MatterSim accelerates materials discovery with 3-5x faster simulation and new multi-task model

Microsoft Research has released significant updates to MatterSim, its machine learning framework for accelerating materials discovery. The advances include experimental validation of predictions, substantial speed improvements, and a new multi-task foundation model.

Experimental breakthrough in thermal conductors#

Researchers used MatterSim-v1 to screen over 240,000 candidate materials for high thermal conductivity. The effort, conducted in partnership with the University of Texas Dallas, University of Illinois Urbana-Champaign, and UC Davis, identified tetragonal tantalum phosphorus (TaP) as a promising thermal conductor.

The team synthesized TaP and measured its thermal conductivity at 152 W/m/K—comparable to silicon. While tetragonal TaP had been synthesized before, it was never previously recognized as a thermal conductor candidate. "For the first time, we can test conventional understanding of what controls thermal conductivity at scale," said Prof. David Cahill of University of Illinois Urbana-Champaign.

Faster performance and LAMMPS integration#

MatterSim-v1 inference speed has increased dramatically: 3x faster for the 5M variant and 5x faster for the 1M variant. The improvements stem from optimized graph construction, ahead-of-time compilation, and reduced data conversion overhead.

The team has also integrated MatterSim-v1 directly into LAMMPS, the widely used molecular dynamics software, enabling researchers to scale simulations across multiple GPUs within their existing workflows.

New multi-task foundation model#

MatterSim-MT, a new multi-task model, extends beyond traditional potential energy surface predictions to simulate complex phenomena simultaneously. The model is pretrained on over 35 million first-principles-labeled structures covering 89 elements across temperatures up to 5000 K and pressures up to 1000 GPa.

MatterSim-MT can predict energies, forces, stress, Bader charges, magnetic moments, Born effective charges, and dielectric matrices. The team demonstrated three key applications:

• Vibrational spectroscopy: Predictions of Born effective charges and dielectric properties enabled accurate phonon spectra calculations in 3c-silicon carbide under extreme pressure, with LO-TO splitting accurate to within 0.06 THz of ab initio results.