SF_generative

This repository contains the scripts, configuration files, and software components used for training generative models and performing reinforcement learning. It includes pre-RL and post-RL generative models, as well as the Chemprop model for excited-state property prediction.

Repository Structure

• configs/: Configuration files for various stages of the project.
• data/: Data files used in the project.
• model/: Model checkpoints and related files.
• script/: Scripts for various tasks.

1. Setup the Environment: Use the environment.yml file to set up the required conda environment.

   conda env create -f environment.yml
   conda activate sf_generative

   If you have the model trained with REINVENT v.3, use convert_r3_model.py to make it compatible with REINVENT v.4.

   To make singlet fission score component visible to REINVENT v.4, bash export PYTHONPATH="${PYTHONPATH}:[]/SF_generative/script"

2. Run REINVENT task: Use TOML files in configs/ directory to optimize the generative model in model/ for singlet fission chromophores

   reinvent rl.toml -l rl.log  

3. Post-optimization analysis: Use python scipts in scipts/ directory to visualize the optimization history, analzye the optimized generative model, and sample for singlet fission candidates

   python rl_analysis.log
   python model_analysis.log
   

Citation

If you use this repository in your research, please cite the following publication:

Worakul T, Laplaza R, Blaskovits JT, Corminboeuf C. Generative Design of Singlet Fission Materials by Revisiting the Use of a Fragment-oriented Database. ChemRxiv. 2025; doi:10.26434/chemrxiv-2025-chmsm