adosar/retnet

A 3D ConvNet for Reticular Chemistry.

💎 RetNet

A 3D ConvNet that takes energy voxels as input and ouputs gas adsorption properties.

This repository contains a PyTorch implementation of RetNet, which can be found on model.py module.

🚀 Training RetNet

The following example is used to train RetNet on the University of Ottawa database¹ for predicting CO₂ uptake.

Important

It is strongly recommended to run all the scripts inside a virtual environment.

Clone the repository

git clone https://github.com/frudakis-research-group/retnet
cd retnet

Dependencies

1. Create a virtual environment:

   python -m venv <venvir_name>

2. Activate it:

   source <venvir_name>/bin/activate

3. Install the dependencies:

   (<venvir_name>) pip install -r requirements.txt

Collect the data

The following directory structure is required prior to training:

data/
├── MOFs
   ├── all_MOFs_screening_data.csv
   ├── batch_train
   │   ├── clean_names.json
   │   └── clean_voxels.npy
   └── batch_val_test
       ├── clean_names.json
       └── clean_voxels.npy

To achieve that:

1. Get the inputs:

   wget -O- 'https://figshare.com/ndownloader/files/43220463' | tar -xzvf-

Warning

If you use any of this data in your research work, you should cite the original work².

2. Get the labels:

   wget -O- 'https://archive.materialscloud.org/record/file?filename=screening_data.tar.gz&record_id=62' | tar -xzvf- -C data/MOFs

Warning

If you use any of this data in your research work, you should cite the original work¹.

Train the model

Check the comments in training.py to customize the training phase on your needs.

(<venvir_name>) python training.py

• GPU training time: 40s per epoch on Nvidia GTX 1650 Super
• CPU training time: 257s per epoch on Intel i5 8400

Tip

If you want to use a GPU but the VRAM is not enough:

1. Decrease the training batch size to a value smaller than 64.
2. Remove some Conv layers or decrease the number of Conv filters.

📰 Citing

If you use the RetNet architecture in your research work or any of the scripts of this repository, please consider citing:

@article{Sarikas2024,
  title = {Gas adsorption meets deep learning: voxelizing the potential energy surface of metal-organic frameworks},
  volume = {14},
  ISSN = {2045-2322},
  url = {http://dx.doi.org/10.1038/s41598-023-50309-8},
  DOI = {10.1038/s41598-023-50309-8},
  number = {1},
  journal = {Scientific Reports},
  publisher = {Springer Science and Business Media LLC},
  author = {Sarikas,  Antonios P. and Gkagkas,  Konstantinos and Froudakis,  George E.},
  year = {2024},
  month = jan 
}

Footnotes

1. Boyd, P.G., Chidambaram, A., García-Díez, E. et al. Data-driven design of metal–organic frameworks for wet flue gas CO₂ capture. Nature 576, 253–256 (2019). https://doi.org/10.1038/s41586-019-1798-7 ↩ ↩²

2. Sarikas, A.P., Gkagkas, K. & Froudakis, G.E. Gas adsorption meets deep learning: voxelizing the potential energy surface of metal-organic frameworks. Sci Rep 14, 2242 (2024). https://doi.org/10.1038/s41598-023-50309-8 ↩