Performance Investigation of Different Optimization Algorithms in Neuro-CMT-Based Intelligent Design of Metasurfaces

Conventional metasurface design methods require a large number of full-wave electromagnetic(EM) simulations to obtain the optimal geometric parameter values, resulting in a low optimization efficiency. Recently, coupled mode theory (CMT) and neural networks have been combined (i.e., neuro-CMT) to rapidly predict the EM response of a metasurface and thus accelerate its design optimization process, in which gradient-based optimization methods (i.e., Quasi-Newton) are used to find the optimal geometric parameter values. However, gradient-based optimization methods may not achieve the optimal design when the initial design is far away from the optimal solution. In this paper, we investigate the performance of four optimization algorithms (i.e., quasi-Newton, genetic algorithm, patternsearch, and surrogateopt) in neuro-CMT-based design optimization of metasurfaces, aiming to further improve the optimization efficiency of the neuro-CMT method.