Machine learning assisted τf value prediction of ABO3-type microwave dielectric ceramics

Abstract

The temperature coefficient of resonance frequency (τ_f or TCF) is the key parameter for evaluating temperature stability of microwave dielectric ceramics. In this work, a machine learning framework was proposed to predict the τ_f values of ABO₃-type microwave dielectric ceramics. Leveraging a curated dataset of 104 single-phase ABO₃-type compounds, we systematically evaluated models based on five machine learning algorithms using 31 structural descriptors as input features. The eXtreme Gradient Boosting (XGB) algorithm emerged as the optimal predictive model, demonstrating robust performance on the test set (R² = 0.7799, RMSE = 15.7494×10^–6 °C^–1). Consistent results on the validation set further confirmed its generalization capability. Critical features contributing to the model's performance include molecular dielectric polarizability (pm), tolerance factor (tt), ionic volume (Vi) and relative molecular mass (m). Structure-property relationship studies revealed that the pm plays an important role in modulating the τ_f value by affecting the permittivity. Quantitative thresholds for these critical descriptors were also derived for identifying materials with near-zero τ_f. This work provides an effective data-driven approach for accelerating the discovery of microwave dielectric ceramics with good temperature stability.