Microsoft Research: MatterSim experimentally synthesized TaP at 152 W/m/K, MatterSim-MT extends output beyond PES

Microsoft Research published MatterSim foundation model results for materials science on May 12, 2026 — including experimental validation of an AI prediction, 3–5× faster inference, and a new multi-task MatterSim-MT model that goes beyond classical potential energy surface (PES) constraints.

What did MatterSim experimentally confirm?

The MatterSim-v1 model identified tetragonal TaP (tantalum phosphide) from a corpus of over 240,000 material candidates as a potentially high thermal conductor. The Microsoft team synthesized TaP and measured its thermal conductivity in the laboratory at 152 W/m/K, close to silicon’s performance. This marks the first identification of TaP as a thermal conductor through AI-driven materials screening, validating ML predictions in real materials discovery.

One project collaborator noted that “we can test conventional understanding of what controls thermal conductivity at scale” while simultaneously discovering functional materials that scientists had overlooked.

How much faster has inference become?

MatterSim-v1 received two speedups: 3× for the 5M-parameter variant and 5× for the 1M-parameter variant. Integration with LAMMPS software enables multi-GPU scaling through existing scientific workflows, making large simulations operationally viable in laboratory settings.

What outputs does MatterSim-MT offer beyond PES?

MatterSim-MT is a multi-task foundation model trained on more than 35 million first-principles structures, across 89 elements, temperatures up to 5,000 K, and pressures up to 1,000 GPa. Classical PES models predict only energies and forces; MatterSim-MT natively outputs stress tensors, magnetic moments, Born effective charges, and dielectric matrices.

Microsoft demonstrated three applications: vibrational spectroscopy predicts phonon spectra in polar crystals with LO-TO splitting within 0.06 THz of ab initio calculations; ferroelectric switching simulates polarization inversion in barium titanate under an electric field; battery chemistry models cationic-to-anionic redox transitions in lithium manganese oxide without task-specific training.

The initiative positions foundation models as a tool that elevates materials science from an experimental to a hybrid AI-experimental workflow.