model_name: Wrenformer
model_key: wrenformer
model_version: v0.1.0 # the aviary version
date_added: '2022-11-26'
date_published: '2021-06-21'
authors:
  - name: Janosh Riebesell
    affiliation: University of Cambridge, Lawrence Berkeley National Laboratory
    email: janosh.riebesell@gmail.com
    orcid: https://orcid.org/0000-0001-5233-3462
  - name: Rhys Goodall
    affiliation: University of Cambridge
    orcid: https://orcid.org/0000-0002-6589-1700
  - name: Rokas Elijošius
    affiliation: University of Cambridge
    email: re344@cam.ac.uk
    orcid: https://orcid.org/0000-0001-6397-0002
trained_by:
  - name: Janosh Riebesell
    affiliation: University of Cambridge, Lawrence Berkeley National Laboratory
    email: janosh.riebesell@gmail.com
    orcid: https://orcid.org/0000-0001-5233-3462

requirements:
  aviary: https://github.com/CompRhys/aviary/releases/tag/v0.1.0
  torch: 1.11.0
  torch-scatter: 2.0.9
  pymatgen: 2022.10.22
  numpy: 1.24.0
  pandas: 1.5.1

repo: https://github.com/CompRhys/aviary
doi: https://doi.org/10.1126/sciadv.abn4117
paper: https://arxiv.org/abs/2106.11132
pr_url: https://github.com/janosh/matbench-discovery/pull/85
checkpoint_url: https://api.wandb.ai/files/janosh/matbench-discovery/2kozbp4q/checkpoint.pth

license:
  code: MIT
  code_url: https://github.com/CompRhys/aviary/blob/3238fb415/LICENSE
  checkpoint: MIT
  checkpoint_url: https://github.com/janosh/matbench-discovery/blob/7c0b089e7/license

openness: OSOD
trained_for_benchmark: true
train_task: RP2RE
test_task: IP2E
targets: E
model_type: Transformer
model_params: 5_166_658
n_estimators: 10

training_set: [MP 2022]
training_cost: missing

notes:
  Description: |
    Wrenformer is a standard PyTorch Transformer Encoder trained to learn material embeddings from composition, space group, Wyckoff positions in a structure.
    ![Model workings](https://science.org/cms/10.1126/sciadv.abn4117/asset/a29e0899-77d1-47c8-82e3-00ab87c3b8d5/assets/images/large/sciadv.abn4117-f1.jpg)
    A ML–powered materials discovery workflow using Wrenformer's Wyckoff string inputs to predict formation energies for candidate materials in an enumerated library of Wyckoff representations (shapes are used to denote different Wyckoff positions and colors to denote different element types). Predicted formation energies are then compared against the known convex hull of stability. Structures satisfying the required symmetries are relaxed for materials predicted to be stable.

  Long: It builds on [Roost](https://doi.org/10.1038/s41467-020-19964-7) and [Wren](https://doi.org/10.1126/sciadv.abn4117), by being a fast structure-free model that is still able to distinguish polymorphs through symmetry.

metrics:
  phonons: not applicable # model doesn't predict forces
  geo_opt: not applicable
  discovery:
    pred_file: models/wrenformer/wrenformer-ens=10/2022-11-15-wbm-IS2RE.csv.gz
    pred_file_url: https://figshare.com/files/52057553
    pred_col: e_form_per_atom_wrenformer_pred_ens
    full_test_set:
      F1: 0.479 # fraction
      DAF: 2.13 # dimensionless
      Precision: 0.365 # fraction
      Recall: 0.693 # fraction
      Accuracy: 0.741 # fraction
      TPR: 0.693 # fraction
      FPR: 0.249 # fraction
      TNR: 0.751 # fraction
      FNR: 0.307 # fraction
      TP: 30566.0 # count
      FP: 53076.0 # count
      TN: 159795.0 # count
      FN: 13526.0 # count
      MAE: 0.105 # eV/atom
      RMSE: 0.182 # eV/atom
      R2: -0.02 # dimensionless
      missing_preds: 7 # count
    most_stable_10k:
      F1: 0.721 # fraction
      DAF: 3.691 # dimensionless
      Precision: 0.564 # fraction
      Recall: 1.0 # fraction
      Accuracy: 0.564 # fraction
      TPR: 1.0 # fraction
      FPR: 1.0 # fraction
      TNR: 0.0 # fraction
      FNR: 0.0 # fraction
      TP: 5642.0 # count
      FP: 4358.0 # count
      TN: 0.0 # count
      FN: 0.0 # count
      MAE: 0.182 # eV/atom
      RMSE: 0.239 # eV/atom
      R2: 0.138 # dimensionless
      missing_preds: 0 # count
    unique_prototypes:
      F1: 0.466 # fraction
      DAF: 2.256 # dimensionless
      Precision: 0.345 # fraction
      Recall: 0.719 # fraction
      Accuracy: 0.745 # fraction
      TPR: 0.719 # fraction
      FPR: 0.25 # fraction
      TNR: 0.75 # fraction
      FNR: 0.281 # fraction
      TP: 23992.0 # count
      FP: 45575.0 # count
      TN: 136539.0 # count
      FN: 9382.0 # count
      MAE: 0.11 # eV/atom
      RMSE: 0.186 # eV/atom
      R2: -0.018 # dimensionless
      missing_preds: 5 # count