Terahertz Metasurface Design with Arbitrary Hexagon Patterns Based on Deep Learning

For the metasurface design, a novel unit cell of arbitrary hexagon patterns and a neural network that can generate these patterns are proposed to get target transmission amplitudes and phases at will. The hexagon patterns are encoded to few design variables but retain high design freedom and fabrication feasibility. Long short-term memory is applied to the neural network to provide meaningful design parameters. A terahertz transmissive metasurface lens that consists of the generated patterns is simulated to further verify the design concept. This work provides a new solution to the inverse design of metasurface and opportunities for the future of multi-functional terahertz devices.