ACEsuit/mace-foundations

What it solves

This project provides pre-trained foundation models for atomistic materials chemistry, enabling the prediction of energies and forces in chemical systems without needing to train a model from scratch for every new material or molecule.

How it works

It offers a suite of MACE models parameterized for up to 89 chemical elements. These models are trained on massive datasets (such as MPTrj, OMAT, and OMOL) using Density Functional Theory (DFT) levels of theory. Depending on the specific model, they are optimized for different targets, such as inorganic crystals, organic systems, molecules with charges and spins, or surfaces.

Who it’s for

Researchers and scientists in materials science, chemistry, and physics who need high-accuracy atomistic simulations and molecular dynamics (MD) simulations.

Highlights

• Broad Elemental Coverage: Models cover up to 89 chemical elements.
• Diverse Model Variants: Includes specialized models for materials (MACE-MP), molecules (MACE-OMOL), and cross-domain performance (MACE-MH).
• Stability Improvements: Second-generation models feature improved stability for high-pressure environments and molecular dynamics simulations.
• State-of-the-Art Accuracy: The MACE-MPA-0 model achieves state-of-the-art accuracy on Matbench benchmarks.