bsgnn is a GNN model for predicting superconductivity based on the CrystalNet[1] framework.
In our Model, we add three modules into bsgnn: nearest-neighbors-only graph represent (NGR), communicative message passing (CMP) and attention (GAT) module. First, CGR module represents the ordered and disordered crystal structures as periodic graphs. Specifically, each node is represented by an initial feature vector that collected from the atom fingerprint, each edge is also represented by a raw feature vector, corresponding to the bond connecting two atoms. Here, only the nearest neighboring nodes were connected with central nodes and lattice distortion were taken into consideration, which allows the graph represents crystal structure correctly. Second, CMP module mimics the complex physical and chemical interactions between atoms and bonds. where the message interactions were strengthened between atoms and bonds through communicative message passing. Third, attention module give different weights to neighboring nodes during message passing.
numpy 1.20.2
pandas 1.2.4
pymatgen 2020.12.18
pyparsing 2.4.7
scikit-learn 0.24.1
scipy 1.6.3
torch 1.5.0+cu101
torch-cluster 1.5.5
torch-geometric 1.5.0
torch-scatter 2.0.5
torch-sparse 0.6.6
torch-spline-conv 1.2.0
torchaudio 0.5.0
torchvision 0.6.0+cu101
tornado 6.1
tqdm 4.60.0
Specified the fowllowing files path in proprecess.py
- cif_files.csv
- data_cif
And then run proprecess.py.
python -u train_all.py --seed 4 --data_path ./data --dataset_type regression --metric r2 --save_dir ./test --epochs 100 --init_lr 1e-5 --max_lr 1e-4 --final_lr 1e-5 --no_features_scaling --show_individual_scores --max_num_neighbors 64
python -u predict.py --test_path ./data/predict.csv --checkpoint_dir ./test --preds_path ./result/predict_result.csv
[1]. Chen P, Chen J, Yan H, et al. Improving Material Property Prediction by Leveraging the Large-Scale Computational Database and Deep Learning[J]. The Journal of Physical Chemistry C, 2022, 126(38): 16297-16305.