Design of Tunable Metasurface Using Deep Neural Networks for Field Localized Wireless Power Transfer

Wireless power transfer (WPT), a convenient method for powering multiple devices, enables a truly wireless connection, eliminating the need for periodic charging and replacing a battery. To further enhance WPT, the unique characteristics of metamaterial, such as its field focusing and evanescent wave amplification, have been successfully utilized. With subwavelength characteristics, computational challenges arise when the number of metamaterial unit cells is increased. In this work, we investigate a deep neural network (DNN)-based design of the tunable metamaterial for WPT. Using structures specifically designed for different tasks, the DNN predicts the frequency spectra and synthesizes the unit cell’s design parameters. When trained using a set of ~23000 randomly selected designs, we achieve an accumulated mean square error (MSE) of less than $1.5\times 10^{-3}$ for 97.3% of the 1929 test set. For synthesizing the unit cell’s design parameters, the MSE is less than $2.5\times 10^{-3}$ for 95.7% of the test set. The data-driven method is further extended to a generative adversarial network (GAN) to create the WPT paths and predict the frequency spectra of them. To achieve high efficiency, we propose a cost function focusing on the spectra’s transmission peak. After training using 80 000 measured data, the GAN can create WPT paths that efficiently connect the transmitter and the receiver on the metasurface. The results show that the DNN provides an alternative and efficient design method for the metamaterial, replacing traditional EM-simulation-based approaches.