The physical layer (PHY) is fundamental to wireless communication systems, enablingrobust signal transmission in complex and dynamic environments. Traditional PHY designs follow ablock-based, model-driven approach where components—such as channel coding, modulation, andchannel estimation—are optimized independently under simplified assumptions, often leading toperformance degradation in real-world scenarios. Deep Learning (DL) offers a data-driven alternativecapable of learning complex, nonlinear mappings directly from data, improving adaptability andaccuracy under diverse channel conditions. This paper reviews recent advances (2020–2025) in applyingDL to enhance key PHY functions, including channel estimation, signal detection, modulationclassification, coding/decoding, beamforming, and physical layer security. We examine various neuralarchitectures—such as CNNs, RNNs, autoencoders, GANs, and reinforcement learning agents—highlighting their roles in next-generation networks (5G, B5G, 6G). While DL demonstrates superiorperformance over conventional methods in adaptability, spectral efficiency, and robustness, challengesremain in computational complexity, interpretability, training data scarcity, and generalization acrossenvironments. The review synthesizes state-of-the-art methods, identifies open issues, and outlinesfuture research trends—such as model-driven DL, transfer/meta-learning, edge intelligence, and crosslayer optimization—towards building intelligent, adaptive, and scalable PHY designs for future wirelesssystems.
Zahhad et al. (Sun,) studied this question.