AI-driven design of multifunctional metasurfaces for wavefront engineering in IRS and antenna systems

This work presents a deep learning (DL)-assisted inverse-design framework for the automated synthesis of multifunctional pixelated metasurfaces. A deep neural network (DNN) is trained to map prescribed electromagnetic responses, specified by amplitude and phase, to corresponding 3-bit encoded unit-cell geometries for both reflection and transmission modes over a broad frequency range of 12–18 GHz. The framework enables rapid generation of unit cells for a range of electromagnetic functionalities, including single- and multi-beam steering, orbital angular momentum (OAM) beam generation, and phase-gradient metasurface (PGM) lens antennas. To directly evaluate the inverse-design capability, the electromagnetic responses of DNN-generated unit cells are independently verified using full-wave simulations and quantitatively compared with their prescribed target amplitude and phase responses, rather than with the geometries contained in the training dataset. This response-based validation accounts for the non-unique nature of the electromagnetic inverse problem and provides a direct assessment of the design accuracy. The practical applicability of the proposed framework is further demonstrated through the fabrication and experimental characterization of a transmissive PGM lens antenna operating at 14 GHz. The prototype achieves a realized gain of 22.6 dBi and an aperture efficiency of $$39.4\%$$ , with good agreement between simulated and measured results. The combination of broadband amplitude–phase inverse design, 3-bit pixelated implementation, reflective and transmissive functionalities, and experimental antenna validation demonstrates the potential of the proposed framework for rapid and physically realizable metasurface design.