【目的】传统电磁场仿真软件主要依赖有限差分、有限元等数值方法进行求解。尽管这些方法能够获得与试验结果较为接近的数值解，但其计算精度严重依赖于网格数量与划分质量。在提高求解精度的同时，也会导致计算时间和成本大幅增加，在大规模优化设计中尤为显著。【方法】为此，本文提出了一种融合人工智能技术的电磁仿真软件开发策略，在前处理、求解和后处理中采用人工神经网络(ANN)模型加速整个求解过程。在建模过程中，使用基于图像、语音和文本的多模态参数化建模技术；在网格划分和矩阵求解中，使用ANN模型进行分类判断或回归预测；在计算结果的处理和可视化阶段，采用机器学习拟合与插值方法对计算结果进行光滑处理，并提高分辨率。【结果】基于电磁仿真软件可以针对特定问题获取大量的有限元数据。在数据驱动环境下，能够实现对电磁场分布情况的预测、基于代理模型的交流铜耗预测、多输入输出和多工况的电机全性能预测、借助分类器的多目标加速优化以及完全基于代理模型的多目标优化和电机建模等。【结论】本研究通过数据驱动方法构建电磁产品的数字孪生体，为其状态监测、预测性维护与性能优化提供有效支持。

[Objective] In traditional electromagnetic field simulation software, numerical calculation methods such as finite difference and finite element method are mainly used to solve problems. Although these methods can obtain numerical solutions closer to the experimental results, their computational accuracy heavily depends on the number of meshes and the quality of the dividing. While improving the solution accuracy, it also leads to a significant in computation time and cost, especially when using software for large-scale optimization design. [Methods] Therefore, this paper proposed a development strategy for electromagnetic simulation software that integrated artificial intelligence technology. Artificial neural network (ANN) models were used in pre-processing, solving, and post-processing to accelerate the entire solving process. In the modeling process, multimodal parametric modeling techniques based on images, speech, and text were used. In the mesh dividing and matrix solving, ANN models were used for classification judgment or regression prediction. In the processing and visualization stages of calculation results, machine learning fitting and interpolation methods were used for smoothing the computational results and improving the resolution. [Results] Based on electromagnetic simulation software, a large amount of finite element data could be obtained for specific problems. In a data-driven environment, it was possible to achieve the prediction of electromagnetic field distribution, the prediction of AC copper consumption based on surrogate models, the full performance prediction of motors with multiple input/output and operating conditions, multi-objective accelerated optimization with the help of classifiers, as well as multi-objective optimization and motor modeling based entirely on surrogate models. [Conclusion] This study constructs digital twins of electromagnetic products through data-driven approaches, providing effective support for their status monitoring, predictive maintenance and performance optimization.

国家自然科学基金(52407010)