MatterSim

MatterSim is a deep-learning atomistic simulator that acts as a universal machine-learning interatomic potential. It covers metals, oxides, sulfides, and halides across crystalline, amorphous, and liquid phases, with predictions valid from 0 to 5,000 K and pressures up to 10⁷ atm. Recent updates deliver 3–5× inference speed-ups via improved graph construction and tighter LAMMPS integration for multi-GPU molecular dynamics. The MatterSim-MT extension predicts not only energies and forces but also stress, magnetic moments, Born effective charges, and dielectric matrices — enabling simulation of ferroelectric switching, electrochemical redox, and vibrational spectroscopy.

MatterSim is the simulation half of Microsoft’s materials flywheel: it scores and validates the candidate structures that MatterGen invents, at orders-of-magnitude lower cost than first-principles calculations while remaining accurate enough for screening. The multi-task architecture brings phenomena previously inaccessible to simple interatomic potentials into reach, supporting in-silico workflows across batteries, semiconductors, and energy applications. Adopted by academic and national-lab groups with public checkpoints, MatterSim is one of the central infrastructure components of Microsoft AI for Science’s materials program.