Abstract: Orthogonal Frequency Division Multiplexing (OFDM) remains the dominant waveform for broadband wireless systems such as Wi-Fi, LTE/5G NR variants, and IoT links. Accurate channel state information (CSI) and robust symbol detection are crucial for optimal OFDM receiver performance. In recent years, deep learning (DL) techniques both purely data driven and model-driven (hybrid) have shown significant potential in improving channel estimation (CE) and signal detection, especially under severe channel impairments, nonlinearities, and hardware imperfections. This paper presents a comprehensive review of the state of the art in deep-learning-based CSI estimation and signal detection for OFDM systems. It summarizes conventional estimation techniques (LS, MMSE), explains various DL architectures (CNNs, RNNs/LSTMs, Transformers, and unfolded networks), and examines model-driven hybrid approaches such as COMNET, CSNET, and unfolded DETNET-style frameworks. Additionally, it discusses commonly used datasets and evaluation metrics, compares reported performances, highlights practical deployment challenges (including computational complexity, generalization ability, dataset bias, and model interpretability), and outlines promising future research directions such as meta-learning, domain adaptation, lightweight inference, and end-to-end joint estimation–detection learning.
The review consolidates and analyses 25 representative references encompassing foundational studies, methodological papers, surveys, and recent advancements in this rapidly evolving research area.

Keywords: OFDM, Channel Estimation, Signal Detection, Deep Learning, CNN, RNN, Model-Driven Deep Learning, Unfolded Networks, MIMO

Downloads: | DOI: 10.17148/IJIREEICE.2026.14202