MAG-Vision: A Vision Transformer Backbone for Magnetic Material Modeling

Abstract

The neural network-based method for modeling magnetic materials enables the estimation of hysteresis B-H loop and core loss across a wide operation range. Transformers are neural networks widely used in sequence-to-sequence tasks. The classical Transformer modeling method suffers from high per-layer complexity and long recurrent inference time when dealing with long sequences. While down-sampling methods can mitigate these issues, they often sacrifice modeling accuracy. In this study, we propose MAG-Vision, which employs a vision Transformer (ViT) as the backbone for magnetic material modeling. It can shorten waveform sequences with minimal loss of information. We trained the network using the open-source magnetic core loss dataset MagNet. Experimental results demonstrate that MAG-Vision performs well in estimating hysteresis B-H loop and magnetic core losses. The average relative error of magnetic core losses for most materials is less than 2%. Experiments are designed to compare MAG-Vision with different network structures to validate its advantages in accuracy, training speed, and inference time.