RLPolyG

Getting Started

This is the code for De Novo Design of Polymers with Specified Properties Using Reinforcement Learning.

The version of python:

python==3.9.19

You can use pip to install the required packagess.

pip install requirements.txt -r

Forward Model

You can find the forward model in the forward_model_0919.ipynb. The forward model predicts the target property (yield strength) of polymers based on their structural representation. It is a critical component of RLPolyG, enabling property evaluation during the reinforcement learning process.
Inverse Model

The inverse model aims to generate polymers with specified properties. You can find the inverse model in the RLPolyG_final.ipynb.
1. Unbiased Generation Model: The inverse model is initially trained using the PI1M dataset to generate polymers without property bias. This model learns the underlying polymer structure distribution.
2. Biased Generation Model: The ReLease algorithm is then applied to refine the model for property-biased generation. Reinforcement learning is used to guide the model toward generating polymers with the desired properties, such as high yield strength.
Screening for Synthetic Accessibility and Degradability
1. Synthetic Accessibility: You can find the related model in the SA score folder.
2. Degradability: You can find the related model in the degradability prediction folder. We first collected degradability data from Yuan et al.’s work. Using this data, we constructed a random forest model to predict the degradability score of the generated polymers, which also took MFF as input features.

Name		Name	Last commit message	Last commit date
Latest commit History 57 Commits
Figure		Figure
SA score		SA score
data		data
degradability prediction		degradability prediction
release		release
README.md		README.md
RLPolyG_final.ipynb		RLPolyG_final.ipynb
forward_model_0919.ipynb		forward_model_0919.ipynb
requirements.txt		requirements.txt