ChemFlow is implemented using PyTorch and runs on Ubuntu with NVIDIA GeForce RTX 4090 GPUs.
The framework also relies on PyTorch Geometric.
Please ensure the following libraries are installed:
- numpy
- pandas
- rdkit
- scikit-learn
- ncps
- torch
- torch_geometric
In the Concentration dependent directory, we provide:
- Detailed implementation of each module
- Data generation scripts
- Example usage based on the activity coefficient dataset
These examples help illustrate data preprocessing procedures and the training workflow.
During training, we did not perform hyperparameter tuning on the validation set for each individual property before making predictions on the test set.
Note:
In practical applications, tuning hyperparameters for each property may further improve predictive performance. However, we did not perform extensive tuning due to the long training time, and because ChemFlow already outperforms models reported in existing literature.
We will continue to update datasets and models, and regularly check the correctness and completeness of the code.
If you encounter any errors while running the code or have any questions, please contact:
