Significant Bandwidth Enhancement of LowReflection Checkerboard Metasurfaces Using Machine Learning

A method of machine learning is proposed for enhancing the low-reflection bandwidth of checkerboard metasurfaces. Instead of iterative optimization to a satisfying performance, a conditional deep neural network (cDNN) is trained to establish a mapping relation between reflection phase and the unit-cell structures first, with which new types of unit cells are massively generated to explore the largest bandwidth limit. As a result, the checkerboard metasurface designed by the trained cDNN achieves a 10-dB bandwidth of scattering reduction 22% higher than the best in the training dataset, from 62% to 84%. The proposed method demonstrates the ability of machine learning for electromagnetic discovery and performance breakthrough.