materialyzeai/matgl

Graph deep learning library for materials

What it solves

MatGL provides a standardized, extensible platform for developing and using graph deep learning models in materials science. It addresses the challenge of predicting materials properties (such as formation energy and band gaps) and creating surrogate models for potential energy surfaces (PES) to replace computationally expensive simulations.

How it works

The library represents collections of atoms as mathematical graphs, where atoms are nodes and bonds are edges. It implements several advanced graph neural network (GNN) architectures, including:

  • QET: A charge-aware, equivariant architecture for systems involving charge transfer.
  • TensorNet: An O(3)-equivariant message-passing network using Cartesian tensor representations.
  • CHGNet: A network that captures atomic bond and angular information to predict DFT atomic magnetic moments.
  • M3GNet: An invariant network incorporating 3-body interactions, capable of acting as a foundation potential across the periodic table.
  • MEGNet: A general graph network for property prediction in molecules and crystals.

To optimize performance, MatGL supports multiple backends: PyTorch Geometric (PyG) by default, NVIDIA Warp for accelerated training and inference on NVIDIA GPUs, and JAX for portable, high-speed inference across CPU, CUDA, and Apple Silicon.

Who it’s for

It is designed for materials scientists and researchers who need to predict material properties, perform structural relaxations, or run molecular dynamics (MD) simulations using machine learning interatomic potentials.

Highlights

  • Diverse Architectures: Supports a wide range of equivariant and invariant GNNs (QET, TensorNet, M3GNet, etc.).
  • High-Performance Backends: Optional acceleration via NVIDIA Warp (2-3x speedup) and JAX (2-3.5x speedup).
  • LAMMPS Integration: Includes a native pair_style matgl for evaluating TensorNet or M3GNet potentials within LAMMPS simulations.
  • Hugging Face Integration: Easy loading and publishing of pre-trained models via the Hugging Face Hub.
  • CLI Tooling: Built-in command-line interface for quick property predictions and structural relaxations.

Related

  • Project
    atomistic-machine-learning/schnetpack

    A deep learning toolbox for predicting quantum-chemical properties of molecules and materials using atomistic neural networks.

  • Project
    microsoft/mattersim

    A deep learning atomistic model for simulating bulk materials across various elements, temperatures, and pressures to predict energy, forces, and stress.

  • Project
    materialyzeai/maml

    A Python package that provides high-level interfaces for materials science machine learning, enabling the conversion of materials data into features for property prediction and energy surface modeling.

  • Project
    CederGroupHub/chgnet

    A pretrained universal neural network potential for charge-informed atomistic modeling that predicts energy, forces, and stress with near-DFT accuracy.

  • Dispatch
    Introduction to Graph Machine Learning

    Hugging Face provides a comprehensive overview of Graph Machine Learning, detailing how graphs are represented and the evolution from pre-neural features to Graph Neural Networks and Graph Transformers.