A Neural-network-based Inverse Design of the Microwave Photonic Filter Using Multiwavelength Laser

Abstract

In light of the great importance of the microwave photonic filter (MPF) in communications technology, an MPF using multiwavelength laser was proposed previously. Although the closed-form formula for the frequency response of the MPF was successfully derived, it is too complicated to use intuitively. To simplify the design flow, this research investigates building the inverse design for the MPF with machine learning technology. The inverse design, implemented by neural networks (NN), aims to predict the appropriate values of the systemic design parameters that can cause the MPF to generate the desired frequency response. Specifically, we explore a traditional NN, a recurrent NN (RNN) with Long Short-Term Memory (LSTM), and an RNN with Gated Recurrent Unit (GRU). Additionally, four types of optimizers are tentatively applied to the three NNs, and the Adam optimizer is ultimately chosen for its best efficiency. When properly trained, all the three models are able to work well with the GRU-based RNN achieving the minimal error in the loss function and the tiniest deviation in the predicted frequency response. Hence, this research recommends that the GRU-based RNN be used for realizing the inverse design of the MPF among all the candidates being studied.