Engineering Metasurfaces with Tandem Neural Networks: Efficient Inverse Design, Functionality-Preserving Repatterning and Accelerated Simulation

Metasurfaces allow for molding electromagnetic waves in ways unattainable with conventional materials, enabling compact and efficient devices across the entire spectrum. While deep learning-based inverse design has boosted progress in this field, two challenges remain: the need for massive simulation data sets and the tendency of generated patterns to be fragmented and difficult to fabricate. This work introduces a tandem neural network (TNN) framework that addresses these issues and provides additional capabilities by simultaneously implementing three functionalities: (1) a spectrum-to-pattern inverse designer that, combining needle-drop initialization and data-augmentation techniques, can be reliably trained with a data set of order ∼7 × 10 3, smaller than in many inverse-design studies with high degrees of freedom, (2) a redesigner that, given a metasurface pattern, generates a different one with the same functionality and new desired geometrical characteristics (e.g., simplified, fabrication-ready unit cells), and (3) a forward predictor that can calculate a metasurface response much faster than a full-wave simulator with similar accuracy. These features rely on the ability of the TNN to exploit the so-called nonuniqueness problem, as well as on the introduction of custom loss functions and other distinctive features. The proposed architecture is verified through full-wave simulations and has applications in stealth, underwater robotics, Internet of things, and RF, 5G, and laser communications.