A Comparison Study of Deep Learning Algorithms for Metasurface Harvester Designs

Abstract

The paper compares three deep learning artificial intelligence algorithms used for metasurface design. In-house design code for designing metasurface structures was developed with Python, the NumPy library. To facilitate the study, the three algorithms used are AdaBelief, Adam, and Yogi. According to the numerical comparison study, Adam has a better performance in terms of model generalization with a large dataset (in our case 7000 samples), while Adabelief and Yogi show a better performance in terms of a low dataset (in our case 4,000 samples), and Yogi has a better performance with a lower dataset correlation between the predicted performance of the energy harvester obtained from three algorithms. Yogi and Adablief performance could be improved by manipulating the hyper-parameters.