Flow Matching-based Sparse Bayesian Learning for AI-Enabled Channel Estimation in Underwater Acoustic OFDM Systems

Abstract

Accurate channel estimation is critical for orthogonal frequency-division multiplexing (OFDM) systems in underwater acoustic (UWA) communications, yet it faces significant challenges from long delay spreads, sparse multipath structures, and nonGaussian noise. Existing sparse Bayesian learning (SBL)-based methods often suffer from model mismatch and computationally intensive posterior inference in complex underwater environments. Meanwhile, diffusion-based generative models, despite their powerful distribution modeling capability, incur high latency due to iterative stochastic sampling, which limits their applicability in real-time cross-domain marine networks. To address these issues, this paper proposes a flow matching (FM)-based SBL framework for efficient and accurate UWA channel estimation. Specifically, we develop a generative-analytic hybrid inference strategy, where FM constructs deterministic probability flows to approximate the posterior distribution of channel coefficients, significantly reducing inference steps while preserving estimation fidelity. To exploit structural sparsity, a gated U-Net with layer-wise gating mechanism is designed to embed SBL hyperparameters, enabling adaptive feature selection aligned with the automatic relevance determination (ARD) principle. Experimental results on real-world measured datasets demonstrate that the proposed method outperforms conventional SBL and deep learning (DL)based channel estimators, achieving lower latency and higher robustness.