MatGL: Accelerating Materials Science with Graph Deep Learning

MatGL is an open-source, PyTorch-based graph deep learning framework designed specifically for materials science, offering powerful tools for property prediction, structure relaxation, and simulation through advanced architectures like M3GNet, CHGNet, and SO3Net. Built on the Deep Graph Library (DGL) and maintained by the Materials Virtual Lab in collaboration with Intel Labs, MatGL provides a modular, intuitive API with multi-GPU support via PyTorch Lightning, pre-trained models through PyTorch Hub, and seamless model versioning. Its ecosystem includes the MatML suite, datasets like MatPES, and integration with LAMMPS via Docker, making it ideal for researchers and developers in computational materials science. Extensive tutorials in Jupyter notebook format guide users through model training, property prediction, and benchmarking, all supported by comprehensive documentation and a strong focus on code quality and usability.

Key Features:

• Advanced Architectures: Implements MEGNet, M3GNet, CHGNet, TensorNet, and SO3Net for accurate materials property prediction.
• M3GNet with 3-Body Interactions: Enables tensorial property prediction (e.g., forces, stresses) using atomic coordinates and lattice matrices.
• Modular Design: Offers dedicated modules for materials-to-graph conversion (matgl.ext), dataset loading (matgl.graphs), model implementations (matgl.models), and specialized applications like interatomic potentials (matgl.apps).
• Pre-Trained Models: Easily load models via matgl.load_model() and PyTorch Hub for rapid deployment.
• Model Versioning: Supports version control with __version__ class variables to ensure model integrity and compatibility.
• GPU Acceleration: Full support for CUDA via PyTorch for high-performance computing.
• Multi-GPU Training: Leverages PyTorch Lightning for efficient multi-GPU model training and inference.
• Command-Line Tools: Enables structure relaxation and property prediction directly from the command line.
• Docker Integration: Provides Docker images with LAMMPS support for scalable simulations and deployment.
• Comprehensive Tutorials: Interactive Jupyter notebook tutorials for training, prediction, benchmarking, and combining models.
• Robust API & I/O: All models inherit IOMixIn for easy saving, loading, and serialization.
• Developer-Friendly: Uses Google-style documentation, includes pytest fixtures, and encourages contributions with rigorous testing.
• Web3-Ready Ecosystem: Part of the broader MatML initiative, supporting future integration with decentralized data and AI workflows.

Pricing: MatGL is free to use and open-source, with no paid tiers or subscriptions required. The project is funded by the U.S. Department of Energy and supported by NSF through the Expanse supercomputing cluster.

Conclusion: MatGL is a cutting-edge, open-source framework that empowers materials scientists and AI developers with state-of-the-art graph neural networks, seamless GPU support, and rich educational resources—making it an essential tool for advancing computational materials research.