rankings / model / M3GNet

M3GNet

M3GNet is a machine-learning interatomic potential — version 1.2 · from Open community · BSD-3 license · 5.1M parameters. ScoreMol scores it on 8 benchmark suites covering Atomization energy, Formation energy, Force MAE, Lattice parameters, Polymorph ranking (ρ), Stable-set discovery (F1), Γ-point frequency error, False imaginary-mode rate, Missed-instability rate, Elastic tensor Cij, NVE energy drift, MD stability 1000 K, OOD degradation, Reaction energy, each run under the same protocol on held-out reference data with cluster-bootstrap confidence intervals.

Suites: QM9-X, FormE-40k, Hull-5k, Phon-1k, Elas-600, MD-Electro-800, OOD-X, MTK-55.