Generative AI for Planar Substrate Integrated Waveguide Microwave Filters: Autonomous Circuit Synthesis With Coupling-Controlled Topology Generation

As microwave device designs become increasingly complex, traditional full-wave electromagnetic (EM) simulations and manual parameter tuning have become time-consuming and computationally prohibitive. This work introduces an intelligent framework for the automated inverse design of passive microwave components, using substrate-integrated waveguide (SIW) filters as a validation platform. Our approach establishes a versatile parametric model that classifies structural features into fundamental geometric and complex cross-coupling parameters. This structured design space is explored by a deep reinforcement learning (DRL) agent, which is guided by a fast and accurate$S$-parameter prediction model based on a multihead attention-enhanced convolutional neural network (CNN). The multihead attention-enhanced CNN excels in performance prediction and generalization, while the DRL agent adaptively generates optimal structural parameters. We experimentally demonstrate that our framework achieves high-fidelity performance prediction and provides a novel, highly efficient pathway for the automated synthesis and optimization of advanced microwave systems.