This project is dedicated to the prediction of critical properties of small organic molecules from their structure, using Neural Networks. Three critical parameters - Tc, Pc, omega can be calculated with workflow, presented in this project
Project has several folders. Every folder contains README which guides thourough the content of the folder:
- Datasets - here datasets from Chemicals library and datasets from paper are curated and digested;
- Fully connected NNs - here descriptors are caclulated and fully-connected NNs are generated. Here the final calculator to get critical values can also be found
- Graph NNs - here Graph Neural Networks, that uses only molecular structure, are created. Again, final calculator to get critical values can be found.
To use models, several libraris must be installed:
- PyTorch 2.6, it would be better that CUDA is present.
- RDKit 2024.09.5
- ChemLib 2.2.4
- Pytorch Geometric 2.7.0 is obligatory for GNNs
- SciKit-learn 1.6.1
Standard libraries - Pandas, Numpy, MatPlotLib, Pickle should also be present :)
During model preparation some other librarires are used and they may be found in the beginning of each Jupyter Notebook in the import section
The best models of Fully-connected Neural Networks have folowing performance:
-
Tc: RMSE for test dataset is 29 K,
$R^{2}$ is 0.89 -
Pc: RMSE for test dataset is 2.19 bar,
$R^{2}$ is 0.95 -
omega: RMSE for test dataset is 0.078,
$R^{2}$ is 0.79
However, there is a small problem that you need to keep a lot of files: column transofrmers (for each separate model), imputer, list of descriptors to drop, etc., so it's recommended to try Graph Neural Networks, which require only molecular structure ^:-))
The best models of Graph Neural Networks have folowing performance:
-
Tc: RMSE for test dataset is 28 K,
$R^{2}$ is 0.91 -
Pc: RMSE for test dataset is 2.02 bar,
$R^{2}$ is 0.96 -
omega: RMSE for test dataset is 0.10,
$R^{2}$ is 0.67
If you have any problems or just want to say some feedback words, you can write me to timyun96@gmail.com