Phase-to-pattern inverse design paradigm for fast realization of functional metasurfaces via transfer learning

TLDR

Metasurfaces enable unprecedented control of electromagnetic waves, yet designing them requires optimizing many meta‑atoms—a time‑consuming and sometimes prohibitive process. The study proposes a fast, accurate inverse design method that uses transfer learning to generate metasurface patterns directly from desired phase profiles. A transfer‑learning network built on GoogLeNet‑Inception‑V3 predicts the phases of 2 × 8×8 meta‑atoms with ~90% accuracy and is employed to monolithically generate patterns for 2D focusing and abnormal reflection. Simulations and experiments confirm high design accuracy, demonstrating a fast inverse design paradigm that can readily build a full‑phase‑span meta‑atom library.

Abstract

Abstract Metasurfaces have provided unprecedented freedom for manipulating electromagnetic waves. In metasurface design, massive meta-atoms have to be optimized to produce the desired phase profiles, which is time-consuming and sometimes prohibitive. In this paper, we propose a fast accurate inverse method of designing functional metasurfaces based on transfer learning, which can generate metasurface patterns monolithically from input phase profiles for specific functions. A transfer learning network based on GoogLeNet-Inception-V3 can predict the phases of 2 8×8 meta-atoms with an accuracy of around 90%. This method is validated via functional metasurface design using the trained network. Metasurface patterns are generated monolithically for achieving two typical functionals, 2D focusing and abnormal reflection. Both simulation and experiment verify the high design accuracy. This method provides an inverse design paradigm for fast functional metasurface design, and can be readily used to establish a meta-atom library with full phase span.