Fig. 2

Spectral tuning for meta-atoms. The dark blue and orange arrows with texts represent the close interplays among three parts of the schematics, namely, meta-atoms (left, as input), learning algorithms (middle) and spectra (right, as output). Left: four typical geometric shapes of meta-atoms situated on a substrate. Middle: the traditional optimizer (e.g., GA) versus deep learning model (e.g., MLP of deep neural network). Right: spectrum modifications from the original line-shape in blue towards the target response in red. Note that the layout of mete-atoms could be represented as image matrices (see Fig. 3) or physical parameter vectors (see Figs. 4 & 5) for the further treatment in the learning algorithms