Machine learning approaches for permittivity prediction and rational design of microwave dielectric ceramics

Low permittivity microwave dielectric ceramics (MWDCs) are attracting great interest because of their promising applications in the new era of 5G and IoT. Although theoretical rules and computational methods are of practical use for permittivity prediction, unsatisfactory predictability and universality impede rational design of new high-performance materials. In this work, based on a dataset of 254 single-phase microwave dielectric ceramics (MWDCs), machine learning (ML) methods established a high accuracy model for permittivity prediction and gave insights of quantitative chemistry/structure-property relationships. We employed five commonly-used algorithms, and introduced 32 intrinsic chemical, structural and thermodynamic features which have correlations with permittivity for modeling. Machine learning results help identify the permittivity decisive factors, including polarizability per unit volume, average bond length, and average cell volume per atom. The feature-property relationships were discussed. The optimal model constructed by support vector regression with radial basis function kernel was validated its superior predictability and generalization by verification dataset. Low permittivity material systems were screened from a dataset of ∼3300 materials without reported microwave permittivity by high-throughput prediction using optimal model. Several predicted low permittivity ceramics were synthesized, and the experimental results agree well with ML prediction, which confirmed the reliability of the prediction model.