Deep Learning-Based Prediction of Electrical Properties of Polymers with Feature Extraction of Process Conditions

Polymers exhibit a variety in properties depending on molecular composition, molecular weight, polymer structure, additives, and polymerization conditions. Materials Informatics (MI) is effective as a material design method with handling a huge amount of polymer data. However, the use of MI for polymer design is relatively limited. This is because the polymerization conditions, including temperature, pressure, and solvents, are not adequately extracted as features, even though they have a large influence on material properties. In addition, the material properties are influenced by many parameters such as polymer morphology, impurities, and temperature in the measurement. Therefore, the detailed process conditions in manufacturing and property-measuring are required for the data-driven property prediction.<br/><br/>In this study, we extract monomer structures, polymerization conditions, and property-measuring conditions for hundreds of polymers registered in PoLyInfo [1], and develop a prediction model of electrical conductivity and dielectric breakdown strength. For composite materials, the additives data, including chemistry composition, content rate, and mixing condition, are added to the features. Physical quantities computed by first-principles calculations are added as necessary to improve the prediction accuracy. The molecular structures of the base monomers and additives are extracted using a graph convolutional neural network. The polymerization and property-measuring conditions are extracted using both a bag-of-words model and a recurrent neural network, which differ in whether the numerical scale is taken into account. The prediction accuracies of the electrical conductivity and dielectric breakdown strength are compared between with and without the polymerization conditions, property-measuring conditions, and their numerical scale, to discuss their contribution as features.<br/><br/>[1] https://polymer.nims.go.jp/en/