A Novel Metasurface Inverse Design Based on Back Propagation Neural Network

This paper proposes a novel reflective meta-surface inverse design by utilizing a back propagation neural network. A reflective meta-surface, measuring 0.92 m in width, 0.92 m in height, and 0.508 em in thickness, is synthesized. This meta-surface is composed of twelve distinct unit types, each possessing unique phase-shifting characteristics. When illuminated by a multi-mode waveguide horn employing the offset design, the meta-surface demonstrates a gain of 31.65 dB at a frequency of 5.8 GHz. Furthermore, the simulated design achieves a side lobe level of 23 dB in the far-field region, accompanied by a system efficiency of 36% and a relative 3-dB bandwidth of 7%. By incorporating more training data and enhancing the machine learning algorithms, this design methodology could be applied to generate complex meta-surface structures with multi-frequencies and multi-polarization responses, demonstrating significant potential in multi-functional meta-surface integration.