Super-Resolution Deep Learning Network for Finite-Difference Time-Domain Simulations

When selecting the spatial grid size in finite-difference time-domain (FDTD) simulations, it is necessary to balance numerical dispersion and computational resources. Coarser grids are less computationally intensive but decrease simulation accuracy, while finer grids can reduce numerical dispersion and enhance accuracy at the cost of increased simulation time. To address this challenge, this letter introduces a super-resolution deep learning network (SR-FDTDNet), which aims to improve FDTD simulation methods. SR-FDTDNet seeks to uncover the implicit mapping between coarse-grid and fine-grid simulation fields, enabling the reconstruction of fine-grid simulations from coarse-grid FDTD simulations. It can capture local information in the spatial and temporal dimensions of simulation data and model long-term dependencies through parallel computing, effectively addressing the complexity of FDTD simulation data. Results from practical electromagnetic simulation scenarios demonstrate that SR-FDTDNet significantly improves simulation efficiency while maintaining high accuracy in reconstructing fine-grid simulation fields.