Machine-Learning-Based Optimization Method for Wideband Metasurface Antenna

A machine-learning-based method is proposed for the design of wideband metasurface antenna. The artificial neural network (ANN) algorithm is utilized to build an accurate and efficient neural network model for synthesizing antenna geometry parameters. The proposed metasurface antenna evolves from the Mosaic antenna with uniform patch cells. By dividing the patch cells into fractional pieces, the impedance bandwidth of the proposed antenna is improved. In the proposed neural network, the input data is the target reflection coefficients of the metasurfaces antenna, and the neural network predicts the geometry of patch pieces that satisfy the target performance. A prototype antenna is fabricated and measured to verify the design strategy. Measurement results show that the |S11|<−10-dB impedance bandwidth of the proposed antenna is 32.3% or ranging from 4.98 GHz to 6.90 GHz. Compared to the original Mosaic antenna, the impedance bandwidth of the proposed metasurface antenna improves by 21.5%.