Deep Learning-Assisted Inverse Design of Pixelated Two-Port RF Matching Networks

This letter proposes a deep learning-based inverse design framework for two-port electromagnetic(EM) structures, which synergistically integrates deep residual neural networks (ResNet) with Direct Binary Search (DBS) optimization. The method targets S-parameter prediction and automatic layout optimization of pixel-level EM structures. A ResNet model is trained on a large-scale dataset to enable fast and accurate prediction of S-parameters, thereby significantly accelerating the optimization process. By discretizing the structure layout into binary pixels, the optimization capability of the DBS algorithm is compared with that of the genetic algorithm (GA) under identical initial conditions, it is shown that DBS can explore the design space more efficiently while maintaining fabrication feasibility. The proposed method is validated on a 5GHz two-port matching network for high-impedance transformation, showing strong potential for achieving end-to-end automation in radio-frequency (RF) circuit design. Moreover, the approach is extendable to multi-port devices with more complex topologies.