Autoformer-Driven Convolutional Perfectly Matched Layer for 2D HIE-FDTD Method

An innovative deep-learning driven convolutional perfectly matched layer (CPML) integrated into the hybrid implicit-explicit finite-difference time-domain (HIE-FDTD) method is proposed to improve the efficiency of open-region electromagnetic simulations. The Autoformer neural network is introduced to replace the conventional multi-layer CPML structure. Both the computational domain size and algorithmic complexity are reduced since only a single-layer boundary layer is involved in the new model. Benefiting from the time series decomposition and sparse attention mechanism, the wave absorption efficacy of the proposed model is significantly improved without backward cumulative errors. Through a column-stacked data acquisition approach, the Autoformer-based CPML is compatible with both the FDTD and HIE-FDTD frameworks. The time step size of this proposed method is only determined by the coarse grid size, thereby extending the applicability of intelligent absorption boundaries beyond traditional FDTD limits. Numerical examples demonstrate that this method markedly improves computational efficiency while maintaining excellent wave absorption performance. Additionally, results confirm the method's robustness in complex scenarios, including multi-material, multi-source and multi-scale environments.