MACE

Official Resources

• Homepage: https://mace-docs.readthedocs.io/
• Documentation: https://mace-docs.readthedocs.io/
• Source Repository: https://github.com/ACEsuit/mace
• License: MIT License

Overview

MACE creates fast and accurate machine learning interatomic potentials using higher-order equivariant message passing. It combines the strengths of the Atomic Cluster Expansion (ACE) with message passing neural networks (MPNNs). MACE achieves state-of-the-art accuracy and is designed to be scalable for large simulations.

Scientific domain: Machine learning potentials, equivariant neural networks
Target user community: MD users, ML researchers

Capabilities (CRITICAL)

• Higher Order Message Passing: Captures many-body interactions efficiently.
• Foundation Models: Pre-trained "universal" potentials (MACE-MP-0) available for the periodic table.
• Speed: Optimized for GPU execution.
• LAMMPS/ASE: Interfaces for MD.

Sources: MACE GitHub, arXiv:2206.07697

Inputs & Outputs

• Input formats: XYZ training data
• Output data types: PyTorch models

Interfaces & Ecosystem

• PyTorch: Backend.
• ASE: Calculator.
• LAMMPS: Production MD.

Workflow and Usage

1. Download pre-trained model or train your own.
2. calc = MACECalculator(model_path='MACE.model', device='cuda')
3. Run MD with ASE or LAMMPS.

Performance Characteristics

• Very fast inference for an equivariant model.
• High accuracy across diverse chemical spaces.

Application Areas

• General purpose MD
• Chemistry (reactions)
• Materials discovery

Community and Support

• Developed by Kovacs, Batatia, et al. (Cambridge/EPFL)
• Rapidly growing community

Verification & Sources

Primary sources:

1. GitHub: https://github.com/ACEsuit/mace
2. Publication: I. Batatia et al., NeurIPS 2022

Confidence: VERIFIED

Verification status: ✅ VERIFIED

• Website: ACTIVE
• Documentation: AVAILABLE
• Source: OPEN (GitHub)
• Development: ACTIVE
• Applications: SOTA ML potentials