Molecular Structure Design of Heat-Resistant Cycloolefin Polymer Based on Machine Learning

Abstract

Abstract Traditional biaxially oriented polypropylene (BOPP) is widely used in capacitors of converter valves, but its high-temperature performance is gradually approaching its design limit. Cycloolefin copolymer/polymer (COC/COP) materials are explored as promising alternatives due to their high glass transition temperature, wide band gap, thermal stability, and low conduction loss. To overcome the lack of large-scale COC/COP polymer databases and the limitations of experimental validation, molecular dynamics, and conventional machine-learning screening, this work proposes a graph convolutional neural network (GCN)-based workflow for inverse design and property prediction of cycloolefin polymer monomers. An initial database of nearly 200,000 COC/COP candidates was constructed through library skeleton design and converted into molecular graph representations for model training, latent-space sampling, and molecular generation. Using a GCN-based forward prediction model, we screened for 12 potential heat-resistant polymers. Combined with subsequent calculations of molecular dynamics (MD) and retrosynthetic analysis, we ultimately identified P9 and P11 as the final candidate heat-resistant polymers. This work provides a case-study approach for the design and screening of high-temperature-resistant polymer dielectrics.