ACEsuit/mace

MACE - Fast and accurate machine learning interatomic potentials with higher order equivariant message passing.

What it solves

MACE provides a way to create fast and accurate machine learning interatomic potentials. It allows researchers to model the energy and forces of atomic systems (like materials and organic molecules) using machine learning, avoiding the high computational cost of traditional quantum mechanical calculations while maintaining high accuracy.

How it works

MACE uses higher-order equivariant message passing neural networks to process 3D point clouds of atoms. It can be trained from scratch on atomic configuration data (XYZ files) or fine-tuned from existing foundation models. The system supports various training techniques such as EMA (Exponential Moving Average), amsgrad, and a two-stage training process to reduce energy errors.

Who it’s for

It is designed for computational chemists, materials scientists, and physicists who need to perform molecular dynamics or evaluate the properties of atomic structures.

Highlights

  • Foundation Models: Includes pretrained models for materials (MACE-MP), organic chemistry (MACE-OFF), and electrostatics (MACE-Polar).
  • Scalability: Supports multi-GPU training via PyTorch DistributedDataParallel and on-line data loading for datasets that exceed GPU memory.
  • ** uma CUDA Acceleration**: Integration with the cuEquivariance library for faster performance on NVIDIA GPUs.
  • Flexible Training: Supports various precision levels (float32/float64) and allows for the estimation of atomic reference energies (E0s) during training.

Related

  • Project
    mir-group/nequip

    An open-source framework for building E(3)-equivariant interatomic potentials using graph neural networks to accurately predict atomic energies and forces.

  • Project
    microsoft/mattersim

    A deep learning atomistic model for simulating bulk materials across various elements, temperatures, and pressures to predict energy, forces, and stress.

  • Project
    deepmodeling/deepmd-kit

    A deep learning package that turns quantum-mechanical reference data into fast, scalable interatomic potentials for molecular and materials science simulations.

  • Project
    ACEsuit/mace-foundations

    A collection of pre-trained foundation models for materials chemistry that predict atomistic properties across 89 chemical elements.

  • Project
    google-deepmind/acme

    Acme is a research framework for reinforcement learning that provides scalable building blocks and baseline agents for developing and testing RL algorithms.