This repository provides M3GNet, a graph neural network for materials science, capable of acting as a surrogate for DFT calculations and predicting material properties. It is targeted at researchers and engineers in materials science and chemistry. The project offers a universal interatomic potential (IAP) for structural relaxations and property predictions across the periodic table.

How It Works

M3GNet incorporates 3-body interactions and includes atomic coordinates and the 3x3 lattice matrix, enabling tensorial property predictions like forces and stresses via auto-differentiation. This design allows for flexibility across diverse chemical spaces and the development of universal IAPs.

Quick Start & Requirements

• Install via pip: pip install m3gnet
• For model training, GPU memory > 18 GB is recommended.
• Apple Silicon installation requires specific steps involving Conda, tensorflow-macos, and tensorflow-metal.
• API documentation: https://materialsvirtuallab.github.io/matgl/ (Note: This links to MatGL, the successor, as M3GNet repo is archived).

Highlighted Details

• Achieves <1% error in lattice constants for most cubic crystals compared to DFT values.
• Offers a universal IAP trained on Materials Project relaxations, applicable across the periodic table.
• Supports molecular dynamics simulations using the trained IAP.
• Enables custom model training with user-provided datasets of structures, energies, forces, and stresses.

Maintenance & Community

This repository has been archived and is no longer maintained. Users are directed to use the successor implementation, MatGL.

Licensing & Compatibility

The README does not explicitly state the license. However, the project is associated with the Materials Project, which typically uses permissive licenses. Compatibility for commercial use would require explicit license confirmation.

Limitations & Caveats

The project is archived and no longer maintained. A successor implementation, MatGL, is recommended. The README notes potential accuracy issues with specific materials like EuTiO3, iodides, and noble gases, possibly due to limited training data for these systems.