The terahertz region is a special region of the electromagnetic spectrum that incorporates the advantages of both microwaves and infrared light waves. In the past decade, metamaterials with effective medium parameters or gradient phases have been studied to control terahertz waves and realize functional devices. Here, we present a new approach to manipulate terahertz waves by using coding metasurfaces that are composed of digital coding elements. We propose a general coding unit based on a Minkowski closed-loop particle that is capable of generating 1-bit coding (with two phase states of 0 and 1806), 2-bit coding (with four phase states of 0, 906, 1806, and 2706), and multi-bit coding elements in the terahertz frequencies by using different geometric scales. We show that multi-bit coding metasurfaces have strong abilities to control terahertz waves by designing-specific coding sequences. As an application, we demonstrate a new scattering strategy of terahertz waves-broadband and wide-angle diffusion-using a 2-bit coding metasurface with a special coding design and verify it by both numerical simulations and experiments. The presented method opens a new route to reducing the scattering of terahertz waves.

Because of their exceptional capability to tailor the effective medium parameters, metamaterials have been widely used to control electromagnetic waves, which has led to the observation of many interesting phenomena, for example, negative refraction, invisibility cloaking, and anomalous reflections and transmissions. However, the studies of metamaterials or metasurfaces are mainly limited to their physical features; currently, there is a lack of viewpoints on metamaterials and metasurfaces from the information perspective. Here we propose to measure the information of a coding metasurface using Shannon entropy. We establish an analytical connection between the coding pattern of an arbitrary coding metasurface and its far-field pattern. We introduce geometrical entropy to describe the information of the coding pattern (or coding sequence) and physical entropy to describe the information of the far-field pattern of the metasurface. The coding metasurface is demonstrated to enhance the information in transmitting messages, and the amount of enhanced information can be manipulated by designing the coding pattern with different information entropies. The proposed concepts and entropy control method will be helpful in new information systems (for example, communication, radar and imaging) that are based on the coding metasurfaces.

Metasurfaces provide unprecedented routes to manipulations on electromagnetic waves, which can realize many exotic functionalities. Despite the rapid development of metasurfaces in recent years, the design process of metasurface is still time‐consuming and computational resource‐consuming. Moreover, it is quite complicated for layman users to design metasurfaces as plenty of specialized knowledge is required. In this work, a metasurface design method named REACTIVE is proposed on the basis of deep learning, as deep learning method has shown its natural advantages and superiorities in mining undefined rules automatically in many fields. REACTIVE is capable of calculating metasurface structure directly through a given design target; meanwhile, it also shows the advantage in making the design process automatic, more efficient, less time‐consuming, and less computational resource‐consuming. Besides, it asks for less professional knowledge, so that engineers are required only to pay attention to the design target. Herein, a triple‐band absorber is designed using the REACTIVE method, where a deep learning model computes the metasurface structure automatically through inputting the desired absorption rate. The whole design process is achieved 200 times faster than the conventional one, which convincingly demonstrates the superiority of this design method. REACTIVE is an effective design tool for designers, especially for laymen users and engineers.

Artificial intelligence is facilitating human life in many aspects. Previous artificial intelligence has been mainly focused on computer algorithms (e.g. deep-learning and extremelearning) and integrated circuits. Recently, all-optical diffractive deep neural networks (D 2 NN) were realized by using passive structures, which can perform complicated functions designed by computer-based neural networks at the light speed. However, once a passive D 2 NN architecture is fabricated, its function will be fixed. Here, we propose a programmable artificial intelligence machine (PAIM) that can execute various intellectual tasks by realizing hierarchical connections of brain neurons via a multi-layer digital-coding metasurface array. Integrated with two amplifier chips in each meta-atom, its transmission coefficient covers a dynamic range of 35 dB (from -40 dB to -5 dB), which is the basis to construct the reprogrammable physical layers of D 2 NN, in which the digital meta-atoms make the artificial neurons alive. We experimentally show that PAIM can handle various deep-learning tasks for wave sensing, including image classifications, mobile communication coder-decoder, and real-time multi-beam focusing. In particular, we propose a reinforcement learning algorithm for on-site learning and discrete optimization algorithm for digital coding, making PAIM have autonomous intelligence ability and perform self-learning tasks without the support of extra computer.