A graph neural network assisted reverse polymers engineering to design low bandgap benzothiophene polymers for light harvesting applications

Abstract

In this study, we present a novel approach to reverse polymer engineering utilizing a Graph Neural Network (GNN) framework to design low bandgap benzothiophene (BT) polymers for light harvesting applications. We have curated an extensive dataset comprising 57,556 structure-property pairs of BT-based compounds, leveraging expert knowledge to enhance the quality and relevance of the data. Our Transformer-Assisted Oriented pretrained model for on-demand polymer generation (TAO) demonstrates exceptional performance, achieving a chemical validity rate of 99.27 % in top-1 generation mode across a test set of 6000 generated polymers, marking the highest success rate reported among polymer generative models to date. Throughout the training process, the loss steadily decreased with each epoch, indicating that the model was learning effectively from the data. The model predictive accuracy is further validated by an impressive average R² value of 0.96 for 15 defined properties, highlighting the TAO with its robust capabilities in polymer design. The newly designed polymers exhibit a bandgap range of 1.5–3.40 eV, making them promising candidates for light harvesting applications. Additionally, their highest Synthetic Accessibility Likelihood Index (SALI) scores reach up to 17 and also indicates that the majority of these polymers are amenable to synthesis. This work not only advances the field of polymer design but also provides a powerful tool for the targeted development of materials with specific electronic properties.