Introducing materials informatics to dielectrics design

Abstract

It has become increasingly common to use machine learning techniques for discovering and designing novel materials. Big data enables machine learning techniques to make accurate predictions. However, experimental data are not abundant especially in the case of dielectric properties. In addition, the properties of polymers depend not only on the structure of monomers but also on the length of polymers, morphology, additives, and so on which further complicates the problem. Here, we review our latest research outcomes that are related to computational and data-driven dielectric materials design. First, we show how we were able to accurately predict the dielectric properties of gases using a small data set and further discover new molecules that can potentially outperform existing SF6 alternative gases. Then we show that by proper feature engineering it is possible to predict the thermal and electrical properties of polymer/inorganic filler composites. The main findings are as follows: (1) in line with our intuition, in general, accurate prediction of dielectric properties is difficult compared to the prediction of thermal or mechanical properties, (2) the process condition has an especially great impact on the electric properties of polymers, (3) with the knowledge of the underlying physics affecting the macroscopic properties, one can predict the dielectric properties of various materials with reasonable accuracy, (4) machine learning helps us understand the factors that control the dielectric properties, and (5) it can also be used to guide experiments or to provide testing standards.