EM optimization of microwave filter based on long short-term memory–feedforward neural network and transfer functions

Abstract An electromagnetic optimization technique based on a long short-term memory–feedforward neural network (LSTM-FNN) and transfer functions is proposed for microwave filter design. The proposed optimization method addresses the situation where a neuro-transfer function model repeatedly trains at each optimization iteration process. The proposed surrogate model combines the LSTM-FNN and polynomial model to map nonlinear relationships between geometric variables and transfer functions. Firstly, by combining the gate mechanism of LSTM with the high generalization ability of an FNN, the proposed LSTM-FNN effectively learns nonlinear relationships between geometric variables and frequency responses at specific frequencies. Secondly, the transfer functions can be accurately approximated via polynomial fitting. Frequency responses at any interesting frequency range can be accurately expressed using the transfer functions. Finally, the trained surrogate model, exploiting the trust-region algorithm, can accurately and efficiently achieve optimization convergence. An example of a low-pass filter (LPF) is adopted to validate the proposed optimization method.