Speaker
Description
The analysis and calculation of magnetic fields serve as the cornerstone for magnet design. In scenarios requiring intricate spatial magnetic field configuration optimization, high-throughput analytical computations are often necessary, posing significant computational cost challenges to existing finite element method (FEM)-based approaches. In recent years, surrogate model computations leveraging artificial intelligence algorithms have emerged and been applied across various domains, including fluid dynamics, materials science, and biology, offering potential breakthroughs to the aforementioned challenges. This paper introduces a magnetic field surrogate model methodology based on the deep operator network (deepONet). By leveraging low-throughput FEM model data, it establishes a precise mapping between the magnet coil design space and the spatial distribution of magnetic fields, laying the foundation for rapid coil design optimization methodologies. Specifically, the effectiveness of the proposed method is validated through two types of cases with varying complexities. The first type comprises coil structure models featuring axial symmetry, while the second type encompasses coil structures with three-dimensional asymmetric configurations. The constructed model takes the coil design space and physical three-dimensional spatial coordinates as inputs, outputting the spatial magnetic field. This paper systematically explores model accuracy across different network structures and provides guiding insights. Additionally, it presents engineering applications based on this model methodology, applying it to the analysis and calculation of electromagnetic-mechanical multi-physics coupling problems in the field of electromagnetic forming manufacturing, and quantitatively assesses its effectiveness and the enhancement in coupling analysis efficiency.
