Infrastructure · Materials science · Chemistry

M3GNet scales graph neural interatomic potentials across the periodic table

A graph neural interatomic potential used for relaxation, dynamics, property prediction, and large candidate screening.

Summary

The Nature Computational Science paper presents M3GNet, a graph neural-network interatomic potential trained on Materials Project structural relaxations. It reports applications to relaxation, dynamics and property prediction, plus screening of 31 million hypothetical crystals with DFT verification of many low-energy candidates.

AI role

Approximated expensive first-principles calculations for materials modeling across many chemistries.

Narrative role

M3GNet belongs in the timeline as infrastructure: faster potentials can change the feasible scale of materials search.

Caveat

Machine-learned potentials can fail outside their training distribution and still require higher-fidelity checks.