Multiband Metamaterial absorber with absorption prediction by assisted machine learning

Abstract

This article presents a machine learning-based multiband metamaterial absorber (MMA) for terahertz applications that considerably reduces simulation time while ensuring accuracy. The designed MMA has absorption peaks at 2.93, 3.34, 3.88, 4.30, and 5.43 THz with absorption coefficients of 99.35, 87.12, 99.70, 99.70, and 99.71%, respectively. The proposed ultrathin and compact structure has a thickness of 0.031 λ and a periodic dimension of 0.199 λ corresponding to the lowest absorption frequency. Furthermore, the effect of changing the geometrical features of the MMA is investigated to realize the absorption phenomenon. The absorption coefficients at intermediate frequencies with substrate thickness, periodic dimension, and angles are predicted using extreme randomized tree (ERT) model. Regression models are assessed using an adjusted R2 score as an evaluation metric utilizing various values of nmin and test cases; TC-30, TC-40, and TC-50. The adjusted R2 score approaches 1 for a lower value of nmin, indicating that absorption values can be predicted with a high degree of accuracy. The regression analysis results suggest that the resource requirements can be reduced by 70% using the proposed ERT-based absorber model. The proposed highly efficient multi-band MMA design with machine learning behavior prediction capability can be used for sensing applications.