On-demand design of materials with enhanced dielectric properties via a machine learning-assisted materials genome approach

Abstract

Materials with specific dielectric properties possess immense potential in the fields of wave-transparent materials and energy storage. Thermoset silicon-containing arylacetylene (PSA) shows promise for exhibiting excellent dielectric performance. However, designing materials that can meet multiple property requirements using a trial-and-error method is time-consuming and labor-intensive. Herein, a machine learning (ML)-assisted materials genome approach was developed to enable the on-demand design of PSAs with tailored dielectric and thermal performance. The high-precision ML prediction of dielectric constant and loss, particularly the dependence on frequency and temperature, is achieved by leveraging data migration and applying principal component analysis and correlation-based feature engineering to identify the key structural features for training models. Additionally, an ML model was constructed for predicting 5% thermal decomposition temperature. By defining, collecting, and combining the genes of PSAs, virtual PSAs are generated. Using established ML models, we predict the properties of candidate PSAs and identify promising PSAs with desirable dielectric and thermal performance. Experimental results validate the reliability of the proposed ML-based design strategy. Finally, to enhance the interpretability of ML models, we conducted gene contribution analysis and summarized the design rules for heat-resistant polymer dielectrics. This study, which is based on the materials genome approach, enables the rational design of materials with specific dielectric properties, thereby accelerating the discovery and innovation of advanced materials.