Transfer learning for data-efficient modeling of composite metasurface and FSS unit cells

Abstract This work presents a transfer learning approach for the forward design of composite unit cells of metasurfaces and frequency-selective surfaces (FSS). As a specific application, we target reconfigurable intelligent surfaces (RIS) for wireless communication. We demonstrate that forward models for dual-dipole unit cells can be trained using significantly fewer data by reusing pre-trained models of simpler, single-dipole structures. Our architecture reduces the required training data by up to a factor of 25 while maintaining a mean squared error (MSE) on the order of $$10^{-2}$$ . After establishing the forward model for the dual-dipole RIS, we use it in an inverse design framework to synthesize a composite 2-bit RIS unit cell operating at 26.5 GHz. The resulting RIS provides phase modulation in a 270 $$^\circ$$ range by tuning the reverse bias voltage of integrated varactor diodes. Numerical simulations confirm the validity of the proposed approach, establishing transfer learning as a data-efficient and practical method for the design of composite unit cell structures.