Prediction of High-Frequency Dielectric Properties of Polymers Using Molecular Dynamics Simulations

Polymeric materials with high or low dielectric constants show great potential for industrial applications, such as in the manufacturing of capacitors, insulators, photovoltaics, and substrates for printed circuit boards. Tailoring the dielectric properties of polymers to meet specific engineering requirements is crucial for developing energy-efficient electronic devices. Since electronic devices incorporating polymers often operate under electric fields at various frequencies, predicting the frequency dependence of dielectric properties is indispensable. For example, Ramprasad’s group predicted frequency-dependent dielectric constants based on experimental data from past studies [npj Comput. Mater. 6, 61 (2020)]. However, a limitation of using experimental data is that the frequency range is typically restricted to up to ≈ 1 GHz due to experimental constraints, while next-generation electronic devices, including 5G-enabled devices, are expected to operate at much higher frequencies, around 100 GHz.

In this study, we developed a simulation package for the automated calculation of various polymer properties, including dielectric constant, dielectric loss, and linear thermal expansion coefficient (LTEC), using molecular dynamics (MD) simulations. The MD method has the advantage of allowing relatively easy calculations in high-frequency regions (>100 MHz), although it requires a longer time to calculate lower frequencies. Using the developed automation package, we calculated the dielectric properties and LTEC for approximately 400 polymers. We first confirmed that the dielectric constants calculated in this work agreed with experimental values at the overlapping frequency ranges. We then applied transfer learning using larger datasets provided by the RadonPy project, which does not include frequency-dependent dielectric properties [npj Comput. Mater. 8:222 (2022)]. The transfer learning significantly reduced prediction errors for dielectric properties. This study demonstrates the potential to predict different properties with small prediction errors by fine-tuning a prediction model based on a large database, such as the RadonPy database, even with a smaller dataset.