Machine Learning Model for Predicting Dielectric Constant of Epoxy Resin with Additional Data Selection and Design of Monomer Structures for Low Dielectric Constant

Abstract

The demand for materials with high insulation and low dielectric loss in the electronic material market has led to a growing need for low dielectric constant (DC) materials. Researchers have repeatedly designed, synthesized, and measured materials to develop low DC materials by utilizing their knowledge, necessitating long development periods and high costs in terms of personnel, reagents, and equipment. This study aims to propose monomer structures for epoxy resins with low DC because they are reactive small molecules that offer good processability and moldability. To this end, a DC prediction model was constructed using machine learning, and then a large number of virtual chemical structures of epoxy resins, which were 612 739 and 430 044 structures, were generated using a method based on the connection of the main and side chains and virtual chemical reactions, respectively. Subsequently, the properties of generated structures were predicted with constructed models to search for the structures of epoxy resins with low DC. Further, the predictive ability of the DC model was improved from 0.270 to 0.371 of r² in double cross-validation by appropriately selecting samples from the official database and adding them to the training data.