Integration of Multiscale Simulations and AI Models for the Design of High-Performance Dielectric Polymers

Polymers are widely used as electrical insulating materials in the power and energy sector. As data-driven approaches to materials development are attracting increasing attention, two major challenges remain: the limited availability of polymer data and the complex hierarchical structures of polymers. In this study, we conduct high-throughput quantum chemistry calculations and molecular dynamics (MD) simulations using a single force field for approximately 5,500 homopolymers to construct a dataset of computed polymer properties. We develop machine learning models to predict MD-calculated polymer properties using quantum chemistry-derived physical quantities and MD force field parameters as descriptors. We also compare their performance with models using structural descriptors. While interpolation performance is comparable between molecular structure–based models, including a fingerprint model and a graph convolutional neural network, and models using quantum chemistry and MD force-field descriptors, the latter shows superior extrapolation performance in our evaluated setting for MD-computed density. In addition, we confirm that increasing the amount of data improves the accuracy of both interpolation and extrapolation predictions. The polymer dataset and predictive approach proposed in this study may help inform computational screening toward the development of high-performance polymer materials; experimental validation remains necessary.