Small-Sample Machine Learning-Based Design Framework for Pixelated Reflective Metasurfaces

Metasurface design involves time-consuming and computationally expensive optimization of multiple unit cells required to manipulate electromagnetic (EM) waves as desired. This article proposes a small-sample machine learning-based design framework to accelerate the design of pixelated reflective metasurface unit cells. An innovative small-sample data acquisition algorithm is developed to optimize the sampling procedure for improved representativeness. By integrating a hybrid neural network, the proposed design framework significantly reduces the dataset size required for designing pixelated metasurfaces. It simultaneously achieves high design freedom and mitigates the data requirement. Implementation results indicate that the proposed approach reduces data and computational costs significantly compared to conventional machine learning-based methods. For further validation, this approach is applied to design six metasurfaces with various capabilities, including single and dual-beam steering for single- and dual-polarized illumination sources. A good agreement between measurement and simulation results demonstrates the effectiveness of the proposed approach.