metatensor and metatomic: Foundational libraries for interoperable atomistic machine learning

Area of Science:

• Computational Materials Science
• Machine Learning in Physics
• Scientific Software Development

Background:

• Machine learning (ML) significantly enhances atomic-scale simulations by improving accuracy and reducing computational costs.
• Integrating ML with traditional atomistic modeling faces challenges due to disparate mathematical foundations and software ecosystems.
• A need exists for tools that facilitate seamless data exchange and model compatibility between ML frameworks and established simulation packages.

Purpose of the Study:

• To introduce novel software libraries, metatensor and metatomic, designed to overcome integration challenges between ML and atomistic simulations.
• To provide a common framework for data handling and model storage, promoting wider adoption of ML in materials modeling.
• To enable efficient sharing and utilization of ML models across diverse simulation software.

Main Methods:

• Development of `metatensor`: a multi-platform, multi-language library for storing and manipulating arrays with metadata, specifically for atomistic ML.
• Implementation of `metatomic`: a library providing a portable interface for storing atomistic ML models and associated metadata.
• Demonstration of an integrated ecosystem of tools, including libraries, training utilities, and interfaces with existing simulation packages.

Main Results:

• `metatensor` enables unified data representation and manipulation, facilitating interoperability between Python-based ML software and Fortran/C/C++-based simulation tools.
• `metatomic` ensures portable storage and distribution of ML models, simplifying their implementation and use across different simulation environments.
• The developed ecosystem showcases the practical effectiveness of `metatensor` and `metatomic` in bridging the gap between traditional and modern computational approaches.

Conclusions:

• `metatensor` and `metatomic` provide essential infrastructure for advancing ML applications in atomistic simulations.
• These libraries effectively address the challenges of combining different software ecosystems, fostering collaboration and innovation.
• The demonstrated ecosystem accelerates the integration of ML into mainstream materials modeling workflows.