Fast Design of Multi-Beam Metasurface Antennas Using Generative Machine Learning

A method of machine-learning aided design (MLAD) is proposed for accelerated design and synthesis of multi-beam metasurface antennas (MAs). By training a deep-learning network with 2400 samples of unit cells, two designs of a single-beam MA and a dual-beam MA out of the training dataset are automatically predicted, where the former is simulated and measured with both results in line with the network prediction. Different from conventional methods of iterative numerical optimization, the proposed MLAD method is based on generative machine learning. It allows for efficient generation of a large number of new metasurfaces beyond the training dataset for achieving multiple beams in given directions, opening a new avenue for further exploring the new performance boundary of metasurface devices in an accelerated way.