High-Throughput Screening Modeling for Exploring Low-k Dielectric Constant Crystals

Dielectric materials with desired mechanical, thermal and electrical functions play an important role in applications of electronic packaging. Even though there has been well developed first-principles computational tool to obtain accurate dielectric property, it is still challenging to efficiently figure out materials with desired dielectric property from tremendous number of candidates in databases. In this work, we have proposed a high-throughput screening framework for identifying low dielectric crystals. Three types of descriptors for collected 7290 crystals from Materials Project database were systematically tested, including Modnet, Xenonpy, and Magpie, among which the Magpie descriptor shows the highest prediction accuracy. The neural network was adopted to train prediction models. By using the factor analysis technique, the number of dominant factors is determined to be 40 based on the Scree plot. The value of the common factors corresponding to each material is evaluated, and the load coefficients between the common factors and the material characteristics is formed with the hot image. Following the above regression model and feature engineering, the decision coefficient of 0.80 was obtained for the testing set, indicating that the trained model has good predictive performance. The results provide useful insights for the following screening of low-k dielectric constant crystals.