Data-driven prediction of insulation properties in dielectric composites by machine learning

Abstract Materials exhibiting superior thermal and electrical insulation characteristics are highly sought after for the development of electrical power systems and electronic equipment. Nevertheless, the composition and architecture complicatedly influence the insulation performance due to the vast array of design variables. Conventional design methods approaches typically involve extensive trial-and-error experiments, necessitating substantial investment and effort to achieve optimal results. The simulation of electrical trees using the finite element method (FEM) serves as a crucial metric for assessing insulation characteristics. A prolonged evolution period of the electrical tree is indicative of superior insulation performance. This approach provides accurate simulations, albeit at a significant computational cost. In this study, we introduce a machine learning (ML) model that enables the precise and efficient prediction of the development time of electrical trees in dielectric composites. The linear and nonlinear ML models are thoroughly trained, including both ElasticNet and a multi-layer perceptron (MLP), leveraging a dataset constructed from FEM simulations. The ElasticNet model demonstrates minimal prediction errors, offering a viable substitute for the FEM model and significantly improving the accuracy of predictions. In addition, an analysis is conducted to determine the influence of each input feature on the predictive outcomes, providing critical insights that facilitate and improve the design process of dielectric composites. To substantiate its effectiveness, the ElasticNet model has been employed to forecast existing experimental outcomes, thereby confirming its practicality and dependability.