Deep Long Short-Term Memory Networks-Based Solving Method for the FDTD Method: 2-D Case

In this letter, a novel finite difference time domain (FDTD) solving method is proposed based on the deep long short-term memory (LSTM) networks. The field data in the object domain of traditional FDTD method are applied to train the newly proposed LSTM-based FDTD model, termed as LSTM-FDTD. Distinguished from the traditional FDTD method, the proposed method is not limited by Courant-Friedrichs-Levy (CFL) condition and does not need the conventional absorbing boundary conditions (ABCs). Thus, the proposed method conveniently decreases both the size of computation domain and the algorithm's complexity. In addition, LSTM-FDTD could reach higher accuracy due to the sequence dependence of LSTM networks. Numerical benchmarks illustrate the efficiency and accuracy of the proposed method.

Automated summary

In this article, a novel finite difference time domain (FDTD) solving method based on the deep long short-term memory (LSTM) networks is proposed. The method, named LSTM-FDTD, uses the field data from traditional FDTD method to train the LSTM-based model. Unlike the traditional FDTD method, this proposed method is not restricted by Courant-Friedrichs-Levy (CFL) condition and does not require conventional absorbing boundary conditions (ABCs). As a result, it conveniently reduces the size of computation domain and complexity of the algorithm while achieving higher accuracy due to the sequence dependence of LSTM networks. Numerical benchmarks demonstrate the efficiency and accuracy of this proposed method.