CHGNet

CHGNet (Crystal Hamiltonian Graph neural Network) is a pretrained universal neural network potential for charge-informed atomistic modeling. It is trained on over 1.5 million structures from the Materials Project and explicitly accounts…

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

Overview

CHGNet (Crystal Hamiltonian Graph neural Network) is a pretrained universal neural network potential for charge-informed atomistic modeling. It is trained on over 1.5 million structures from the Materials Project and explicitly accounts for magnetic moments and charge states.

Reference Papers

Reference papers are not yet linked for this code.

Full Documentation

Official Resources

  • Homepage: https://chgnet.lbl.gov/
  • Documentation: https://github.com/CederGroupHub/chgnet
  • Source Repository: https://github.com/CederGroupHub/chgnet
  • License: BSD-3-Clause

Overview

CHGNet (Crystal Hamiltonian Graph neural Network) is a pretrained universal neural network potential for charge-informed atomistic modeling. It is trained on over 1.5 million structures from the Materials Project and explicitly accounts for magnetic moments and charge states.

Scientific domain: Universal ML potentials, materials science, charge-informed modeling
Target user community: Materials scientists needing universal potentials with charge information

Theoretical Methods

  • Graph neural networks
  • Charge-informed features
  • Magnetic moment prediction
  • Universal potential training
  • Materials Project data

Capabilities (CRITICAL)

  • Pretrained universal potential
  • Charge state prediction
  • Magnetic moment prediction
  • Oxidation state awareness
  • ASE calculator
  • Structure relaxation
  • MD simulations

Key Strengths

Charge Information:

  • Oxidation states
  • Magnetic moments
  • Charge transfer
  • Redox reactions

Universal Coverage:

  • Most periodic table
  • Materials Project training
  • Ready to use

Inputs & Outputs

  • Input formats:

    • ASE Atoms
    • Pymatgen structures
  • Output data types:

    • Energies
    • Forces
    • Stresses
    • Charges
    • Magnetic moments

Interfaces & Ecosystem

  • ASE: Calculator
  • Pymatgen: Structure handling
  • Materials Project: Training data

Advanced Features

  • Charge prediction: Oxidation states
  • Magnetism: Magnetic moments
  • Universal: Broad element coverage
  • Fine-tuning: Transfer learning
  • Relaxation: Structure optimization

Performance Characteristics

  • Fast inference
  • GPU acceleration
  • Good accuracy for materials
  • Pretrained ready to use

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 if needed
  • Check charge predictions

Limitations & Known Constraints

  • Materials focus
  • May need fine-tuning
  • Charge accuracy varies
  • Active development

Application Areas

  • Battery materials
  • Catalysis
  • Oxidation/reduction
  • Magnetic materials
  • Materials screening

Comparison with Other Codes

  • vs M3GNet: CHGNet charge-aware, M3GNet 3-body focus
  • vs MACE: CHGNet charge/magnetism, MACE higher accuracy
  • vs DeepMD-kit: CHGNet pretrained universal, DeepMD custom training
  • Unique strength: Charge and magnetic moment prediction, Materials Project training

Community and Support

  • Active development (Berkeley/Ceder group)
  • GitHub issues
  • Good documentation
  • Materials Project integration

Verification & Sources

Primary sources:

  1. GitHub: https://github.com/CederGroupHub/chgnet
  2. B. Deng et al., Nat. Mach. Intell. 5, 1031 (2023)

Secondary sources:

  1. CHGNet tutorials
  2. Materials Project documentation
  3. Battery materials applications

Confidence: VERIFIED

Verification status: ✅ VERIFIED

  • Source code: OPEN (GitHub, BSD-3)
  • Published in Nature Machine Intelligence
  • Academic citations: >200 (rapid growth)
  • Active development: Ceder group

Related Tools in 6.3 Machine Learning Potentials