MLFF Geometry Optimization

This task measures how closely machine-learning force-field relaxations reproduce DFT-relaxed crystal structures across the 257k-material WBM test set. It compares normalized structure-matching RMSD, retained symmetry, and the relaxation settings used by each model.

Not all models relaxed every structure, but each reported model was evaluated on at least 249k relaxations. Symmetry detection uses moyopy, a Rust successor to spglib.

Leaderboard

The table ranks current model-YAML results and exposes training-data, openness, output, and column filters.

Symmetry-tolerance caveat: RMSD is symprec-invariant. The leaderboard shows symmetry metrics at both symprec=1e-2 and symprec=1e-5, while Aggregate Diagnostics use symprec=1e-5. Compare symmetry values within a single view.

Training data
Openness
Targets
Presets
SymmetryHyperparams
Model RMSD Σ= 10-2 Σ 10-2 Σ 10-2 Σ= 10-5 Σ 10-5 Σ 10-5 Optimizer Steps fmax (eV/Å) Cell filter
TECE-OAM-RRA-1.00.05881.83%4.98%12.45%59.07%18.46%21.02%FIRE5000.02Frechet
GRACE-3L-OAM-L0.05882.4%3.84%13%69.29%4.27%25.94%FIRE5000.01Frechet
EquFlashV20.05882.02%4.74%12.51%70.05%5.07%24.34%FIRE5000.02Frechet
EquiformerV3+DeNS-OAM0.05974.24%12.9%11.37%37.11%45.54%12.99%FIRE5000.02Frechet
PET-OAM-XL0.06080.7%6.67%11.85%13.08%86.68%0.09%FIRE5000.02Frechet
MatRIS-10M-OAM0.06081.53%5.47%12.27%59.55%19.5%19.96%FIRE5000.02Frechet
EquFlash0.06081.98%4.68%12.64%70.19%4.89%24.4%FIRE5000.02Frechet
TACE-OAM-L0.06181.88%4.79%12.6%69.22%6.03%24.15%FIRE5000.02Frechet
eSEN-30M-OAM0.06171.03%16.16%10.85%36.36%47.39%11.88%FIRE5000.02Frechet
SevenNet-Omni-i12*0.06281.87%4.69%12.72%70.56%4.39%24.54%FIRE8000.02Frechet
Nequip-OAM-XL0.06382.23%2.53%14.37%68.34%9.75%21.17%GOQN5000.005Frechet
GRACE-2L-OAM-L0.06481.66%5.37%12.26%73.43%5.39%20.67%FIRE5000.03Frechet
Nequip-OAM-L0.06581.6%5.43%12.26%71.25%7.92%20.18%GOQN5000.05Frechet
Allegro-OAM-L0.06581.6%5.42%12.26%71.63%7.53%20.23%GOQN5000.05Frechet
GRACE-2L-OAM0.06781.59%5.63%12.09%73.3%3.25%23.01%FIRE5000.03Frechet
DPA-4.0.1-Pro-MPtrj0.06966.31%20.8%10.37%38.99%41%14.22%FIRE5000.02Frechet
eqV2 M0.06977.25%11.18%10.52%14.22%85.2%0.45%FIRE5000.02Frechet
DPA-3.1-3M-FT0.06981.09%5.02%13.17%72.74%5.39%21.38%FIRE5000.05Exp
EquiformerV3+DeNS-MP0.07064.13%23.01%10.22%39.19%42.23%13.08%FIRE5000.02Frechet
MatRIS-10M-MP0.07279.92%7.89%11.45%57.66%23.95%17.2%FIRE5000.05Frechet
GRACE-1L-OAM0.07281.4%5.76%12.16%73.36%3.29%22.92%FIRE5000.03Frechet
MACE-MPA-00.07381.44%5.56%12.31%73.24%3.28%23.05%FIRE5000.05Frechet
MatterSim v1 5M0.07381.48%5.24%12.54%68.74%7.55%23.05%FIRE5000.02Exp
ORB v30.07572.8%15.9%10.1%13%86.9%0.01%FIRE5000.02Frechet
eSEN-30M-MP0.07561.06%26.04%9.8%38.59%42.81%12.82%FIRE5000.02Frechet
eqV2 S DeNS0.07653.96%37.24%6.53%13.22%86.71%0.06%FIRE5000.02Frechet
AlphaNet-v1-OAM*0.07974.31%14.46%10.37%14.16%85.05%0.67%FIRE5000.03Frechet
HIENet0.08080.02%7.55%11.63%45.21%34.93%15.32%FIRE5000.05Frechet
Allegro-MP-L0.08281.25%5.72%12.3%72.11%7.12%20.15%GOQN5000.05Frechet
Eqnorm MPtrj0.08481.13%5.89%12.24%69.83%6.39%23.16%FIRE5000.02Frechet
SevenNet-l3i5*0.08578.8%8.95%11.34%44.12%36.67%14.51%FIRE5000.05Frechet
Nequix MP0.08581.11%5.54%12.61%69.99%6.07%23.35%FIRE5000.02Frechet
Nequip-MP-L0.08681.09%5.92%12.27%71.63%7.88%19.82%GOQN5000.05Frechet
Nequix MP PFT0.08781.09%5.4%12.76%69.76%6.02%23.62%FIRE5000.02Frechet
GRACE-2L-MPtrj0.09081.07%5.96%12.28%73.15%3.49%22.92%FIRE3000.03Frechet
MACE-MP-00.09181.12%5.99%12.2%73.85%3.35%22.39%FIRE5000.05Frechet
CHGNet0.09579.21%9.22%10.76%60.62%22.41%16.01%FIRE5000.05Frechet
ORB v2 MPA0.09749.26%45.47%4.53%14.33%85.31%0.34%FIRE5000.02Frechet
ORB v2 MPtrj0.10143.32%52.13%3.87%13.35%86.55%0.09%FIRE5000.02Frechet
M3GNet0.11280.31%6.81%12.17%74.08%6.71%18.69%FIRE5000.05Exp

