Pattern-to-Absorption Prediction for Multilayered Metamaterial Absorber Based on Deep Learning

Abstract

Metamaterial absorbers (MMAs) allow for a wider range of applications than single-layer ones for their multilayered nature, especially in ultrabroadband absorption. However, the design of multilayered MMAs is extremely complicated. Employed deep learning (DL), using a surrogate model to replace the time-consuming full-wave simulations during the design process can greatly improve the design efficiency. In this letter, an efficient approach for constructing the surrogate model of multilayered MMA is proposed. The coding frequency selective surfaces (FSSs) are converted into multichannel images and then amplified to enhance the efficiency of dataset utilization and model training. A convolutional neural network (CNN) is developed as the surrogate model to achieve pattern-to-absorption prediction for the multilayered MMA with a high degree of freedom. Trained on only 18 000 instances with $2^{108}$ total permutations, the CNN can predict the absorption of the meta-atoms within the frequency range of 1.00–20.00 GHz in 0.05 s with a mean deviation of 0.02144. Our letter provides an efficient way to construct surrogate models for multilayered MMA in the DL-based design process.