Usage of Physics-Informed Neural Network to Extract Physical Parameters From High Voltage Experiments

Computer simulations based on partial differential equations (PDEs) describing physical phenomena, are widely used for analyzing the performance of high voltage insulation. Such simulation models require access to physical model parameters, which are often hard to obtain. The purpose of the present paper is to explore how Physics-Informed Neural Network (PINN) tailored for specific PDEs, can be used to extract relevant physical parameters from synthetic data. Electric potential or charge distributions in 1-dimensional coaxial and Cartesian domains were used as training data for PINN models. The following physical parameters were successfully extracted: (i) space charge profiles, (ii) mobilities of ions, and (iii) the radius of the inner electrode in the coaxial geometry. The impact of noisy training data on the accuracy was also studied for the different cases. The main finding is that PINN models can successfully be used to extract model parameters for electrical charge transport problems using synthetic data as input. This approach has large potential to strengthen the research on charge dynamics in gas, liquid, and solid insulation as well as other topics related to high voltage insulation systems.