Deep Learning Based Modelling of Three-Dimensional Magnetic Field

Abstract

Computation of the magnetic field generated by permanent magnets is essential in the design and optimization of a wide range of applications. However, the existing methods to calculate the magnetic field can be time-consuming or ungeneralised. In this research, a deep learning-based fast-computed and generalised model of three-dimensional (3D) magnetic field is studied. The volumetric deep neural network model (V-Net) which consists of a contracting part to learn the geometrical context and an expanding part to enable the concise localization was applied. We synthetically generated the ground truth datasets from permanent magnets of different 3D shapes to train the V-Net. The accuracy and efficiency of this deep learning model are validated. Predicting on 50 random samples, the V-Net took 4.6 s with a GPU T4 and 23.2 s with the CPU whereas the others took a few hundreds to thousands of seconds. Therefore, the deep learning model can be potentially utilised to replace the other methods in the computation and study of the magnetic field for the design and optimization of magnetic devices. (the codes used in this research will be published openly in a Github repository)

Keywords: 3D magnetic field, Deep learning, V-Net, Magnetic devices