Neural network assisted design of a high-performance broadband all-dielectric infrared metasurface absorber with high fabrication tolerance

Neural networks have become a prominent approach for inverse design of metasurfaces. However, the on-demand design of metasurface devices remains challenging due to the high-dimensionality of structural parameter spaces and the nonlinear relationship between these parameters and optical responses. In this paper, we propose a tandem neural network (TNN) framework for the inverse design of a high-performance broadband all-dielectric infrared metasurface (BADIM) absorber. The TNN integrates a reverse design network (RDN) with a forward prediction network (FPN), exhibiting strong generalization capability, as confirmed by the high agreement between its predicted, simulated, and experimental values. To ensure high absorption efficiency in fabricated devices, a fabrication tolerance criterion is incorporated. Four distinct BADIM samples were fabricated using advanced micro-nano manufacturing techniques, each demonstrating an average spectral absorption ( A spe ) exceeding 0.9. Owing to its high geometric symmetry, the BADIM absorber exhibits polarization-independent absorption performance and maintains strong angular insensitivity for both TE and TM polarizations. Specifically, the A spe remains above 0.8 for incident angles up to 60°. In the 8–14 μm wavelength region, our BADIM absorber exhibits exceptional optical performance, demonstrating significant potential for applications in infrared camouflage and thermal imaging.