Benchmarks/MatterGen/MatterGen Base (Alex-MP)

MatterGen Base (Alex-MP)

MatterGen — Crystal Generation

Unconditional generation trained on the full Alex-MP dataset (608k stable materials).

Sources

Performance Metrics

MetricValueDescription
Structure Validity100%No atom-atom distances < 0.5 Å or volume < 0.1 ų
Composition Validity88.5%Valid by SMACT charge neutrality check
Uniqueness99.9%Fraction of non-duplicate structures in batch
Novelty61.6%Fraction not found in MP2020 training reference
Stability Rate73.3%E_above_hull < 0.1 eV/atom (MP2020 convex hull)
S.U.N. Rate38.9%Stable + Unique + Novel (the key composite metric)
Avg E_above_hull0.075 eV/atomMean energy above convex hull
Avg RMSD0.063 ÅMean displacement from DFT-relaxed geometry

Model Comparison

ModelS.U.N.StableNovelUniqueRMSD (Å)E_hull (eV)
MatterGen (Alex-MP)38.9%73.3%61.6%99.9%0.063 Å0.075
MatterGen (MP-20)25.8%47.1%76.0%99.8%0.147 Å0.201
DiffCSP (MP-20)20.1%41.3%75.8%99.4%0.413 Å0.189
CDVAE18.3%24.3%92.3%99.9%0.479 Å0.243
P-G-SchNet——————
S.U.N. = Stable + Unique + Novel. Stability threshold: E_above_hull < 0.1 eV/atom (MP2020 convex hull). Source: MatterGen GitHub benchmark JSONs.
Test Set
MP2020correction convex hull reference (Materials Project)
Notes
Primary checkpoint trained on Alex-MP. S.U.N. rate is 2.1x higher than the next best model (DiffCSP at 18.3%). RMSD is 7.6x lower than CDVAE (0.479 Å), indicating structures are much closer to local energy minima.
← All Benchmarks