Recent Advances in Physics-Driven Machine Learning Approaches to Intelligent Design of Metasurfaces

Metasurfaces are widely used to manipulate electromagnetic (EM) waves. Traditional methods to metasurfaces design demand extensive full-wave EM simulations. Machine-learning-based approaches have been used to create fast and intelligent metasurfaces design. However, most of these approaches are data-driven methods that require a vast number of data produced by full-wave EM simulations. Physics-driven machine learning approaches can achieve higher accuracy and better physical interpretability than data-driven machine learning approaches. In this paper, we review the recent advances in physics-driven machine learning approaches that incorporate temporal coupled mode theory (CMT) especially the neuro-CMT model to improve the design efficiency of metasurfaces. Three examples of metasurface absorbers are provided to illustrate the benefits of this method.