Digital Coded Metasurface Design Based on Deep Learning Technique: A Study

Abstract

This paper presents a comprehensive study of advanced techniques for designing digitally coded metasurfaces, highlighting their ability to manipulate electromagnetic wave propagation. Five key methodologies are investigated and compared: Back Projection (BP), Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Superposition, and Deep Learning (DL). While the computationally efficient BP method is suitable for single-beam configurations, it is inadequate for multi-beam or complex radiation scenarios. GA and PSO, both optimizationbased, provide improved pattern accuracy yet suffer from high computational complexity and slow convergence, making them unsuitable for real-time applications. The superposition method offers design simplicity but does not accurately manage side-lobe suppression or beam interference. In contrast, the DL approach outperforms the others by effectively learning intricate spatial relationships between target beam profiles and the required metasurface codes, delivering high prediction accuracy and realtime computational efficiency. Through extensive analysis, this study shows that deep learning surpasses traditional methods, making it the most effective and scalable approach for the rapid, high-precision design of digitally programmable metasurfaces.