Compact Neural-Network Digital-Twin Models and Material Comparison for Power Magnetics

Accurate modeling of magnetic behaviors, such as the hysteresis loop and core loss, is critical in the design and optimization of power magnetic components. However, these behaviors can be significantly influenced by various factors, such as excitation waveforms and operating conditions, which makes their precise modeling challenging. Neural-network-based power magnetics modeling has emerged as an effective approach for capturing complicated multi-variable non-linearity. This paper uses transformer-based encoder-projector-decoder neural network models to study the material characteristics under ideal excitations, which cannot be done with traditional methods. The neural network model functions as a digital-twin of the original material database with over 10000x smaller data size (e.g., from 3.8 GB to 204 kB) while accurately and rapidly reproducing the material characteristics under complex operating conditions. An online platform - MagNet-AI - is developed as an example to demonstrate the effectiveness of using neural networks as the digital-twins to study and compare power magnetics materials.