In this paper, we introduce a new hybrid optimization method for the inverse design of metasurfaces, which combines the original Harris hawks optimizer (HHO) with a gradient-based optimization method. The HHO is a population-based algorithm that mimics the hunting process of hawks tracking prey. The hunting strategy is divided into two phases: exploration and exploitation. However, the original HHO algorithm performs poorly in the exploitation phase and may get trapped and stagnate in a basin of local optima. To improve the algorithm, we propose pre-selecting better initial candidates obtained from a gradient-based-like (GBL) optimization method. The main drawback of the GBL optimization method is its strong dependence on initial conditions. However, like any gradient-based method, GBL has the advantage of broadly and efficiently spanning the design space at the cost of computation time. By leveraging the strengths of both methods, namely GBL optimization and HHO, we show that the proposed hybrid approach, denoted as GBL-HHO, is an optimal scenario for efficiently targeting a class of unseen good global optimal solutions. We apply the proposed method to design all-dielectric meta-gratings that deflect incident waves into a given transmission angle. The numerical results demonstrate that our scenario outperforms the original HHO.
Keywords: inverse design; metasurfaces; optimization.