Prediction of dielectric constants using a combination of first principles calculations and machine learning

Abstract

This study reports the method of exploring new dielectric materials by combining a large set of first principles calculations and machine learning. A database of dielectric constants was constructed using the first principles calculations. Crystal structures of 3382 candidate compounds were obtained from the Materials Project database. Harmonic phonon calculations were made to select the compounds showing no imaginary phonon modes. The dielectric constants were then calculated using the density function perturbation theory resulting in 2504 compounds to be constructed in the database. Machine learning methods were adopted to correct the calculated dielectric constants for the systematic errors found between the calculated and the experimental dielectric constants. A random forest model with 68 feature variables successfully predicted dielectric constants within the 50% error range of the logarithmic of the dielectric constant. The predicted dielectric constants for most of the compounds were in the range 3-100.