Medium-Frequency Transformers (MFTs) represent a promising technology for future power systems. However, MFT design is limited by the computational cost of Finite Element Method (FEM) simulations, which are essential for detailed magnetic field analysis. While generative models show promise, the need for deterministic outputs and the computational burden of standard models limits their application to this domain. This paper introduces Geo2Field, a novel framework that pioneers the application of a Denoising Diffusion Probabilistic Model (DDPM) for Medium-Frequency Transformer (MFT) magnetic field simulation. Geo2Field features a novel deterministic conditioning strategy and an optimized neural network architecture designed for high computational efficiency. Evaluation demonstrates Geo2Field’s accuracy ( < 1%) with a 2-3x speed-up over FEM. We also characterize the performance trade-offs of model size reduction under data-scarce conditions. Geo2Field offers a high-speed alternative for generating detailed magnetic field maps, laying critical groundwork for rapid and accurate magnetic component design.
.jpg "Xiao Yang (he/him/his) photo")
