Description
MACE (Machine Learning Interatomic Potentials) is a powerful Python library designed for the development and deployment of fast and accurate machine learning interatomic potentials. It leverages higher-order equivariant message passing, a sophisticated neural network architecture, to model interactions between atoms in materials and molecules with high fidelity. This approach allows MACE to capture complex atomic behaviors, leading to more reliable simulations in fields like materials science, chemistry, and condensed matter physics.
Read More
The MACE repository provides a comprehensive reference implementation, developed by a team of researchers. It includes tools for training custom MACE models from scratch, evaluating their performance, and utilizing pre-trained foundation models. The library supports various installation methods, including PyPI for ease of use and from source for developers. For users requiring accelerated computations, MACE offers CUDA acceleration through the cuEquivariance library and supports Apple Silicon GPU acceleration via MPS.
Key capabilities of MACE include its ability to handle large datasets through on-line data loading and preprocessing scripts. It also integrates with experiment tracking tools like Weights and Biases for robust research workflows. The library offers a range of pre-trained foundation models, such as MACE-MP for materials science and MACE-OFF for organic chemistry, which can be used directly or fine-tuned for specific applications. Documentation is available to guide users through installation, usage, and advanced features.
MACE is particularly valuable for researchers and engineers in computational materials science, quantum chemistry, and molecular dynamics. Its focus on accuracy and speed makes it suitable for tasks ranging from predicting material properties to simulating chemical reactions. The availability of foundation models and fine-tuning capabilities democratizes access to advanced interatomic potentials, enabling faster discovery and design of new materials and molecules.
- CategoryScientific Discovery & Lab Tools
- Pricing ModelOpen source
- DeploymentCloud, Self-hosted
- PlatformsWindows, macOS, Linux
MACE Machine Learning Potentials's Core Features
Fast and accurate machine learning interatomic potentials
Higher order equivariant message passing architecture
Tools for training custom MACE models
Model evaluation capabilities
Support for CUDA acceleration with cuEquivariance
Apple Silicon GPU acceleration (MPS)
Pre-trained foundation models for materials and organic chemistry
Fine-tuning support for foundation models
On-line data loading for large datasets
Integration with Weights and Biases for experiment tracking
Support for PyTorch >= 1.12
MIT License for the MACE code
Getting Started with MACE Machine Learning Potentials
Install MACE: Use pip install mace-torch or install from source.
Set up environment: Ensure Python >= 3.9 and PyTorch >= 1.12 are installed.
Train a model: Utilize the mace_run_train script with specified parameters.
Evaluate a model: Use the mace_eval_configs script for performance assessment.
Utilize foundation models: Load and fine-tune pre-trained models for specific tasks.
Enable GPU acceleration: Specify '--device=cuda' or '--device=mps' for faster computations.
Integrate with ASE: Use MACE calculators within the Atomic Simulation Environment.
MACE Machine Learning Potentials's Use Cases
- Materials Discovery
- Molecular Dynamics
- Chemical Reaction Simulation
- Predicting Material Properties
- Organic Chemistry Simulations
- Force Field Development
- Computational Chemistry
- Active Learning
