New machine learning model quickly and accurately predicts dielectric function

Researchers Tomohito Amano and Shinji Tsuneyuki of the University of Tokyo with Tamio Yamazaki of CURIE (JSR-UTokyo Collaboration Hub) have developed a new machine learning model to predict the dielectric function of materials, rather than calculating from first-principles. The dielectric function measures the polarization of negative and positive charges within materials, the phenomenon underlying dielectric materials. Thus, the fast and accurate prediction of dielectric function facilitates the development of novel dielectric materials, an ingredient of many cutting-edge technologies such as 6G networks. The findings were published in the journal Physical Review B.

Although they might not be as widely known as semiconductors, dielectric materials have great potential to improve modern electronic systems. Dielectric materials do not conduct electricity well, but they are not insulators either. Instead, when placed in an electric field, positive charges within the material shift toward the field and negative charges away from it, resulting in dielectric polarization. The dielectric function is the measure of the strength of polarization. However, calculating the dielectric function meant calculating it from first- principles using quantum mechanics, a computationally slow and heavy process.

“The study of dielectrics is important both for fundamental and applied science,” says Amano, the first author. “On the fundamental side, dielectrics can help elucidate the microscopic origin of how materials respond to electric fields. On the applied side, low-dielectric polymer materials for high-speed communication have garnered attention recently.”

So, the researchers set out to develop a machine learning model that could help address these challenges. They generated training data for the algorithm by doing first-principle calculations of the electronic state of various materials. Moreover, instead of the conventional calculations based on individual molecules, they based their model on the chemical bonds between atoms. The researchers then checked the accuracy of the model by comparing its results to empirical data of simple molecules such as methanol and ethanol.

The model was a success. It not only described the electronic state of various materials close to the accuracy of first-principle calculations, it did so with a fraction of the computational burden. The model also proved to be useful for large-scale and long-time simulations. As computational costs have been the limiting factor, this model makes it possible for the first time to address the macroscopic origin of the dielectric properties of many-molecule systems.

Despite these achievements, Amano is already looking ahead.

“In this work, we investigated the dielectric properties of simple molecules, but application to more complex molecules, including polymers, is yet to be done. So, we are also planning to construct a universal neural network that can be used in industry.”