An Informatics Approach for Polymer Dielectric Constants

Polymers with low dielectric constants, referred to as low-k polymers, are critical components for flexible electronic devices, effectively preventing signal crosstalk and RC delay. However, given the vast chemical space of polymer candidates, discovering low-k polymers through empirical measurements or computational simulation is resource-intensive. In this work, we utilized a data-driven approach to establish a predictive platform for polymer dielectric constants. To ensure the practical applicability of the prediction model, we constructed a database of experimentally measured polymer dielectric constants by mining scientific literature, which was then used to train the model. By further applying transfer learning and integrating 7 molecular descriptors derived from the polymer’s constitutional repeating units, our model achieved a high prediction accuracy (R2 of 0.89), outperforming the baseline model that uses motif-based fingerprints. Using the trained predictive model, a downstream screening task on 26,852 virtual polyacrylates was performed, and 18 new low-k candidates were successfully identified, whose predicted properties were further validated by physics-based simulation.