Machine learning enhanced prediction of permittivity of spinel microwave dielectric ceramics compared to traditional C-M calculation

Abstract

Abstract Microwave dielectric ceramic (MWDC) is crucial in advancing the development of 5G technology and the communication field. The prediction or calculation of its properties is of great significance for accelerating the design and development of MWDCs. Therefore, the prediction of permittivity of spinel MWDCs based on machine learning was investigated in this work. Firstly, we collected 327 single-phase spinel MWDC entries and constructed feature engineering, which includes feature generation and feature selection (five dominant features, including Mpo, Dar, Mmbe, Aose and Dgnve, were selected from 208 generated features). Next, seven commonly used algorithms were utilized during the training process of machine learning models. The extreme gradient boosting (XGBoost) model shows the best performance, achieving R-squared (R 2) of 0.9095, mean absolute error of 1.02 and root mean square error of 1.96 on the train and test dataset. In addition, the machine learning models, especially the XGBoost model, show enhanced prediction (calculation accuracy) of the permittivity of spinel MWDCs compared to the traditional Clausius–Mossotti equation, which can provide a guide for the design and development of spinel MWDCs applied for wireless communication.