Cross-Band Channel Impulse Response Prediction: Leveraging 3.5 GHz Channels for Upper Mid-Band

TL;DR

The paper tackles cross-band channel prediction by transferring multipath information from the abundant 3.5 GHz data to the upper mid-band 7 GHz, addressing measurement and RT bottlenecks. It introduces CIR-UNext, combining Gain-UNext for path gains and Phase-UNext for path phases within an Attention U-Net framework, and leverages geometry-invariant multipath descriptions via a per-path transfer matrix $\mathbf{T}_l(f)$. A delay-informed phase approach and auxiliary inputs enhance accuracy, with AU-Net-Aux achieving a median gain error of $0.58$ dB and phase error $0.27$ rad on unseen environments; Channel2ComMap extends CIR-UNext to throughput prediction in MIMO-OFDM, achieving a median error of $24.49$ Mbps over $0$--$1900$ Mbps. The work demonstrates a scalable pipeline from cross-band channel reconstruction to link-layer performance, enabling improved localization, beam management, digital twins, and resource allocation for future 6G networks.

Abstract

Accurate cross-band channel prediction is essential for 6G networks, particularly in the upper mid-band (FR3, 7--24 GHz), where penetration loss and blockage are severe. Although ray tracing (RT) provides high-fidelity modeling, it remains computationally intensive, and high-frequency data acquisition is costly. To address these challenges, we propose CIR-UNext, a deep learning framework designed to predict 7 GHz channel impulse responses (CIRs) by leveraging abundant 3.5 GHz CIRs. The framework integrates an RT-based dataset pipeline with attention U-Net (AU-Net) variants for gain and phase prediction. The proposed AU-Net-Aux model achieves a median gain error of 0.58 dB and a phase prediction error of 0.27 rad on unseen complex environments. Furthermore, we extend CIR-UNext into a foundation model, Channel2ComMap, for throughput prediction in MIMO-OFDM systems, demonstrating superior performance compared with existing approaches. Overall, CIR-UNext provides an efficient and scalable solution for cross-band prediction, enabling applications such as localization, beam management, digital twins, and intelligent resource allocation in 6G networks.

Figures (5)

• Figure 1: Architecture of CIR-UNext: Gain-UNext is an AU-Net-Aux with a $(4\times L + 1)$-channel input for efficient and accurate per-path gain prediction, where $L$ is the number of resolved multipath components. Phase-UNext shares the simliar AU-Net backbone (without the red auxiliary part), with a $(2\times L)$-channel input, but is trained separately for per-path phase prediction. In our experiments, we consider $L=20$.
• Figure 2: Building Maps used in the experiments. (a) BM1, characterized by a prominent open space. (b) BM2, featuring densely clustered buildings.
• Figure 3: Boxplot comparison of path gain prediction errors: (a) between different building maps, and (b) with and without delay input (in both building maps). Each box represents the IQR of the absolute error, with the red line inside indicating the median value. The whiskers extend to the furthest data points within the range defined as Q1$-1.5\cdot$IQR to Q3$+1.5\cdot$IQR, where Q1 and Q3 denote the first and third quartiles, respectively.
• Figure 4: Regions of high error in BM1 and corresponding ray-tracing visualization. (a) Top view of top $5\%$ error distribution. The red dot denotes the BS, and the green dot denotes the UE. (b) Ray-tracing result of the BS-UE link in (a), where the yellow lines represent traced rays.
• Figure 5: Boxplot comparison of Tput prediction errors for Channel2ComMap and Geo2ComMap. AU-Net represents Channel2ComMap with the standard AU-Net backbone, while AU-Net-Aux denotes Channel2ComMap with the AU-Net-Aux architecture.