Advancing Materials AI with MatterSim: Experimental Synthesis, High-Speed Simulation, and a Multi-Task Model

MatterSim-v1 validated a high thermal conductivity material candidate through actual experiments. Together with UT Dallas, UIUC, and UC Davis, it screened over 240,000 candidates to discover tetragonal TaP, whose experimental thermal conductivity reached 152 W/m/K, close to that of silicon.

High thermal conductivity materials are important for computing, power electronics, and aerospace cooling. MatterSim rapidly narrows down candidate pools through phonon-based thermal conductivity prediction, enabling exploration at a scale that conventional methods struggle to handle.

• Performance improvements: MatterSim-v1.0.0-5M became 3x faster, and v1.0.0-1M became 5x faster. Graph construction optimization, ahead-of-time compilation, and reduced atomic representation conversion drove the speed gains, and LAMMPS integration now enables large-scale simulations across multiple GPUs.
• MatterSim-MT: The multi-task foundation model predicts not only energy, force, and stress but also magnetic moments, Born effective charges, and dielectric matrices. It was pretrained on over 35 million first-principles labeled structures spanning 89 elements, up to 5000 K, and 1000 GPa, covering phenomena such as vibrational spectroscopy, ferroelectric switching, and electrochemical redox. The full manuscript also presents scaling with more data and parameters, fine-tuning toward higher levels of theory, and extensions to active learning.