Deep Learning for Channel Estimation and CSI Feedback in FDD Massive MIMO: A Critical Review and Future Directions

Abstract

Accurate channel state information (CSI) is essential for achieving high spectral efficiency in frequency-division duplex (FDD) massive multiple-input multiple-output (MIMO) systems. However, the increasing dimensionality of massive-MIMO channels creates substantial overhead in downlink channel estimation and uplink CSI feedback. This paper presents a structured critical review of deep-learning-based channel estimation and CSI-feedback methods for FDD massive MIMO. A two-level taxonomy first distinguishes channel estimation from CSI feedback and then categorises the reviewed methods according to their functional objectives and neural architectures. Quantitative comparisons are restricted to studies with sufficiently compatible tasks, channel models or datasets, compression ratios, feedback-bit budgets, and evaluation metrics. Channel extrapolation, beam and blockage prediction, vehicular channel prediction, end-to-end transceiver learning, and over-the-air federated learning are discussed separately as adjacent tasks rather than CSI-feedback methods. The article identifies persistent challenges in benchmark standardisation, finite-bit quantisation, feedback-channel reliability, hardware-aware encoder design, computational complexity, cross-domain generalisation, and integration with beam management. Future research directions include physics-informed learning, online and continual adaptation, channel foundation models, and carefully validated generative approaches for CSI acquisition and feedback.