Feasibility of Deep Neural Network Surrogate Models for Simulations of High Voltage Equipment Design

Abstract

With the recent advances in Machine Learning, there is a natural interest in investigating potential application areas where it could contribute to more efficient solutions. This thesis investigates the feasibility of the Deep Neural Network (DNN) approach implemented in the COMSOL Multiphysics software, for simulations of high voltage components by comparing respective results with traditional finite element solutions. The project encompasses a best practice study for DNN model parameter optimization, by considering electrostatic problems for two cases, utilizing different representative geometries. In addition, the result from one of the study cases is further applied for establishing best practices for streamer breakdown detection problems. The DNN approach was found feasible for applications with relatively low-complexity geometries and simple physics. However, there are areas of study where it did not outperform traditional finite element solutions. Thus, when simulating complex geometries or physics with DNN, the improvement in efficiency compared to using finite element methods diminishes. From the obtained results, it is apparent that further studies in the area are necessary to explore advantages and limitations of the method.