krict-data/configstabilityml

ConfigStabilityML

** All geometries are obtained from IS2RE data (available in OC20 dataset web-page)

** For data labeling, "oc20_data_mapping.pkl" should exist (available in here)

1. Code for ML model

• cgcc.py (CGCNN)
• schnet.py (SchNet)
• mpnn.py (MPNN)
• mpnn_a_mha.py (MPNN-A)
• matformer_mha.py (Matformer)

2. Code for model training/prediction

** All trained model parameters could be found at here.

• script_train.py [model_type] [num_dat] [random_seed]; it will generate directory name of [model_type] under directory RandomSeed_[random_seed], same as [path_pref] used in script_predict.py
• script_predict.py [model_type] [is_test] [path_pref] [num_dat]; it will generate Predicted_[num_dat]_{id, ood_ads, ood_cat or ood_both}.{test or val}.pkl file under the path of [pref_path]/[model_type]

** List of arguments

• [model_type]; CGCNN,SchNet,MPNN,MPNN_A,Matformer
• [num_dat]; the number of training set (maximum # of training data = 460328)
• [random_seed]; seed number used for random sampling of training set with size of [num_dat]
• [is_test]; 1 for predicting test set data and 0 for predicting validation set data
• [path_pref]; directory path including directory with name of [model_type] (i.e. RandomSeed_[random_seed])

3. 'ReproduceData' directory contains all scripts to reproduce all data plotted in this work

4. System information

• python == 3.9
• sklearn == 1.0.2
• numpy == 1.21.5
• pytorch == 1.13.0 (CUDA verison == 11.7)
• torch_scatter == 2.0.9 (for scatter operation)
• V100 32GB

** For visualization (with Spyder)

• matplotlib == 3.5.1
• seaborn == 0.11.2