model_name: EquiformerV3+DeNS-MP
model_key: equiformer_v3_mp
model_version: v2026.04.07
date_added: '2026-04-07'
date_published: '2026-04-07'
authors:
  - name: Yi-Lun Liao
    affiliation: Massachusetts Institute of Technology
    email: ylliao@mit.edu
    github: https://github.com/yilunliao
    corresponding: true
  - name: Alexander J. Hoffman
    affiliation: Mirror Physics
  - name: Sabrina C. Shen
    affiliation: Mirror Physics
  - name: Sam Walton Norwood
    affiliation: Mirror Physics
    email: sam@mirrorphysics.com
    corresponding: true
  - name: Tess Smidt
    affiliation: Massachusetts Institute of Technology
    email: tsmidt@mit.edu
    url: https://blondegeek.github.io/
    orcid: https://orcid.org/0000-0001-5581-5344
    corresponding: true


trained_by:
  - name: Yi-Lun Liao
    affiliation: Massachusetts Institute of Technology
    email: ylliao@mit.edu


repo: https://github.com/atomicarchitects/equiformer_v3
doi: https://doi.org/10.48550/arXiv.2604.09130
paper: https://arxiv.org/abs/2604.09130
url: https://github.com/atomicarchitects/equiformer_v3
pr_url: https://github.com/janosh/matbench-discovery/pull/320
checkpoint_url: https://github.com/atomicarchitects/equiformer_v3


license:
  code: MIT
  code_url: https://github.com/atomicarchitects/equiformer_v3
  checkpoint: MIT
  checkpoint_url: https://github.com/atomicarchitects/equiformer_v3


openness: OSOD
trained_for_benchmark: true
train_task: S2EFS
test_task: IS2RE-SR
targets: EFS_G
model_type: UIP
model_params: 30_278_113
n_estimators: 1


training_set: [MPtrj]


training_cost:
  Nvidia H100 GPUs: {amount: 16, hours: 66.875}


hyperparams:
  max_force: 0.02
  max_steps: 500
  ase_optimizer: FIRE
  cell_filter: FrechetCellFilter
  graph_construction_radius: 6.0
  max_neighbors: .inf

  num_blocks: 7
  lmax: 4
  mmax: 2
  embedding_dimension: 128
  attention_value_hidden_size: 32
  num_attention_heads: 8
  attention_alpha_hidden_size: 64
  attention_value_size_per_head: 16
  ffn_hidden_size_after_swiglu_s2_activation: 512
  grid_resolution_attention: (14, 8)
  grid_resolution_ffn: (14, 14)

  DeNS_probability: 0.5
  DeNS_coefficient: 10
  DeNS_noise_std: 0.025
  DeNS_corruption_ratio: 0.5


notes:
  Description: EquiformerV3 is the third generation of SE(3)-equivariant graph attention Transformers. Please refer to our paper for more details.


requirements:
  ase: 3.25.0
  ase-db-backends: 0.10.0
  e3nn: 0.5.6
  tqdm: 4.67.1
  wandb: 0.21.0
  pyyaml: 6.0.2
  torch: 2.7.1


metrics:
  phonons:
    kappa_103:
      pred_file: models/equiformer_v3/equiformer-v3-mp/2026-04-07-kappa-103-FIRE-dist=0.03-fmax=1e-4-symprec=1e-5.json.gz
      pred_file_url: https://figshare.com/files/63515175
      κ_SRME: 0.2753
      κ_SRE: 0.1136
  geo_opt:
    pred_file: models/equiformer_v3/equiformer-v3-mp/2026-04-07-wbm-IS2RE-FIRE.json.gz
    pred_file_url: https://figshare.com/files/63515178
    struct_col: structure
    symprec=1e-5:
      rmsd: 0.0701 # unitless
      n_sym_ops_mae: 3.3334 # unitless
      symmetry_decrease: 0.4223 # fraction
      symmetry_match: 0.3919 # fraction
      symmetry_increase: 0.1308 # fraction
      n_structures: 256963 # count
      analysis_file: models/equiformer_v3/equiformer-v3-mp/2026-04-07-wbm-IS2RE-FIRE-symprec=1e-5-moyo=0.7.9.csv.gz
      analysis_file_url: https://figshare.com/files/63520452
    symprec=1e-2:
      rmsd: 0.0701 # unitless
      n_sym_ops_mae: 2.4068 # unitless
      symmetry_decrease: 0.2301 # fraction
      symmetry_match: 0.6413 # fraction
      symmetry_increase: 0.1022 # fraction
      n_structures: 256963 # count
      analysis_file: models/equiformer_v3/equiformer-v3-mp/2026-04-07-wbm-IS2RE-FIRE-symprec=1e-2-moyo=0.7.9.csv.gz
      analysis_file_url: https://figshare.com/files/63520449
  discovery:
    pred_file: models/equiformer_v3/equiformer-v3-mp/2026-04-07-wbm-IS2RE.csv.gz
    pred_file_url: https://figshare.com/files/63515172
    pred_col: e_form_per_atom_fairchem
    full_test_set:
      F1: 0.848 # fraction
      DAF: 4.778 # dimensionless
      Precision: 0.82 # fraction
      Recall: 0.878 # fraction
      Accuracy: 0.946 # fraction
      TPR: 0.878 # fraction
      FPR: 0.04 # fraction
      TNR: 0.96 # fraction
      FNR: 0.122 # fraction
      TP: 38706.0 # count
      FP: 8508.0 # count
      TN: 204363.0 # count
      FN: 5386.0 # count
      MAE: 0.028 # eV/atom
      RMSE: 0.073 # eV/atom
      R2: 0.836 # dimensionless
      missing_preds: 2 # count
    unique_prototypes:
      F1: 0.863 # fraction
      DAF: 5.479 # dimensionless
      Precision: 0.838 # fraction
      Recall: 0.89 # fraction
      Accuracy: 0.956 # fraction
      TPR: 0.89 # fraction
      FPR: 0.032 # fraction
      TNR: 0.968 # fraction
      FNR: 0.11 # fraction
      TP: 29708.0 # count
      FP: 5759.0 # count
      TN: 176355.0 # count
      FN: 3666.0 # count
      MAE: 0.029 # eV/atom
      RMSE: 0.074 # eV/atom
      R2: 0.84 # dimensionless
      missing_preds: 0 # count
    most_stable_10k:
      F1: 0.989 # fraction
      DAF: 6.396 # dimensionless
      Precision: 0.978 # fraction
      Recall: 1.0 # fraction
      Accuracy: 0.978 # fraction
      TPR: 1.0 # fraction
      FPR: 1.0 # fraction
      TNR: 0.0 # fraction
      FNR: 0.0 # fraction
      TP: 9777.0 # count
      FP: 223.0 # count
      TN: 0.0 # count
      FN: 0.0 # count
      MAE: 0.021 # eV/atom
      RMSE: 0.069 # eV/atom
      R2: 0.895 # dimensionless
      missing_preds: 0 # count