A generalized deep learning framework for inverse design of multilayer spatial filters under wide-angle incidence

This research introduces an effective inverse design approach for Multilayer Spatial Filters (MSFs), employing a Residual Neural Network (ResNet) informed by transmission line (TL) models. Traditional MSF designs experience interlayer coupling and require substantial optimization calculations. To tackle these issues, we reconfigure the MSF design as a corresponding electronic filter synthesis issue. The suggested method disaggregates the MSF design into autonomous single-layer modules grounded in transmission line theory, thereby successfully attaining interlayer decoupling. A ResNet is subsequently trained to swiftly deduce structural properties of metallic array-dielectric modules directly from tar-get electromagnetic (EM) responses. This expedites the design process, attaining inference durations of approximately one second per layer. MSFs demonstrate superior frequency selectivity, low insertion loss, and better angular stability within the X-band (8–12 GHz). Experimental validation corroborates that the manufactured samples closely align with theoretical predictions. Furthermore, the incorporation of the suggested MSF with a conical radome exhibits robust frequency-selective properties, advantageous for stealth technology. The method provides an efficient and adaptable alternative for the rapid and precise design of high-performance MSFs.