Search

In recent years, significant strides have been made in the inverse design of metamaterial perfect absorbers (MPAs) using deep learning techniques. However, this progress has been hindered by the functional homogeneity arising from the structural uniformity of the inverse-designed MPAs. In this paper, we address this limitation by designing reconfigurable MPAs (RMPAs) with three distinct structures and propose a residual parallel neural network (RPNN) that incorporates the optimized residual fully connected neural network (RFC-NN) for the inverse design of multifunctional MPAs. The trained RPNN accurately predicts the structural parameters and their corresponding absorption spectra with remarkable precision, yielding R2 values of 0.9981 and 0.9928, respectively. With this model, we successfully inverted the design of MPAs with three functions: broadband absorption, dual-band absorption, and triple-band absorption properties. A particularly noteworthy achievement was the realization of absorption bandwidth shifts using liquid crystal (LC) materials. Our RPNN showcases its proficiency in designing RMPAs with multifunctionality, all within a single network model. This marks a significant advancement over previous research methodologies. The proposed methodology holds great promise in diverse applications such as solar energy harvesting, detection, and filtration. [ABSTRACT FROM AUTHOR]

Deep learning-based research on metamaterial absorbers (MAs) has received increasing attention. However, the problem of homogeneity of structure and material of MAs has constrained their further development. In this paper, we designed MA with a top metal layer consisting of eight rectangular nano-rods, and adjusting their lengths can form various structures. In addition, we formed a material database for constructing MAs with the results of random combinations of eight materials and represented them in a coded manner. Meanwhile, we design MAs with ultra-wideband and dual absorption bandwidths using a dual-channel tandem neural network (DTNN). Compared with the existing methods, our method not only simplifies the steps of selecting materials and structures but also enables the design of MAs with different absorption properties. [ABSTRACT FROM AUTHOR]