M3GNet / MatGL

M3GNet (Materials 3-body Graph Network) is a universal graph neural network interatomic potential that incorporates 3-body interactions. MatGL is the PyTorch implementation that provides pretrained universal potentials for the periodic t…

6. DYNAMICS 6.3 Machine Learning Potentials VERIFIED
Back to Directory Official Website

Overview

M3GNet (Materials 3-body Graph Network) is a universal graph neural network interatomic potential that incorporates 3-body interactions. MatGL is the PyTorch implementation that provides pretrained universal potentials for the periodic table, enabling rapid materials simulations without system-specific training.

Reference Papers

Reference papers are not yet linked for this code.

Full Documentation

Official Resources

  • Homepage: https://matgl.ai/
  • Documentation: https://matgl.ai/
  • Source Repository: https://github.com/materialsvirtuallab/matgl
  • License: BSD-3-Clause

Overview

M3GNet (Materials 3-body Graph Network) is a universal graph neural network interatomic potential that incorporates 3-body interactions. MatGL is the PyTorch implementation that provides pretrained universal potentials for the periodic table, enabling rapid materials simulations without system-specific training.

Scientific domain: Universal ML potentials, materials science, graph neural networks
Target user community: Materials scientists needing universal potentials

Theoretical Methods

  • Graph neural networks
  • 3-body interactions
  • Many-body descriptors
  • Universal potential training
  • Materials Project data

Capabilities (CRITICAL)

  • Pretrained universal potential
  • Full periodic table coverage
  • Structure relaxation
  • MD simulations
  • Property prediction
  • ASE calculator
  • Fine-tuning support

Key Strengths

Universal Coverage:

  • 89 elements
  • Materials Project training
  • Ready to use
  • No training needed

3-body Interactions:

  • Beyond pairwise
  • Accurate geometries
  • Better transferability

Inputs & Outputs

  • Input formats:

    • ASE Atoms
    • Pymatgen structures
  • Output data types:

    • Energies
    • Forces
    • Stresses
    • Model files

Interfaces & Ecosystem

  • ASE: Calculator
  • Pymatgen: Structure handling
  • PyTorch: Backend (MatGL)
  • Materials Project: Training data

Advanced Features

  • Universal potential: Full periodic table
  • MEGNet: Property prediction
  • Fine-tuning: Transfer learning
  • Relaxation: Structure optimization
  • MD: Molecular dynamics

Performance Characteristics

  • Fast inference
  • GPU acceleration
  • Good accuracy
  • Pretrained ready

Computational Cost

  • Pretrained: No training needed
  • Inference: Fast
  • Fine-tuning: Hours
  • Overall: Very efficient

Best Practices

  • Use pretrained model first
  • Validate for your system
  • Fine-tune for accuracy
  • Check against DFT

Limitations & Known Constraints

  • Materials focus
  • May need fine-tuning
  • Accuracy varies by system
  • Active development

Application Areas

  • Materials discovery
  • High-throughput screening
  • Structure prediction
  • Property prediction
  • Catalysis

Comparison with Other Codes

  • vs CHGNet: M3GNet 3-body focus, CHGNet charge-aware
  • vs MACE: M3GNet universal pretrained, MACE higher accuracy
  • vs DeepMD-kit: M3GNet pretrained ready, DeepMD custom training
  • Unique strength: 3-body interactions, universal coverage, Materials Project training

Community and Support

  • Active development (Materials Virtual Lab)
  • GitHub issues
  • Good documentation
  • Pymatgen integration

Verification & Sources

Primary sources:

  1. GitHub: https://github.com/materialsvirtuallab/matgl
  2. C. Chen et al., Nat. Comput. Sci. 2, 718 (2022)

Secondary sources:

  1. MatGL tutorials
  2. Materials Project documentation
  3. MEGNet property prediction papers

Confidence: VERIFIED

Verification status: ✅ VERIFIED

  • Source code: OPEN (GitHub, BSD-3)
  • Published in Nature Computational Science
  • Academic citations: >500
  • Active development: Materials Virtual Lab

Related Tools in 6.3 Machine Learning Potentials