Efficient design method for terahertz broadband metasurface patterns via deep learning

The traditional design method of metasurfaces is the trial-and-error process with full-wave electromagnetic simulation. Recently, as an effective method, deep learning has been widely used in a variety of fields to solve complex problems. Here, forward prediction and inverse design methods for terahertz (THz) random metasurfaces are proposed based on deep Convolutional Neural Networks (CNN) and Genetic Algorithms (GA). According to the metasurface pattern, the forward prediction model accurately obtains the reflection amplitude and phase response. Compared with the Full-wave solver, the calculation speed is increased by 40,000 times. Furthermore, the THz random structure can be accurately and quickly derived from the target response, operating in a broadband range of 0.2–2 THz. Then, we discuss the advantages of single-objective and multi-objective optimization in the inverse design of metasurface patterns. By combining with the GA, the design efficiency is greatly improved. This can serve as an efficient method for global optimization in complex designs. Finally, we obtain two meta-atoms used to encode metasurface in only 10 min and built a three-beam splitter. The model proposed in this paper provides a new approach to metasurface design at THz frequencies.