Generative and Explainable AI for High-Dimensional MIMO-OFDM Channel Estimation in Time–Frequency–Space Domain
Nghia Thinh Nguyen, Tri Nhu Do · IEEE Transactions on Green Communications and Networking · 2026
In this paper, we address a high-dimensional (HD) geometry-based channel model (GCM) characterized in the time-frequency-space (TFS) domain for Multiple-Input Multiple-Output Orthogonal Frequency-Division Multiplexing (MIMO-OFDM) transmission systems. Specifically, HD-TFS data tensors, collected via realistic ray-tracing channel models, are sensitive to geometric-based parameter changes, i.e., varying rays and clusters in multipath, due to user mobility. Therefore, we propose an adversarial training framework where the deep learning channel generator learns an accurate variational posterior distribution (VPD) that is robust to geometry changes of the channel model. By extracting features from the GCM, a novel loss function is introduced that incorporates the third moment of the GCM power in the HD-TFS domain. Besides, a stochastic gradient estimator based on the reparameterization trick is introduced, leveraging the distribution of GCM to enhance VPD learning. The numerical results on realistic scenarios show that the learned VPD is not only accurate, with superior HD-TFS channel estimation in low signal-to-noise ratio (SNR) regimes, but also improves runtime efficiency through an optimized channel generator architecture design. Finally, an explanation mechanism is proposed to highlight high channel gain within the HD-TFS tensor, which is interpreted as attention-relevant areas during model inference in realistic scenarios.