A magnetic field prediction method based on UNet++ with the addition of inception module

抄録

To solve the problem of long simulation time and inability to reuse in finite element method for magnetic field analysis, this paper proposes a magnetic field prediction method based on UNet++ with the addition of Inception module. To enhance the model’s capability in capturing multi-scale features of complex geometric structures and reinforcing key features, the UNet++ model was improved by integrating Inception modules, establishing the I-UNet++ model. By predicting the pixels of the image, the corresponding magnetic field distribution is obtained, and the magnetic field prediction results are represented by the image. Use transformers as validation cases to train and test the model. Compared to the prediction results of the UNet++ model, the I-UNet++ model has higher accuracy and can achieve magnetic field prediction on small sample datasets. This provides ideas for real-time calculation of multiple physical fields in transformer.