Dataset Shrinking for Accelerated Deep Learning-Based Metamaterial Absorber Design
Qimin Ding, Guobin Wan, Nan Wang, Xin Ma
Abstract
The design of metamaterial absorbers (MMAs) using deep learning (DL) models typically involves the construction of a large dataset through time-consuming full-wave simulations (FSs). In this letter, an efficient dataset shrinking method is proposed that employs both electromagnetic (EM) theory and DL theory to enhance design efficiency. The impact of each MMA on training the DL model is estimated through the equivalent circuit model (ECM) of the MMA by analyzing the gradients of the MMA parameters with respect to DL model training. Impact sampling is then used to construct a smaller dataset while still maintaining DL model performance and reducing time on FS. The proposed approach is validated through the design of single- and multilayer absorbers using the inverse-DL algorithm and DL integrated into the genetic algorithm (GA). Numeric results demonstrate that the proposed method achieves a much faster design than those without optimized dataset construction processes.