Transformer-Based PINN for Semisupervised Electromagnetic Forward Simulations

Traditional numerical methods face serious efficiency challenges for solving complex electromagnetic problems. In this paper, a Transformer-based physical information neural network (PINN), which is highly efficient in capturing long-range dependencies of the sequence, is proposed to accelerate electromagnetic forward simulations. To improve model performance on limited data, semi-supervised learning is further introduced by using pseudo-label strategy. Numerical examples show good prediction accuracy of the proposed framework when compared to other PINNs. Through numerical experiments under different electromagnetic scenarios, its generalization ability can also be verified. The proposed approach provides an effective solution for addressing challenging electromagnetic forward problems in practical applications