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

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.

💡 Research Summary

The paper addresses a critical challenge for upcoming 6 G networks: predicting high‑frequency (upper‑mid‑band, FR3, 7–24 GHz) channel impulse responses (CIRs) without costly measurements or exhaustive ray‑tracing simulations. The authors propose CIR‑UNext, a deep‑learning framework that leverages abundant low‑frequency (3.5 GHz) CIR data to infer the corresponding 7 GHz CIRs.

Physical Motivation
Based on the 3GPP TR 38.901 channel model, the geometric parameters of each multipath component—delays, angles of arrival/departure, and direction vectors—are frequency‑invariant. Frequency dependence resides only in the per‑path transfer matrix Tₗ(f), which captures path loss, material interaction, and phase rotation. Consequently, low‑band measurements provide accurate geometry, while the high‑band response can be obtained by learning how Tₗ(f) varies with frequency.

Dataset Generation
The authors construct paired 3.5 GHz and 7 GHz datasets using a ray‑tracing pipeline. Real‑world building maps are extracted from OpenStreetMap, converted to detailed 3‑D meshes with Blender, and simulated with Sionna’s Shooting‑and‑Bouncing‑Rays algorithm. Antenna patterns differ between bands (6.3 dBi/65° vs 12.3 dBi/32.5°), and random orientations are applied to capture diverse conditions. Each sample contains up to L = 20 resolved paths with associated delays, interaction counts, AoDs, and antenna pattern values, yielding a rich set of paired features for supervised learning.

Model Architecture
CIR‑UNext consists of two specialized attention U‑Net (AU‑Net) models:

• Gain‑UNext predicts per‑path magnitude at 7 GHz. The AU‑Net‑Aux variant incorporates auxiliary inputs (antenna pattern, AoD) both in the encoder and via skip connections, mitigating information loss.
• Phase‑UNext predicts per‑path phase using a standard AU‑Net backbone.

Both networks share a nine‑block convolutional encoder‑decoder with four down‑sampling/up‑sampling stages and attention gates that focus on salient regions. Training employs mean‑square error loss over 200 epochs on an RTX 5080 GPU, with an 80/10/10 train‑validation‑test split at the building‑map level. Data augmentation includes rotation and mirroring.

Experimental Findings
Four Gain‑UNext variants were compared: baseline U‑Net, AU‑Net, AU‑Net‑Parallel (separate LoS/NLoS branches), and AU‑Net‑Aux. AU‑Net‑Aux achieved the lowest median absolute gain error (0.70 dB) and the smallest inter‑quartile range (1.34 dB). Adding antenna pattern reduced median error from 1.04 dB to 0.95 dB; including both pattern and AoD further lowered it to 0.70 dB, confirming the importance of directional features. Performance was slightly better in the dense‑urban map (BM2) than in the open‑area map (BM1), likely because richer multipath provides stronger learning signals. Phase‑UNext attained a median phase error of 0.27 rad, sufficient for accurate MIMO‑OFDM channel reconstruction.

Extension to Throughput Prediction
Building on CIR‑UNext, the authors introduce Channel2ComMap, a foundation model that maps predicted CIRs to MIMO‑OFDM throughput. Compared with Geo2ComMap, which relies on sparse throughput samples and geographic metadata, Channel2ComMap leverages the full per‑path gain and phase information, delivering a >15 % reduction in throughput prediction error and enabling real‑time resource allocation and beam management.

Contributions and Limitations
The work demonstrates that (1) frequency‑invariant geometric information can be extracted from low‑band measurements, (2) attention‑augmented U‑Nets can learn the frequency‑dependent re‑weighting of multipath components, and (3) the learned model can be generalized to downstream tasks such as throughput prediction. Limitations include validation only at 7 GHz (no tests at higher FR3 frequencies), reliance on simulated ray‑tracing data without extensive field measurements, and a fixed path count L = 20 that may not capture extreme propagation scenarios. Future research directions suggested are multi‑band joint training, incorporation of physics‑informed loss functions, and integration into live digital‑twin platforms for 6 G network optimization.