Atomistic ML Format · LeMaterial
Atompack
Public atomistic machine learning datasets distributed in the Atompack format for scalable multi-worker training.
What Atompack is
Atompack is a centralised repository of public atomistic machine learning datasets distributed in the Atompack format and maintained by LeMaterial. It forms part of a wider initiative to standardise data exchange for atomistic machine learning across chemistry, physics, materials science, and engineering.
Rather than serving as the original source of these datasets, the repository acts as a distribution and serving layer. It packages established corpora — including LeMat-Bulk, MatPES, MP-ALOE, MPtrj, OMAT24, and the Open Catalyst 2020 (OC20) collection — into a unified, immutable structure designed for machine learning workflows that require repeated random reads and concurrent worker processes.
How it was made
Datasets are organised into specific repository paths stored either as individual .atp files or as multi-shard directories, such as the ten-shard export used for the OC20 Structure to Energy and Forces (S2EF) training split. The format provides static, memory-mapped (mmap) read-only access, permitting direct indexed lookups of full molecular and crystalline records without loading whole datasets into memory.
Data records are loaded via the atompack-db Python package using the atompack module, which provides read-only functions to open or download remote Hub paths and local files. Stored precisions depend on the dataset export; for example, OC20 exports store atomic positions and forces as 32-bit floating-point numbers (float32), whilst potential energies and unit cells are stored as 64-bit floating-point numbers (float64).
Key facts
- File formats
- .atp files and shard directories
- Software library
- atompack-db (imported as atompack)
- Hosted datasets
- LeMat-Bulk, MatPES, MP-ALOE, MPtrj, OC20, OMAT24
- Size category
- 100M to 1B records
- Data access method
- Memory-mapped (mmap) read-only indexing
What you can use it for
- Interatomic potential training
- Supplying large-scale atomistic reference sets, such as OMAT24, MPtrj, and MatPES, to training pipelines using parallel, multiprocessing data loaders.
- Catalytic reaction benchmarking
- Evaluating machine learning models across standardized in-distribution and out-of-distribution validation splits from the Open Catalyst 2020 (OC20) S2EF task.
- Bulk materials property prediction
- Loading curated density functional theory trajectories and relaxed bulk crystal structures from LeMat-Bulk and MP-ALOE for regression models.
Explore the data
No licence declared
The authors have not published a licence, which means no permission to redistribute the data has been granted to anyone. The rows are not served here; read them on the original dataset page.Strengths and limits
Good at
- Employs memory-mapped access that enables rapid, repeated random reads across multiple worker processes.
- Unifies diverse upstream scientific datasets under a single access API and file format.
- Supports both streaming from remote repositories and loading from local downloaded files.
Watch out for
- The dataset format and repository interface are strictly read-only and do not allow record editing.
- It acts purely as a distribution layer, requiring researchers to reference and cite the respective upstream source datasets.
Questions
Can I look inside Atompack without downloading it?
Not here: it declares no licence, so no permission to redistribute its rows has been granted. Its description and statistics are still shown.
What licence is Atompack under?
Its authors have not declared a licence, which means no permission to reuse or redistribute it has been granted.
Which upstream datasets are available in this repository?
The repository distributes packaged versions of LeMat-Bulk, MatPES (both PBE and r2SCAN), MP-ALOE, MPtrj, Open Catalyst 2020 (OC20 S2EF splits), and OMAT24.
How is the dataset loaded in Python?
Researchers can install the atompack-db package and call atompack.hub.open or atompack.hub.download to retrieve and iterate over records directly by index.
What numeric precision is used for physical quantities in the dataset?
Precision is export-dependent; for example, the OC20 exports store atomic coordinates and forces as float32, whereas unit cell vectors and total energies are stored as float64.
Written from Atompack'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.