From Graph Machine Learning to Furnaces: Permittivity Exploring of Dielectric Ceramics

The exploration of potential dielectric materials with specific permittivity remains a critical scientific challenge. Facing microwave band applications, dielectric ceramics with suitable permittivity are widely studied to meet the requirements of passive components such as antennas, filters, and resonators in 5G communication. However, discovering dielectric ceramics with desired permittivity often relies on traditional trial-and-error methods. Classic physical theories for predicting permittivity are often limited by inherent assumptions and theoretical constraints in accurate prediction, while traditional machine learning methods are stuck in complex feature engineering, restricting extension to new material systems. Herein, a new graph neural network (Res-GCN) is developed to directly predict permittivity from the connections of atoms. The model accurately predicts permittivity with root mean squared error (RMSE) of 1.873 and determination coefficient (R2) of 0.80 in extrapolation tests. The high-throughput screening based on pattern recognition is conducted over 6,000 material entries in few minutes, recommending a series of low-permittivity dielectric ceramics. Eight phases were experimented (predicted permittivity RMSE as low as 1.49), discovering two novel ceramics with expected permittivity and excellent dielectric properties. This work provides an efficient and effective method for discovery of low-permittivity materials, immensely shortening the research cycle.