Metasurfaces, known as ultra‐thin and planar structures, are widely used in optical components with their excellent ability to manipulate the wavefront of the light. The key function of the metasurfaces is the spatial phase modulation, originated from the meta‐atoms. Thus, to find the relation between the phase modulation and the parameters of an individual meta‐atom, including the sizes, shapes, and material's optical properties, is the most important but also time‐consuming part in the metasurface design. Here by developing a backpropagation neural network based machine learning tool, the design process of a high performance achromatic metalens can be greatly simplified and accelerated. A library of the phase modulation data from 15 753 meta‐atoms can be generated in less than 1 s by our backpropagation neural network. In the experiment, it is demonstrated that the designed metalens shows an excellent achromatic focusing and imaging ability in the visible wavelengths from 420 to 640 nm without the polarization dependence.

We report on a high-efficiency cross-polarization conversion metamaterial design consisting of novel spiral split-ring resonators (SRRs). Numerical simulations on the resonant electric field and surface current distributions demonstrate that the cross-polarization response is attributed to the charge accumulation in the horizontal SRR gap. The dependencies of resonance frequency on the structural parameters of the SRR reveal that an inductive-capacitive resonance dominates the SRR. We further show that the polarization conversion efficiency can be significantly enhanced by integrating two orthogonal gratings, which enable a linear polarization wave to be rotated to its orthogonal direction with a high efficiency of ∼90%. These results offer a way to engineer novel high-performance metamaterial polarization devices.

We used the backpropagation neural network to design a high performance achromatic metalens at visible range. Experimental demonstration showed that the fabricated achromatic metalens can operate from 420 to 640 nm without the polarization dependence.

Metalenses with both achromatic performance and high focusing efficiency are always challenging, especially in visible range. In this work, a deep learning model is developed to accelerate the design of achromatic metalenses based on the geometric phase theory. During the building process of the phase response library and selection of the nano‐structures, converted transmission coefficients including both phase and amplitude are considered in order to ensure the achromatic focusing, as well as a high focusing efficiency. To test the performance of the design developed from the deep learning model, numerical simulations are performed in the visible wavelengths from 428 to 652 nm, which show a focal length of 266 µm with the deviation under 5%, and the average focusing efficiency reaches 52%.

The determination of the relation between the phase modulation and the geometric parameters of a single meta-atom, is the most important but also time-consuming part in a meta-surface design. Here, by developing a machine learning tool, the design process of a high performance achromatic metalens can be greatly simplified and accelerated. The backpropagation neural network is used to build a library of the phase modulation data with 15753 meta-atoms in less than 1 s. In the experiment, designed metalens has been demonstrated to show a high performance of achromatic focusing and imaging ability in the visible wavelengths from 420 to 640 nm without the polarization dependence.