MatterSim

MatterSim is Microsoft's deep-learning atomistic model, a machine-learning force field trained on millions of first-principles computations. It predicts energies, forces, and stresses across the periodic table under realistic thermodynamic conditions, covering temperatures up to 5000 K and pressures reaching 1000 GPa, which makes it a practical in silico alternative to slow density functional theory calculations for many workflows.

Two pretrained checkpoints ship on GitHub: MatterSim-v1.0.0-1M, a fast general-purpose version, and MatterSim-v1.0.0-5M, a more accurate one for complex configurations. Both build on the M3GNet architecture and plug into the ASE calculator interface and LAMMPS for large-scale molecular dynamics, with a fine-tuning script for custom datasets. Caveats: the public checkpoints are research-oriented, so check the repo's license file before commercial deployment, and like all ML potentials it should be validated against first-principles data for chemistries far from its training distribution.