model_name: EquFlash
model_key: equflash-29M-oam
model_version: v2025.06.23
date_added: '2025-06-23'
date_published: '2025-06-23'
authors:
  - name: Hyuntae Cho
    affiliation: Materials AI Lab at Samsung Electronics
    email: robert.cho@samsung.com
    orcid: https://orcid.org/0000-0001-9379-6877
  - name: Saerom Choi
    affiliation: Materials AI Lab at Samsung Electronics
    email: sincere.choi@samsung.com
  - name: Heejae Kim
    affiliation: Materials AI Lab at Samsung Electronics
  - name: Jaehee Jang
    affiliation: Materials AI Lab at Samsung Electronics
  - name: Gunhee Kim
    affiliation: Materials AI Lab at Samsung Electronics
  - name: Heesun Lee
    affiliation: Materials AI Lab at Samsung Electronics
  - name: Hyunwoo Lee
    affiliation: Materials AI Lab at Samsung Electronics
  - name: Yongdeok Kim
    affiliation: Materials AI Lab at Samsung Electronics
    email: yd.mlg.kim@samsung.com

repo: https://github.com/SNU-ARC/flashTP
doi: TBD
paper: https://openreview.net/pdf?id=wiQe95BPaB
pr_url: https://github.com/janosh/matbench-discovery/pull/269
checkpoint_url: missing
license:
  code: MIT
  code_url: https://github.com/SNU-ARC/flashTP/blob/main/LICENSE
  checkpoint: unreleased

requirements:
  flashTP_e3nn: 0.1.0
  torch: 2.8.0+cu126
  fairchem-core: 1.10.0

openness: CSOD
trained_for_benchmark: false
train_task: S2EFS
test_task: IS2RE-SR
targets: EFS_G
model_type: UIP
model_params: 28_741_188
n_estimators: 1

training_set: [OMat24, MPtrj, sAlex]
training_cost: missing
hyperparams:
  max_force: 0.02
  max_steps: 500
  ase_optimizer: FIRE
  cell_filter: FrechetCellFilter
  graph_construction_radius: 6.0 # Å
  max_neighbors: .inf
notes:
  Description: |
    EquFlash is an E(3)-equivariant model based on the SevenNet-0 architecture, with tensor products accelerated by FlashTP.
    FlashTP achieves up to 41.6× and 60.8× kernel speedups over e3nn and NVIDIA cuEquivariance, respectively, while reducing memory usage by 6×.
    Leveraging these gains, we scaled EquFlash to a larger capacity than the original SevenNet-0.
  Training: EquFlash, a scaled-up model derived from SevenNet-0 and       accelerated with FlashTP, was pretrained on OMat24 and finetuned on MPtrj and sAlex.

metrics:
  phonons:
    kappa_103:
      pred_file: models/equflash/2025-06-23-kappa-103-FIRE-dist=0.03-fmax=1e-4-symprec=1e-5.json.gz
      pred_file_url: https://figshare.com/files/57277382
      κ_SRME: 0.1583
      κ_SRE: 0.0772
  geo_opt:
    pred_file: models/equflash/2025-06-23-wbm-geo-opt.jsonl.gz
    pred_file_url: https://figshare.com/files/57277370
    struct_col: pred_relaxed_struct
    symprec=1e-5:
      rmsd: 0.0602 # unitless
      n_sym_ops_mae: 2.0595 # unitless
      symmetry_decrease: 0.0489 # fraction
      symmetry_match: 0.7019 # fraction
      symmetry_increase: 0.244 # fraction
      n_structures: 256963 # count
      analysis_file: models/equflash/2025-06-23-wbm-geo-opt-symprec=1e-5-moyo=0.4.4.csv.gz
      analysis_file_url: https://figshare.com/files/57277376
    symprec=1e-2:
      rmsd: 0.0602 # unitless
      n_sym_ops_mae: 1.6926 # unitless
      symmetry_decrease: 0.0468 # fraction
      symmetry_match: 0.8198 # fraction
      symmetry_increase: 0.1264 # fraction
      n_structures: 256963 # count
      analysis_file: models/equflash/2025-06-23-wbm-geo-opt-symprec=1e-2-moyo=0.4.4.csv.gz
      analysis_file_url: https://figshare.com/files/57277379
  discovery:
    pred_file: models/equflash/2025-06-23-wbm-IS2RE.csv.gz
    pred_file_url: https://figshare.com/files/57277367
    pred_col: pred_formation_energy_per_atom
    full_test_set:
      F1: 0.9 # fraction
      DAF: 5.214 # dimensionless
      Precision: 0.895 # fraction
      Recall: 0.906 # fraction
      Accuracy: 0.966 # fraction
      TPR: 0.906 # fraction
      FPR: 0.022 # fraction
      TNR: 0.978 # fraction
      FNR: 0.094 # fraction
      TP: 39932.0 # count
      FP: 4699.0 # count
      TN: 208172.0 # count
      FN: 4160.0 # count
      MAE: 0.019 # eV/atom
      RMSE: 0.066 # eV/atom
      R2: 0.864 # dimensionless
      missing_preds: 2 # count
    unique_prototypes:
      F1: 0.919 # fraction
      DAF: 5.983 # dimensionless
      Precision: 0.915 # fraction
      Recall: 0.922 # fraction
      Accuracy: 0.975 # fraction
      TPR: 0.922 # fraction
      FPR: 0.016 # fraction
      TNR: 0.984 # fraction
      FNR: 0.078 # fraction
      TP: 30786.0 # count
      FP: 2875.0 # count
      TN: 179239.0 # count
      FN: 2588.0 # count
      MAE: 0.019 # eV/atom
      RMSE: 0.066 # eV/atom
      R2: 0.871 # dimensionless
      missing_preds: 0 # count
    most_stable_10k:
      F1: 0.985 # fraction
      DAF: 6.353 # dimensionless
      Precision: 0.971 # fraction
      Recall: 1.0 # fraction
      Accuracy: 0.971 # fraction
      TPR: 1.0 # fraction
      FPR: 1.0 # fraction
      TNR: 0.0 # fraction
      FNR: 0.0 # fraction
      TP: 9712.0 # count
      FP: 288.0 # count
      TN: 0.0 # count
      FN: 0.0 # count
      MAE: 0.017 # eV/atom
      RMSE: 0.067 # eV/atom
      R2: 0.902 # dimensionless
      missing_preds: 0 # count