Few-Sample Information-Enhanced Inverse Design Framework for Customizing Transmission-Modulated Elastic Metasurfaces

Abstract

The burgeoning field of metamaterials and metasurfaces has been significantly propelled by the integration of deep learning (DL) techniques, enabling a rapid artificial design with tailored exotic properties. However, the DL-based inverse design strategies frequently face reliability issues when dealing with limited sample datasets. To overcome this challenge, we propose a few-sample information-enhanced inverse design framework specifically developed for the efficient design of columnar elastic metasurfaces, to fulfill customized transmission modulation requirements. The novelty of our approach lies in developing an information-enhanced convolutional neural network (IECNN) integrating substructure combinations, stacking effects, and CBAM, which provide more comprehensive and refined input data to substantially improve the prediction performance and generalization capability. So, the IECNN can precisely replicate FEM transmission calculations with about the 105 times computational speedup using the few-sample, significantly reducing computational time and resources. By integrating IECNN with a genetic algorithm, an automated inverse design framework is established to yield the metasurface structure with specified target transmission performance in only 3.5 minutes. Various numerical simulations and experimental measurements demonstrate its practicality and effectiveness. Furthermore, the physical mechanism behind the customized transmission properties is elucidated to offer deeper insights into the design process. Our approach not only ensures reliable and superior design outcomes but also diminishes the dependence on extensive labeled datasets, presenting a pragmatic framework for metasurface inverse design, particularly valuable in few-sample scenarios.