m3gnet

Official Resources

• Homepage: https://materialsvirtuallab.github.io/m3gnet/
• Documentation: https://materialsvirtuallab.github.io/m3gnet/
• Source Repository: https://github.com/materialsvirtuallab/m3gnet
• License: BSD 3-Clause License

Overview

M3GNet is a graph neural network (GNN) potential trained on the massive Materials Project trajectory dataset (>187,000 materials). It serves as a "universal" interatomic potential for the periodic table, capable of relaxing structures and performing MD for diverse chemistries. It can also be used as a surrogate model for property prediction.

Scientific domain: Universal ML potentials, graph neural networks
Target user community: Materials scientists needing quick relaxation/properties

Capabilities (CRITICAL)

• Universal Potential: Covers 89 elements.
• Relaxation: Can relax arbitrary crystal structures (often replacing DFT relaxation).
• MD: Run molecular dynamics.
• Surrogate: Predicts formation energy, band gap, elastic moduli directly from structure.
• Integration: Works with Pymatgen and ASE.

Sources: M3GNet GitHub, Nat. Comput. Sci. 2, 718 (2022)

Inputs & Outputs

• Input formats: Pymatgen Structure / ASE Atoms
• Output data types: Energy, forces, stress, properties

Interfaces & Ecosystem

• TensorFlow: Backend.
• Pymatgen: Native object support.
• ASE: Calculator interface.

Workflow and Usage

1. model = M3GNet.load()
2. relaxer = Relaxer()
3. relax_results = relaxer.relax(structure)

Performance Characteristics

• Fast compared to DFT (seconds vs hours).
• Accuracy reasonably close to DFT PBE for stability prediction.

Application Areas

• High-throughput screening (filtering unstable structures before DFT).
• Phonon calculations (approximate).
• MD of complex multi-component systems.

Community and Support

• Developed by Ong Group (UCSD) / Materials Virtual Lab
• Active development (MatGL is the successor)

Verification & Sources

Primary sources:

1. GitHub: https://github.com/materialsvirtuallab/m3gnet
2. Publication: C. Chen and S. P. Ong, Nat. Comput. Sci. 2, 718 (2022)

Confidence: VERIFIED

Verification status: ✅ VERIFIED

• Website: ACTIVE
• Documentation: COMPREHENSIVE
• Source: OPEN (GitHub)
• Development: ACTIVE
• Applications: Universal potential, GNN