model_name: GNoME
model_key: gnome
model_version: n/a
date_added: '2024-02-03'
date_published: '2023-11-29'
authors:
  - name: Amil Merchant
    affiliation: Google DeepMind
    email: amilmerchant@google.com
    orcid: https://orcid.org/0000-0001-5262-6599
  - name: Simon Batzner
    affiliation: Google DeepMind
  - name: Samuel S. Schoenholz
    affiliation: Google DeepMind
  - name: Muratahan Aykol
    affiliation: Google DeepMind
  - name: Gowoon Cheon
    affiliation: Google DeepMind
  - name: Ekin Dogus Cubuk
    affiliation: Google DeepMind
    email: cubuk@google.com
    orcid: https://orcid.org/0000-0003-0524-2837

repo: https://github.com/google-deepmind/materials_discovery
doi: https://doi.org/10.1038/s41586-023-06735-9
paper: https://nature.com/articles/s41586-023-06735-9
pr_url: https://github.com/janosh/matbench-discovery/pull/84
checkpoint_url: missing

license:
  code: Apache-2.0
  code_url: https://github.com/google-deepmind/materials_discovery/blob/be1fa43a9/LICENSE
  checkpoint: unreleased
  checkpoint_url: missing

openness: OSCD
trained_for_benchmark: false
train_task: S2EFS
test_task: IS2RE-SR
targets: EF_G
model_type: UIP
model_params: 16_240_000
n_estimators: 1

requirements:
  e3nn-jax: 0.20.3
  flax: 0.7.5
  jax-md: 0.2.8
  jax: 0.4.20
  numpy: 1.26.2
  pymatgen: 2023.11.12

training_set: [GNoME]
training_cost: missing

hyperparams:
  optimizer: Adam
  learning_rate: 2e-3
  batch_size: 32
  n_layers: 5
  n_features: 128 l=0 scalars, 64 l=1 vectors, 32 l=2 tensors
  graph_construction_radius: 5.0 # Å
  max_neighbors: .inf

notes:
  Description: |
    GNoME is an equivariant Nequip-type graph neural network implemented in e3nn-jax.
  Training: Using pre-trained model released with "Scaling deep learning for materials discovery" paper. Training set unpublished as of 2024-02-03. The model was trained 1.5 years prior to submission to Matbench Discovery according to private communication.
  Missing Preds: According to the authors, the 1734 missing WBM predictions are mostly due out-of-memory (OOM) errors. The model was evaluated on A100s but without neighbor lists. The plan is to backfill the missing predictions once H100s are available or neighbor list implementation goes live.

metrics:
  phonons: not available # model is closed source, original GnoME submission predates phonon tasks and DeepMind did not resubmit later
  geo_opt: not available # author's declined to share model-relaxed structures and can't be reproduced without model access
  discovery:
    pred_file: models/gnome/2023-11-01-gnome-50076332-wbm-IS2RE.csv.gz
    pred_file_url: https://figshare.com/files/52057556
    pred_col: e_gnome_after_relax
    full_test_set:
      F1: 0.81 # fraction
      DAF: 4.81 # dimensionless
      Precision: 0.825 # fraction
      Recall: 0.796 # fraction
      Accuracy: 0.936 # fraction
      TPR: 0.796 # fraction
      FPR: 0.035 # fraction
      TNR: 0.965 # fraction
      FNR: 0.204 # fraction
      TP: 35082.0 # count
      FP: 7421.0 # count
      TN: 205450.0 # count
      FN: 9010.0 # count
      MAE: 0.034 # eV/atom
      RMSE: 0.083 # eV/atom
      R2: 0.786 # dimensionless
      missing_preds: 1744 # count
    most_stable_10k:
      F1: 0.967 # fraction
      DAF: 6.127 # dimensionless
      Precision: 0.937 # fraction
      Recall: 1.0 # fraction
      Accuracy: 0.937 # fraction
      TPR: 1.0 # fraction
      FPR: 1.0 # fraction
      TNR: 0.0 # fraction
      FNR: 0.0 # fraction
      TP: 9366.0 # count
      FP: 634.0 # count
      TN: 0.0 # count
      FN: 0.0 # count
      MAE: 0.035 # eV/atom
      RMSE: 0.089 # eV/atom
      R2: 0.836 # dimensionless
      missing_preds: 0 # count
    unique_prototypes:
      F1: 0.829 # fraction
      DAF: 5.523 # dimensionless
      Precision: 0.844 # fraction
      Recall: 0.814 # fraction
      Accuracy: 0.948 # fraction
      TPR: 0.814 # fraction
      FPR: 0.028 # fraction
      TNR: 0.972 # fraction
      FNR: 0.186 # fraction
      TP: 27178.0 # count
      FP: 5009.0 # count
      TN: 177105.0 # count
      FN: 6196.0 # count
      MAE: 0.035 # eV/atom
      RMSE: 0.085 # eV/atom
      R2: 0.785 # dimensionless
      missing_preds: 1517 # count