Transfer-Learning-Assisted Inverse Metasurface Design for 30% Data Savings

Deep learning is found to be a powerful data-driven force to transform the way we discover, design, and utilize photonics and metasurfaces. More recently, there has been growing interest in deep-learning-enabled on-demand structural design, as it can ease the limitations of low efficiency, time-consuming, and experience navigation in conventional design. However, training data is a valuable source, especially for high-dimensional scatterers. It is extremely challenging and costly to keep the pace of data collection with the increasing degrees of freedom. Here, we propose a transfer-learning-assisted inverse-metasurface-design method to relieve the data dilemma. A flexible transferrable neural network composed of an encoder-decoder network and a physical assistance network is constructed, the latter of which is attached to solve the nonuniqueness problem. Starting from the 5 \ifmmode\times\else\texttimes\fi{} 5 metasurface, we successfully migrate the inverse design to a 20 \ifmmode\times\else\texttimes\fi{} 20 metasurface, with a Pearson correlation coefficient that reaches 97%. Compared with direct learning, the data requirement is reduced by over 30%. In the experiment, we validate the concept via wave-front customization. Our work constitutes a green and efficient inverse-design paradigm for fast far-field customization and provides a key advance for the next generation of large-scale intelligent metadevices.