RMSD measures the normalized, unitless structure-matching RMSD between ML- and DFT-relaxed ground state structures, as returned by pymatgen’s StructureMatcher after matching. Optimizer, Steps, fmax, and Filter show the ASE optimizer, maximum relaxation steps, force convergence criterion (eV/Å), and cell filter used during structure relaxation. Σ= / Σ / Σ denote the fraction of structures that retain, increase, or decrease the symmetry of the DFT-relaxed structure during MLFF relaxation. The match criterion is for the ML ground state to have identical spacegroup as DFT. For Σ / Σ, the number of symmetry operations for a structure decreased / increased during MLFF relaxation. Note that the symmetry metrics are sensitive to the symprec value passed to spglib so we show results for multiple values. See the spglib docs and paper for details.

Model Comparison

Use the axis, color, and size controls to compare the models with geometry-optimization metrics directly from the current model YAML data.

Σ= 10-2 vs RMSD

The default view compares structure-matching RMSD (lower is better) with the fraction of matching spacegroups at symprec=1e-2 (higher is better). Marker size defaults to model parameters and color to training-set size.

  • Params 41 models
Log Scale

Aggregate Diagnostics

These views use the per-structure symprec=1e-5 analyses. The model picker applies to every view below.

  • TECE-OAM-RRA-1.0
  • GRACE-3L-OAM-L
  • EquFlashV2
  • DPA-4.0.1-Pro-MPtrj
  • TACE-OAM-L

Cumulative Distribution of RMSD

Cumulative distribution of RMSD between ML and DFT-relaxed structures.

Difference in Number of Symmetry Operations vs DFT

EquFlashV2 (σ=6.36)
GRACE-3L-OAM-L (σ=6.38)
TACE-OAM-L (σ=6.42)
TECE-OAM-RRA-1.0 (σ=7.44)
DPA-4.0.1-Pro-MPtrj (σ=8.44)

Difference in number of symmetry operations of ML vs DFT-relaxed structures. Models are sorted by the standard deviation σ of ΔNsym ops = Nsym ops,ML - Nsym ops,DFT.

Sankey Diagrams for ML vs DFT Spacegroups

The Sankey diagrams show corresponding spacegroups of DFT-relaxed and MLFF-relaxed structures at symprec=1e-5. For visual clarity, only the 10 most common pairs of (DFT, MLFF) spacegroups are shown.

Relaxation-protocol caveat

The WBM DFT references were generated with MPRelaxSet: ISYM=2, ISIF=3, and IBRION=2 (conjugate gradient). Most MLFF relaxations instead use FIRE. Different symmetry constraints and optimizers can reach different minima, so a lower symmetry-match rate can occasionally indicate a valid symmetry-broken structure rather than a model error. See MPRelaxSet.yaml. Thanks to Alex Ganose for highlighting this distinction.