Deep Learning-Driven Hybrid Architecture for Accurate IRS-Assisted Channel Estimation

Intelligent reflecting surfaces (IRS) have emerged as a promising technology to enhance signal propagation in sixth-generation (6G) wireless communication systems. However, acquiring accurate channel state information (CSI) in IRS-assisted networks remains a significant challenge due to high pilot overhead and the passive nature of IRS elements. This paper presents a novel deep learning framework, termed U-BiTRCE (U-Net, bidirectional long short-term memory, and transformer-based cascaded estimator), which combines U-Net-based spatial encoding, bidirectional long short-term memory (BiLSTM) for temporal modeling, and a Transformer block for contextual refinement. The model predicts cascaded IRS channels using partially observed least squares (LS) estimates of known element groups, thereby minimizing training overhead while achieving high estimation accuracy. A key novelty of this work lies in the unified integration of spatial, sequential, and attention-driven feature learning within a single architecture for IRS channel estimation. In contrast to conventional methods such as LS, Linear Minimum Mean Squared Error (LMMSE), and orthogonal matching pursuit (OMP), the proposed scheme does not require graph modeling or sparsity assumptions. To support real-time feasibility, a detailed computational complexity analysis is presented, demonstrating the model’s scalability and low inference latency. Experimental results show that U-BiTRCE consistently outperforms baseline methods including convolutional deep residual network (CDRN) and graph transformer IRS channel estimator (G-TIRC), achieving up to 152.4% improvement in normalized mean squared error (NMSE) over LS. Additionally, the model exhibits robust convergence across varying signal-to-noise ratio (SNR) levels, IRS sizes, batch sizes, and multi-user configurations. U-BiTRCE maintains an average inference time below 18 milliseconds per frame, validating its practicality for real-time deployment in large-scale IRS-assisted 6G systems.