Metaheuristic-Assisted Inverse Design of Elastic Metasurfaces with Parametric Polygonal Pillars

This study develops an inverse design framework for elastic metasurfaces with parametric polygonal pillars to manipulate transmitted flexural waves. The proposed unit cell is characterized by five geometric parameters, including three independently adjustable parameters defining its cross-sectional shape, allowing flexible control of the resonance characteristics. To address the increased complexity of identifying target modes with different polarizations, an automated modal screening procedure is introduced to construct a finite element dataset. The multilayer perceptron is trained as an efficient forward surrogate for predicting the characteristic frequencies of the target bending and compression modes. Based on the pretrained MLP, a tandem neural network enables on-demand retrieval of geometric parameters for prescribed resonance frequencies, while particle swarm optimization further explores alternative feasible configurations satisfying similar resonance conditions. Finally, phase gradient metasurfaces are constructed to demonstrate anomalous refraction, wave splitting, and self-bending flexural wave propagation through full-wave numerical simulations.