3D Convolutional Neural Networks Utilizing Molecular Topological Features for Accurate Atomization Energy Predictions

Requirements

• python 3.8
• numpy
• natsort
• torch-summary
• scikit-learn
• PyTorch Lightning (version 1.2.0rc0)
• Multiwfn
• Gaussian 16

Note: We recommend using Anaconda for easy installation of libraries.

Data

The QM9-G4MP2 dataset is publicly available through Materials Data Facility. The geometries (xyz files) and energies of the molecules were extracted from the dataset, which were then processed using the scripts in this repository.

Data Preparation

1. Extract xyz files of the molecules from the QM9-G4MP2 dataset. Create two directories to store training and test set data. Move the training and test set coordinate files to their respective directories. Now, carry out the following steps inside the training and test set directories.
2. Generate a unique molecular orientation and prepare a corresponding Gaussian input file.
   python gen_unique_xyz.py --wfx_path "/wfx_file_directory_path/"
   The user defined string argument represents the path of the directory, where the wfx files would be stored. For the sake of convenience, keep the wfx directory path the same as in step-1 for all the training/test set molecules.
3. Run the input files generated in step-2 using Gaussian to obtain the wfx files.
4. Generate requisite volumetric properties from wfx files using Multiwfn.
   ./calc_3dprop.sh
5. Prepare training and test datasets using the property files generated in step-4.
   python make_inp.py --data_split "train"
python make_inp.py --data_split "test"
6. Place the training and test datasets (pickle files) generated in step-5 in a directory of your choice. Also, place the target labels (g4mp2_b3lyp_diff_labels.pickle) provided in the data folder in the same directory.

Note: For the sake of convenience, all the processing steps need to be carried out separately for training and test set molecules.

Model Training

Train the model using the following command.
python train_pl.py --data_path "/datasets_directory_path/" --channel 2 --grid_length 14 --batch_size 32 --epochs 250 --dense1 16 --dense2 16 > results.txt 2> errors.txt

Arguments:
data_path: path of the directory where input and output data are stored
channel: one of 0, 1, 2, or 3
0: Nuclear Electrostatic potential (NEP)
1: Electron Localization Function (ELF)
2: Localized Orbital Locator (LOL)
3: Electrostatic Potential (ESP)
grid_length: Voxel grid length of the cubic volume
dense1: Depth of the first dense block
dense2: Depth of the second dense block

Note: We recommend using GPUs for model training.

References

1. DenseNet model adapted from https://github.com/gpleiss/efficient_densenet_pytorch