Recurrent Neural Network-Assisted Truncation of Convolutional Perfectly Matched Layers for FDTD

To improve the efficiency of the finite-difference time-domain (FDTD) method for unbounded electromagnetic problems, a truncation method for the convolutional perfectly matched layers (CPMLs) is proposed based on a recurrent neural network (RNN). In this work, the RNN is used to capture the time-marching information in the FDTD, complete the offline training of field components at the truncated CPML interfaces, and predict the field components on the cells of the truncated interfaces at each time step. By using this method, the unknowns are well reduced, which results in an improvement in computational efficiency. Additionally, since some of the CPMLs are preserved, the late-time electromagnetic response is still stable without error accumulation. The proposed method can balance accuracy and efficiency by an optimal truncation position for CPMLs, which is validated by two numerical examples.