Scientists from Southeast University, China, have developed an innovative method that uses deep learning to make controlling electromagnetic waves faster and easier. This advancement focuses on programmable metasurfaces—ultra-thin materials engineered to manipulate waves like light and radio waves. The findings, led by Professor Tie Jun Cui, are published in iScience.
Programmable metasurfaces are known for their ability to shape electromagnetic waves, but designing the patterns to control them has always been a slow and challenging task. Electromagnetic waves are forms of energy, such as light or radio signals, that travel through space. “Our model can calculate these patterns almost instantly by simply specifying what the wave should look like,” explained Professor Cui. Their method combines a cutting-edge deep learning technique, a type of artificial intelligence that trains computers to recognize patterns and make decisions, with a physical model of how electromagnetic waves behave. This approach ensures that the results are both accurate and practical. The researchers demonstrated its success in tests, showing it works well for simple and complex wave patterns.
Metasurfaces are essentially ultra-thin materials engineered to interact with electromagnetic waves in precise ways. Unlike traditional materials, they can be programmed to perform specific tasks, such as focusing signals or filtering unwanted frequencies. Previously, these surfaces could not be adjusted after being made, limiting their usefulness. The invention of programmable metasurfaces allowed engineers to make these adjustments electronically, enabling a much broader range of applications. However, figuring out how to create the best wave patterns required complex, time-consuming processes, often involving iterative methods where solutions are refined step by step. These methods were impractical for real-world applications. The new method solves this by using deep learning to bypass these traditional challenges.
The new system has several benefits. It works almost in real-time, meaning it can respond quickly to changing needs. By incorporating physical principles into deep learning models, it also doesn’t require the huge amounts of labeled training data or simulations that older methods needed. Training data refers to the examples that artificial intelligence systems learn from to make accurate predictions. The process uses a deep learning system to calculate the arrangement of surface components and a simplified physical model to predict how the waves will behave. Tests showed the system could produce patterns in moments, a significant improvement over traditional techniques, which could take hours.
To understand how well their method works, the researchers compared it with an older approach called binary swarm optimization, a computational technique inspired by the collective behavior of animals like birds or fish when searching for food. They found that their deep learning model not only worked much faster but also created more effective wave patterns. By eliminating unnecessary data preparation and using a faster process, this approach is both more practical and powerful than earlier solutions.
“The results of our experiments show that this method can reliably design wave patterns for real-time applications, such as scanning objects or improving wireless communication,” said Professor Cui.
The technology has the potential for a wide range of uses, including intelligent sensors, devices that collect and respond to environmental data, tracking systems, and other applications where quick adjustments to electromagnetic waves are needed. The team also showed how it could be used in scenarios requiring constant updates, such as following a moving target with a focused beam of energy.
Despite its success, the researchers acknowledged there are still areas for improvement. For example, their model assumes that changes in the surface’s components only affect the wave’s phase, or timing, without considering more complex interactions. This simplification works well but might overlook some details that could improve accuracy. The researchers are exploring ways to refine their system for even better performance.
This breakthrough represents a significant step forward in technology that controls electromagnetic waves. By making the process faster, easier, and more adaptable, this research opens the door to new applications in communication, sensing, and beyond.
Journal Reference
Jianghan Bao, Weihan Li, Siqi Huang, Wen Ming Yu, Che Liu, and Tie Jun Cui. “Physics-driven unsupervised deep learning network for programmable metasurface-based beamforming.” iScience, 2024. DOI: https://doi.org/10.1016/j.isci.2024.110595
About the Authors
Tie Jun Cui (Fellow, IEEE) received the B.Sc., M.Sc., and Ph.D. degrees from Xidian University, Xi’an, China, in 1987, 1990, and 1993, respectively. In March 1993, he joined the Department of Electromagnetic Engineering, Xidian University, and was promoted to an Associate Professor in November 1993. From 1995 to 1997, he was a ResearchFellow with the Institut fur Hochstfrequenztechnikund Elektronik (IHE), University of Karlsruhe, Karlsruhe, Germany. In July 1997, he joined the Center for Computational Electromagnetics, Departmentof Electrical and Computer Engineering, University of Illinois at UrbanaChampaign, Champaign, IL, USA, first as a Postdoctoral Research Associateand then as a Research Scientist. In September 2001, he was a Cheung-Kong Professor with the Department of Radio Engineering, Southeast University, Nanjing, China. In January 2018, he became the Chief Professor of Southeast University. He is an Academician of the Chinese Academy of Science. He is the first author of the books Metamaterials: Theory, Design, and Applications (Springer, November 2009), Metamaterials: Beyond Crystals, Noncrystals, and Quasicrystals (CRC Press, March 2016), and Information Metamaterials (Cambridge University Press, 2021). He has authored or coauthored morethan 600 peer-reviewed journal articles, which have been cited by more than 62,000 times (H-Factor 122), and licensed more than 150 patents. His researchhas been selected as one of the most exciting peer-reviewed optics research Optics in 2016 by Optics and Photonics News Magazine, ten Breakthroughs of China Science in 2010, and many Research Highlights in a series of journals. His work has been widely reported by Nature News, MIT Technology Review, Scientific American, Discover, and New Scientists. He was the recipient of the Research Fellowship from Alexander von Humboldt Foundation, Bonn, Germany, in 1995, Young Scientist Award from the International Union of Radio Science in 1999.
Che Liu received the B.Eng. degree in information science and technology and the Ph.D. degree from Southeast University, Nanjing, China, in 2015 and 2022, respectively. He is currently a Zhishan Postdoctor with Southeast University. He has been selected for the 2024 annual World’s Top 2% Scientists list (Networking & Telecommunications) published by Elsevier publishing group. His research interests include computational electromagnetic, meta-material, and deep learning. He is committed to use artificial intelligence technology solving electromagnetic issues, including ISAR imaging, holographic imaging, inverse scattering imaging, automatic antenna design, and diffraction neural network.
