Abstract
This paper presents a framework for inverse modeling of microwave filters based on a conditional diffusion model developed to address the intrinsic non-uniqueness of reconstructing coupling matrices from specified S-parameter responses. In the forward diffusion process, Gaussian noise is progressively added to the filter design variables, and a denoising network conditioned on the target electrical responses is trained to predict the injected noise at arbitrary diffusion steps. At inference, we initialize with Gaussian noise and execute the learned reverse denoising dynamics process; independent seeds yield diverse sets of physically feasible design-variable solutions that satisfy identical electrical-response constraints. Experiments on fourth- and sixth-order filters show that the proposed method outperforms multivalued neural networks (MVNNs) and conditional generative adversarial networks (CGANs) in prediction accuracy, solution diversity, and cumulative training cost, thereby providing a robust and efficient framework for inverse microwave-filter modeling.
Funding Information
  • National University of Defense Technology (BHCX-2023060101)
  • Tianjin Municipal Science and Technology Program (23ZGZNGX00040)

This publication has 24 references indexed in Scilit: