Transfer learning and MEMS-integrated design of metasurfaces with dynamic electromagnetic properties

Our work aims to accelerate metasurface design using machine learning and to experimentally implement tunable metasurfaces. We explored transfer learning to optimize combined unit cell structures with fewer simulations, reducing computational costs. For a dipole-dipole structure, transfer learning required only 29 simulations to reach an MSE of 0.05, compared to 583 in direct learning. We also implemented MEMS-based tunable metasurfaces, demonstrating a microwave SSPP filter with voltage-tunable attenuation and simulating a MEMS-tuned beam-steering antenna (57.5 GHz, >60° deflection). This approach readily scales to terahertz frequencies.