model_name: ESNet
model_key: esnet
model_version: 2025.03.14
date_added: '2025-06-20'
date_published: '2025-06-20'
authors:
  - name: Chao Huang
    affiliation: Institute of Computing Technology, Chinese Academy of Science, Beijing
    email: chuang@ict.ac.cn
  - name: Chunyan Chen
    affiliation: Ningbo Institute of Artificial Intelligence Industry, Ningbo, China
  - name: Ling Shi
    affiliation: Ningbo Institute of Artificial Intelligence Industry, Ningbo, China

trained_by:
  - name: Ling Shi
    affiliation: Ningbo Institute of Artificial Intelligence Industry, Ningbo, China
    orcid: https://orcid.org/0000-0002-7185-5044
    github: https://github.com/zzz-sl/

repo: https://github.com/zzz-sl/ESNet
doi: https://doi.org/10.21203/rs.3.rs-5979703/v1
paper: https://www.researchsquare.com/article/rs-5979703/v1
pr_url: https://github.com/janosh/matbench-discovery/pull/254
checkpoint_url: https://figshare.com/files/55540055

license:
  code: Apache-2.0
  code_url: https://github.com/zzz-sl/ESNet/blob/main/LICENSE
  checkpoint: CC-BY-4.0
  checkpoint_url: https://figshare.com/files/55540055

requirements:
  torch: 2.1.0+cu121
  torch_sparse: 0.6.18
  torch_geometric: 2.6.1
  torch_scatter: 2.1.2
  pandarallel: 1.6.5
  pydantic_settings: 2.9.1
  e3nn: 0.5.6
  numpy: 1.26.4
  jarvis-tools: 2022.9.26
  einops: 0.8.1
  pymatgen: 2025.4.24
  pytorch-ignite: 0.5.2
  scikit-learn: 1.6.1

model_params: 5429033

openness: OSOD
model_type: Transformer
train_task: RS2RE
test_task: IS2E
targets: E
trained_for_benchmark: true

n_estimators: 1

training_set: [MP 2022]
training_cost:
  Nvidia A100 GPUs: { amount: 1, hours: 96 }

hyperparams:
  graph_construction_radius: 8.0
  max_neighbors: 25

notes:
  Description: |
    ESNet is a graph neural network model designed for predicting the energy.
    The model builds on existing models based on crystal structure graph, to provide an in-depth analysis of how material elemental composition and crystal structure work together to influence material properties.

metrics:
  phonons: not applicable
  geo_opt: not applicable
  discovery:
    pred_file: models/esnet/2025-06-20-esnet-wbm-IS2RE.csv.gz
    pred_file_url: https://figshare.com/files/55546568
    pred_col: e_form_per_atom_esnet
    full_test_set:
      F1: 0.568 # fraction
      DAF: 3.151 # dimensionless
      Precision: 0.541 # fraction
      Recall: 0.597 # fraction
      Accuracy: 0.844 # fraction
      TPR: 0.597 # fraction
      FPR: 0.105 # fraction
      TNR: 0.895 # fraction
      FNR: 0.403 # fraction
      TP: 26327.0 # count
      FP: 22360.0 # count
      TN: 190511.0 # count
      FN: 17765.0 # count
      MAE: 0.107 # eV/atom
      RMSE: 0.193 # eV/atom
      R2: -0.148 # dimensionless
      missing_preds: 1 # count
    most_stable_10k:
      F1: 0.879 # fraction
      DAF: 5.132 # dimensionless
      Precision: 0.784 # fraction
      Recall: 1.0 # fraction
      Accuracy: 0.784 # fraction
      TPR: 1.0 # fraction
      FPR: 1.0 # fraction
      TNR: 0.0 # fraction
      FNR: 0.0 # fraction
      TP: 7845.0 # count
      FP: 2155.0 # count
      TN: 0.0 # count
      FN: 0.0 # count
      MAE: 0.076 # eV/atom
      RMSE: 0.113 # eV/atom
      R2: 0.754 # dimensionless
      missing_preds: 0 # count
    unique_prototypes:
      F1: 0.572 # fraction
      DAF: 3.498 # dimensionless
      Precision: 0.535 # fraction
      Recall: 0.614 # fraction
      Accuracy: 0.857 # fraction
      TPR: 0.614 # fraction
      FPR: 0.098 # fraction
      TNR: 0.902 # fraction
      FNR: 0.386 # fraction
      TP: 20485.0 # count
      FP: 17823.0 # count
      TN: 164291.0 # count
      FN: 12889.0 # count
      MAE: 0.109 # eV/atom
      RMSE: 0.194 # eV/atom
      R2: -0.114 # dimensionless
      missing_preds: 0 # count