ColabFit exchange: Open-access datasets for data-driven interatomic potentials.

Area of Science:

• Computational Materials Science
• Computational Chemistry
• Data Science

Background:

• Data-driven interatomic potentials (IPs) are crucial for atomic-scale simulations in materials science and chemistry.
• A significant lack of organized, publicly accessible datasets hinders the development and benchmarking of IPs.
• This deficiency limits the creation of more general and transferable interatomic potentials.

Purpose of the Study:

• To introduce the ColabFit Exchange, the first database designed for interatomic potential development.
• To provide open access to a large, systematically organized collection of datasets from diverse domains.
• To facilitate community-driven contributions and advancements in IP research.

Main Methods:

• Developed a web-based platform (https://colabfit.org) for dataset exploration, download, and contribution.
• Aggregated 139 datasets covering nearly 70,000 unique chemistries from literature and community submissions.
• Implemented a software framework for dataset construction and access, including diversity analysis tools.

Main Results:

• The ColabFit Exchange offers a centralized resource for interatomic potential development datasets.
• The database currently houses 139 diverse datasets, with plans for continuous growth.
• Analysis quantified the database's diversity and proposed metrics for assessing dataset diversity.

Conclusions:

• The ColabFit Exchange addresses a critical need for standardized datasets in interatomic potential research.
• It enables robust benchmarking, model comparison, and the development of more universal IPs.
• The platform supports an end-to-end IP development pipeline, integrating fitting and validation tools.