Polymers are widely used as dielectrics in devices ranging from capacitors to multi-layered IC’s and as insulation for high voltage cables. In these applications, the dielectric breakdown strength is a critical parameter affecting device performance [1]. In this work, we study the dielectric breakdown properties of common polymers using machine learning. The complex nature of the dielectric breakdown process makes it analytically intractable for real systems and thus ideal for application of machine learning.

The input to the machine learning model consists of material property values and the model outputs dielectric breakdown. For our training data, dielectric breakdown of polymer thin films is experimentally measured using small area breakdown measurements [2] at multiple temperatures ranging from -100 °C to 100 °C. The input to our model are property data that are known to influence dielectric breakdown. Selecting a subset of our original feature space using variable selection algorithms, we train a machine learning model using gaussian process regression.

We test our model on unseen experimental data and find that it predicts dielectric breakdown of unseen polymers accurate to the experimental error of measurement. In addition we find that the underlying physics of the temperature dependence of dielectric breakdown is captured well in the training domain. Our model can be used to systematically search polymer spaces to accelerate discovery of high breakdown strength polymers.

References

[1] Zulkifli Ahmad (October 3rd 2012). Polymer Dielectric Materials, Dielectric Material, Marius Alexandru Silaghi, IntechOpen, DOI:10.5772/50638
[2] Rytoluoto, I., et al. ”Large-area dielectric breakdown performance of polymer films-part i: measurement method evaluation and statistical considerations on area-dependence.” IEEE Transactions on Dielectrics and Electrical Insulation 22.2 (2015): 689-700.