Retrosynthetic and Synthetic Reaction Prediction Model Based on Sequence-to-Sequence with Attention for Polymer Designs

Abstract

Polymer designs, especially monomer designs, can be performed with machine learning and artificial intelligence using a polymer dataset, however, it is meaningless if the designed monomer structures cannot be synthesized and the polymer compound cannot be polymerized. In this study, a retrosynthesis prediction model based on sequence-to-sequence (Seq2Seq) with attention is developed, which is originally used in language transformation, to predict reactants from monomer structures corresponding to polymers. In addition, Seq2Seq with an attention-based synthetic reaction prediction model that predicts monomer structures from reactants is also developed to propose monomer structures with free bonds for polymer design. Through case studies using an actual polymer dataset, it is confirmed that appropriate polymer designs can be achieved by using the proposed method, including the generation of valid monomer structures, the selection of the monomer structures with promising polymer properties, and the prediction of reactants for the monomer structures.