kwargs))

# atoms objects can be retrieved by index
atoms = dataset.get_atoms(0)

To read more than one subdataset you can simply pass a list of subdataset paths,

from fairchem.core.datasets import AseDBDataset

config_kwargs = {} # see tutorial on additional configuration
dataset_paths = [
      "/path/to/omat24/train/rattled-relax",
      "/path/to/omat24/train/rattled-1000-subsampled",
      "/path/to/omat24/train/rattled-1000"
]
dataset = AseDBDataset(config=dict(src=dataset_paths, **config_kwargs))

To read all of the OMat24 training or validations splits simply pass the paths to all subdatasets.

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
• Discussion Board

Citation

The OMat24 dataset is licensed under a Creative Commons Attribution 4.0 License. If you use this work, please cite:

@misc{barroso_omat24,
      title={Open Materials 2024 (OMat24) Inorganic Materials Dataset and Models}, 
      author={Luis Barroso-Luque and Muhammed Shuaibi and Xiang Fu and Brandon M. Wood and Misko Dzamba and Meng Gao and Ammar Rizvi and C. Lawrence Zitnick and Zachary W. Ulissi},
      year={2024},
      eprint={2410.12771},
      archivePrefix={arXiv},
      primaryClass={cond-mat.mtrl-sci},
      url={https://arxiv.org/abs/2410.12771}, 
}

### We hope to move our datasets and models to the Hugging Face Hub in the near future to make it more accessible by the community. ###