Deep learning enhanced prediction of microwave dielectric constant of spinel ceramics eliminating manual feature engineering

Abstract

Machine learning (ML) has demonstrated considerable promise and superiority in the prediction of microwave dielectric ceramic (MWDC) properties. Nonetheless, applying traditional ML models often requires processing numerous features, making their implementation complex and unwieldy. In this work, a deep learning (DL) model named CRANCNN-M2V that predicts the dielectric constant of spinel-MWDCs eliminating manual feature engineering has been constructed. The model can identify essential features directly from chemical compositions using the data embedding method. Our model further improved the network based on a Compositionally Restricted Attention-based Neural Network (CrabNet) and showed the enhanced prediction performance in the dielectric constant of spinel-MWDCs, achieving an RMSE of 1.52, an MAE of 0.938, and an R² of 0.954, and it also outperformed commonly used traditional ML models (e.g., XGBoost, Random Forest (RF), Decision Tree (DT), etc.). Furthermore, the contribution of different elements in the dielectric constant of spinel MWDCs has been analyzed via our CRANCNN-M2V model. Highly accurate and efficient prediction of our model will effectively promote the design and development of spinel-MWDCs applied for wireless communication.