model_name: CGCNN model_key: cgcnn model_version: v0.1.0 # the aviary version date_added: '2022-12-28' date_published: '2017-10-27' authors: - name: Tian Xie email: txie@csail.mit.edu affiliation: Massachusetts Institute of Technology url: https://txie.me - name: Jeffrey C. Grossman affiliation: Massachusetts Institute of Technology url: https://dmse.mit.edu/people/faculty/jeffrey-grossman 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 repo: https://github.com/CompRhys/aviary doi: https://doi.org/10.1103/PhysRevLett.120.145301 paper: https://arxiv.org/abs/1710.10324 pr_url: https://github.com/janosh/matbench-discovery/pull/85 # submission used an ensemble of 10 models, URL is just the first checkpoint checkpoint_url: https://api.wandb.ai/files/janosh/matbench-discovery/cvrqqjf1/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 requirements: aviary: https://github.com/CompRhys/aviary/releases/tag/v0.1.0 torch: 1.11.0 torch-scatter: 2.0.9 numpy: 1.24.0 pandas: 1.5.1 openness: OSOD train_task: RS2RE test_task: IS2E targets: E model_type: GNN model_params: 128_450 trained_for_benchmark: true n_estimators: 10 training_set: [MP 2022] training_cost: missing hyperparams: graph_construction_radius: 5.0 # Å, from https://github.com/CompRhys/aviary/blob/451f5739/aviary/cgcnn/data.py#L28 max_neighbors: .inf # CGCNN paper benchmarks on 6.0 Å against 12 NNs graph construction notes: Description: | Published in 2018, CGCNN was the first crystal graph convolutional neural network to directly learn 8 different DFT-computed material properties from a graph representing the atoms and bonds in a crystal. ![Illustration of the crystal graph convolutional neural networks](https://researchgate.net/profile/Tian-Xie-11/publication/320726915/figure/fig1/AS:635258345119746@1528468800829/Illustration-of-the-crystal-graph-convolutional-neural-networks-a-Construction-of-the.png) Aviary CGCNN model is based on the original implementation in https://github.com/txie-93/cgcnn. Long: CGCNN was among the first to show that just like in other areas of ML, given large enough training sets, neural networks can learn embeddings that reliably outperform all human-engineered structure features directly from the data. metrics: phonons: not applicable # model doesn't predict forces geo_opt: not applicable # model doesn't predict forces discovery: pred_file: models/cgcnn/2023-01-26-cgcnn-ens=10-wbm-IS2RE.csv.gz pred_file_url: https://figshare.com/files/51607271 pred_col: e_form_per_atom_mp2020_corrected_pred_ens full_test_set: F1: 0.51 # fraction DAF: 2.631 # dimensionless Precision: 0.451 # fraction Recall: 0.587 # fraction Accuracy: 0.807 # fraction TPR: 0.587 # fraction FPR: 0.148 # fraction TNR: 0.852 # fraction FNR: 0.413 # fraction TP: 25895.0 # count FP: 31474.0 # count TN: 181397.0 # count FN: 18197.0 # count MAE: 0.135 # eV/atom RMSE: 0.229 # eV/atom R2: -0.624 # dimensionless missing_preds: 4 # count most_stable_10k: F1: 0.745 # fraction DAF: 3.88 # dimensionless Precision: 0.593 # fraction Recall: 1.0 # fraction Accuracy: 0.593 # fraction TPR: 1.0 # fraction FPR: 1.0 # fraction TNR: 0.0 # fraction FNR: 0.0 # fraction TP: 5931.0 # count FP: 4069.0 # count TN: 0.0 # count FN: 0.0 # count MAE: 0.165 # eV/atom RMSE: 0.23 # eV/atom R2: 0.181 # dimensionless missing_preds: 0 # count unique_prototypes: F1: 0.507 # fraction DAF: 2.855 # dimensionless Precision: 0.436 # fraction Recall: 0.605 # fraction Accuracy: 0.818 # fraction TPR: 0.605 # fraction FPR: 0.143 # fraction TNR: 0.857 # fraction FNR: 0.395 # fraction TP: 20191.0 # count FP: 26073.0 # count TN: 156041.0 # count FN: 13183.0 # count MAE: 0.138 # eV/atom RMSE: 0.233 # eV/atom R2: -0.603 # dimensionless missing_preds: 2 # count