De novo Design of Polyimides Leveraging Deep Reinforcement Learning Agent

---

DAPiGen: De novo Design of Polyimides Leveraging Deep Reinforcement Learning Agent

• https://doi.org/10.1002/adma.202511099

Background

---

This code is the basis of our work submitted to Advanced Materials, which aims to de novo design of polyimides using PPO algorithm of reinforcement learning based on fragment generation to bring more insights into polymer design.

Prerequisites and dependencies

$ env.yml

Usage

---

Raw data of PI structure and properties

The relevant files are kept in './raw_data'

GNN model

The relevant files and code for the GNN models are in './QSPR/GNN/'

A command line for training a AFP model and tuning the splitting seed:

$ python Seed_Tuning.py 

A command line for training a FraGAT model and tuning the splitting seed:

$ python Seed_Tuning_Frag.py 

The results of every attempt containing metrics are automatically printed in './QSPR/GNN/output/'.

To carry out a single-objective bayesian optimization on a AFP model, do:

$ python Optimization.py

To carry out a single-objective bayesian optimization on a FraGAT model, do:

$ python Optimization_frag.py

To generate ensemble models with random initializations on AFP models, do:

$ python Ensembling.py

To generate ensemble models with random initializations on FraGAT models, do:

$ python Ensembling_frag.py

A command line for predicting each PI in datasets:

$ python Ensemble_PI.py
$ python Ensemble_frag_PI.py

Descriptor model

The relevant files and code for the descriptor model are in './QSPR/Descriptors'

A command line for preprocessing descriptor conversion and dimensionality reduction:

$ python data_process.py 

The result files after propocessing are in './QSPR/Descriptors/datasets/'.

To carry out a single-objective bayesian optimization on a descriptor model, do:

$ python Optimization.py

To generate ensemble models with random initializations on descriptor models, do:

$ python Ensembling.py

polyBERT model

The relevant files and code for the polyBERT model are in './QSPR/polyBERT'

Before running, download polyBERT model on https://huggingface.co/kuelumbus/polyBERT

To carry out a single-objective bayesian optimization on a polyBERT model, do:

$ python Optimization.py

To generate ensemble models with random initializations on polyBERT models, do:

$ python Ensembling.py

A command line for predicting each PI in datasets:

$ python Ensemble_PI.py

RL_PPO

To decompose dianhydride and diamine, do:

$ python Decompose.py

The building blocks of dianhydride and diamine are in './RL_PPO/outpus/building_blocks/'

The agent takes an action from building blocks obtained from a decomposition step. The agent is trained by PPO with RLlib.

Before running, download polyBERT model on https://huggingface.co/kuelumbus/polyBERT

$ python train.py

Generated polyimides are sampled through the trained agent. Select and rewrite the model path that you want to use. To generate polyimides randomly, do:

$ python env_test.py

To the cluster analysis of the generate polyimides, run './RL_PPO/outpus/postprocess.py'