Physics‐Aware Bidirectional Mamba Network for Microwave Filter Inverse Design

ABSTRACT

Inverse modeling is crucial for microwave filter design, yet traditional neural networks struggle with long‐range electromagnetic (EM) response dependencies and severe multi‐parameter coupling. To address these challenges, a physics‐aware bidirectional Mamba network (PABM) is proposed for microwave filter inverse design. Specifically, a multi‐dimensional EM response preprocessing and enhancement (MSPE) module extracts first‐order and second‐order EM response derivatives to enhance physical feature perception. Crucially, a four‐layer bidirectional Mamba (Bi‐Mamba) backbone captures non‐local broadband EM dependencies via concurrent forward and backward scanning with linear computational complexity. Furthermore, a parallel decoupled prediction (PDGP) network, regularized by an adaptive weight soft‐label (AWSL) loss, balances multi‐parameter geometric parameters (GPs) pre‐dictions. Empirical results on cross‐coupled cavity and hairpin ring resonator filter demonstrate that PABM significantly improves geometric prediction accuracy and ensures stable model convergence over traditional inverse modeling methods. Furthermore, the practical applicability of the proposed PABM framework is successfully validated through the fabrication and measurement of the cross‐coupled cavity filter.