model_name: ALIGNN
model_key: alignn
model_version: 2023.01.10
date_added: '2023-06-02'
date_published: '2021-02-22'
authors:
  - name: Kamal Choudhary
    affiliation: National Institute of Standards and Technology
    email: kamal.choudhary@nist.gov
    orcid: https://orcid.org/0000-0001-9737-8074
  - name: Brian DeCost
    affiliation: National Institute of Standards and Technology
    orcid: https://orcid.org/0000-0002-3459-5888
trained_by:
  - name: Philipp Benner
    affiliation: German Federal Institute of Materials Research and Testing (BAM)
    orcid: https://orcid.org/0000-0002-0912-8137
    github: https://github.com/pbenner

repo: https://github.com/usnistgov/alignn
url: https://jarvis.nist.gov/jalignn
doi: https://nature.com/articles/s41524-021-00650-1
paper: https://arxiv.org/abs/2106.01829
pypi: https://pypi.org/project/alignn
pr_url: https://github.com/janosh/matbench-discovery/pull/85
# ALIGNN model 2023-02-07-alignn-checkpoint.pth trained on mp_computed_structure_entries
checkpoint_url: https://figshare.com/files/40344436

license:
  code: MIT
  code_url: https://github.com/usnistgov/alignn/blob/408bb6e996/LICENSE.rst
  checkpoint: CC-BY-4.0
  checkpoint_url: https://figshare.com/articles/dataset/Matbench_Discovery_v1_0_0/22715158?file=41233560

requirements:
  ase: 3.22.0
  dgl-cu111: 0.6.1
  numpy: 1.24.3
  pandas: 2.0.1
  scikit-learn: 1.2.2
  torch: 1.9.0+cu111

openness: OSOD
trained_for_benchmark: true
model_type: GNN
train_task: RS2RE
test_task: IS2E
targets: E
model_params: 4_026_753 # pre-trained 'mp_e_form_alignn' and our custom MBD checkpoint have the same size
# for other hyperparams, see alignn-config.json
n_estimators: 1

# model trained on Materials Project energies specifically for this submission
training_set: [MP 2022]
training_cost: missing

hyperparams:
  # taken from alignn/alignn-mp22/alignn-config.json
  graph_construction_radius: 8.0 # `cutoff` in Å, based on original implementation
  max_neighbors: 12

metrics:
  phonons: not applicable # reason: ALIGNN does not predict forces
  geo_opt: not applicable # reason: ALIGNN does not predict forces
  discovery:
    pred_file: models/alignn/alignn-mp22/2023-06-02-wbm-IS2RE.csv.gz
    pred_file_url: https://figshare.com/files/51607262
    pred_col: e_form_per_atom_alignn
    full_test_set:
      F1: 0.565 # fraction
      DAF: 2.921 # dimensionless
      Precision: 0.501 # fraction
      Recall: 0.649 # fraction
      Accuracy: 0.829 # fraction
      TPR: 0.649 # fraction
      FPR: 0.134 # fraction
      TNR: 0.866 # fraction
      FNR: 0.351 # fraction
      TP: 28598.0 # count
      FP: 28464.0 # count
      TN: 184407.0 # count
      FN: 15494.0 # count
      MAE: 0.092 # eV/atom
      RMSE: 0.154 # eV/atom
      R2: 0.274 # dimensionless
      missing_preds: 1 # count
    most_stable_10k:
      F1: 0.748 # fraction
      DAF: 3.905 # dimensionless
      Precision: 0.597 # fraction
      Recall: 1.0 # fraction
      Accuracy: 0.597 # fraction
      TPR: 1.0 # fraction
      FPR: 1.0 # fraction
      TNR: 0.0 # fraction
      FNR: 0.0 # fraction
      TP: 5969.0 # count
      FP: 4031.0 # count
      TN: 0.0 # count
      FN: 0.0 # count
      MAE: 0.176 # eV/atom
      RMSE: 0.247 # eV/atom
      R2: 0.081 # dimensionless
      missing_preds: 0 # count
    unique_prototypes:
      F1: 0.567 # fraction
      DAF: 3.206 # dimensionless
      Precision: 0.49 # fraction
      Recall: 0.672 # fraction
      Accuracy: 0.841 # fraction
      TPR: 0.672 # fraction
      FPR: 0.128 # fraction
      TNR: 0.872 # fraction
      FNR: 0.328 # fraction
      TP: 22436.0 # count
      FP: 23346.0 # count
      TN: 158768.0 # count
      FN: 10938.0 # count
      MAE: 0.093 # eV/atom
      RMSE: 0.154 # eV/atom
      R2: 0.297 # dimensionless
      missing_preds: 0 # count