ANNs-Based Surrogate-Assisted Optimization Design of Microwave Filters With SA-PSO

In conventional microwave filter design, extensive full-wave electromagnetic (EM) simulations often result in high computational cost and long design cycles. This paper proposes an automated optimization design framework, termed E-NASOD, which integrates surrogate modeling based on artificial neural networks (ANNs) with simulated annealing-particle swarm optimization (SA-PSO). ANNs are developed to model the mapping from structural parameters to S-parameter responses, and a weighted loss function is introduced to improve prediction accuracy in critical response regions, especially near the minima of the reflection response, while also improving the optimization landscape. In addition, a composite fitness function is formulated for filter synthesis and incorporated into SA-PSO for goal-oriented parameter search. A dielectric waveguide filter example using a cross-coupled structure is employed to validate the proposed framework. The results demonstrate that the proposed method is effective for both conventional and highly nonlinear microwave filter synthesis problems, while offering good stability, robustness, and improved design efficiency.