The interaction of machine learning and optics/photonics is transforming the way we design new photonic structures, unearth latent physical laws, and develop intelligent photonic devices. Despite certain achievements, a major impediment persistently exists; datasets and networks are only disposable. Thus, for each new state or task, all datasets and networks have to be discarded, and it is imperative to reconstruct new datasets and networks, leading to an enormous waste of resources. In machine learning-based metamaterial designs, much effort has been inaugurated to enlarge the training dataset or construct specific networks. Either way, each metamaterial is physically separated, and the data utilization efficiency is very low. Therefore, it is highly desirable to exploit whether there are any physical connections or network correlations among various metamaterials that can robustly handle a broad range of metamaterials.
In our recent work published in Light Science & Application, a knowledge-inherited neural network for metasurface inverse design and wireless applications is developed in a “green” manner. As its name implies, such a network can inherit the knowledge from “parent” metasurfaces, and then freely assemble for “offspring” metasurfaces. This inheritance-to-assembly mixture scheme is analogous to building a container-type house. In other words, this method bridges the metasurface assembly in physical space to the synthesis of the neural network. Such a paradigm breaks a long-standing stereotype that neural networks only work for predefined and shape-bound objects. Further, they benchmark the universality of this paradigm by one aperiodic metasurface and three periodic stretchable origami metasurfaces with an accuracy of over 86.7%, in stark contrast to 20.0% that was achieved by conventional neural network. More interestingly, the successful design and solid experiment of intelligent origami metasurfaces usher in an innovative spaceborne antenna for future satellite communication. The reported method and technique will open up a new avenue for automatic metasurface design and leverages the assemblability to broaden the adaptability of intelligent metadevices.
Different from conventional “brick-by-brick” neural network whose input–output parameters are predetermined and fixed, the reported “panel-by-panel” method endows the network recyclability and flexible assemblability, analogous to building a container-type house with high flexibility and free assembly (Fig. 1a). The knowledge-inherited paradigm is composed of two functional networks, i.e., an inherited neural network (labelled INN) and an assembled neural network (labelled SNN). The INN is responsible for the inverse design of each “panel” metasurface, and the SNN functions as a deployer to assign the task for each INN. The operational principle of the network is as follows:
“The database consists of seven “panel” metasurfaces, each of which has its own INN (Fig. 1b). For a given metasurface, such as a rectangle and a diamond-shape, we first construct it with these seven “panel” metasurfaces in physical space, and then synthesize the holistic neural network by using the handy-prepared INN. In this procedure, the INN is completely inherited and reserved, and instead, we only need to dynamically adjust the SNN, enabling a green and data-efficient metasurface inverse design.”
In practice, by applying our inheritance-to-assembly strategy, we can also achieve multi-stage network assembly for larger or more complex models. Another point we would like to emphasize is that the advantages of “panel-by-panel” are extremely maximized in the application of geometrical periodic metasurfaces, such as origami metasurfaces. Attached with external force stimulation, origami structures can smoothly dominate their folding/unfolding movement to form a “modular” metasurface with excellent structural stiffness and adjustable periodicity, which firmly coincides with our “panel-by-panel” policy. Its compatibility and lightweight characteristics also appeal to underlying applications in satellite communications (Fig. 1c). As a high coverage communication system, satellite communication can flexibly peruse multiple access communication and channel on-demand allocation, offering terrific signals for every corner of the world, even in remote mountainous areas or Mount Everest.
Figure 1. Schematic of the knowledge-inherited neural network. a, Knowledge-inherited paradigm for a metasurface inverse design. Similar to “brick-by-brick” masonry buildings, conventional neural networks are inseparable, fossilized, and single-functional once built up. In contrast, the proposed knowledge-inherited (“panel-by-panel”) neural network is oriented for multi-object and shape-unbound metasurfaces. b, Flowchart of the knowledge-inherited paradigm. To match the inheritance-to-assembly scheme, two networks (INN and SNN) are established, where the INN is responsible for the inverse design of each “panel” metasurface, and the SNN aims to explore the relationship between the global target EM response and the local EM response provided by each metasurface panel. For a given “offspring” metasurface, the holistic neural network can be synthesized by assembling the INN and dynamically adjusting the SNN. c, Schematic of satellite communication based on intelligent origami metasurfaces. As a flexible, ultrathin and low-cost competitor for free beam forming, the origami metasurfaces can be conformed on the wings of satellites, which can be freely folded and stretched. Two satellites and one earth station are depicted to illustrate intersatellite communication (case 1) and satellite-earth communication (case 2).
At the first glance, our method and transfer learning may seem similar. However, we would like to particularly emphasize that the underlying mechanism and performance are fundamentally different. And our method has not been mentioned in any previous study and cannot be readily duplicated from/to computer science. Their differences are explained below, including operation principle and performance comparison.
Transfer learning is a mature algorithm extended from computer science. The basic operation process is to transfer the pre-trained neural network in the source task to assist the training of target task. However, the performance of transfer learning cannot be guaranteed. Sometimes, the performance even becomes worse compared with that without transfer learning. In other word, transfer learning is like a ‘black box’ without revealing its internal mechanism, which heavily relies on the brute-force attack of features and lacks reasonable explanation (Fig. 2a).
By contrast, our “knowledge-inherited learning” is a unique and exclusive method, which can be regarded as a ‘white box’ with physical connection between the internal transferred knowledge (Fig. 2b). This is the first time to propose this method and it is not extended from computer science. Due to the inimitable physical character of metasurfaces, our knowledge-inherited network is associated with the complex spatial information of structures, which can further inherit the knowledge from “parent” metasurfaces, and then freely assemble for “offspring” metasurfaces. In other words, the synthesis of networks in the virtual space is inseparably correlation to the metasurface assembly in physical space. Further, combined with a physical auxiliary module in each INN, we can obtain the phase distribution of each assembled panel easily and accurately without the non-uniqueness issue (namely, a nearly identical far field can be induced by multiple phase distributions).
Figure 2. Comparison between (a) transfer learning and (b) our knowledge-inherited learning.
See the article:
Yuetian Jia, Chao Qian, Zhixiang Fan, Tong Cai, Er-Ping Li, and Hongsheng Chen
A knowledge-inherited learning for intelligent metasurface design and assembly, Light: Science & Applications, 12, 82 (2023). https://doi.org/10.1038/s41377-023-01131-4