MACE - Machine Learning Force Fields

Overview

MACE is a a machine learning software for predicting many-body atomic interactions and generating force fields. It utilizes higher order equivariant message passing for fast and accurate predictions.

Features

• Predicts many-body atomic interactions with high accuracy

• Generates highly accurate force fields for use in molecular dynamics simulations

• Utilizes higher order equivariant message passing

• Fast and efficient predictions

• Out of the box foundation models for various applications

• Interface to LAMMPS and openMM for molecular dynamics simulations

Getting Started

To get quickly started with MACE, please go to the Quick Start section of the documentation.

Documentation

For detailed information on how to use MACE, please refer to the following documentation:

User Guide

• Quick Start
• User Guide
  • Introduction
  • Installation
  • Troubleshooting and Q&A Guide
  • Training
  • Evaluation
  • Heterogeneous Data Training
  • ASE calculator
  • MACE descriptors
  • Analytical Hessians
  • CUDA Acceleration with cuEquivariance Library
  • OpenMM Interface
  • MACE in LAMMPS
  • Foundation models
  • Fine-tuning Foundation Models
  • Multihead Training for MACE
  • Large Dataset Pre-processing
  • Multi-GPUs Training

Examples

• Example scripts for training MACE models
  • MD22: large molecules
  • ANI-1x dataset: H, C, N, O transferable FF
  • Liquid water
• Foundation model to run NVT simulation
• Tutorials on MACE training and architecture
  • Tutorial 1: Introduction to MACE training and evaluation
  • Tutorial 2: MACE active learning and fine-tuning
  • Tutorial 3: MACE theory and code (advanced)

Support

If you need help using MACE, please use GitHub Discussions or send an email to ib467@cam.ac.uk.