Inverse Design of Pixelated Differential Bandpass Filters Using a Proximal Policy Optimization Framework Assisted by a Convolutional Neural Network Surrogate Model

A convolutional neural network (CNN) surrogate model (SM)-assisted proximal policy optimization (PPO) inverse design methodology is proposed for pixelated differential bandpass filters (DBPFs). A dumbbell-shaped defected ground structure (DGS) is encoded as a symmetric binary-pixel matrix, in which only one quarter of the pixelated encoding region is independently optimized to reduce the complexity of the design space, and the complete structure is reconstructed through mirroring. Mixed-mode S-parameters are used to jointly optimize differential-mode (DM) matching and common-mode (CM) noise suppression. A single-frequency-point surrogate model (SF-SM) predicts the complex responses and is embedded into PPO for rapid candidate evaluation, reducing repeated full-wave electromagnetic (EM) simulations. The optimized first-order unit is cascaded into a third-order DBPF. The simulated and measured results show good agreement and validate the feasibility of the proposed inverse design methodology for DBPF applications.