LAM-LSTM assisted inverse design of hexagonal honeycomb metasurface

Abstract Traditional design of frequency selective surface (FSS) absorbers is time-intensive, requiring numerous electromagnetic (EM) simulations for parameter optimization. For certain metasurfaces with complex structures, forward modeling presents significant challenges. This paper introduces a target prediction network to generate target S 11 curves directly from design specifications, and proposes an inverse design model based on a long short-term memory network with a local attention mechanism (LAM-LSTM). The method establishes a robust relationship between geometric structures and EM parameters, enabling effective optimization of structural parameters while addressing the ‘uniqueness’ problem. The proposed approach is demonstrated through the design of a hexagonal close-packed metasurface absorber with an 8–17 GHz bandwidth. Simulations and experiments validate the model’s effectiveness, achieving a mean squared error below 1.6 × 10 −4 on the validation set. The results also confirm that the LAM-LSTM based model can rapidly and accurately perform inverse design for hexagonal close-packed absorbers, which can be applied to the design of EM devices with polygonal configurations.